lang
stringclasses
1 value
s2FieldsOfStudy
listlengths
0
8
url
stringlengths
78
78
fieldsOfStudy
listlengths
0
5
lang_conf
float64
0.8
0.98
title
stringlengths
4
300
paperId
stringlengths
40
40
venue
stringlengths
0
300
authors
listlengths
0
105
publicationVenue
dict
abstract
stringlengths
1
10k
text
stringlengths
1.94k
184k
openAccessPdf
dict
year
int64
1.98k
2.03k
publicationTypes
listlengths
0
4
isOpenAccess
bool
2 classes
publicationDate
timestamp[us]date
1978-02-01 00:00:00
2025-04-23 00:00:00
references
listlengths
0
958
total_tokens
int64
509
40k
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Business", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00c5185c5a5c68761ff4b22fb523d734af9c5c16
[ "Computer Science", "Business" ]
0.81413
Down with the #Dogefather: Evidence of a Cryptocurrency Responding in Real Time to a Crypto-Tastemaker
00c5185c5a5c68761ff4b22fb523d734af9c5c16
Journal of Theoretical and Applied Electronic Commerce Research
[ { "authorId": "1435375867", "name": "Michael Cary" } ]
{ "alternate_issns": null, "alternate_names": [ "J Theor Appl Electron Commer Res" ], "alternate_urls": [ "http://www.jtaer.com/" ], "id": "890beb40-ba59-4681-9bb6-88ed97b7decb", "issn": "0718-1876", "name": "Journal of Theoretical and Applied Electronic Commerce Research", "type": "journal", "url": "http://www.scielo.cl/scielo.php?lng=en&pid=0718-1876&script=sci_serial" }
Recent research in cryptocurrencies has considered the effects of the behavior of individuals on the price of cryptocurrencies through actions such as social media usage. However, some celebrities have gone as far as affixing their celebrity to a specific cryptocurrency, becoming a crypto-tastemaker. One such example occurred in April 2021 when Elon Musk claimed via Twitter that “SpaceX is going to put a literal Dogecoin on the literal moon”. He later called himself the “Dogefather” as he announced that he would be hosting Saturday Night Live (SNL) on 8 May 2021. By performing sentiment analysis on relevant tweets during the time he was hosting SNL, evidence is found that negative perceptions of Musk’s performance led to a decline in the price of Dogecoin, which dropped 23.4% during the time Musk was on air. This shows that cryptocurrencies are affected in real time by the behaviors of crypto-tastemakers.
_Article_ # Down with the #Dogefather: Evidence of a Cryptocurrency Responding in Real Time to a Crypto-Tastemaker **Michael Cary** Division of Resource Economics and Management, West Virginia University, Morgantown, WV 26506, USA; macary@mix.wvu.edu **Abstract: Recent research in cryptocurrencies has considered the effects of the behavior of indi-** viduals on the price of cryptocurrencies through actions such as social media usage. However, some celebrities have gone as far as affixing their celebrity to a specific cryptocurrency, becoming a crypto-tastemaker. One such example occurred in April 2021 when Elon Musk claimed via Twitter that “SpaceX is going to put a literal Dogecoin on the literal moon”. He later called himself the “Dogefather” as he announced that he would be hosting Saturday Night Live (SNL) on 8 May 2021. By performing sentiment analysis on relevant tweets during the time he was hosting SNL, evidence is found that negative perceptions of Musk’s performance led to a decline in the price of Dogecoin, which dropped 23.4% during the time Musk was on air. This shows that cryptocurrencies are affected in real time by the behaviors of crypto-tastemakers. **Keywords: cryptocurrency; crypto-tastemaker; Dogecoin; price dynamics; sentiment analysis** [����������](https://www.mdpi.com/article/10.3390/jtaer16060123?type=check_update&version=1) **�������** **Citation: Cary, M. Down with the** #Dogefather: Evidence of a Cryptocurrency Responding in Real Time to a Crypto-Tastemaker. J. Theor. _Appl. Electron. Commer. Res. 2021, 16,_ [2230–2240. https://doi.org/10.3390/](https://doi.org/10.3390/jtaer16060123) [jtaer16060123](https://doi.org/10.3390/jtaer16060123) Academic Editor: Arcangelo Castiglione Received: 13 August 2021 Accepted: 2 September 2021 Published: 3 September 2021 **Publisher’s Note: MDPI stays neutral** with regard to jurisdictional claims in published maps and institutional affil iations. **Copyright: © 2021 by the author.** Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **JEL Classification: G41; G10** **1. Introduction** The number of cryptocurrencies has grown rapidly over the past decade. With such diversity, choosing a specific cryptocurrency to use can be a daunting task, especially for more casual cryptocurrency users. While some users are concerned with price dynamics, others are concerned with the popularity of the cryptocurrency [1]. In fact, herding behavior in cryptocurrency markets has become a well documented phenomenon in the literature [2], and even cryptocurrencies such as Bitcoin are traded at least in part due to emotional cues [3]. Herding behavior occurs in cryptocurrency markets for many different reasons and is commonly observed during periods where higher levels of risk aversion are exhibited [4]. Herding behavior is particularly strong in smaller cryptocurrencies [5]. Such behavior is a market inefficiency and can lead to market destabilization, particularly in the case of smaller cryptocurrencies [6]. Combined with the effects of the ongoing COVID-19 global pandemic on cryptocurrency markets, the potential for market destabilization among smaller cryptocurrencies is only exacerbated [7]. It is important to note, however, that the choice of empirical framework can potentially impact whether or not evidence of herding is found [8]. On the other side of this phenomenon are the cryptocurrency tastemakers (cryptotastemakers) who attach their notoriety to a particular cryptocurrency, advocating for its growth. There is evidence that social influences can affect cryptocurrencies [9]. However, current research on the impact of crypto-tastemakers is extremely limited, with no papers looking at the real time effects of the actions of a major celebrity on the price of a cryptocurrency to which the celebrity has affixed themselves as a crypto-tastemaker. The literature that does exist considers the impact of social media on cryptocurrencies, which, while extremely valuable, analyzes the impact of pre-planned, low risk activities such as sending a single Tweet, e.g., the impact of a president’s tweets on Bitcoin [10], ----- _J. Theor. Appl. Electron. Commer. Res. 2021, 16_ 2231 predicting the price of a cryptocurrency using social media data [11,12], and predicting bubbles in cryptocurrency markets with social media data [13]. This is in contrast to what is studied in this paper, an extended period of heavily scrutinized, riskier actions performed live, for a public audience. One recent example of a celebrity becoming a crypto-tastemaker is Elon Musk, who affixed his celebrity status to Dogecoin. On 1 April 2021, Elon Musk claimed via Twitter that “SpaceX is going to put a literal Dogecoin on the literal moon”. Shortly after making this hyperbolized claim, it was announced that Musk would be hosting the 8 May 2021 episode of Saturday Night Live (SNL). Musk confirmed this in a personal announcement on his Twitter account on 28 April 2021 in which he called himself the “Dogefather”. All of these specific examples of Musk’s Twitter activity are part of a much larger corpus of crypto-tastemaking, dating back to January 2021 when Musk starting giving Dogecoin attention on Twitter during the GameStop short squeeze [14]. Musk also went on to call Dogecoin mining “fun” in order to increase the popularity of the cryptocurrency [14]. Elon Musk makes for a great example of a crypto-tastemaker since he has been a public figure for decades, largely due to his business ventures and immense wealth. Moreover, during this time he has become a rather divisive figure. He has both an ardent core of followers and currently has 59.5 million followers on Twitter, but he also has many detractors as well—a common nickname for Musk (which appears hundreds of time in our data set) is “Muskrat”. This level of notoriety and divisiveness, along with his longstanding interest in cryptocurrencies, means that once Musk coupled his name to Dogecoin, he was indeed a crypto-tastemaker. In this paper we test for evidence of the real time impact of the highly publicized actions of a crypto-tastemaker by performing sentiment analysis on real time data from twitter during the time that Musk was hosting SNL and finding its effect on the price of Dogecoin. Using standard VAR techniques, we document for the first time in the literature a definite instance of the price of a cryptocurrency responding in real time to the actions of a crypto-tastemaker. Specifically, we find that Elon Musk’s performance on SNL significantly and negatively affected the price of Dogecoin. **2. Dogecoin** Dogecoin is a cryptocurrency alternative to Bitcoin, or an altcoin, that was created in 2013 [15]. Originally created as a joke currency with a randomized reward for mining [14], for most of its history Dogecoin was a niche cryptocurrency that had some degree of cultural relevance due to the peculiarity of its name, but was not a target of significant investment [15]. Prior to 2021, the price of Dogecoin had never been above $0.02 [14]. The technical development of Dogecoin was also underwhelming, with the most recent consistent activity on its main branch on GitHub as of the writing of Young [15] occurring in 2015 (the rise in popularity experienced by Dogecoin in 2021 has led to renewed development, [per the commit history found at https://github.com/dogecoin accessed on 13 August](https://github.com/dogecoin) 2021). However, Dogecoin users have performed some noteworthy, attention grabbing events including sponsoring an American stock car race in 2013 and the Jamaican bobsled team in the 2014 Winter Olympics [15]. Functionally, Dogecoin is based on the Scrypt algorithm and is a derivative of Litecoin, another cryptocurrency derived from Bitcoin [15]. However, unlike Bitcoin and most other cryptocurrencies, there is no limit to the amount of Dogecoin that can theoretically exist [15]. Consequentially, mining Dogecoin remains a quicker and easier process than mining other cryptocurrencies. From a research perspective, Dogecoin remains essentially unstudied in the literature. This is likely due to its effective irrelevance as a potential investment prior to 2021. In fact, in the case of this paper, Dogecoin is studied not for anything intrinsic to Dogecoin itself, but rather for the fact that a crypto-tastemaker affixed themselves to Dogecoin. ----- _J. Theor. Appl. Electron. Commer. Res. 2021, 16_ 2232 **3. Data and Methodology** The ultimate goal of this paper is to test whether the price of Dogecoin responded in real time to the public perception of Musk’s performance on SNL using a standard vector autoregression (VAR) approach. To do this, we need data on the price of Dogecoin as well as a measure of the public perception of Musk’s performance. While the former data set is easily obtained, in this case from CoinDesk.com, the latter data requires some effort to obtain. Twitter is an excellent source of public opinions and tweets are widely used in the quantitative social sciences, e.g., [16–19], thus we will use data collected from Twitter as the basis for measuring public opinion of Musk’s performance. To create the final data on the public’s perception of Musk’s performance, two primary steps were performed. First, relevant tweets from the time period of Musk’s performance must be collected from Twitter. A window of one hour before and after the event was included in our sample to account for delayed responses since there was no a prioi knowledge of the lag time from trade-causing-opinion to the trade itself. Tweets containing any of the following key words as text, hashtags, and/or cashtags were collected: {SNL, SNLmay8, Dogefather, tothemoon, Elon, Musk, Dogecoinrise, Doge, Dogecoin}. Once these tweets were collected, sentiment analysis was performed on the tweets. Sentiment analysis is a form of textual analysis which assigns quantitative values to subjective statements [20]. Positive values are assigned to tweets with a positive opinion, and negative values are assigned to tweets with a negative opinion. In our case, whenever Musk’s performance was well received by the public we obtain positive scores from sentiment analysis, while poorly received portions of Musk’s performance received negative scores from sentiment analysis. To obtain these scores, individual tweets were assigned their own, unique score using the nltk module in Python. Once every tweet had been assigned a score via sentiment analysis, two time series measuring the overall public perception of the performance were created—one for positive opinions and one for negative opinions—by aggregating the tweets during each minute of the performance. The rationale for the two distinct times series is that positive and negative opinions may have asymmetric affects on the price of Dogecoin. Asymmetric effects in time series regressions have proven significant in many cases, e.g., [21–24]. In our case, risk averse investors/users of Dogecoin may sell their holdings if they fear that a poor performance by Musk is actively lowering the price of Dogecoin, but positive opinions of Musk’s performance may have a more muted positive effect. Furthermore, by aggregating all tweets during each minute of the event, we allow for a weighted time series where larger magnitudes for the positive and negative sentiment analysis scores indicate a greater degree of public consensus regarding the performance. The granularity of one minute intervals was chosen because this matches the frequency of the price data for Dogecoin obtained from CoinDesk.com. Summary statistics of the three time series are presented in Table 1 and the three times series are plotted together in Figure 1. **Table 1. Summary statistics of the three time series. The negative sentiment scores are in absolute** value for convenience of use/interpretation. **Time Series** **Mean** **SD** **Min** **Max** Price of Dogecoin in USD 0.584 0.054 0.471 0.700 Total Positive Sentiment 29.67 37.58 4.79 188.35 Total Negative Sentiment 25.23 30.75 2.73 99.53 As can be seen in Figure 1, the general trends of positive and negative sentiment were similar. Twitter activity pertaining to Musk’s performance spiked just as the episode began to air, reaching a peak around 15 min into the episode. From there, a steady decline in both positive and negative sentiment, driven by a decrease in the volume of tweets pertaining to Musk’s performance, was observed. A small spike in both positive and negative sentiment occurs shortly after the conclusion of the episode, likely driven by summary reviews of the ----- _J. Theor. Appl. Electron. Commer. Res. 2021, 16_ 2233 episode, but once the episode had finished airing, the volume of tweets steadily declined to pre-episode levels. The sharp, early decline in the price of Dogecoin, a loss which was never recouped, coincides with the outburst of opinions on Twitter pertaining to Musk’s performance on SNL. A Comparison of the price of dogecoin and public perception of Musk's performance Price of dogecoin Positive Sentiment Negative Sentiment 0.7 175 0.6 150 0.5 125 0.4 100 0.3 75 0.2 50 0.1 25 0.0 0 −50 0 50 100 150 Time (Minutes Relative to Start of Episode) **Figure 1. The three time series. The price of Dogecoin is measured in USD on the left hand axis while** the positive and negative sentiment scores are measured in aggregate values on the right hand axis. The x axis represents time relative to the start of the episode, and the two vertical black lines denote the start and end of the episode. Finally, to run the VAR, we first-difference the data to transform the times series and ensure that they are stationary. As can be seen in Figure 1, the original time series are clearly non-stationary. Augmented Dickey-Fuller tests confirmed that the three first-differenced time series are indeed stationary. Once the time series were first-differenced, the optimal lag length for VAR was determined to be 15 periods (minutes). Once the optimal lag length was determined, VAR was performed. The VAR model takes the standard specification in vector notation found in Equation (1) where p = 15 is the optimal lag length. _Yt = a +_ _p_ ## ∑ ΦkYt−k + ϵt (1) _k=1_ **4. Results and Discussion** Predicting the price of cryptocurrencies, even established ones such as Bitcoin, is no easy feat. From a pure predictive standpoint, myriad machine learning techniques have been applied to this problem with only limited success [25]. Data from social media have been used to aid in this endeavor in various forms including search trends data [26] and sentiment analysis performed on developers comments [27]. However, these types of studies have historically relied on discrete events such as tweets by a crypto-tastemaker, or have used more continuous data from groups of people rather than from individual crypto-tastemakers. This rule extends to causal inference settings as well, e.g., [28]. The VAR results indicate that increases in the magnitude of negative public perception of Musk’s performance had a negative effect on the price of Dogecoin. This can be seen in the upper-rightmost subplot in the cumulative effects plot from our VAR in Figure 2. ----- _J. Theor. Appl. Electron. Commer. Res. 2021, 16_ 2234 Changes in the positive public perception of Musk’s performance had no significant long run effect on the price of Dogecoin. Full VAR results can be found in Tables A1–A3, along with the corresponding impulse response function plots and autocorrelation plots in Figures A1 and A2, respectively. **Figure 2. Cumulative effects plots from VAR.** Looking at the impulse response functions, we see that negative sentiment had a delayed but significant effect on the price of Dogecoin. There was a steady decline in the impulse response function from 5 min to 12 min, and this effect can also be seen in the point estimates from the VAR model for the price of Dogecoin, where the lagged values of negative sentiment for 11 and 12 lags (L11 Negative Sentiment and L12 Negative Sentiment) were negative and statistically significant (Table A1). What this shows is evidence that increases in negative sentiment led to Dogecoin users selling their holdings. Trades began to be finalized in earnest approximately five minutes after an event occurred that led to an increase in negative sentiment, and this behavior continued until the cumulative effect of these sales led to a statistically significant decrease in the price of Dogecoin, occurring approximately at the 12 min mark. These results indicate that investors/users of cryptocurrencies who are interested in the popularity of the cryptocurrency are influenced by the actions of crypto-tastemakers, but that crypto-tastemakers, once thoroughly affixed to a specific cryptocurrency, may only be able to harm the popularity of the coin. Given the fact that this is the first such study, it is possible that a “better performance” (perhaps, e.g., a humanitarian action involving a crypto-tastemaker or a more convincing performance on SNL) could have a positive effect on the price of that cryptocurrency. However, it is entirely possible that when a crypto-tastemaker affixes themselves to a cryptocurrency, that cryptocurrency enters a high risk, low reward state. This is different than some previous, related results on cryptocurrencies, such as [29] who found that Bitcoin responded positively to unscheduled news, whether that news was positive or negative. However, our results do align with [30] who found that certain news ----- _J. Theor. Appl. Electron. Commer. Res. 2021, 16_ 2235 from authorities led to declines and increased volatility in cryptocurrency markets in the largest cryptocurrency exchange in China. Granger causality testing confirms that changes in the level of aggregate negative sentiment Granger-causes changes in the level of the price of Dogecoin, but no other instances of Granger causality exist in this study. Finally, a stability analysis shows that the results are indeed stable. The roots of the characteristic polynomial of the VAR are presented in Figure A3 and are clearly all within the unit circle, a sufficient condition for stability. **5. Conclusions** Cryptocurrencies are used in part based on their popularity; this much is an observed reality of cryptocurrencies. Consequentially, cryptocurrencies are being endorsed by cryptotastemakers. This analysis has shown for the first time that cryptocurrency price dynamics are subject to the real time behaviors of a crypto-tastemaker. Since less mature cryptocurrencies are more likely to be influenced by a crypto-tastemaker, this suggests that less mature cryptocurrencies may have a more complex nature to their price variance. Future research on the relationship between cryptocurrencies and crypto-tastemakers should investigate the direct impact of crypto-tastemakers on the volatility of cryptocurrencies, and if there are spillover effects across cryptocurrencies due to the action of crypto-tastemakers. **Funding: This research received no external funding.** **Data Availability Statement: Final data for the econometric analyses and code for this project is** [available at: https://github.com/cat-astrophic/dogefather accessed on 13 August 2021. The raw](https://github.com/cat-astrophic/dogefather) twitter data set is not stored in the repository due to its size, but it is available from the author upon request. **Conflicts of Interest: The authors declare no conflict of interest.** **Appendix A. VAR Results** **Figure A1. The impulse response function plots from the VAR.** ----- _J. Theor. Appl. Electron. Commer. Res. 2021, 16_ 2236 **Table A1. VAR results for the regressions on the price of Dogecoin. Optimal lag length was selected** using a built in function in the VAR submodule of the statsmodels module in Python. Lx before a variable name denotes that a variable was lagged × times. **Variable** **Coefficient** **Std. Err.** **t-Stat** **_p_** Constant _−0.0003_ 0.0007 _−0.4481_ 0.6541 L1 Price _−0.0565_ 0.0811 _−0.6970_ 0.4858 L1 Positive Sentiment _−0.0444_ 0.1131 _−0.3924_ 0.6947 L1 Negative Sentiment _−0.1338_ 0.1489 _−0.8985_ 0.3689 L2 Price _−0.0426_ 0.0767 _−0.5554_ 0.5787 L2 Positive Sentiment 0.2365 0.1108 2.1341 0.0328 L2 Negative Sentiment _−0.1537_ 0.1534 _−1.0018_ 0.3165 L3 Price _−0.0364_ 0.0783 _−0.4652_ 0.6418 L3 Positive Sentiment _−0.0358_ 0.1132 _−0.3164_ 0.7517 L3 Negative Sentiment 0.0680 0.1548 0.4395 0.6603 L4 Price _−0.1887_ 0.0786 _−2.3993_ 0.0164 L4 Positive Sentiment 0.2558 0.1109 2.3064 0.0211 L4 Negative Sentiment _−0.1971_ 0.1614 _−1.2208_ 0.2221 L5 Price _−0.0250_ 0.080 _−0.3131_ 0.7542 L5 Positive Sentiment _−0.0253_ 0.1160 _−0.2181_ 0.8273 L5 Negative Sentiment 0.0061 0.1637 0.0375 0.9701 L6 Price _−0.0542_ 0.0753 _−0.7191_ 0.4721 L6 Positive Sentiment _−0.1438_ 0.1207 _−1.1918_ 0.2334 L6 Negative Sentiment 0.1467 0.1614 0.9093 0.3632 L7 Price _−0.0498_ 0.0778 _−0.6396_ 0.5224 L7 Positive Sentiment _−0.0709_ 0.1213 _−0.5844_ 0.5590 L7 Negative Sentiment _−0.0612_ 0.1615 _−0.3789_ 0.7047 L8 Price _−0.1269_ 0.0761 _−1.6678_ 0.0954 L8 Positive Sentiment _−0.1935_ 0.1207 _−1.6028_ 0.1090 L8 Negative Sentiment _−0.0623_ 0.1615 _−0.3859_ 0.6996 L9 Price 0.0476 0.0766 0.6217 0.5341 L9 Positive Sentiment 0.0915 0.1214 0.7537 0.4510 L9 Negative Sentiment _−0.0143_ 0.1597 _−0.0897_ 0.9285 L10 Price 0.0742 0.0785 0.9449 0.3447 L10 Positive Sentiment _−0.1283_ 0.1166 _−1.1002_ 0.2712 L10 Negative Sentiment _−0.2174_ 0.1576 _−1.3795_ 0.1678 L11 Price 0.0761 0.0786 0.9681 0.3330 L11 Positive Sentiment 0.0118 0.1153 0.1025 0.9183 L11 Negative Sentiment _−0.2561_ 0.1550 _−1.6521_ 0.0985 L12 Price _−0.0401_ 0.0779 _−0.5144_ 0.6069 L12 Positive Sentiment 0.0812 0.1134 0.7165 0.4737 L12 Negative Sentiment _−0.2535_ 0.1523 _−1.6641_ 0.0961 L13 Price 0.0825 0.0797 1.0350 0.3007 L13 Positive Sentiment _−0.1346_ 0.1113 _−1.2099_ 0.2263 L13 Negative Sentiment _−0.1352_ 0.1476 _−0.9157_ 0.3598 L14 Price 0.0305 0.0785 0.3886 0.6976 L14 Positive Sentiment _−0.3995_ 0.1120 _−3.5670_ 0.0004 L14 Negative Sentiment _−0.0198_ 0.1447 _−0.1369_ 0.8911 L15 Price 0.1195 0.0793 1.5071 0.1318 L15 Positive Sentiment 0.1268 0.1178 1.0766 0.2817 L15 Negative Sentiment _−0.2441_ 0.1429 _−1.7083_ 0.0876 **Table A2. VAR results for the regressions on aggregate positive sentiment. Optimal lag length** was selected using a built in function in the VAR submodule of the statsmodels module in Python. Lx before a variable name denotes that a variable was lagged × times. **Variable** **Coefficient** **Std. Err.** **t-Stat** **_p_** Constant 0.0002 0.0006 0.4140 0.6789 L1 Price _−0.0462_ 0.0667 _−0.6918_ 0.4891 L1 Positive Sentiment _−0.1333_ 0.0930 _−1.4332_ 0.1518 L1 Negative Sentiment 0.3832 0.1226 3.1268 0.0018 L2 Price 0.1605 0.0631 2.5431 0.0110 L2 Positive Sentiment _−0.1773_ 0.0912 _−1.9441_ 0.0519 L2 Negative Sentiment 0.2909 0.1262 2.3052 0.0212 L3 Price _−0.0991_ 0.0644 _−1.5386_ 0.1239 L3 Positive Sentiment _−0.1636_ 0.0932 _−1.7553_ 0.0792 L3 Negative Sentiment 0.3428 0.1274 2.6913 0.0071 ----- _J. Theor. Appl. Electron. Commer. Res. 2021, 16_ 2237 **Table A2. Cont.** **Variable** **Coefficient** **Std. Err.** **t-Stat** **_p_** L4 Price 0.1309 0.0647 2.0233 0.0430 L4 Positive Sentiment 0.2863 0.0913 3.1380 0.0017 L4 Negative Sentiment 0.0617 0.1328 0.4646 0.6422 L5 Price 0.0697 0.0658 1.0597 0.2893 L5 Positive Sentiment _−0.2438_ 0.0954 _−2.5543_ 0.0106 L5 Negative Sentiment 0.1894 0.1347 1.4055 0.1599 L6 Price 0.1146 0.0620 1.8488 0.0645 L6 Positive Sentiment _−0.1225_ 0.0993 _−1.2332_ 0.2175 L6 Negative Sentiment 0.2241 0.1328 1.6881 0.0914 L7 Price _−0.0369_ 0.0640 _−0.5758_ 0.5648 L7 Positive Sentiment _−0.1004_ 0.0998 _−1.0062_ 0.3143 L7 Negative Sentiment 0.1467 0.1329 1.1035 0.2698 L8 Price _−0.0459_ 0.0626 _−0.7336_ 0.4632 L8 Positive Sentiment _−0.1302_ 0.0993 _−1.3104_ 0.1900 L8 Negative Sentiment _−0.0109_ 0.1329 _−0.0818_ 0.9348 L9 Price _−0.0835_ 0.0630 _−1.3252_ 0.1851 L9 Positive Sentiment _−0.0479_ 0.0999 _−0.4797_ 0.6314 L9 Negative Sentiment _−0.2419_ 0.1314 _−1.8405_ 0.0657 L10 Price 0.0016 0.0646 0.0243 0.9806 L10 Positive Sentiment _−0.1976_ 0.0960 _−2.0589_ 0.0395 L10 Negative Sentiment _−0.0600_ 0.1297 _−0.4626_ 0.6437 L11 Price _−0.0624_ 0.0647 _−0.9645_ 0.3348 L11 Positive Sentiment _−0.0680_ 0.0949 _−0.7167_ 0.4735 L11 Negative Sentiment 0.0541 0.1275 0.4241 0.6715 L12 Price 0.1659 0.0641 2.5862 0.0097 L12 Positive Sentiment 0.1125 0.0933 1.2060 0.2278 L12 Negative Sentiment _−0.0633_ 0.1253 _−0.5050_ 0.6136 L13 Price _−0.0338_ 0.0656 _−0.5156_ 0.6061 L13 Positive Sentiment _−0.0902_ 0.0916 _−0.9855_ 0.3244 L13 Negative Sentiment 0.2246 0.1215 1.8485 0.0645 L14 Price 0.1669 0.0646 2.5841 0.0098 L14 Positive Sentiment 0.1857 0.0921 2.0157 0.0438 L14 Negative Sentiment _−0.1116_ 0.1191 _−0.9369_ 0.3488 L15 Price 0.1164 0.0653 1.7831 0.0746 L15 Positive Sentiment 0.0173 0.0969 0.1783 0.8585 L15 Negative Sentiment _−0.0698_ 0.1176 _−0.5938_ 0.5527 **Table A3. VAR results for the regressions on aggregate negative sentiment. Optimal lag length** was selected using a built in function in the VAR submodule of the statsmodels module in Python. Lx before a variable name denotes that a variable was lagged × times. **Variable** **Coefficient** **Std. Err.** **t-Stat** **_p_** Constant 0.0001 0.0004 0.2079 0.8353 L1 Price _−0.1288_ 0.0499 _−2.5803_ 0.0099 L1 Positive Sentiment _−0.1070_ 0.0696 _−1.5359_ 0.1246 L1 Negative Sentiment 0.0033 0.0917 0.0365 0.9709 L2 Price 0.0980 0.0472 2.0762 0.0379 L2 Positive Sentiment _−0.1651_ 0.0682 _−2.4197_ 0.0155 L2 Negative Sentiment 0.1167 0.0945 1.2353 0.2167 L3 Price _−0.0589_ 0.0482 _−1.2207_ 0.2222 L3 Positive Sentiment 0.0219 0.0697 0.3143 0.7533 L3 Negative Sentiment _−0.1159_ 0.0953 _−1.2163_ 0.2239 L4 Price 0.0611 0.0484 1.2624 0.2068 L4 Positive Sentiment 0.2521 0.0683 3.6914 0.0002 L4 Negative Sentiment _−0.1634_ 0.0994 _−1.6432_ 0.1003 L5 Price _−0.0091_ 0.0493 _−0.1853_ 0.8530 L5 Positive Sentiment 0.0732 0.0714 1.0247 0.3055 L5 Negative Sentiment 0.0968 0.1008 0.9604 0.3368 L6 Price 0.1543 0.0464 3.3251 0.0009 L6 Positive Sentiment 0.0828 0.0743 1.1142 0.2652 L6 Negative Sentiment _−0.0941_ 0.0994 _−0.9472_ 0.3435 L7 Price _−0.0023_ 0.0479 _−0.0489_ 0.9610 L7 Positive Sentiment 0.0496 0.0747 0.6640 0.5067 L7 Negative Sentiment 0.0919 0.0995 0.9237 0.3557 ----- _J. Theor. Appl. Electron. Commer. Res. 2021, 16_ 2238 **Table A3. Cont.** **Variable** **Coefficient** **Std. Err.** **t-Stat** **_p_** L8 Price _−0.0279_ 0.0469 _−0.5960_ 0.5512 L8 Positive Sentiment _−0.0365_ 0.0743 _−0.4914_ 0.6232 L8 Negative Sentiment _−0.0170_ 0.0995 _−0.1709_ 0.8643 L9 Price _−0.1617_ 0.0472 _−3.4295_ 0.0006 L9 Positive Sentiment 0.0526 0.0748 0.7029 0.4821 L9 Negative Sentiment _−0.2434_ 0.0984 _−2.4742_ 0.0134 L10 Price _−0.0501_ 0.0484 _−1.0358_ 0.3003 L10 Positive Sentiment _−0.1080_ 0.0718 _−1.5032_ 0.1328 L10 Negative Sentiment _−0.1346_ 0.0971 _−1.3868_ 0.1655 L11 Price 0.0459 0.0484 0.9478 0.3432 L11 Positive Sentiment 0.0101 0.0710 0.1421 0.8870 L11 Negative Sentiment 0.0189 0.0955 0.1985 0.8427 L12 Price 0.1147 0.0480 2.3908 0.0168 L12 Positive Sentiment 0.0729 0.0698 1.0439 0.2965 L12 Negative Sentiment _−0.0868_ 0.0938 _−0.9253_ 0.3548 L13 Price _−0.0609_ 0.0491 _−1.2404_ 0.2148 L13 Positive Sentiment _−0.0424_ 0.0685 _−0.6182_ 0.5365 L13 Negative Sentiment 0.0477 0.0909 0.5241 0.6002 L14 Price 0.0462 0.0483 0.9568 0.3387 L14 Positive Sentiment 0.1419 0.0690 2.0582 0.0396 L14 Negative Sentiment 0.0179 0.0891 0.2011 0.8406 L15 Price 0.0871 0.0488 1.7832 0.0745 L15 Positive Sentiment _−0.1290_ 0.0725 _−1.7775_ 0.0755 L15 Negative Sentiment _−0.0307_ 0.0880 _−0.3485_ 0.7275 **Figure A2. The autocorrelation plots from the VAR.** ----- _J. Theor. Appl. Electron. Commer. Res. 2021, 16_ 2239 Roots of the VAR Characteristic Polynomial 1.00 0.75 0.50 0.25 0.00 −0.25 −0.50 −0.75 −1.00 −1.00 −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 1.00 Real **Figure A3. This plot shows the unit roots from the VAR characteristic polynomial. Since all unit roots** lie inside the unit circle (in red), the VAR process is stable. **References** 1. Al Shehhi, A.; Oudah, M.; Aung, Z. Investigating factors behind choosing a cryptocurrency. In Proceedings of the 2014 IEEE International Conference on Industrial Engineering and Engineering Management, Selangor, Malaysia, 9–12 December 2014; pp. 1443–1447. 2. [Bouri, E.; Gupta, R.; Roubaud, D. Herding behaviour in cryptocurrencies. Financ. Res. Lett. 2019, 29, 216–221. [CrossRef]](http://doi.org/10.1016/j.frl.2018.07.008) 3. [Ahn, Y.; Kim, D. Emotional trading in the cryptocurrency market. Financ. Res. Lett. 2020, 101912. [CrossRef]](http://dx.doi.org/10.1016/j.frl.2020.101912) 4. Da Gama Silva, P.V.J.; Klotzle, M.C.; Pinto, A.C.F.; Gomes, L.L. Herding behavior and contagion in the cryptocurrency market. _[J. Behav. Exp. Financ. 2019, 22, 41–50. [CrossRef]](http://dx.doi.org/10.1016/j.jbef.2019.01.006)_ 5. Vidal-Tomás, D.; Ibáñez, A.M.; Farinós, J.E. Herding in the cryptocurrency market: CSSD and CSAD approaches. Financ. Res. _[Lett. 2019, 30, 181–186. [CrossRef]](http://dx.doi.org/10.1016/j.frl.2018.09.008)_ 6. [Kallinterakis, V.; Wang, Y. Do investors herd in cryptocurrencies–and why? Res. Int. Bus. Financ. 2019, 50, 240–245. [CrossRef]](http://dx.doi.org/10.1016/j.ribaf.2019.05.005) 7. Vidal-Tomás, D. Transitions in the cryptocurrency market during the COVID-19 pandemic: A network analysis. Financ. Res. Lett. **[2021, 101981. [CrossRef]](http://dx.doi.org/10.1016/j.frl.2021.101981)** 8. Stavroyiannis, S.; Babalos, V. Herding behavior in cryptocurrencies revisited: novel evidence from a TVP model. J. Behav. Exp. _[Financ. 2019, 22, 57–63. [CrossRef]](http://dx.doi.org/10.1016/j.jbef.2019.02.007)_ 9. Aggarwal, G.; Patel, V.; Varshney, G.; Oostman, K. Understanding the social factors affecting the cryptocurrency market. arXiv **2019, arXiv:1901.06245.** 10. [Huynh, T.L.D. Does Bitcoin React to Trump’s Tweets? J. Behav. Exp. Financ. 2021, 31, 100546. [CrossRef]](http://dx.doi.org/10.1016/j.jbef.2021.100546) 11. Lamon, C.; Nielsen, E.; Redondo, E. Cryptocurrency price prediction using news and social media sentiment. SMU Data Sci. Rev. **2017, 1, 1–22.** 12. [Philippas, D.; Rjiba, H.; Guesmi, K.; Goutte, S. Media attention and Bitcoin prices. Financ. Res. Lett. 2019, 30, 37–43. [CrossRef]](http://dx.doi.org/10.1016/j.frl.2019.03.031) 13. Phillips, R.C.; Gorse, D. Predicting cryptocurrency price bubbles using social media data and epidemic modelling. In Proceedings of the 2017 IEEE Symposium Series on Computational Intelligence (SSCI), Honolulu, HI, USA, 27 November–1 December 2017; pp. 1–7. 14. [Chohan, U.W. A History of Dogecoin. Discussion Series: Notes on the 21st Century. 2017. Available online: https://ssrn.com/](https://ssrn.com/abstract=3091219) [abstract=3091219 (accessed on 13 August 2021).](https://ssrn.com/abstract=3091219) 15. [Young, I. Dogecoin: A Brief Overview & Survey. 2018. Available online: https://ssrn.com/abstract=3306060 (accessed on 13](https://ssrn.com/abstract=3306060) August 2021). ----- _J. Theor. Appl. Electron. Commer. Res. 2021, 16_ 2240 16. [Ante, L. How Elon Musk’s Twitter Activity Moves Cryptocurrency Markets. 2021. Available online: https://ssrn.com/abstract=](https://ssrn.com/abstract=3778844) [3778844 (accessed on 13 August 2021).](https://ssrn.com/abstract=3778844) 17. López, M.; Sicilia, M.; Moyeda-Carabaza, A.A. Creating identification with brand communities on Twitter: The balance between [need for affiliation and need for uniqueness. Internet Res. 2017, 27, 21–51. [CrossRef]](http://dx.doi.org/10.1108/IntR-12-2013-0258) 18. Saura, J.R.; Reyes-Menéndez, A.; deMatos, N.; Correia, M.B. Identifying Startups Business Opportunities from UGC on Twitter [Chatting: An Exploratory Analysis. J. Theor. Appl. Electron. Commer. Res. 2021, 16, 1929–1944. [CrossRef]](http://dx.doi.org/10.3390/jtaer16060108) 19. Mohammadi, A.; Hashemi Golpayegani, S.A. SenseTrust: A Sentiment Based Trust Model in Social Network. J. Theor. Appl. _[Electron. Commer. Res. 2021, 16, 2031–2050. [CrossRef]](http://dx.doi.org/10.3390/jtaer16060114)_ 20. Liu, B. Sentiment analysis and subjectivity. Handb. Nat. Lang. Process. 2010, 2, 627–666. 21. Maiti, M.; Vyklyuk, Y.; Vukovi´c, D. Cryptocurrencies chaotic co-movement forecasting with neural networks. Internet Technol. _[Lett. 2020, 3, e157. [CrossRef]](http://dx.doi.org/10.1002/itl2.157)_ 22. Maiti, M.; Grubisic, Z.; Vukovic, D.B. Dissecting Tether’s Nonlinear Dynamics during Covid-19. J. Open Innov. Technol. Mark. _[Complex. 2020, 6, 161. [CrossRef]](http://dx.doi.org/10.3390/joitmc6040161)_ 23. Vukovic, D.; Maiti, M.; Grubisic, Z.; Grigorieva, E.M.; Frömmel, M. COVID-19 Pandemic: Is the Crypto Market a Safe Haven? [The Impact of the First Wave. Sustainability 2021, 13, 8578. [CrossRef]](http://dx.doi.org/10.3390/su13158578) 24. Yue, W.; Zhang, S.; Zhang, Q. Asymmetric news effects on cryptocurrency liquidity: An Event study perspective. Financ. Res. _[Lett. 2021, 41, 101799. [CrossRef]](http://dx.doi.org/10.1016/j.frl.2020.101799)_ 25. Ortu, M.; Uras, N.; Conversano, C.; Destefanis, G.; Bartolucci, S. On Technical Trading and Social Media Indicators in Cryptocurrencies’ Price Classification Through Deep Learning. arXiv 2021, arXiv:2102.08189. 26. Matta, M.; Lunesu, I.; Marchesi, M. Bitcoin Spread Prediction Using Social and Web Search Media. In Proceedings of the UMAP 2015—23rd Conference on User Modeling, Adaptation and Personalization, Dublin, Ireland, 29 June 2015–3 July 2015; pp. 1–10. 27. Bartolucci, S.; Destefanis, G.; Ortu, M.; Uras, N.; Marchesi, M.; Tonelli, R. The Butterfly “Affect”: Impact of development practices [on cryptocurrency prices. EPJ Data Sci. 2020, 9, 21. [CrossRef]](http://dx.doi.org/10.1140/epjds/s13688-020-00239-6) 28. Mai, F.; Shan, Z.; Bai, Q.; Wang, X.; Chiang, R.H. How does social media impact Bitcoin value? A test of the silent majority [hypothesis. J. Manag. Inf. Syst. 2018, 35, 19–52. [CrossRef]](http://dx.doi.org/10.1080/07421222.2018.1440774) 29. Rognone, L.; Hyde, S.; Zhang, S.S. News sentiment in the cryptocurrency market: An empirical comparison with Forex. Int. Rev. _[Financ. Anal. 2020, 69, 101462. [CrossRef]](http://dx.doi.org/10.1016/j.irfa.2020.101462)_ 30. Zhang, S.; Zhou, X.; Pan, H.; Jia, J. Cryptocurrency, confirmatory bias and news readability–evidence from the largest Chinese [cryptocurrency exchange. Account. Financ. 2019, 58, 1445–1468. [CrossRef]](http://dx.doi.org/10.1111/acfi.12454) -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/jtaer16060123?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/jtaer16060123, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/0718-1876/16/6/123/pdf?version=1630647783" }
2,021
[ "JournalArticle" ]
true
2021-09-03T00:00:00
[ { "paperId": "c0d2f057b8a1080c85ea56ccd6f37abcdde172ab", "title": "Does Bitcoin React to Trump’s Tweets?" }, { "paperId": "fe9399330887ccadd3b7556656545deb0c842fb2", "title": "COVID-19 Pandemic: Is the Crypto Market a Safe Haven? The Impact of the First Wave" }, { "paperId": "128238e97e16c20b610deef9f5ff6a12e893edb3", "title": "SenseTrust: A Sentiment Based Trust Model in Social Network" }, { "paperId": "59d6741b709b907f019ae5ebe41947d74a614c68", "title": "Identifying Startups Business Opportunities from UGC on Twitter Chatting: An Exploratory Analysis" }, { "paperId": "5d7b6f303ebee8c1bff3fc69f9c16e0a176404b9", "title": "On Technical Trading and Social Media Indicators in Cryptocurrencies' Price Classification Through Deep Learning" }, { "paperId": "5e1c8359169dd62e5fbd16f46612034dbeb561f4", "title": "How Elon Musk's Twitter Activity Moves Cryptocurrency Markets" }, { "paperId": "903b914604e1d28bd130e7837871352134acf24e", "title": "Transitions in the cryptocurrency market during the COVID-19 pandemic: A network analysis" }, { "paperId": "c0871a4de73861e7ea55e16bec1c0ea5c8faa264", "title": "Emotional trading in the cryptocurrency market" }, { "paperId": "27d2b03b3021128d0a4363dc456076b9352d27f5", "title": "Dissecting Tether’s Nonlinear Dynamics during Covid-19" }, { "paperId": "03327a2ef5a32ce7a71e69259774fb920811168b", "title": "Asymmetric News Effects on Cryptocurrency Liquidity: an Event Study Perspective" }, { "paperId": "adeb0a2c26f6cc9c0f953d3e9d372682449962f3", "title": "The Butterfly “Affect”: impact of development practices on cryptocurrency prices" }, { "paperId": "50f316ef0b1c3ac87b117c94f7286712cf61ec94", "title": "News sentiment in the cryptocurrency market: An empirical comparison with Forex" }, { "paperId": "4683a10d803eded36443d45dcce394b2a0fca0f5", "title": "Cryptocurrencies chaotic co‐movement forecasting with neural networks" }, { "paperId": "a75d70b1f007319ccee4807ef76170a7a38058b6", "title": "Do investors herd in cryptocurrencies – and why?" }, { "paperId": "70927b9d6f3e4fd45409e4f75d7cc6bc04f7a73c", "title": "Herding in the cryptocurrency market: CSSD and CSAD approaches" }, { "paperId": "bc7ae10f0bcc1c11d5a2dcf3d09b9453f21a8706", "title": "Herding behavior in cryptocurrencies revisited: Novel evidence from a TVP model" }, { "paperId": "c13adb42f983a1c9553cd0ea0826f896936faa4f", "title": "Herding behaviour in cryptocurrencies" }, { "paperId": "29810afb6fa758911760eca141344d70ef199695", "title": "Herding behavior and contagion in the cryptocurrency market" }, { "paperId": "8fab40b779469a4b7497c5aebbb59d34c6ec1d22", "title": "Cryptocurrency, Confirmatory Bias and News Readability – Evidence from the Largest Chinese Cryptocurrency Exchange" }, { "paperId": "b0f2741f846fe08040a77c6bd59f0f9e8d6eff34", "title": "Understanding the Social Factors Affecting the Cryptocurrency Market" }, { "paperId": "8c0a5bb73f6e03c0d7abebebe37efcef2fac940b", "title": "Media Attention and Bitcoin Prices" }, { "paperId": "7a5ce021e8e2231a70e1c4a990b34a32a2e0308b", "title": "Dogecoin: A Brief Overview & Survey" }, { "paperId": "7677a0070f604e97bc7544c168f84d20ebb26791", "title": "How Does Social Media Impact Bitcoin Value? A Test of the Silent Majority Hypothesis" }, { "paperId": "097fa70b534568d46f665219f101e94a79f21dee", "title": "Predicting cryptocurrency price bubbles using social media data and epidemic modelling" }, { "paperId": "4a610c43cb02610535facfb87f139f6ca6cb51c4", "title": "Creating identification with brand communities on Twitter: The balance between need for affiliation and need for uniqueness" }, { "paperId": "5d21f76ff461fce03ec0e539eb1b9edd21a407c2", "title": "Investigating factors behind choosing a cryptocurrency" }, { "paperId": "c3b80de058596cee95beb20a2d087dbcf8be01ea", "title": "Cryptocurrency Price Prediction Using News and Social Media Sentiment" }, { "paperId": "1345a50edee28418900e2c1a4292ccc51138e1eb", "title": "Bitcoin Spread Prediction Using Social and Web Search Media" }, { "paperId": "29366f39dc4237a380f20e9ae9fbeffa645200c9", "title": "Sentiment Analysis and Subjectivity" }, { "paperId": null, "title": "A Brief Overview & Survey" }, { "paperId": null, "title": "Emotional trading in the cryptocurrency" }, { "paperId": null, "title": "A History of Dogecoin. Discussion Series: Notes on the 21st Century" }, { "paperId": null, "title": "Notes on the 21st Century" }, { "paperId": null, "title": "A History of Dogecoin . Discussion Series : Notes on the 21 st Century . 2017" } ]
10,960
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00cb943e73fce88d768f3891ad67fa48c6b70ed2
[ "Computer Science" ]
0.885892
Name enhanced SDN framework for service function chaining of elastic Network functions
00cb943e73fce88d768f3891ad67fa48c6b70ed2
Conference on Computer Communications Workshops
[ { "authorId": "153673738", "name": "Sameer G. Kulkarni" }, { "authorId": "2913151", "name": "M. Arumaithurai" }, { "authorId": "40593388", "name": "Argyrious G. Tasiopoulos" }, { "authorId": "3456310", "name": "Yiaonis Psaras" }, { "authorId": "145922660", "name": "K. Ramakrishnan" }, { "authorId": "1799074", "name": "Xiaoming Fu" }, { "authorId": "1680313", "name": "G. Pavlou" } ]
{ "alternate_issns": null, "alternate_names": [ "INFOCOM WKSHPS", "Conf Comput Commun Work" ], "alternate_urls": null, "id": "be267cb9-6411-4126-8b64-4847025171ee", "issn": null, "name": "Conference on Computer Communications Workshops", "type": "conference", "url": null }
null
# Name Enhanced SDN Framework for Service Function Chaining of Elastic Network Functions ## Sameer G Kulkarni[∗], Mayutan Arumaithurai[∗], Argyrious Tasiopoulos[‡], Yiaonis Psaras[‡], K.K. Ramakrishnan[†], Xiaoming Fu[∗], George Pavlou[‡] _∗University of G¨ottingen, Germany, ‡University College London, †University of California, Riverside._ **_Abstract—Middleboxes have become an integral part of Inter-_** **net infrastructure, providing additional flow processing for policy** **control, security, and performance optimization. Network Func-** **tion Virtualisation (NFV) proposes the deployment of software-** **based middleboxes on top of commercial off-the-shelf (COTS),** **enabling the dynamic adjustment of Virtual Network Functions** **(VNFs), both in terms of instance numbers and computational** **power. The performance of Data center and Enterprise networks** **depend strongly on efficient scaling of VNFs and the traffic load** **balance across VNF instances. To this end, we present Name** **enhanced SDN framework for service function chaining of elastic** **Network functions (NSN) that extends the Function-Centric** **Service Chaining (FCSC) with load balancing functionalities to** **achieve efficient network utilization while reducing the switch** **flow rules by 2-4x compared to traditional SDN approaches.** I. INTRODUCTION Software Defined Networking (SDN) enables to realize the policy enforcement by providing greater flexibility and control in steering the packets through desired function chain. With logically centralized controller, it is easier to enforce heterogeneous policies for different flows and to steer the traffic across the network. In addition, with the global view of network topology, it is easier to monitor the resource utilization. Network Function Virtualization (NFV) has caused the paradigm shift towards deploying the soft middleboxes that provide flexible realization of network services with greater cost optimization [1]. SDN and NFV greatly augment to provide flexible and dynamic software-based network environment. On the other hand, Information-Centric Networks (ICN) and Named Data Networking (NDN) architectures introduce the naming layer to the network architecture that decouple the content/name from the location. This offers greater flexibility in routing the flows based on service types, without actually knowing the exact location in the network. Traffic dynamics often trigger for reallocation and reconfiguration of network resources. In case of high demands, some resources end up being over-utilized, resulting into higher latency and SLA degradation, while on other occasions, end up being underutilized. In such circumstances, in order to meet the performance and energy objectives, the NF instances (NFIs) need to be dynamically instantiated or decommissioned or even relocated/migrated. However, to make it happen, several key decisions need to be made in terms of knowing when to instantiate, decommission or migrate the instance, which network instances need to be scaled, where in the network to place the instances and how to redistribute the Fig. 1: NSN Architecture load among the available instances. Several recent works [1], [2], [3] have tried to address these aspects within the hood of traditional or SDN framework. FCSC [4] exploits the benefits of NDN in combination with SDN to provide a more flexible, scalable and reliable framework to realize service function chaining. However, it falls short of incorporating a reliable mechanism for applying load balancing over NFIs. Load balancing is fundamental to ensure efficient utilization of resources and to meet the SLA requirements. Herein, we present Name enhanced SDN framework for service function chaining of elastic Network functions(NSN) that exploits named service instances and compliments the SDN framework by providing the capabilities of efficient load balancing and elastic scaling of VNF’s via service instantiation, consolidation, while supporting flow redirections for achieving higher VNF utilization. II. RELATED WORK Slick [1] provides a programming model abstraction, where the SDN controller employs heuristic based approaches for estimating the dynamic placement, steering and consolidation of VNFs. However, load balancing is not explicitly addressed and the routing does not take into consideration the network load upon the load steering decisions. E2 [2] presents a NFV scheduling framework that supports affinity based NF placement while trying to minimize the traffic across switches as well as deploying dynamic scaling of NF instances. SIMPLE [3] primarily addresses the SDN based traffic steering approach that tries to optimize on the total rules. It relies on the ILP solver to provide online load balancing. III. NSN ENHANCED ARCHITECTURE We present the high level architecture and design of NSN, that incorporates name based network function instances and enhances SDN’s capability to handle placement, routing and flow redirections. ----- _A. Name based Routing_ NSN enhances FCSC’s name based routing mechanism to perform NFI based routing, wherein all the NFIs are uniquely identified by a name. Policy enforcement on the flows is performed by the controller by encoding the names of the sequence of network functions that are required for the flow to pass through. This aspect is essentially the concept of Information-Centric Networking, wherein the function that a flow requests is decoupled from the location where this function/service is going to be executed. That is, packets indicate through their headers the service function they require and the network is responsible for routing those packets towards the right location. This notion of location-independence can support real-time flow steering and redirection to dynamically instantiated/re-located services out of the box. Once a packet goes through a NF and the corresponding service is executed, the header is modified to remove this service from the chain of required services. A key difference compared to current IP based SDN solutions is that the intermediate switches do not need to maintain per flow forwarding information or similar finegrained forwarding rules, but only need to store forwarding information to reach the named instances. The switches that only route packets at specific service instances, have to keep a single forwarding rule for each service instance. Thus, the state maintained at intermediate routers is proportional to the number of instances and not to the number of flows. Moreover, these rules can be set in a proactive manner as soon as an NFI is instantiated, removed or re-located. Only the ingress switches and the edge switches connected to NFI that are servicing the flow, keep a per flow state forwarding table to ensure that the right labels (i.e., the next hops service instances) are placed on the flow’s header. Another advantage compared to existing IP based solutions is that when an NFI is removed or re-located, in the case of NSN, only the forwarding entries to these instances need to be changed, whereas in the case of current solutions, all entries pertaining to flows that are being serviced by this NFI needs to be modified. Similarly, in case of flow redirection, the proposed scheme provides a notion of atomic rule update as it needs one rule update. NFI node just needs to change the NFI tag to another instance. 1,400 1,200 1,000 800 600 400 200 0 FCSC NSN S-SDN 677 1,264 100 80 60 40 20 0 FCSC NSN S-SDN Svc-A Svc-B Svc-C 1,035 408 313 371 407 53 55109 86102231 112183 140 151 167 5 10 25 50 75 100 No. of Flows (a) Total Switch rules for flows. Network Functions (b) Load across different NFs. _B. VNF Placement and Elastic Scaling_ NSN supports placement, instantiation, removal and relocation of instances to better support the dynamic requirements of flows. The NSN architecture can facilitate for different heuristic based placement mechanisms. NSN enables SDN framework to make quicker heuristic based placement decisions, and allows for finer and quicker course corrections to redistribute the load by either redirecting flows via other instances and/or by instantiating, removing or relocating the NFIs, and thereby enables to overcome the disadvantages of making such heuristic based decisions instead of time consuming and complex ILP [3] decisions. SDN controller can periodically monitor the utilization of the NFIs and network Fig. 2: Evaluation Results indicating Flow rule optimization and load balancing characteristics. link utilization, that can assist in identifying the optimal placement of NFI’s that need to be dynamically instantiated. One such approach is to compute the optimal location by accounting available resources with greater affinity for the flows. On the same lines, the under-utilized NF instances can be decommissioned. NSN estimates the load on each instance based on the gathered link statistics, flows in the system, and the explicit notifications from instances. Based on the load thresholds, it can then dynamically instantiate and decommission specific network instances to ensure the instance utilization rates are kept within the optimal levels. IV. IMPLEMENTATION AND EVALUATION We implemented NSN, as set of modules on top of POX SDN contoller in Python (around 2500 lines of code). We use the Open vSwitch (2.4.0) for SDN switches and implemented custom network modules in linux using python scapy utility. We evaluate NSN using the Mininet network emulator and use the data center tree topology. As in [3], each flow has a policy chain of 3 distinct NFs that originate and terminate at random nodes. Key focus of our evaluation is to measure and quantify the benefits in terms of overall number of flow rules required for steering and load across active network instances. We compare NSN with Standard SDN (S-SDN) that employs IP 5-tuple based rule setting and with FCSC that employs rule setting purely based on the named network functions. Figure 2(a) depicts the total number of flow rules installed across all switches in the network for different number of flows. Figure 2(b) indicates the normalized average load observed across all the active instances of services A, B and C. We can see that NSN provides significant reduction in the total number of rules stored at switches compared to S-SDN. Moreover, NSN ensures load balancing capability identical to that of the S-SDN solution. ACKNOWLEDGEMENT This work was supported by EU FP7 Marie Curie Actions CleanSky ITN project Grant No. 607584, and NICT EUJAPAN GreenICN project Grant No. 608518. REFERENCES [1] B. Anwer, T. Benson, N. Feamster et al., “Programming slick network functions,” in ACM SOSR. ACM, 2015. [2] S. Palkar, C. Lan, S. Han et al., “E2: A framework for nfv applications,” in ACM SOSP, 2015. [3] Z. A. Qazi, C.-C. Tu, L. Chiang et al., “Simple-fying middlebox policy enforcement using sdn,” in ACM SIGCOMM, 2013. [4] M. Arumaithurai, J. Chen, E. Monticelli et al., “Exploiting icn for flexible management of software-defined networks,” in ACM ICN, 2014. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/INFCOMW.2016.7562043?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/INFCOMW.2016.7562043, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "other-oa", "status": "GREEN", "url": "https://discovery.ucl.ac.uk/1536699/1/Pavlou_Kulkarni-2016-infocom-workshop.pdf" }
2,016
[ "JournalArticle", "Conference" ]
true
2016-04-10T00:00:00
[]
2,649
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00cf3d116538b3fbc71c113df890d54299369a61
[ "Computer Science" ]
0.891446
Universal Randomized Guessing With Application to Asynchronous Decentralized Brute–Force Attacks
00cf3d116538b3fbc71c113df890d54299369a61
IEEE Transactions on Information Theory
[ { "authorId": "1734871", "name": "N. Merhav" }, { "authorId": "1796627", "name": "A. Cohen" } ]
{ "alternate_issns": null, "alternate_names": [ "IEEE Trans Inf Theory" ], "alternate_urls": [ "https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?puNumber=18", "http://ieeexplore.ieee.org/servlet/opac?punumber=18" ], "id": "748e730b-add9-47ee-819d-8ae54e504ef9", "issn": "0018-9448", "name": "IEEE Transactions on Information Theory", "type": "journal", "url": "http://www.comm.utoronto.ca/trans-it/" }
Consider the problem of guessing the realization of a random vector <inline-formula> <tex-math notation="LaTeX">${X}$ </tex-math></inline-formula> by repeatedly submitting queries (guesses) of the form “Is <inline-formula> <tex-math notation="LaTeX">${X}$ </tex-math></inline-formula> equal to <inline-formula> <tex-math notation="LaTeX">${x}$ </tex-math></inline-formula>?” until an affirmative answer is obtained. In this setup, a key figure of merit is the number of queries required until the right vector is identified, a number that is termed the <italic>guesswork</italic>. Typically, one wishes to devise a guessing strategy which minimizes a certain guesswork moment. In this work, we study a universal, decentralized scenario where the guesser does not know the distribution of <inline-formula> <tex-math notation="LaTeX">${X}$ </tex-math></inline-formula>, and is not allowed to use a strategy which prepares a list of words to be guessed in advance, or even remember which words were already used. Such a scenario is useful, for example, if bots within a Botnet carry out a brute–force attack in order to guess a password or decrypt a message, yet cannot coordinate the guesses between them or even know how many bots actually participate in the attack. We devise universal decentralized guessing strategies, first, for memoryless sources, and then generalize them for finite–state sources. In each case, we derive the guessing exponent, and then prove its asymptotic optimality by deriving a compatible converse bound. The strategies are based on randomized guessing using a universal distribution. We also extend the results to guessing with side information. Finally, for all above scenarios, we design efficient algorithms in order to sample from the universal distributions, resulting in strategies which do not depend on the source distribution, are efficient to implement, and can be used asynchronously by multiple agents.
## Universal Randomized Guessing with Application to Asynchronous Decentralized Brute–Force Attacks #### Neri Merhav[∗] Asaf Cohen[†] #### November 13, 2018 Abstract Consider the problem of guessing the realization of a random vector X by repeatedly submitting queries (guesses) of the form “Is X equal to x?” until an affirmative answer is obtained. In this setup, a key figure of merit is the number of queries required until the right vector is identified, a number that is termed the guesswork. Typically, one wishes to devise a guessing strategy which minimizes a certain guesswork moment. In this work, we study a universal, decentralized scenario where the guesser does not know the distribution of X, and is not allowed to use a strategy which prepares a list of words to be guessed in advance, or even remember which words were already used. Such a scenario is useful, for example, if bots within a Botnet carry out a brute–force attack in order to guess a password or decrypt a message, yet cannot coordinate the guesses between them or even know how many bots actually participate in the attack. We devise universal decentralized guessing strategies, first, for memoryless sources, and then generalize them for finite–state sources. In each case, we derive the guessing exponent, and then prove its asymptotic optimality by deriving a compatible converse bound. The strategies are based on randomized guessing using a universal distribution. We also extend the results to guessing with side information. Finally, for all above scenarios, we design efficient algorithms in order to sample from the universal distributions, resulting in strategies which do not depend on the source distribution, are efficient to implement, and can be used asynchronously by multiple agents. Index Terms: guesswork; universal guessing strategy; randomized guessing; decentralized guessing; guessing with side information; Lempel–Ziv algorithm; efficient sampling from a distribution. ∗N. Merhav is with the Andrew and Erna Viterbi Faculty of Electrical Engineering, Technion – Israel Institute of Technology, Haifa 32000, Israel. E-mail: merhav@technion.ac.il †A. Cohen is with the Department of Communication System Engineering, Ben Gurion University of the Negev, Beer Sheva 84105, Israel. E-mail: coasaf@bgu.ac.il ----- ### 1 Introduction Consider the problem of guessing the realization of a random n–vector X using a sequence of yes/no queries of the form: “Is X = x1?”, “Is X = x2?” and so on, until an affirmative answer is obtained. Given a distribution on X, a basic figure of merit in such a guessing game is the guesswork, defined as the number of trials required until guessing the right vector. Devising guessing strategies to minimize the guesswork, and obtaining a handle on key analytic aspects of it, such as its moments or its large deviations rate function, has numerous applications in information theory and beyond. For example, sequential decoding [1, 2] or guessing a codeword which satisfies certain constraints [3]. In fact, since the ordering of all sequences of length n in a descending order of probabilities is, as expected, the optimal strategy under many optimality criteria[1], the guessing problem is intimately related to fixed to-variable source coding without the prefix constraint, or one-shot coding, where it is clear that one wishes to order the possible sequences in a descending probability of appearance before assigning them codewords [4, 5, 6]. Contemporary applications of guesswork focus on information security, that is, guessing passwords or decrypting messages protected by random keys. E.g., one may use guessing strategies and their guesswork exponents while proactively trying to crack passwords, as a mean of assessing password security within an organization [7, 8]. Indeed, it is increasingly important to be able to assess password strength [9], especially under complex (e.g., non i.i.d.) password composition requirements. While the literature includes several studies assessing strength by measuring how hard it is for common cracking methods to break a certain set of passwords [8, 9] or by estimating the entropy of passwords created under certain rules [10], the guesswork remains a key analytic tool in assessing password strength for a given sequence length and distribution. As stated in [11], “we are yet to see compelling evidence that motivated users can choose passwords which resist guessing by a capable attacker”. Thus, analyzing the guesswork is useful in assessing how strong a key–generation system is, how hard will it be for a malicious party to break it, or, from the malicious side point of view, how better is one guessing strategy compared to the other. Arguably, human–created passwords may be of a finite, relatively small length, rather 1Specifically, if G(X) is the guesswork, this order minimizes E{F [G(X)]} for any monotone nondecreasing function F . ----- than long sequences which justify asymptotic analysis of the guesswork. Yet, as mentioned above, the guesswork, as a key figure of merit, may be used to aid in assessing computer generated keys [12] or passwords as well. For example, a random key might be of tens or even hundreds of bits long, and passwords saved on servers are often salted before being hashed, resulting in increased length [13]. Moreover, experiments done on finite block lengths agree with the insights gained from the asymptotic analysis [14]. As a result, large deviations and asymptotic analysis remain as key analytic tools in assessing password strength [15, 16, 14, 17, 18]. Such asymptotic analysis provides us, via tractable expressions, the means to understand the guesswork behaviour, the effect various problem parameters have on its value, and the fundamental information measures which govern it. E.g., while the entropy is indeed a relevant measure for “randomness” in passwords [10], via asymptotic analysis of the guesswork [2] we now know that the R´enyi entropy is the right measure when guessing or even a distributed brute–force attack [18, 19]. Non–asymptotic results, such as the converse result in [20], then give us finer understanding of the dependence on the sequence length. Keeping the above applications in mind, it is clear the vanilla model of a single, all capable attacker, guessing a password X drawn from an i.i.d. source of a known distribu tion, is rarely the case of interest. In practical scenarios, several intricacies complicate the problem. While optimal passwords should have maximum entropy, namely, be memoryless and uniformly distributed over the alphabet, human-created passwords are hardly ever such. They tend to have memory and a non–uniform distribution [21], due to the need to remem ber them as well as many other practical considerations (e.g., keyboard structure or the native language of the user) [22, 23]. Thus, the ability to efficiently guess non-memoryless passwords and analyze the performance of such guessing strategies is crucial. Moreover, the underlying true distribution is also rarely known. In [21], the authors investigated the distribution of passwords from four known databases, and tried to fit a Zipf distribution[2]. While there was no clear match, it was clear that a small parameter s is required, to account for a heavy tail. Naturally, [21] also stated that “If the right distribution of passwords can be identified, the cost of guessing a password can be reduced”. Last but not least, from the attacker’s side, there might be additional information which 2In this model, the probability of password with rank i is Pi = Kis, where s is a parameter and K is a normalizing constant ----- facilitates the guessing procedure on the one hand, yet there might be restrictions that prevent him/her from carrying out the optimal guessing strategy. That is, on the one hand, the attacker might have side information, e.g., passwords for other services which are correlated with the one currently attacked, and thereby significantly decrease the guesswork [2, 24, 15, 18]. On the other hand, most modern systems will limit the ability of an attacker to submit too many queries from a single IP address, hence to still submit a large amount, these must be submitted from different machines. Such machines may not be synchronized[3], namely, one may not know which queries were already submitted by the other. Moreover, storing a large (usually, an exponentially large) list of queries to be guessed might be a too heavy burden, especially for small bots in the botnet (e.g., IoT devices). The attacker is thus restricted to distributed brute force attacks, where numerous devices send their queries simultaneously, yet without the ability to synchronize, without knowing which queries were already sent, or which bots are currently active and which ones failed [19]. #### Main Contributions In this paper, we devise universal, randomized (hence, decentralized) guessing strategies for a wide family of information sources, and assess their performance by analyzing their guessing moments, as well as exponentially matching converse bounds, thereby proving their asymptotic optimality. Specifically, we begin from the class of memoryless sources, and propose a guessing strategy. The strategy is universal both in the underlying source distribution and in the guesswork moment to be optimized. It is based on a randomized approach to guessing, as opposed to an ordered list of guesses, and thus it can be used by asynchronous agents that submit their guesses concurrently. We prove that it achieves the optimal guesswork exponent, we provide an efficient implementation for the random selection of guesses, and finally, extend the results to guessing with side information. Next, we broaden the scope to a wider family of non–unifilar finite–state sources, namely, hidden Markov sources. We begin with a general converse theorem and then provide a simple matching direct theorem, based on deterministic guessing. We then provide an alternative direct theorem, that employs a randomized strategy, which builds on the Lempel–Ziv (LZ) algorithm [27]. Once again, both results are tight in terms of the guesswork exponent, and 3When bots or processes working in parallel are able to be completely synchronized, they may use a pre-compiled list of usernames and passwords - “hard-coding” the guessing strategy [25, 26]. ----- are universal in the source distribution and the moment. A critical factor in guessing strategies is their implementation. Deterministic approaches require hard-coding long lists (exponentially large in the block length), and hence are mem ory consuming, while in a randomized approach, one needs to sample from a specific dis tribution, which might require computing an exponential sum. In this paper, we give two efficient algorithms to sample from the universal distribution we propose. The first algorithm is based on (a repeated) random walk on a growing tree, thus randomly and independently generating new LZ phrases, to be used as guesses. The second algorithm is based on feeding a (slightly modified) LZ decoder with purely random bits. Finally, the results and algorithms are extended to the case with side information. The rest of this paper is organized as follows. In Section 2, we review the current literature with references both to information–theoretic results and to key findings regarding brute force attacks on passwords. In Section 3, we formally define the problem and our objectives. Section 4 describes the results for memoryless sources, while Section 5 describes the results for sources with memory. Section 6 concludes the paper. ### 2 Related Work The first information–theoretic study on guesswork was carried out by Massey [28]. Arikan [2] showed, among other things, that the exponential rate of the number of guesses required for memoryless sources is given by the R`enyi entropy of order [1] 2 [. Guesswork under a distor-] tion constraint was studied by Arikan and Merhav [29], who also derived a guessing strategy for discrete memoryless sources (DMS’s), which is universally asymptotically optimal, both in the unknown memoryless source and the moment order of the guesswork being analyzed. Guesswork for Markov processes was studied by Malone and Sullivan [30], and extended to a large class of stationary measures by Pfister and Sullivan [3]. In [31], Hanawal and Sundaresan proposed a large deviations approach. They derived the guesswork exponent for sources satisfying a large deviations principle (LDP), and thereby generalized the re sults in [2] and [30]. In [16], again via large deviations, Christiansen et al. proposed an approximation to the distribution of the guesswork. In [32], Sundaresan considered guess ing under source uncertainty. The redundancy as a function of the radius of the family of possible distributions was defined and quantified in a few cases. For the special class of discrete memoryless sources, as already suggested in [29], this redundancy tends to zero as ----- the length of the sequence grows without bound. In [33], Christiansen et al. considered a multi-user case, where an adversary (inquisitor) has to guess U out of V strings, chosen at random from some string-source µ[n]. Later, Beirami et al. [34] further defined the inscrutability S[n](U, V, µ[n]) of a string-source, the inscrutability rate as the exponential rate of S[n](U, V, µ[n]), and gave upper and lower bounds on this rate by identifying the appropriate string-source distributions. They also showed that ordering strings by their type-size in ascending order is a universal guessing strategy. Note, however, that both [33, 34] considered a single attacker, with the ability to create a list of strings and guess one string after the other. Following Weinberger et al. [35], ordering strings by the size of their type-class before assigning them codewords in a fixed-to-variable source coding scheme was also found useful by Kosut and Sankar in [6] to minimize the third order term of the minimal number of bits required in lossless source coding (the first being the entropy, while the second is the dispersion). A geometric approach to guesswork was proposed by Beirami et al. in [36], showing that indeed the dominating type in guesswork (the position of a given string in the list) is the largest among all types whose elements are more likely than the given string. Here we show that a similar ordering is also beneficial for universal, decentralized guessing, though the sequences are not ordered in practice, and merely assigned probabilities to be guessed based on their type or LZ complexity. Guesswork over a binary erasure channel was studied by Christiansen et al. in [15]. While the underlying sequence to be guessed was assumed i.i.d., the results therein apply to channels with memory as well (yet satisfying an LDP). Interestingly, it was shown that the guesswork exponent is higher than the noiseless exponent times the average fraction of erased symbols, and one pays a non-negligible toll for the randomness in the erasures pattern. In [18], Salamatian et al. considered multi-agent guesswork with side information. The effect of synchronizing the side information among the agents was discussed, and its affect on the exponent was quantified. Multi-agent guesswork was then also studied in [19], this time devising a randomized guessing strategy, which can be used by asynchronous agents. The strategy in [19], however, is hard to implement in practice, as it depends on both the source distribution and the moment of the guesswork considered, and requires computing ----- an exponential sum.[4] Note that besides the standard application of guessing a password for a certain service, while knowing a password of the same user to another service, guessing with side information may also be applicable when breaking lists of hashed honeywords [38, 39]. In this scenario, an attacker is faced with a list of hashed sweatwords, where one is the hash of the true password while the rest are hashes of decoy honeywords, created with strong correlation to the real password. If one is broken, using it as side information can significantly reduce the time required to break the others. Furthermore, guessing with side information is also related the problem of guessing using hints [40]. In this scenario, a legitimate decoder should be able to guess a password (alternatively, a task to be carried out) using several hints, in the sense of having a low expected conditional guesswork, yet an eavesdropper knowing only a subset of the hints, should need a large number of guesses. In that case, the expected conditional guesswork generalizes secret sharing schemes by quantifying the amount of work Bob and Eve have to do. From a more practical viewpoint, trying to create passwords based on real data, Weir et al. [41] suggested a context-free grammar to create passwords at a descending order of probabilities, where the grammar rules as well as the probabilities of the generalized letters (sequences of English letters, sequences of digits or sequences of special characters) were learned based on a given training set. In [8], Dell’Amico et al. evaluated experimentally the probability of guessing passwords using dictionary-based, grammar-free and Markov chain strategies, using existing data sets of passwords for validation. Not only was it clear that complex guessing strategies, which take into account the memory, perform better, but moreover, the authors stress out the need to fine-tune memory parameters (e.g., the length of sub-strings tested), strengthening the necessity for a universal, parameter-free guessing strategy. In [11] Bonneau also implicitly mentions the problems coping with passwords from an unknown distribution, or an unknown mixture of several known distributions. 4An algorithm which produces a Markov chain, whose distribution converges to this sum was suggested in the context of randomized guessing tools in [37], yet only asymptotically and only for fixed, known distributions. ----- ### 3 Notation Conventions, Problem Statement and Objectives #### 3.1 Notation Conventions Throughout the paper, random variables will be denoted by capital letters, specific values they may take will be denoted by the corresponding lower case letters, and their alphabets will be denoted by calligraphic letters. Random vectors and their realizations will be de noted, respectively, by capital letters and the corresponding lower case letters, both in the bold face font. Their alphabets will be superscripted by their dimensions. For example, the random vector X = (X1, . . ., Xn), (n – positive integer) may take a specific vector value x = (x1, . . ., xn) in X [n], the n–th order Cartesian power of X, which is the alphabet of each component of this vector. Sources and channels will be denoted by the letter P or Q with or without some subscripts. When there is no room for ambiguity, these sub scripts will be omitted. The expectation operator will be denoted by E . The entropy {·} of a generic distribution Q on X will be denoted by HQ(X) where X designates a random variable drawn by Q. For two positive sequences an and bn, the notation an = bn will stand for equality in the exponential scale, that is, limn→∞ n[1] [log][ a]bn[n] [= 0. Similarly,][ a][n] ≤ bn means that lim supn→∞ n1 [log][ a]bn[n] [≤] [0, and so on. When both sequences depend on a vector,] x ∈X [n], namely, an = an(x) and bn = bn(x), the notation an(x) =· bn(x) means that the asymptotic convergence is uniform, namely, limn→∞ maxx∈X [n] | n[1] [log][ a]bn[n]([(]x[x])[)] [|][ = 0. Likewise,] an(x) ≤ bn(x) means lim supn→∞ maxx∈X [n] n[1] [log][ a]bn[n]([(]x[x])[)] [≤] [0, and so on.] The empirical distribution of a sequence x ∈X [n], which will be denoted by P[ˆ]x, is the vector of relative frequencies P[ˆ]x(x) of each symbol x ∈X in x. The type class of x ∈X [n], denoted T (x), is the set of all vectors x[′] with P[ˆ]x′ = P[ˆ]x. Information measures associated with empirical distributions will be denoted with ‘hats’ and will be subscripted by the sequences from which they are induced. For example, the entropy associated with Pˆx, which is the empirical entropy of x, will be denoted by ˆHx(X). Similar conventions will apply to the joint empirical distribution, the joint type class, the conditional empirical distributions and the conditional type classes associated with pairs of sequences of length n. Accordingly, P[ˆ]xy would be the joint empirical distribution of (x, y) = {(xi, yi)}i[n]=1[,][ T][ (][x][,][ y][)] or T ( P[ˆ]xy) will denote the joint type class of (x, y), T (x|y) will stand for the conditional type class of x given y, H[ˆ]xy(X|Y ) will be the empirical conditional entropy, and so on. In Section IV, the broader notion of a type class, which applies beyond the memoryless ----- case, will be adopted: the type class of x w.r.t. a given class of sources, will be defined P as (x) = � x[′] : P (x[′]) = P (x) . (1) T { } P ∈P Obviously, the various type classes, {T (x)}x∈X [n], are equivalence classes, and therefore, form a partition of [n]. Of course, when is the class of memoryless sources over, this X P X definition of (x) is equivalent to the earlier one, provided in the previous paragraph. T #### 3.2 Problem Statement and Objectives In this paper, we focus on the guessing problem that is defined as follows. Alice selects a secret random n–vector X, drawn from a finite alphabet source P . Bob, which is unaware of the realization of X, submits a sequence of guesses in the form of yes/no queries: “Is X = x1?” “Is X = x2?”, and so on, until receiving an affirmative answer. A guessing list, Gn, is an ordered list of all members of X [n], that is, G = {x1, x2, . . ., xM }, M = |X|[n], and it is associated with a guessing function, G(x), which is the function that maps [n] onto X {1, 2, . . ., M } by assigning to each x ∈X [n] the integer k for which xk = x, namely, the k–th element of Gn. In other words, G(x) is the number of guesses required until success, using Gn, when X = x. The guessing problem is about devising a guessing list Gn that minimizes a certain moment of G(X), namely, E G[ρ](X), where ρ > 0 is a given positive real (not necessarily { } a natural number). Clearly, when the source P is known and ρ is arbitrary, the optimal guessing list orders the members of [n] in the order of non–increasing probabilities. When P X is unknown, but known to belong to a given parametric class, like the class of memoryless P sources, or the class of finite–state sources with a given number of states, we are interested in a universal guessing list, which is asymptotically optimal in the sense of minimizing the guessing exponent, namely, achieving log E G[ρ](X) { } E(ρ) = lim sup, (2) n→∞ [min]Gn n uniformly for all sources in and all positive real values of ρ. P Motivated by applications of distributed, asynchronous guessing by several agents (see Introduction), we will also be interested in randomized guessing schemes, which have the advantages of: (i) relaxing the need to consume large volumes of memory (compared to deterministic guessing which needs the storage of the guessing list Gn) and (ii) dropping ----- the need for synchronization among the various guessing agents (see [19]). In randomized guessing, the guesser sequentially submits a sequence of random guesses, each one dis tributed independently according to a certain probability distribution P[˜](x). We would like the distribution P[˜] to be universally asymptotically optimal in the sense of achieving (on the average) the optimal guessing exponent, while being independent of the unknown source P and independent of ρ. Another desirable feature of the random guessing distribution P[˜] is that it would be easy to implement in practice. This is especially important when n is large, as it is not trivial to implement a general distribution over [n] in the absence of any X structure to this distribution. We begin our discussion from the case where the class of sources,, is the class of P memoryless sources over a finite alphabet of size α. In this case, some of the results X we will mention are already known, but it will be helpful, as a preparatory step, before we address the more interesting and challenging case, where is the class of all non–unifilar, P finite–state sources, a.k.a. hidden Markov sources (over the same finite alphabet ), where X even the number of states is unknown to the guesser, let alone, the parameters of the source for a given number of states. In both cases, we also extend the study to the case where the guesser is equipped with side information (SI) Y, correlated to the vector X to be guessed. ### 4 Guessing for Memoryless Sources #### 4.1 Background Following [28], Arikan [2] has established some important bounds associated with guessing moments with relation the R´enyi entropy, with and without side information, where the main application he had in mind was sequential decoding. Some of Arikan’s results set the stage for guessing n–vectors emitted from memoryless sources. Some of these results were later extended to the case of lossy guessing[5] [29] with a certain emphasis on universality issues. In particular, narrowing down the main result of [29] to the case of lossless guessing considered here, it was shown that the best achievable guessing exponent, E(ρ), is given by the following single–letter expression for a given memoryless source P : E(ρ) = max (3) Q [[][ρH][Q][(][X][)][ −] [D][(][Q][∥][P] [)] =][ ρH] [1][/][(1+][ρ][)][(][X][)][,] 5Here, by “lossy guessing” we mean that a guess is considered successful if its distance (in the sense of a given distortion function) from the underlying source vector, does not exceed a given distortion level. ----- where Q is an auxiliary distribution over to be optimized, and H [α](X) designates the X R´enyi entropy of order α, 1 H [α](X) = 1 α [ln] − �� P [α](x)� x∈X , (4) which is asymptotically achieved using a universal deterministic guessing list, Gn, that orders the members of [n] according to a non–decreasing order of their empirical entropies, X namely, Hˆx1(X) ≤ Hˆx2(X) ≤ . . . ≤ HˆxM (X). (5) In the presence of correlated side information Y, generated from X by a discrete memoryless channel (DMC), the above findings continue to hold, with the modifications that: (i) HQ(X) is replaced by HQ(X|Y ), (ii) D(Q∥P ) is understood to be the divergence between the two joint distributions of the pair (X, Y ) (which in turn implies that H [1][/][(1+][ρ][)](X) is replaced by the corresponding conditional R´enyi entropy of X given Y ), and (iii) H[ˆ]xk (X) is replaced by H[ˆ]xky(X|Y ), k = 1, 2, . . ., M . #### 4.2 Randomized Guessing and its Efficient Implementation For universal randomized guessing, we consider the following guessing distribution 2[−][n][ ˆ]Hx(X) P˜(x) = (6) �x[′][ 2][−][n][ ˆ]Hx′ (X) [.] We then have the following result: Theorem 1 Randomized guessing according to eq. (6) achieves the optimal guessing expo nent (3). Proof. We begin from the following lemma, whose proof is deferred to the appendix. Lemma 1 For given a 0 and ρ > 0, ≥ ∞ � k[ρ](1 e[−][na])[k][−][1] e[(1+][ρ][)][na]. (7) − ≤ k=1 Denoting by Pn the set of probability distributions over X with rational letter proba bilities of denominator n (empirical distributions), we observe that since 1 � 2[−][n][ ˆ]Hx(X) ≤ x ----- = � max Q∈P [Q][(][x][)] x = � max Q∈Pn [Q][(][x][)] x � ≤ x � Q(x) Q∈Pn = � Q∈Pn = |Pn| � Q(x) x it follows that (n + 1)[|X|−][1], (8) ≤ P˜(x) = 2[·] [−][n][ ˆ]Hx(X). (9) Given that X = x, the ρ–th moment of the number of guesses under P[˜] is given by ∞ � k[ρ][1 P (x)][k][−][1][ ˜]P (x) − [˜] k=1 ∞ = P˜(x) � k[ρ][1 P (x)][k][−][1] − [˜] k=1 ∞ =· 2[−][n][ ˆ]Hx(X) � k[ρ][1 2[−][n][ ˆ]Hx(X)]k−1 − k=1 2[−][n][ ˆ]Hx(X)2n(1+ρ) H[ˆ]x(X) ≤ = 2[nρ]H[ ˆ]x(X), (10) where in the inequality, we have used Lemma 1 with the assignment a = H[ˆ]x(X) ln 2. Taking the expectation of 2[nρ]H[ ˆ]x(X) w.r.t. P (x), using the method of types [42], one easily obtains (see also [29]) the exponential order of 2[nE][(][ρ][)], with E(ρ) as defined in (3). This completes the proof of Theorem 1. ✷ Remark: It is easy to see that the random guessing scheme has an additional important feature: not only the expectation of G[ρ](x) (w.r.t. the randomness of the guesses) has the optimal exponential order of 2[nρ]H[ ˆ]x(X) for each and every x, but moreover, the probability that G(x) would exceed 2[n][[ ˆ]Hx(X)+ǫ] decays double–exponentially rapidly for every ǫ > 0. This follows from the following simple chain of inequalities: Hx(X)+ǫ] Pr �G(x) 2[n][[ ˆ]Hx(X)+ǫ][�] =· �1 2[−][n][ ˆ]Hx(X)[�][2][n][[ ˆ] ≥ − = exp �2[n][[ ˆ]Hx(X)+ǫ] ln �1 2[−][n][ ˆ]Hx(X)[��] − ----- exp � 2[n][[ ˆ][H][x][(][X][)+][ǫ][]] 2[−][n][ ˆ]Hx(X)[�] ≤ − = exp 2[nǫ] . (11) {− } A similar comment will apply also to the random guessing scheme of Section 5. The random guessing distribution (6) is asymptotically equivalent (in the exponential scale) to a class of mixtures of all memoryless sources over, having the form X � M (x) = µ(Q) Q(x)dQ (12) S where µ( ) is a density defined on the simplex of all distributions on, and where it is - X assumed that µ( ) is bounded away from zero and from infinity, and that it is independent of n. As mentioned in [43], one of the popular choices of µ( ) is the Dirichlet distribution, parametrized by λ > 0, where µ(Q) = [Γ(][λ][|X|][)] Γ[|X|](λ) [·] λ−1 �� Q(x)�, (13) x∈X ∞ - � Γ(s) = x[s][−][1]e[−][x]dx, (14) 0 and we remind that for a positive integer n, Γ(n) = (n 1)! (15) − � 1 � √ Γ = π (16) 2 � Γ n + [1] 2 � √ � = π 1 n - − [1] 2 2 2 [·][ 3] [· · ·] � , n 1. (17) ≥ For example, with the choice λ = 1/2, the mixture becomes �x∈X [Γ(][n][ ˆ][P]x[(][x][) + 1][/][2)] . (18) π[|X|][/][2]Γ(n + /2) |X| � |X| M (x) = Γ 2 � This mixture distribution can be implemented sequentially, as M (x) = n−1 � M (xt+1|x[t]), (19) t=0 where M (xt+1|x[t]) = [M] [(][x][t][+1][)] = [t][ ˆ][P][x][t][(][x][t][+1][) + 1][/][2], (20) M (x[t]) t + /2 |X| where P[ˆ]xt (x) is the relative frequency of x ∈X in x[t] = (x1, . . ., xt). So the sequential implementation is rather simple: draw the first symbol, X1, according to the uniform ----- distribution. Then, for t = 1, 2, . . ., n − 1, draw the next symbol, Xt+1, according to the last equation, taking into account the relative frequencies of the various letters drawn so far. #### 4.3 Side Information All the above findings extend straightforwardly to the case of a guesser that is equipped with SI Y, correlated to the random vector X to be guessed, where it is assumed that (X, Y ) is a sequence of n independent copies of a pair of random variables (X, Y ) jointly distributed according to PXY . The only modification required is that the universal randomized guessing distribu Hxy (X|Y ) tion will now by proportional (and exponentially equivalent) to 2[−][n][ ˆ] instead of 2[−][n][ ˆ]Hx(X), and in the sequential implementation, the mixture and hence also the relative frequency counts will be applied to each SI letter y separately. Consequently, the ∈Y conditional distribution M (xt+1|x[t]) above would be replaced by M (xt+1|x[t], y[t]) = [t][ ˆ][P][x][t][y][t][(][x][t][+1][, y][t][+1][) + 1][/][2] . (21) tP[ˆ]yt (yt+1) + |X|/2 ### 5 Guessing for Finite–State Sources We now extend the scope to a much more general class of sources – the class of non–unifilar finite–state sources, namely, hidden Markov sources [44]. Specifically, we assume that X is drawn by a distribution P given by P (x) = � z n � P (xi, zi+1|zi), (22) i=1 where {xi} is the source sequence as before, whose elements take on values in a finite alphabet X of size α, and where {zi} is the underlying state sequence, whose elements take on values in a finite set of states, Z of size s, and where the initial state, z1, is assumed to be a fixed member of . The parameter set P (x, z[′] z), x, z, z[′] is unknown Z { | ∈X ∈Z} the guesser. In fact, even the number of states, s, is not known, and we seek a universal guessing strategy. ----- #### 5.1 Converse Theorem Let us parse x into c = c(x) distinct phrases, by using, for example, the incremental parsing procedure[6] of the Lempel–Ziv (LZ) algorithm [27] (see also [45, Subsection 13.4.2]). The following is a converse theorem concerning the best achievable guessing performance. Theorem 2 Given a finite–state source (22), any guessing function satisfies the following inequality: E{G[ρ](X)} ≥ 2[−][n][∆][n]E [exp2{ρc(X) log c(X)}], (23) where ∆n is a function of s, α and n, that tends to zero as n →∞ for fixed s and α. Proof. Without essential loss of generality, let ℓ divide n and consider the segmentation of x = (x1, . . ., xn) into n/ℓ non–overlapping sub–blocks, xi = (xiℓ+1, xiℓ+2, . . ., x(i+1)ℓ), i = 0, 1, . . ., n/ℓ − 1. Let z[ℓ] = (z1, zℓ+1, z2ℓ+1, . . ., zn+1) be the (diluted) state sequence pertaining to the boundaries between neighboring sub–blocks. Then, P (x, z[ℓ]) = n/ℓ−1 � P (xi, z(i+1)ℓ+1|ziℓ+1). (24) i=0 For a given z[ℓ], let (x z[ℓ]) be the set of all sequences x[′] that are obtained by permuting T | { } different sub–blocks that both begin at the same state and end at the same state. Owing to the product form of P (x, z[ℓ]), it is clear that P (x[′], z[ℓ]) = P (x, z[ℓ]) whenever x[′] (x z[ℓ]). ∈T | It was shown in [46, Eq. (47) and Appendix A] that |T (x|z[ℓ])| ≥ exp2{c(x) log c(x) − nδ(n, ℓ)}, (25) independently of z[ℓ], where δ(n, ℓ) tends to C/ℓ (C > 0 – constant) as n for fixed ℓ. →∞ Furthermore, by choosing ℓ = ℓn = [√]log n, we have that δ(n, ℓn) = O(1/[√]log n). We then have the following chain of inequalities: E G[ρ](X) = � { } z[ℓn] = � z[ℓn] = � z[ℓn] � P (x, z[ℓ][n]) � G[ρ](x) {T (x|z[ℓn] )} x[′]∈T (x|z[ℓn] ) � P (x, z[ℓ][n])G[ρ](x) x � {T (x|z[ℓn] )} � P (x, z[ℓ][n])G[ρ](x) x[′]∈T (x|z[ℓn] ) 6The incremental parsing procedure is a sequential procedure of parsing a sequence, such that each new parsed phrase is the shortest string that has not been obtained before as a phrase. ----- G[ρ](x) (x z[ℓ][n]) |T | | = � z[ℓn] � ≥ z[ℓn] � P (x, z[ℓ][n]) (x z[ℓ][n]) � - |T | | · {T (x|z[ℓn] )} x[′]∈T (x|z[ℓn] ) � P (x, z[ℓ][n]) (x z[ℓ][n]) - |T | | · [|T][ (][x][|][z][ℓ][n][)][|][ρ] 1 + ρ {T (x|z[ℓn] )} 1 ≥ 1 + ρ � z[ℓn] � P (x, z[ℓ][n]) · |T (x|z[ℓ][n])| · exp2{ρ[c(x) log c(x) − nδ(n, ℓn)]} {T (x|z[ℓn] )} 2[−][nδ][(][n,ℓ][n][)] = E [exp2{ρ · c(X) log c(X)}] 1 + ρ = 2[−][n][∆][n]E [exp2{ρ · c(X) log c(X)}], (26) where the first inequality follows from the following genie–aided argument: The inner–most summation in the fourth line of the above chain can be viewed as the guessing moment of a guesser that is informed that X falls within a given (x z[ℓ][n]). Since the distribution within T | (x z[ℓ][n]) is uniform, no matter what guessing strategy may be, T | |T (x|z[ℓn] )| � k[ρ] k=1 � x[′]∈T (x|z[ℓn] ) G[ρ](x) 1 ≥ (x z[ℓ][n]) (x z[ℓ][n]) |T | | |T | | 1 (x z[ℓ][n]) ≥ |T | |[ρ] (x z[ℓ][n]) |T | | |T (x|z[ℓn] )| � k=1 � k (x z[ℓ][n]) |T | | ρ � � 1 (x z[ℓ][n]) u[ρ]du ≥ |T | |[ρ] 0 (x z[ℓ][n]) |T | |[ρ] = . (27) 1 + ρ This completes the proof of Theorem 2. ✷ #### 5.2 Direct Theorem We now present a matching direct theorem, which asymptotically achieves the converse bound in the exponential scale. Theorem 3 Given a finite–state source (22), there exists a universal guessing list that satisfies the following inequality: E{G[ρ](X)} ≤ 2[n][∆]n[′] E [exp2{ρc(X) log c(X)}], (28) where ∆[′]n [is a function of][ s][,][ α][ and][ n][, that tends to zero as][ n][ →∞] [for fixed][ s][ and][ α][.] Proof. The proposed deterministic guessing list orders all members of [n] in non–decreasing X order of their Lempel–Ziv code–lengths [27, Theorem 2]. Denoting the LZ code–length of ----- x by LZ(x), we then have G(x) x[′] : LZ(x[′]) LZ(x) ≤ |{ ≤ }| LZ(x) = � x[′] : LZ(x[′]) = i |{ }| i=1 LZ(x) � 2[i] ≤ i=1 < 2[LZ][(][x][)+1] ≤ exp2{[c(x) + 1] log(2α[c(x) + 1]) + 1} = exp2{c(x) log c(x)}, (29) where the inequality x[′] : LZ(x[′]) = i 2[i] is due to the fact that the LZ code is |{ }| ≤ uniquely decipherable (UD) and the last inequality is from Theorem 2 of [27]. By raising this inequality to the power of ρ and taking the expectation of both sides, Theorem 3 is readily proved. An alternative, randomized guessing strategy pertains to independent random guesses according to the following universal distribution, 2[−][LZ][(][x][)] P˜(x) = (30) �x[′][ 2][−][LZ][(][x][′][)] [.] Since the LZ code is UD, it satisfies the Kraft inequality, and so, the denominator cannot be larger than 1, which means that P˜(x) ≥ 2[−][LZ][(][x][)] ≥ exp2{−[c(x) + 1] log(2α[c(x) + 1])}. (31) Similarly as in (10), applying Lemma 1 to (30) (or to (31)), this time with a = [ln 2] n [[][c][(][x][) +] 1] log(2α[c(x) + 1]), we obtain that the ρ–th moment of G(X) given that X = x is up per bounded by an expression of the exponential of exp2{ρ[c(x) + 1] log(2α[c(x) + 1])} = exp2{ρc(x) log c(x)}, and then upon taking the expectation w.r.t. the randomness of X, we readily obtain the achievability result once again. ✷ #### 5.3 Algorithms for Sampling From the Universal Guessing Distribution Similarly as in Section 4, we are interested in efficient algorithms for sampling from the universal distribution, (30). In fact, it is enough to have an efficient implementation of an algorithm that efficiently samples from a distribution P[˜] that satisfies P[˜](x) 2[−][c][(][x][) log][ c][(][x][)]. ≥ We propose two different algorithms, the first is inspired by the predictive point of view ----- associated with LZ parsing [47], [48], and the second one is based on the simple idea of feeding the LZ decoder with purely random bits. The latter algorithm turns out to lend itself more easily to generalization for the case of guessing in the presence of SI. Both algo rithms are described in terms of walks on a growing tree, but the difference is that in the first algorithm, the tree is constructed in the domain of the source sequences, whereas in the second algorithm, the tree is in the domain of the compressed bit-stream. First algorithm. As said, the idea is in the spirit of the predictive probability assignment mechanism proposed in [47] and [48, Sect. V], but here, instead of using the incremental parsing mechanism for prediction, we use it for random selection. As mentioned before, the algorithm is described as a process generated by a repeated walk on a growing tree, beginning, each time, from the root and ending at one of the leaves. Consider a tree which is initially composed of a root connected to α leaves, each one corresponding to one alphabet letter, x . We always assign to each leaf a weight ∈X of 1 and to each internal node – the sum of weights of its immediate off-springs, and so, the initial weight of the root is α. We begin by drawing the first symbol, X1, such that the probability of X1 = x is given by the weight of x (which is 1) divided by the weight of the current node, which is the root (i.e., a weight of α, as said). In other words, we randomly select X1 according to the uniform distribution over X . The leaf corresponding to the outcome of X1, call it x1, will now become an internal node by adding to the tree its α off-springs, thus growing the tree to have 2α 1 leaves. Each one of the leaves of − the extended tree has now weight 1, and then, the weight of the their common ancestor (formerly, the leaf of x1), becomes the sum of their weights, namely, α, and similarly, the weights of all ancestors of x1, all the way up to the root, are now sequentially updated to become the sum of the weights of their immediate off-springs. We now start again from the root of the tree to randomly draw the next symbol, X2, such that the probability that X2 = x, is given by the weight of the node x divided by the weight of the current node, which is again the root, and then we move from the root to its corresponding off-spring pertaining to X2, that was just randomly drawn. If we have reached a leaf, then again this leaf gives birth to α new off-springs, each assigned with weight 1, then all corresponding weights are updated as described before, and finally, we move back to the root, etc. If we are still in an internal node, then again, we draw the next ----- symbol according to the ratio between the weight of the node pertaining to the next symbol and the weight of the current node, and so on. The process continues until n symbols, X1, X2, . . ., Xn have been generated. Note that every time we restart from the root and move along the tree until we reach a leaf, we generate a new LZ phrase that has not been obtained before. Let c(x) be the number of phrases generated. Along each path from the root to a leaf, we implement a telescopic product of conditional probabilities, where the numerator pertaining to the last conditional probability is the weight of the leaf, which is 1, and the denominator of the first probability is the total number of leaves after i rounds, which is α + i(α 1) (because after − every birth of a new generation of leaves, the total number of leaves is increased by α 1). All − other numerators and denominators of the conditional probabilities along the path cancel each other telescopically. The result is that the induced probability distribution along the various leaves is uniform. Precisely, after i phrases have been generated, the probability of each leaf is exactly 1/[α + i(α 1)]. Therefore, − c(x)−1 1 P (x) = � α + i(α 1) i=0 − c(x)−1 1 � ≥ α + [c(x) 1](α 1) i=0 − − 1 = α + [c(x) 1](α 1) { − − }[c][(][x][)] = 2[−][c][(][x][) log][{][[][c][(][x][)][−][1](][α][−][1)+][α][}], (32) which is of the exponential order of 2[−][c][(][x][) log][ c][(][x][)]. Second algorithm. As said, the second method for efficiently generating random guesses according to the LZ distribution is based on the simple idea of feeding purely random bits into the LZ decoder until a decoded sequence of length n is obtained. To describe it, we refer to the coding scheme proposed in [27, Theorem 2], but with a slight modification. Recall that according to this coding scheme, for the i–th parsed phrase, x[n]n[j]j−1+1[, one encodes] two integers: the index 0 π(j) j 1 of the matching past phrase and the index the ≤ ≤ − additional source symbol, 0 ≤ IA(xnj ) ≤ α − 1. These two integers are mapped together bijectively into one integer, I(x[n]n[j]j−1+1[) =][ π][(][j][)][ ·][ α][ +][ I][A][(][x][n]j [), which takes on values in the] set {0, 1, 2, . . ., jα − 1}, and so, according to [27], it can be encoded using Lj = ⌈log(jα)⌉ ----- bits. Here, instead, we will encode I(x[n]n[j]j−1+1[) with a tiny modification that will make the] encoding equivalent to a walk on a complete binary tree [7] from the root to a leaf. Considering the fact that (by definition of Lj), 2[L][j] [−][1] < jα ≤ 2[L][j], we first construct a full binary tree with 2[L][j] [−][1] leaves at depth Lj − 1, and then convert jα − 2[L][j] [−][1] leaves to internal nodes by generating their off-springs. The resulting complete binary tree will then have exactly jα leaves, some of them at depth Lj − 1 and some - at depth Lj. Each leaf of this tree will now correspond to one value of I(x[n]n[j]j−1+1[), and hence to a certain decoded phrase. Let] Lˆj denote the length of the codeword for I(x[n]n[j]j−1+1[). Obviously, either ˆ][L][j][ =][ L][j][ −] [1 or] Lˆj = Lj. Consider now what happens if we feed the decoder of this encoder by a sequence of purely random bits (generated by a binary symmetric source): every leaf at depth L[ˆ]j will be obtained with probability 2[−][L][ˆ][j], and since the tree is complete, these probabilities sum up to unity. The probability of obtaining x at the decoder output is, therefore, equal to the probability of the sequence of bits pertaining to its compressed form, namely, P˜(x) = c(x)+1 � 2[−][L][ˆ][j] j=1       c(x)+1 � Lˆj j=1 c(x)+1 � Lj j=1 c(x)+1  � log(2jα) j=1  = exp2 ≥ exp2 ≥ exp2   [−]   [−]   [−] ≥ exp2 {−[c(x) + 1] log[2c(x)α]} = exp2{−[c(x) + 1] log c(x) − [c(x) + 1] log(2α)}, (33) which is again of the exponential order of 2[−][c][(][x][) log][ c][(][x][)]. #### 5.4 Side Information As we have done at the end of Section 4, here too, we describe how our results extend to the case where the guesser is equipped with SI. The parts that extend straightforwardly will be described briefly, whereas the parts whose extension is non–trivial will be more detailed. 7By “complete binary tree”, we mean a binary tree where each node is either a leaf or has two off-springs. The reason for the need of a complete binary tree is that for the algorithm to be valid, every possible sequence of randomly chosen bits must be a legitimate compressed bit-stream so that it would be decodable by the LZ decoder. ----- Consider the pair process {(Xt, Yt)}, jointly distributed according to a hidden Markov model, P (x, y) = � z n � P (xt, yt, zt+1|zt), (34) t=1 where, as before, zt is the state at time t, taking on values in a finite set of states Z of cardinality s. Here, our objective is to guess x when y is available to the guesser as SI. Most of our earlier results extend quite easily to this case. Basically, the only modification needed is to replace the LZ complexity of x by the conditional LZ complexity of x given y, which is defined as in [49] and [50]. In particular, consider the joint parsing of the sequence of pairs, {(x1, y1), (x2, y2), . . ., (xn, yn)}, let c(x, y) denote the number of phrases, c(y) – the number of distinct y-phrases, y(ℓ) – the ℓ-th distinct y-phrase, 1 ℓ c(y), and finally, let ≤ ≤ cℓ(x|y) denote the number of times y(ℓ) appears as a phrase, or, equivalently, the number of distinct x-phrases that appear jointly with y(ℓ), so that [�]ℓ[c]=1[(][y][)] [c][ℓ][(][x][|][y][) =][ c][(][x][,][ y][). Then,] we define u(x y) = | c(y) � cℓ(x|y) log cℓ(x|y). (35) ℓ=1 For the converse theorem (lower bound), the proof is the same as the proof of Theorem 2, except that here, we need a lower bound on the size of a “conditional type” of x given y. This lower bound turns to be of the exponential order of 2[u][(][x][|][y][)], as can be seen in [51, Lemma 1]. Thus, the lower bound on the guessing moment is of the exponential order of E[exp2{ρu(X|Y )}]. For the direct theorem (upper bound), we can either create a deterministic guessing list by ordering the members of [n] according to increasing order of their conditional LZ code– X length function values, LZ(x y) u(x y), [49, p. 2617], [50, page 460, proof of Lemma 2], | ≈ | or randomly draw guesses according to 2[−][LZ][(][x][|][y][)] P˜(x y) = (36) | �x[′][ 2][−][LZ][(][x][′][|][y][)][ .] Following Subsection 5.C, we wish to have an efficient algorithm for sampling from the dis tribution (36), or, more generally, for implementing a conditional distribution that satisfies P˜(x y) - 2[−][LZ][(][x][|][y][)] = 2· −u(x|y). | ≥ While we have not been able to find an extension of the first algorithm of Subsection 5.C to the case of SI, the second algorithm therein turns out to lend itself fairly easily to such an ----- extension. Once again, generally speaking, the idea is to feed a sequence of purely random bits as inputs to the decoder pertaining to the conditional LZ decoder, equipped with y as SI, and wait until exactly n symbols, x1, . . ., xn, have been obtained at the output of the decoder. We need, however, a few slight modifications in conditional LZ code, in order to ensure that any sequence of randomly drawn bits would be legitimate as the output of the encoder, and hence be also decodable by the decoder. Once again, to this end, we must use complete binary trees for the prefix codes for the various components of the conditional LZ code. As can be seen in [49], [50], the conditional LZ compression algorithm sequentially encodes x phrase by phrase, where the code for each phrase consists of three parts: 1. A code for the length of the phrase, L[y(ℓ)]. 2. A code for the location of the matching x–phrase among all previous phrases with the same y–phrase. 3. A code for the index of the last symbol of the x–phrase among all members of . X Parts 2 and 3 are similar to those of the ordinary LZ algorithm and they in fact can even be united, as described before, into a single code for both indices (although this is not necessary). Part 1 requires a code for the integers, which can be implemented by the Elias code, as described in [49]. However, for the sake of conceptual simplicity of describing the required complete binary tree, consider the following alternative option. Define the following distribution on the natural numbers, 6 Q(i) = i = 1, 2, 3, . . . (37) π[2]i[2][,] and construct a prefix tree for the corresponding Shannon code, whose length function is given by (i) = log Q(i) . (38) L ⌈− ⌉ Next prune the tree by eliminating all leaves that correspond to values of i = L[y(ℓ)] that cannot be obtained at the current phrase: the length L[y(ℓ)] cannot be larger than the maximum possible phrase length and cannot correspond to a string that has not been obtained as a y–phrase before.[8] Finally, shorten the tree by eliminating branches that 8This is doable since both the encoder and the decoder have this information at the beginning of the current phrase. ----- emanate from any node that has one off-spring only. At the end of this process, we have a complete binary tree where the resulting code length for every possible value of L[y(ℓ)] cannot be larger than its original value (38). The probability of obtaining a given x at the output of the above–described conditional LZ decoder is equal to the probability of randomly selecting the bit-stream that generates x (in the presence of y as SI), as the response to this bit-stream. Thus, cℓ(x|y) � exp2{−⌈log(jα)⌉−L(L[y(ℓ)])} j=1 P˜(x y) | ≥ c(y) � ℓ=1 cℓ(x|y) � j=1 c(y) � ℓ=1 ≥ exp2 ≥ exp2   [−]   [−] c(y) cℓ(x|y) � �[] � � log(2jα) + 2 log L[y(ℓ)] + log [π][2]  6 [+ 1] ℓ=1 j=1  c(y) c(y) � cℓ(x|y) log[2αcℓ(x|y)] − 2 � cℓ(x|y) log L[y(ℓ)]− ℓ=1 ℓ=1 �� log [π][2] 6 [+ 1] c(x, y) � = exp2{−u(x|y)}, (39) where the last step follows from the observation [50, p. 460] that c(y) c(y) � cℓ(x|y) log L[y(ℓ)] = c(x, y) � cℓ(x|y) c(x, y) [log][ L][[][y][(][ℓ][)]] ℓ=1 ℓ=1  c(y) � ℓ=1 [c][ℓ][(][x][|][y][)][L][[][y][(][ℓ][)]] c(x, y) log ≤  c(x, y)   n = c(x, y) log (40) c(x, y) [,] and the fact that c(x, y) cannot be larger than O(n/ log n) [27]. ### 6 Conclusion In this work, we studied the guesswork problem under a very general setup of unknown source distribution and decentralized operation. Specifically, we designed and analyzed guessing strategies which do not require the source distribution, the exact guesswork mo ment to be optimized, or any synchronization between the guesses, yet achieve the optimal guesswork exponent as if all this information was known and full synchronization was pos sible. Furthermore, we designed efficient algorithms in order to sample guesses from the universal distributions suggested. We believe such sampling methods may be interesting on their own, and find applications outside the guesswork regime. ----- ### Appendix Proof of Lemma 1. We denote S = ∞ � k[ρ](1 e[−][na])[k][−][1]. (A.1) − k=1 For a given, arbitrarily small ǫ > 0, we first decompose S as follows. ∞ � k[ρ](1 e[−][na])[k][−][1][ △]= A + B. (A.2) − k=e[n][(][a][+][ǫ][)]+1 S = e[n][(][a][+][ǫ][)] � k[ρ](1 e[−][na])[k][−][1] + − k=1 Now, A ≤ e[n][(][a][+][ǫ][)] � e[n][(][a][+][ǫ][)][ρ](1 e[−][na])[k][−][1] − k=1 e[n][(][a][+][ǫ][)] = e[n][(][a][+][ǫ][)][ρ] � (1 e[−][na])[k][−][1] − k=1 ∞ e[n][(][a][+][ǫ][)][ρ] �(1 e[−][na])[k][−][1] ≤ − k=1 1 = e[n][(][a][+][ǫ][)][ρ] 1 (1 e[−][na]) − − = e[na] e[n][(][a][+][ǫ][)][ρ] e[n][(1+][ρ][)(][a][+][ǫ][)]. (A.3) ≤ It remains to show then that B has a negligible contribution for large enough n. Indeed, we next show that B decays double–exponentially rapidly in n for every ǫ > 0. B = = = ≤ = ∞ � k[ρ](1 e[−][na])[k][−][1] − k=e[n][(][a][+][ǫ][)]+1 ∞ � exp (k 1) ln(1 e[−][na]) + ρ ln k { − − } k=e[n][(][a][+][ǫ][)]+1 ∞ � exp k ln(1 e[−][na]) + ρ ln(k + 1) { − } k=e[n][(][a][+][ǫ][)] ∞ � exp k e[−][na] + ρ ln(k + 1) {− - } k=e[n][(][a][+][ǫ][)] ∞ � � �� � exp k e[−][na] (A.4) − - − [ρ][ ln(][k][ + 1)] k k=e[n][(][a][+][ǫ][)] Since {[ln(k + 1)]/k}k≥1 is a monotonically decreasing sequence, then for all k ≥ e[n][(][a][+][ǫ][)], ρ ln(k + 1) = ρe[−][n][(][a][+][ǫ][)] ln[e[n][(][a][+][ǫ][)] + 1]. ≤ [ρ][ ln[][e][n][(][a][+][ǫ][)][ + 1]] k e[n][(][a][+][ǫ][)] ----- Thus, ∞ B � exp k (e[−][na] ρe[−][n][(][a][+][ǫ][)] ln[e[n][(][a][+][ǫ][)] + 1]) ≤ {− - − } k=e[n][(][a][+][ǫ][)] exp e[n][(][a][+][ǫ][)](e[−][na] ρe[−][n][(][a][+][ǫ][)] ln[e[n][(][a][+][ǫ][)] + 1]) {− − } = 1 exp (e[−][na] ρe[−][n][(][a][+][ǫ][)] ln[e[n][(][a][+][ǫ][)] + 1]) − {− − } exp (e[nǫ] ρ ln[e[n][(][a][+][ǫ][)] + 1]) {− − } = 1 exp (e[−][na] ρe[−][n][(][a][+][ǫ][)] ln[e[n][(][a][+][ǫ][)] + 1]) − {− − } [e[n][(][a][+][ǫ][)] + 1][ρ] exp e[nǫ] {− } = (A.5) 1 exp (e[−][na] ρe[−][n][(][a][+][ǫ][)] ln[e[n][(][a][+][ǫ][)] + 1]) − {− − } Now, for small x, we have 1 e[−][x] = x + O(x[2]), and so, the factor − [1 exp (e[−][na] ρe[−][n][(][a][+][ǫ][)] ln[e[n][(][a][+][ǫ][)] + 1]) ][−][1] − {− − } is of the exponential order of e[na], which does not affect the double exponential decay due to the term e[−][e][nǫ]. The proof of the lemma is completed by taking into account the arbitrariness of ǫ > 0 (in particular, one may let ǫ decay sufficiently slowly with n). ### References [1] J. M. Wozencraft, “Sequential decoding for reliable communication,” Ph.D. disserta tion, Research Laboratory of Electronics, Massachusetts Institute of Technology, May 1957. [2] E. Arikan, “An inequality on guessing and its application to sequential decoding,” IEEE Transactions on Information Theory, vol. 42, no. 1, pp. 99–105, January 1996. [3] C. E. Pfister and W. G. Sullivan, “R´enyi entropy, guesswork moments, and large de viations,” IEEE Transactions on Information Theory, vol. 50, no. 11, pp. 2794–2800, November 2004. [4] W. Szpankowski and S. Verd´u, “Minimum expected length of fixed-to-variable lossless compression without prefix constraints,” IEEE Transactions on Information Theory, vol. 57, no. 7, pp. 4017–4025, July 2011. [5] I. Kontoyiannis and S. Verd´u, “Optimal lossless data compression: Non-asymptotics and asymptotics,” IEEE Transactions on Information Theory, vol. 60, no. 2, pp. 777– 795, February 2014. ----- [6] O. Kosut and L. Sankar, “Asymptotics and non-asymptotics for universal fixed-to variable source coding,” IEEE Transactions on Information Theory, vol. 63, no. 6, pp. 3757–3772, June 2017. [7] M. Bishop and D. V. Klein, “Improving system security via proactive password check ing,” Computers & Security, vol. 14, no. 3, pp. 233–249, 1995. [8] M. D. Amico, P. Michiardi, and Y. Roudier, “Password strength: An empirical analy sis,” in Proceedings IEEE Infocom, March 2010, pp. 1–9. [9] P. G. Kelley, S. Komanduri, M. L. Mazurek, R. Shay, T. Vidas, L. Bauer, N. Christin, L. F. Cranor, and J. Lopez, “Guess again (and again and again): Measuring password strength by simulating password-cracking algorithms,” in 2012 IEEE Symposium on Security and Privacy, May 2012, pp. 523–537. [10] S. Komanduri, R. Shay, P. G. Kelley, M. L. Mazurek, L. Bauer, N. Christin, L. F. Cranor, and S. Egelman, “Of passwords and people: Measuring the effect of password composition policies,” in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, ser. CHI ’11. New York, NY, USA: ACM, 2011, pp. 2595–2604. [11] J. Bonneau, “The science of guessing: Analyzing an anonymized corpus of 70 million passwords,” in 2012 IEEE Symposium on Security and Privacy, May 2012, pp. 538–552. [12] J. O. Pliam, “On the incomparability of entropy and marginal guesswork in brute-force attacks,” in Progress in Cryptology —INDOCRYPT 2000, B. Roy and E. Okamoto, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2000, pp. 67–79. [13] P. Gauravaram, “Security analysis of salt——password hashes,” in 2012 International Conference on Advanced Computer Science Applications and Technologies (ACSAT), November 2012, pp. 25–30. [14] A. Rezaee, A. Beirami, A. Makhdoumi, M. M´edard, and K. Duffy, “Guesswork subject to a total entropy budget,” in 2017 55th Annual Allerton Conference on Communica tion, Control, and Computing (Allerton), October 2017, pp. 1008–1015. [15] M. M. Christiansen, K. R. Duffy, F. du Pin Calmon, and M. M´edard, “Guessing a password over a wireless channel (on the effect of noise non-uniformity),” in 2013 Asilomar Conference on Signals, Systems and Computers, November 2013, pp. 51–55. ----- [16] M. M. Christiansen and K. R. Duffy, “Guesswork, large deviations, and shannon en tropy,” IEEE Transactions Information Theory, vol. 59, no. 2, pp. 796–802, February 2013. [17] Y. Yona and S. Diggavi, “The effect of bias on the guesswork of hash functions,” in 2017 IEEE International Symposium on Information Theory (ISIT), June 2017, pp. 2248–2252. [18] S. Salamatian, A. Beirami, A. Cohen, and M. M´edard, “Centralized vs. decentral ized multi-agent guesswork,” in Information Theory (ISIT), 2017 IEEE International Symposium on. IEEE, 2017, pp. 2258–2262. [19] S. Salamatian, W. Huleihel, A. Beirami, A. Cohen, and M. M´edard, “Why botnets work: Distributed brute-force attacks need no synchronization,” in revision, IEEE Transactions on Information Forensics and Security, 2018. [20] T. A. Courtade and S. Verd´u, “Variable-length lossy compression and channel coding: Non-asymptotic converses via cumulant generating functions,” in 2014 IEEE Interna tional Symposium on Information Theory, June 2014, pp. 2499–2503. [21] D. Malone and K. Maher, “Investigating the distribution of password choices,” in Proceedings of the 21st international conference on World Wide Web. ACM, 2012, pp. 301–310. [22] J. Yan, A. Blackwell, R. Anderson, and A. Grant, “Password memorability and secu rity: empirical results,” IEEE Security Privacy, vol. 2, no. 5, pp. 25–31, September 2004. [23] D. Vishwakarma and C. E. V. Madhavan, “Efficient dictionary for salted password analysis,” in 2014 IEEE International Conference on Electronics, Computing and Com munication Technologies (CONECCT), January 2014, pp. 1–6. [24] R. Sundaresan, “Guessing under source uncertainty with side information,” in 2006 IEEE International Symposium on Information Theory, July 2006, pp. 2438–2440. [25] J. Owens and J. Matthews, “A study of passwords and methods used in brute-force SSH attacks,” in USENIX Workshop on Large-Scale Exploits and Emergent Threats (LEET), 2008. ----- [26] E. Tirado, B. Turpin, C. Beltz, P. Roshon, R. Judge, and K. Gagneja, “A new dis tributed brute-force password cracking technique,” in International Conference on Fu ture Network Systems and Security. Springer, 2018, pp. 117–127. [27] J. Ziv and A. Lempel, “Compression of individual sequences via variable-rate coding,” IEEE transactions on Information Theory, vol. 24, no. 5, pp. 530–536, September 1978. [28] J. L. Massey, “Guessing and entropy,” in Proceedings of IEEE International Symposium on Information Theory, 1994, p. 204. [29] E. Arikan and N. Merhav, “Guessing subject to distortion,” IEEE Transactions on Information Theory, vol. 44, no. 3, pp. 1041–1056, May 1998. [30] D. Malone and W. G. Sullivan, “Guesswork and entropy,” IEEE Transactions on In formation Theory, vol. 50, no. 3, pp. 525–526, March 2004. [31] M. K. Hanawal and R. Sundaresan, “Guessing revisited: A large deviations approach,” IEEE Transactions on Information Theory, vol. 57, no. 1, pp. 70–78, January 2011. [32] R. Sundaresan, “Guessing under source uncertainty,” IEEE Transactions on Informa tion Theory, vol. 53, no. 1, pp. 269–287, January 2007. [33] M. M. Christiansen, K. R. Duffy, F. du Pin Calmon, and M. M´edard, “Multi-user guess work and brute force security,” IEEE Transactions on Information Theory, vol. 61, no. 12, pp. 6876–6886, December 2015. [34] A. Beirami, R. Calderbank, K. Duffy, and M. M´edard, “Quantifying computational security subject to source constraints, guesswork and inscrutability,” in 2015 IEEE International Symposium on Information Theory (ISIT), June 2015, pp. 2757–2761. [35] M. J. Weinberger, J. Ziv, and A. Lempel, “On the optimal asymptotic performance of universal ordering and of discrimination of individual sequences,” IEEE Transactions on Information Theory, vol. 38, no. 2, pp. 380–385, March 1992. [36] A. Beirami, R. Calderbank, M. Christiansen, K. Duffy, A. Makhdoumi, and M. Mdard, “A geometric perspective on guesswork,” in 2005 53rd Annual Allerton Conference on Communication, Control, and Computing (Allerton), September 2015, pp. 941–948. ----- [37] M. K. Hanawal and R. Sundaresan, “Randomised attacks on passwords,” DRDO-IISc Programme on Advanced Research in Mathematical Engineering, 2010. [38] A. Juels and R. L. Rivest, “Honeywords: Making password-cracking detectable,” in Proceedings of the 2013 ACM SIGSAC conference on Computer & communications security. ACM, 2013, pp. 145–160. [39] D. Wang, H. Cheng, P. Wang, J. Yan, and X. Huang, “A security analysis of honey [words,” https://tinyurl.com/y87ffmny, NDSS, 2018.](https://tinyurl.com/y87ffmny) [40] A. Bracher, E. Hof, and A. Lapidoth, “Guessing attacks on distributed-storage sys [tems,” arXiv preprint arXiv:1701.01981, 2017.](http://arxiv.org/abs/1701.01981) [41] M. Weir, S. Aggarwal, B. d. Medeiros, and B. Glodek, “Password cracking using prob abilistic context-free grammars,” in 2009 30th IEEE Symposium on Security and Pri vacy, May 2009, pp. 391–405. [42] I. Csisz´ar and J. K¨orner, Information theory: coding theorems for discrete memoryless systems. Cambridge University Press, 2011. [43] R. Krichevsky and V. Trofimov, “The performance of universal encoding,” IEEE Trans actions on Information Theory, vol. 27, no. 2, pp. 199–207, March 1981. [44] Y. Ephraim and N. Merhav, “Hidden markov processes,” IEEE Transactions on infor mation theory, vol. 48, no. 6, pp. 1518–1569, June 2002. [45] T. M. Cover and J. A. Thomas, Elements of information theory, 2nd ed. New York, NY, USA: John Wiley & Sons, 2006. [46] N. Merhav, “Universal coding with minimum probability of codeword length overflow,” IEEE Transactions on Information Theory, vol. 37, no. 3, pp. 556–563, May 1991. [47] M. Feder, “Gambling using a finite state machine,” IEEE Transactions on Information Theory, vol. 37, no. 5, pp. 1459–1465, September 1991. [48] M. Feder, N. Merhav, and M. Gutman, “Universal prediction of individual sequences,” IEEE transactions on Information Theory, vol. 38, no. 4, pp. 1258–1270, July 1992. ----- [49] T. Uyematsu and S. Kuzuoka, “Conditional lempel-ziv complexity and its application to source coding theorem with side information,” IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, vol. 86, no. 10, pp. 2615–2617, 2003. [50] J. Ziv, “Universal decoding for finite-state channels,” IEEE Transactions on Informa tion Theory, vol. IT-31, no. 4, pp. 453–460, July 1985. [51] N. Merhav, “Universal detection of messages via finite-state channels,” IEEE Trans actions on Information Theory, vol. 46, no. 6, pp. 2242–2246, April 2000. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/TIT.2019.2920538?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/TIT.2019.2920538, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "http://arxiv.org/pdf/1811.04363" }
2,020
[ "JournalArticle" ]
true
2020-01-01T00:00:00
[]
20,643
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Physics", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00d0783da6191568a814381a0dc4db49262736e9
[ "Computer Science", "Medicine" ]
0.853884
On Global Quantum Communication Networking
00d0783da6191568a814381a0dc4db49262736e9
Entropy
[ { "authorId": "143857714", "name": "I. Djordjevic" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "http://www.mdpi.com/journal/entropy/", "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-155606", "https://www.mdpi.com/journal/entropy" ], "id": "8270cfe1-3713-4325-a7bd-c6a87eed889e", "issn": "1099-4300", "name": "Entropy", "type": "journal", "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-155606" }
Research in quantum communications networks (QCNs), where multiple users desire to generate or transmit common quantum-secured information, is still in its beginning stage. To solve for the problems of both discrete variable- and continuous variable-quantum key distribution (QKD) schemes in a simultaneous manner as well as to enable the next generation of quantum communication networking, in this Special Issue paper we describe a scenario where disconnected terrestrial QCNs are coupled through low Earth orbit (LEO) satellite quantum network forming heterogeneous satellite–terrestrial QCN. The proposed heterogeneous QCN is based on the cluster state approach and can be used for numerous applications, including: (i) to teleport arbitrary quantum states between any two nodes in the QCN; (ii) to enable the next generation of cyber security systems; (iii) to enable distributed quantum computing; and (iv) to enable the next generation of quantum sensing networks. The proposed QCNs will be robust against various channel impairments over heterogeneous links. Moreover, the proposed QCNs will provide an unprecedented security level for 5G+/6G wireless networks, Internet of Things (IoT), optical networks, and autonomous vehicles, to mention a few.
# entropy _Perspective_ ### On Global Quantum Communication Networking **Ivan B. Djordjevic** Department of Electrical and Computer Engineering, University of Arizona, Tucson, AZ 85721, USA; ivan@email.arizona.edu; Tel.: +1-520-626-5119 Received: 29 June 2020; Accepted: 28 July 2020; Published: 29 July 2020 [����������](https://www.mdpi.com/1099-4300/22/8/831?type=check_update&version=2) **�������** **Abstract: Research in quantum communications networks (QCNs), where multiple users desire to** generate or transmit common quantum-secured information, is still in its beginning stage. To solve for the problems of both discrete variable- and continuous variable-quantum key distribution (QKD) schemes in a simultaneous manner as well as to enable the next generation of quantum communication networking, in this Special Issue paper we describe a scenario where disconnected terrestrial QCNs are coupled through low Earth orbit (LEO) satellite quantum network forming heterogeneous satellite–terrestrial QCN. The proposed heterogeneous QCN is based on the cluster state approach and can be used for numerous applications, including: (i) to teleport arbitrary quantum states between any two nodes in the QCN; (ii) to enable the next generation of cyber security systems; (iii) to enable distributed quantum computing; and (iv) to enable the next generation of quantum sensing networks. The proposed QCNs will be robust against various channel impairments over heterogeneous links. Moreover, the proposed QCNs will provide an unprecedented security level for 5G+/6G wireless networks, Internet of Things (IoT), optical networks, and autonomous vehicles, to mention a few. **Keywords: quantum key distribution (QKD); discrete variable (DV)-QKD; continuous variable** (CV)-QKD; postquantum cryptography (PQC); quantum communications networks (QCNs) **1. Introduction** Quantum communication (QuCom) employs quantum information theory concepts, in particular the no-cloning theorem and the theorem of indistinguishability of arbitrary quantum states, to implement the distribution of keys with verifiable security, commonly referred to as quantum key distribution (QKD), where security is guaranteed by the fundamental laws of physics as opposed to unproven mathematical assumptions employed in computational security-based cryptography [1–3]. Despite the appealing features of QuComs, there are some fundamental and technical challenges that need to be addressed prior to its widespread application. For instance, both the rate and distance of QuCom are fundamentally limited by channel loss, which is specified by the rate-loss tradeoff. To overcome the rate-distance limit of discrete variable (DV)-QKD protocols, two predominant approaches have been pursued recently: (i) the development of quantum relays and (ii) the employment of trusted relays. Quantum relays require the use of long-duration quantum memories and high-fidelity entanglement distillation [4], which are not yet widely available. On the other hand, the trusted-relay methodology assumes that the relay between two users can be trusted [5]; unfortunately, this assumption is difficult to verify in practice. The measurement device independent (MDI)-QKD approach [6] was able to close the detection loopholes; however, its secret-key rate (SKR) is still bounded by O(T)-dependence (with T standing for transmissivity). Recently, twin-field (TF) QKD has been proposed to overcome the rate-distance limit [7], whose SKR scales with the square-root of transmittance, which represents a promising approach to extend the transmission distance. Another key limitation of DV-QKD is the deadtime of single-photon detectors (SPDs), which limits the baud rate and consequently the SKRs. To solve for this problem, a continuous variable (CV)-QKD can be ----- _Entropy 2020, 22, 831_ 2 of 9 used instead [1,8–10], which employs homodyne/heterodyne detection instead and thus does not exhibit the SPDS’ deadtime limitation problem. In particular, the discrete modulation (DM)-based CV-QKD protocols offer much better reconciliation efficiency compared to that of Gaussian modulation (GM)-based CV-QKD protocols. Unfortunately, the security proofs of DM-based CV-QKD schemes for collective and coherent attacks are still incomplete. To overcome key challenges for DV-QKD, such as low SKR values and limited distance, as well as for DM-based CV-QKD, such as incompleteness of security proofs, the following approaches have been proposed in our recent papers: (1) discretized GM (DGM)-CV-QKD [11], (2) optimized CV-QKD [12], and (3) hybrid DV-CV QKD [13]. An alternative approach to QKD is post-quantum cryptography (PQC) [14]. PQC is typically referred to by various cryptographic algorithms that are thought to be secure against any quantum computer-based attack. Unfortunately, PQC is also based on unproven assumptions and some of the PQC algorithms will be broken in the future by developing more sophisticated quantum algorithms. Modern classical communication networks consist of multiple nodes connected by various types of channels, including free-space optical (FSO) links, optical fibers, ground–satellite links, wireless RF, and coaxial cables. Such a heterogeneous architecture would be equally important for QCNs, as quantum nodes may access a QCN via different kinds of channels. Indeed, quantum communications have been individually validated in free-space, optical fibers, and between a satellite and a ground station, but a combined heterogeneous QCN employing multiple types of channels remains elusive. Unlike in the point-to-point communication case, the fundamental quantum communication rate limits are not well known. Several QKD testbeds have been reported so far, including the DARPA QKD network [15], Tokyo QKD network [16], and secure communication based on quantum cryptography (SECOQC) network [17]. The QKD can also be used to establish QKD-based campus-to-campus virtual private networks employing the IPsec protocol [18] as well as to establish the network setup for using transport-layer security (TLS) based on QKD [19]. However, all of these networks employ the dark fiber infrastructure. Quantum communication over satellite links has already been demonstrated; see for example [20,21]. In this Special Issue paper, we propose to implement the multipartite QCN by employing the cluster state-based concept [22]. The proposed quantum network can be used to: (i) perform distributed quantum computing, (ii) teleport quantum states between any two nodes in the network, and (iii) enable the next generation of cyber security systems. The cluster states can be described by using the stabilizer formalism and as such they can easily be certified by simple syndrome measurements. In this formalism, the cluster states can be interpreted as codewords of a corresponding quantum error correction code, while corresponding errors can be corrected for by simple syndrome decoding, among others. By performing simple Y and Z measurements on properly selected nodes we can straightforwardly establish the Einstein–Podolsky–Rosen (EPR) pair between any two nodes in the network. Moreover, multiple EPR pairs can be established simultaneously. We further propose a cluster state-based quantum network of satellites that enables global coverage. The quantum satellite network would be composed of quantum subnetworks comprised of low Earth orbit (LEO) satellites. Some of these LEO satellite-based quantum subnetworks can be connected to a subnetwork of medium Earth orbit (MEO)/ geostationary orbit (GEO) satellites. The LEO satellites should be used to interconnect terrestrial cluster state-based quantum networks. This quantum global network can also be used to distribute the entangled states for quantum sensing applications and to enable distributed quantum computing on a global scale. SDN concepts should be used to reconfigure the proposed QCN. The paper is organized as follows. In Section 2, we describe the proposed cluster states-based QCN concept. In Section 3, we describe potential approaches to extend the transmission distance between QCN nodes. In Section 4, we describe the QCN that is currently under development at the University of Arizona. Finally, in Section 5, we provide some relevant concluding remarks. ----- _Entropy 2020, 22, 831_ 3 of 9 ## N(a) denoting the neighborhood of a ∈ C. To create a 2-D cluster sta ilbert et al. [23] is applicable; it employs linear states, generated 2. Proposed Cluster States-Based Quantum Communications Networks wn conversion (SPDC), local unitaries, and type I fusion to create To enable the next generation of quantum communication networking, we envision a scenario in which disconnected terrestrial cluster states-based QCNs are coupled through the LEO satellite ## e type I fusion is illustrated in Figure 1, based on [23]. The vertical ph(cluster state) quantum network, thus providing global coverage. The proposed quantum network will be highly robust against turbulence encountered by FSO links, as the envisioned quantum ## ion beam splitter (PBS), while the horizontal photon is transmitted tsatellite network will communicate to ground nodes only through the LEO satellite-to-ground links, exhibiting a vertical downlink profile through vacuum followed by a turbulence layer with strength ## abilistic nature of the PBS, with the photons present at both the leftthat is altitude-dependent. The cluster states belong to the class of the graph states, which also include Bell states, ## e four possible outcomes, each occurring with probability 0.25. Greenberger–Horne–Zeilinger (GHZ) states, W-states, and various entangled states used in quantum error correction [22]. When the cluster C is defined as a connected subset on a d-dimensional lattice, it � � ## he desired fusion operators, and the success probability of the fusioobeys the set of eigenvalue equations Sa���φ C [=] ���φ C[,][ S][a][ =][ X][a] b∈N⊗(a) Zb, where Sa are stabilizer operators with N(a) denoting the neighborhood of a ∈ _C. To create a 2-D cluster state, the approach proposed by_ ## detected by the detector, a successful fusion is declared. The procedGilbert et al. [23] is applicable; it employs linear states, generated by spontaneous parametric down conversion (SPDC), local unitaries, and type I fusion to create the desired 2-D cluster state. The type I ## state is described in Figure 2. To create the box-cluster state, we sfusion is illustrated in Figure 1, based on [23]. The vertical photon is reflected by the polarization beam splitter (PBS), while the horizontal photon is transmitted through the PBS. Given the probabilistic ## ster state, re-label the qubits 2 and 3, and apply the Hadamard gatesnature of the PBS, with the photons present at both the left and right input ports, there are four possible outcomes, each occurring with probability 0.25. Two outcomes correspond to the desired ## vely establish the bond between qubits 1 and 4. Namely, relabelifusion operators, and the success probability of the fusion is 0.5. When a single photon is detected by the detector, a successful fusion is declared. The procedure to create the T-shape cluster state is ## e SWAP gate action. To create the box-on-chain cluster state, we stadescribed in Figure 2. To create the box-cluster state, we start with a four-qubit linear cluster state, re-label the qubits 2 and 3, and apply the Hadamard gates to qubits 2 and 3, which effectively establish ## qubits and apply the same approach as in a box-state creation. Twothe bond between qubits 1 and 4. Namely, relabeling the qubits is equivalent to the SWAP gate action. To create the box-on-chain cluster state, we start with a longer linear chain of qubits and apply the ## ed together to get the same approach as in a box-state creation. Two T-shape cluster states can be fused together to get theH-shape cluster state, etc. _H-shape cluster state, etc._ ###### Polarization discriminating detector ##### left right |l PBS |r **Figure 1. Illustrating the type I fusion process. PBS: polarization beam splitter.** ### igure 1. Illustrating the type I fusion process. PBS: polarization beam splitte ### 5 ### 5 ### 5 ----- |l PBS |r _Entropy 2020, 22, 831_ 4 of 9 **Figure 1. Illustrating the type I fusion process. PBS: polarization beam splitter.** _SWAP[(][2,3)]_ _T-shape_ cluster _Z-measurement_ state on qubit 3 1 _Entropy 2020, 22, x FOR PEER REVIEW Figure 2.Figure 2. Gilbert’s approach to create theGilbert’s approach to create the TT-shape cluster state.-shape cluster state._ 4 of 9 Once the 2-D cluster state of nodes is created, we can use properly selectedOnce the 2-D cluster state of nodes is created, we can use properly selected Y and Z measurementsY and _Z_ to create the EPR pair between any two arbitrary nodes in the quantum network. As a reminder, themeasurements to create the EPR pair between any two arbitrary nodes in the quantum network. As role of thea reminder, the role of the Z measurement is to remove the particular node (qubit) from the cluster, whereas the roleZ measurement is to remove the particular node (qubit) from the cluster, ofwhereas the role of Y measurement is to remove a given node and link neighboring nodes. As an illustration, the 2-DY measurement is to remove a given node and link neighboring nodes. As an cluster state with nine nodes is shown in Figureillustration, the 2-D cluster state with nine nodes is shown in Figure 3. Let us assume that we are 3. Let us assume that we are interested in establishing EPR pairs between nodes 3 and 7 as well as nodes 1 and 9. We first performinterested in establishing EPR pairs between nodes 3 and 7 as well as nodes 1 and 9. We first perform Y measurements in the following order:Y measurements in the following order: Y8, Y5, and Y6 to get the intermediate stage. We then performY8, Y5, and Y6 to get the intermediate stage. We then perform Z-measurement on node 2 and Y measurement on node 4 to get the two desired EPR pairs. Given that the 2-D cluster stateZ-measurement on node 2 and Y measurement on node 4 to get the two desired EPR pairs. Given is universal, it is possible to use the same network architecture for both QCN and distributed quantumthat the 2-D cluster state is universal, it is possible to use the same network architecture for both QCN computing. We also imagine the scenario in which each node is equipped with multiple qubits, whereinand distributed quantum computing. We also imagine the scenario in which each node is equipped several layers of 2-D cluster states are active at the same time, which will allow us to simultaneouslywith multiple qubits, wherein several layers of 2-D cluster states are active at the same time, which perform QCN and distributed quantum computing. Moreover, when several 2-D cluster states are runwill allow us to simultaneously perform QCN and distributed quantum computing. Moreover, when in parallel on the same set of network nodes, we will be able to reconfigure the QCN as needed. Thisseveral 2-D cluster states are run in parallel on the same set of network nodes, we will be able to can be done with the help of the SDN concept. The SDN has been introduced to separate the controlreconfigure the QCN as needed. This can be done with the help of the SDN concept. The SDN has plane and data plane, manage network services through the abstraction of higher-level functionality,been introduced to separate the control plane and data plane, manage network services through the and implement new applications and algorithms eabstraction of higher-level functionality, and implement new applications and algorithms efficiently. fficiently. It has already been studied to enable the coexistence of classical and quantum communication channels. Our SDN-based QCN architectureIt has already been studied to enable the coexistence of classical and quantum communication is composed of three layers, namely an application layer, a control layer, and a QCN layer. Userschannels. Our SDN-based QCN architecture is composed of three layers, namely an application layer, send their requests from the application layer with the help of the northbound interface to the SDNa control layer, and a QCN layer. Users send their requests from the application layer with the help controller. The SDN controller allocates the QCN resources with the help of its global map through theof the northbound interface to the SDN controller. The SDN controller allocates the QCN resources southbound interface. The QCN layer would be composed of dense wavelength-division multiplexingwith the help of its global map through the southbound interface. The QCN layer would be composed (DWDM) FSOof dense wavelength-division multiplexing (DWDM) FSO/single-mode fiber (SMF)/few-mode fiber /single-mode fiber (SMF)/few-mode fiber (FMF) links and QCN nodes. Any two nodes in the QCN can communicate through either through a dedicated SMF(FMF) links and QCN nodes. Any two nodes in the QCN can communicate through either through a /FSO/FMF link or through a wavelength channel. The SDN control should also determine sequence of measurements to bededicated SMF/FSO/FMF link or through a wavelength channel. The SDN control should also performed in order to establish desired EPR pairs. To deal with time-varying channel conditions overdetermine sequence of measurements to be performed in order to establish desired EPR pairs. To heterogeneous links, we should adapt the system configuration based on both application requirementdeal with time-varying channel conditions over heterogeneous links, we should adapt the system and link condition.configuration based on both application requirement and link condition. _Y8, Y5, Y6_ _Z2, Y4_ **Figure 3.Figure 3. Establishing EPR pairs between nodes 1 and 9 as well as between nodes 3 and 7.Establishing EPR pairs between nodes 1 and 9 as well as between nodes 3 and 7.** **3. Extending the Distance between Nodes in QCN** ----- _Entropy 2020, 22, 831_ 5 of 9 **3. Extending the Distance between Nodes in QCN** The DV-QKD can be used to build QKD networks, as discussed in the introduction. Unfortunately, the DV-QKD is affected by the deadtime of SPDs. Moreover, even if Eve cannot get the key because DV-QKD is used, she can prevent parties from creating secure keys, which is similar to the Denial of Service (DoS) attack. Further, since SKRs for DV-QKD are low, the quantum key pool, storing the secure keys, will often be empty, hampering the operation of QKD networks. To solve for this problem we propose to use the hybrid QKD-PQC protocols, in which QKD is used for raw key transmission and PQC in information reconciliation to reduce the leakage during the error reconciliation stage, which is illustrated in Figure 4. As mentioned in the introduction, the PQC is typically referred to in various cryptographic algorithms that are thought to be secure against any quantum computer-based attack. Unfortunately, the PQC is also based on unproven assumptions and some of the QPC algorithms might be broken in the future by developing advanced quantum algorithms. For this reason we propose to use the PQC algorithms only in the information reconciliation phase so as to limit the _Entropy leakage due to transmission of parity bits over an authenticated classical channel (in conventional2020, 22, x FOR PEER REVIEW_ 5 of 9 QKD). The quantum algorithms to be developed (not yet known), which will be capable of breaking terms of the number of operations the PQC algorithms, will have certain complexity expressed in terms of the number of operationsL. By ensuring that the number of parity bits N–K is shorter than L. the number of secure PQC bits logBy ensuring that the number of parity bits2L, the proposed cryptographic scheme will be secure. Evidently, N–K is shorter than the number of secure PQC bits log2L, the proposed cryptographic scheme exploits the complexity of corresponding quantum algorithms the proposed cryptographic scheme will be secure. Evidently, the proposed cryptographic scheme used to break the PQC protocols. Given that the McEliece cryptosystem based on quasi cyclic (QC)-exploits the complexity of corresponding quantum algorithms used to break the PQC protocols. Given low-density parity-check (LDPC) coding is straightforward to implement as shown in [24], whereas that the McEliece cryptosystem based on quasi cyclic (QC)-low-density parity-check (LDPC) coding the corresponding LDPC encoders and decoders have been already implemented in field-is straightforward to implement as shown in [24], whereas the corresponding LDPC encoders and programmable gate array (FPGA) [25], it represents an excellent candidate to be used for the decoders have been already implemented in field-programmable gate array (FPGA) [25], it represents transmission of parity bits in the TF-QKD scheme. As an illustration, the secret fraction that can be an excellent candidate to be used for the transmission of parity bits in the TF-QKD scheme. As an achieved with the BB84 protocol is lower bounded by [1]: illustration, the secret fraction that can be achieved with the BB84 protocol is lower bounded by [1]: ###### r r= =q( ) qZ [(][Z]1[)]−[�]1h e −2 ( h(2X�)e)[(][X][)][��]−− qq[(]( )[Z]Z[)]f h efeeh22(�e( )[(]Z[Z][)])[�],, (1) (1) where where qq[(]([Z]Z[)]) denotes the probability of declaring a successful result when Alice sent a single-photon denotes the probability of declaring a successful result when Alice sent a single-photon and Bob detected it in the and Bob detected it in theZ-basis, Z-basis,fe denotes the error correction inefficiency ( fe denotes the error correction inefe _≥ 1), fficiency (e[(X)] [e[(Z)]] denotes fe ≥_ 1), the e[(X)] [eQBER [(Z)]] denotes the QBER in the X-basis (in the X-basis (Z-basis), _Zand -basis), andh2(x)_ _his 2(xthe ) is the binary entropy functionbinary_ entropy function ###### h xh22( )(x) ==− −xlogx log2 ( ) (2x(x−) −1−(1x −)log 1x)2 log( −2x()1. The second term − x). The second termq[(Z)]h2 q[e[(Z)][(X)]h] denotes the amount of information Eve 2[e[(X)]] denotes the amount of information Eve was able to learn during the raw key transmission, and this information can be removed from the was able to learn during the raw key transmission, and this information can be removed from the final key during the privacy amplification phase. The third term final key during the privacy amplification phase. The third termq q[(Z)][(Z)]fe hfe2 h[e2[(Z)][e] represents the amount of [(Z)]] represents the amount of information revealed during the error correction stage. By sending the parity bits over the PQC information revealed during the error correction stage. By sending the parity bits over the PQC channel this term can be effectively eliminated and the SKR can be increased. channel this term can be effectively eliminated and the SKR can be increased. Bob Alice Raw key Sifting Sifted Syndrome PQC Public channel PQC LDPC Corrected input procedure key, x _p=xH[T]_ encryption decryption decoder key **Figure 4. Figure 4. Illustration of post-quantum cryptography-based information reconciliation.Illustration of post-quantum cryptography-based information reconciliation.** By using this approach, as illustrated in Figure 5, the transmission distance between two nodes By using this approach, as illustrated in Figure 5, the transmission distance between two nodes in QCN can be significantly extended. Here we provide comparisons of the joint TF-QKD-McEliece in QCN can be significantly extended. Here we provide comparisons of the joint TF-QKD-McEliece encryption scheme against the phase-matching (PM) TF-QKD protocol introduced in [26], the MDI-QKD encryption scheme against the phase-matching (PM) TF-QKD protocol introduced in [26], the MDI protocol [6], and the decoy-state-based BB84 protocol [27]. The system parameters are selected QKD protocol [6], and the decoy-state-based BB84 protocol [27]. The system parameters are selected as follows: the detector efficiency as follows: the detector efficiencyηd η = 0.25, reconciliation inefficiency d = 0.25, reconciliation inefficiencyfe = 1.15, the dark count rate f e = 1.15, the dark countpd = 8 × 10rate pd =[−8], the misalignment error 8 × 10[−][8], the misalignment errored = 1.5%, and the number of phase slices for PM TF-QKD is set to ed = 1.5%, and the number of phase slices for PM TF-QKD is set to M = 16. Regarding the transmission medium, it is assumed that recently reported _M = 16. Regarding the transmission medium, it is assumed that recently reported ultra-low-loss fiber_ of attenuation 0.1419 dB/km (at 1560 nm) is employed [28]. In the same Figure, the Pirandola– Lauren a Otta iani Banchi (PLOB) bound on a linear key rate is pro ided as ell Both PM TF QKD ----- _Entropy 2020, 22, 831_ 6 of 9 ultra-low-loss fiber of attenuation 0.1419 dB/km (at 1560 nm) is employed [28]. In the same Figure, the Pirandola–Laurenza–Ottaviani–Banchi (PLOB) bound on a linear key rate is provided as well. Both PM TF-QKD and joint TF-QKD-McEliece encryption schemes outperform the decoy-state BB84 protocol for distances larger than 162 km, while simultaneously outperforming the MDI-QKD protocol for all distances, and exceed the PLOB bound at a distance of 322 km. The PM TF-QKD protocol can achieve the maximum distance of 623 km. The proposed joint TF-QKD-McEliece encryption scheme is able to achieve the distance of even 1127 km, thus significantly outperforming all other schemes. Even though the operating wavelength was 1560 nm, other suitable wavelengths such as 2 µm and 3.9 µm can be used as well.Entropy 2020, 22, x FOR PEER REVIEW 6 of 9 10-1 10-3 10-5 10-7 10-9 10-11 10-13 Distance, L [km] **Figure 5.Figure 5. Proposed hybrid QKD-PQC scheme against MDI-QKD and TF-QKD in terms of secret-keyProposed hybrid QKD-PQC scheme against MDI-QKD and TF-QKD in terms of secret-key** rate vs. distance, assuming that ultra-low loss fiber is used.rate vs. distance, assuming that ultra-low loss fiber is used. Now, by connecting the base stations to the nodes in the proposed QCNs, we can provide Now, by connecting the base stations to the nodes in the proposed QCNs, we can provide the the unconditional security to the 5G+/6G wireless networks. By organizing the base stations in a unconditional security to the 5G+/6G wireless networks. By organizing the base stations in a quantum quantum optical mesh network and employing the proposed hybrid QKD-PQC concept we can provide optical mesh network and employing the proposed hybrid QKD-PQC concept we can provide unconditional security to a large number of users. The Internet of Things (IoT) architecture will comprise unconditional security to a large number of users. The Internet of Things (IoT) architecture will widely distributed nodes connected via different types of channels to enable new functionalities in comprise widely distributed nodes connected via different types of channels to enable new communication, sensing, and computing. Communication security in such a giant network is of functionalities in communication, sensing, and computing. Communication security in such a giant paramount importance. Our proposed QCNs will underpin the unconditional physical-layer security network is of paramount importance. Our proposed QCNs will underpin the unconditional physical of the IoT given that it will allow any two arbitrary nodes to securely transmit data at a high rate layer security of the IoT given that it will allow any two arbitrary nodes to securely transmit data at via an optical link. Critically, the security of such a network will not rest upon the trusted-node a high rate via an optical link. Critically, the security of such a network will not rest upon the trusted assumption, and a compromised node will not affect the security of other nodes. As such, the proposed node assumption, and a compromised node will not affect the security of other nodes. As such, the QCNs will lead to a substantially stronger security level for the IoT. To enable security for future 6G proposed QCNs will lead to a substantially stronger security level for the IoT. To enable security for wireless networks at a reasonable cost, the proposed joint satellite–terrestrial QCN can be based on the future 6G wireless networks at a reasonable cost, the proposed joint satellite–terrestrial QCN can be Cubesat satellites. based on the Cubesat satellites. For satellite-to-satellite quantum communications, in addition to the proposed hybrid QKD-PQC For satellite-to-satellite quantum communications, in addition to the proposed hybrid QKD concept, it also possible to employ our recent restricted eavesdropping concept [29], which offers a PQC concept, it also possible to employ our recent restricted eavesdropping concept [29], which significant increase in SKRs. This concept was presented in the ICTON 2020 paper [30]. Alternatively, offers a significant increase in SKRs. This concept was presented in the ICTON 2020 paper [30]. the hybrid QKD can also be applied [13]. Alternatively, the hybrid QKD can also be applied [13]. **4. QCN under Development** **4. QCN under Development** The terrestrial QCN to be developed at the University of Arizona is shown in Figure 6; it will The terrestrial QCN to be developed at the University of Arizona is shown in Figure 6; it will exploit the existing NSF MRI INQUIRE quantum network, representing the quantum hub (QuHub) exploit the existing NSF MRI INQUIRE quantum network, representing the quantum hub (QuHub) to share entangled photons and SPDs among different labs across the campus. The outdoor FSO to share entangled photons and SPDs among different labs across the campus. The outdoor FSO bidirectional link, connecting the Electrical and Computer Engineering and Optical Sciences buildings, bidirectional link, connecting the Electrical and Computer Engineering and Optical Sciences has already been established, with the FSO transceiver shown in Figure 7. We will also create the mesh buildings, has already been established, with the FSO transceiver shown in Figure 7. We will also create the mesh network as well as the hybrid network composed of mesh, optical star, and ring network segments The deployed heterogeneous QCNs will allow us to test novel quantum ----- _py_,, _Entropymuch faster than Gaussian beams for such long-distance applications. Hence, we need to use pure 2020, 22, 831_ 7 of 9 much faster than Gaussian beams for such long-distance applications. Hence, we need to use pure Bessel beams to overcome this problem, as we have shown in our recent paper [32]. To enable Bessel beams to overcome this problem, as we have shown in our recent paper [32]. To enable robustness against turbulence encountered by FSO links, the envisioned quantum satellite QCN network as well as the hybrid network composed of mesh, optical star, and ring network segments. robustness against turbulence encountered by FSO links, the envisioned quantum satellite QCN should communicate to ground nodes only through the LEO satellite-to-ground links, exhibiting a The deployed heterogeneous QCNs will allow us to test novel quantum-networking theories and should communicate to ground nodes only through the LEO satellite-to-ground links, exhibiting a vertical downlink profile through vacuum followed by a turbulence layer with altitude-dependent develop experimental tools for counteracting various channel impairments. To deal with atmospheric vertical downlink profile through vacuum followed by a turbulence layer with altitude-dependent strength. In principle. MEO/GEO satellite QCNs can be created above LEO QCNs to provide the strength. In principle. MEO/GEO satellite QCNs can be created above LEO QCNs to provide the turbulence eplanetary coverage. ffects, the adaptive optics (AO) subsystem, composed of a wavefront sensor (WFS) and deformable mirror will be used. The AO will be combined with adaptive LDPC coding. planetary coverage. User 3 User 2 building MeinelUser 3 building MSE User 2 Meinel MSE building building SMF links User 4User 4 FMF/FMF/MMF linksQuComQuCom(ECE 549) Lab(ECE 549) Lab QuHub(ECE QuHub111) (ECE 111) User 1User 1SMF links Physicsbuilding Physicsbuilding Alice MMF links OCSL(ECE Keatingbuilding (server) OCSL441) Keating **NSF MRI INQUIRE** Alice (ECE building **quantum network** (server) 441) **NSF MRI INQUIRE** **Figure 6. Terrestrial quantum communication network to be developed at the University of** **Figure 6.Figure 6. Terrestrial quantum communication network to be developed at the University of Arizona.Terrestrial quantum communication network to be developed at the University of Arizona.** Arizona. **Figure 7.Figure 7. Free-space optical transceiver used in outdoor FSO link.Free-space optical transceiver used in outdoor FSO link.** **Figure 7. Free-space optical transceiver used in outdoor FSO link.** **5. Concluding Remarks To provide global coverage, we envision a scenario in which disconnected terrestrial QCNs, such** **5. Concluding Remarks as the one shown in Figure 6, are coupled through the LEO satellite quantum network. We have recently** To enable the next generation of quantum-enabled cyber security systems, we proposed a shown that a Bessel–Gaussian (BG) beam, carrying an orbital angular momentum mode, exhibits better quantum network of satellites that will provide the global coverage. The quantum satellite network To enable the next generation of quantum-enabled cyber security systems, we proposed a tolerance to atmospheric turbulence effects compared to Gaussian beams for distances up to a few quantum network of satellites that will provide the global coverage. The quantum satellite network will be composed of quantum subnetworks comprised of LEO satellites. Some of these LEO satellite kilometers [31]. However, for LEO satellite-to-ground QuCom links, BG beams diffract much faster will be composed of quantum subnetworks comprised of LEO satellites. Some of these LEO satellite-based quantum subnetworks will be connected to a subnetwork of MEO satellites. The MEO satellite than Gaussian beams for such long-distance applications. Hence, we need to use pure Bessel beams based quantum subnetworks will be connected to a subnetwork of MEO satellites. The MEO satellite subnetworks will then be interconnected to the global network of GEO satellites. The LEO/MEO to overcome this problem, as we have shown in our recent paper [32]. To enable robustness against subnetworks will then be interconnected to the global network of GEO satellites. The LEO/MEO satellites will also be used to interconnect terrestrial quantum networks. Each quantum turbulence encountered by FSO links, the envisioned quantum satellite QCN should communicate to satellites will also be used to interconnect terrestrial quantum networks. Each quantum communication subnetwork will be based on the cluster state concept. This quantum global network ground nodes only through the LEO satellite-to-ground links, exhibiting a vertical downlink profile communication subnetwork will be based on the cluster state concept. This quantum global network will allow us to establish EPR pairs between any two nodes in the global network. It can also be used through vacuum followed by a turbulence layer with altitude-dependent strength. In principle. will allow us to establish EPR pairs between any two nodes in the global network. It can also be used to distribute the entangled states for quantum-sensing applications and to enable distributed MEO/GEO satellite QCNs can be created above LEO QCNs to provide the planetary coverage. to distribute the entangled states for quantum-sensing applications and to enable distributed quantum computing on a global scale. quantum computing on a global scale. **5. Concluding RemarksFunding: This research received no external funding.** **Funding: This research received no external funding.** **Conflicts of Interest: To enable the next generation of quantum-enabled cyber security systems, we proposed a quantumThe author declares no conflict of interest.** **Conflicts of Interest: network of satellites that will provide the global coverage. The quantum satellite network will beThe author declares no conflict of interest.** composed of quantum subnetworks comprised of LEO satellites. Some of these LEO satellite-basedReferences **References quantum subnetworks will be connected to a subnetwork of MEO satellites. The MEO satellite** 1. Djordjevic, I.B. Physical-Layer Security and Quantum Key Distribution; Springer Nature Switzerland: Cham, subnetworks will then be interconnected to the global network of GEO satellites. The LEO/MEO 1. Djordjevic,Switzerland,I.B.2019. Physical-Layer Security and Quantum Key Distribution; Springer Nature Switzerland: Cham, satellites will also be used to interconnect terrestrial quantum networks. Each quantum communication Switzerland, 2019. subnetwork will be based on the cluster state concept. This quantum global network will allow us to ----- _Entropy 2020, 22, 831_ 8 of 9 establish EPR pairs between any two nodes in the global network. It can also be used to distribute the entangled states for quantum-sensing applications and to enable distributed quantum computing on a global scale. **Funding: This research received no external funding.** **Conflicts of Interest: The author declares no conflict of interest.** **References** 1. Djordjevic, I.B. Physical-Layer Security and Quantum Key Distribution; Springer Nature Switzerland: Cham, Switzerland, 2019. 2. Pljonkin, A.P. Features of the Photon Pulse Detection Algorithm in the Quantum Key Distribution System. In Proceedings of the 2017 International Conference on Cryptography, Security and Privacy (ICCSP ’17), Wuhan, China, 17–19 March 2017; Association for Computing Machinery: New York, NY, USA, 2017; pp. 81–84. [[CrossRef]](http://dx.doi.org/10.1145/3058060.3058078) 3. Pljonkin, A.P. Vulnerability of the synchronization process in the quantum key distribution system. Int. J. _[Cloud Appl. Comput. 2019, 9, 4. [CrossRef]](http://dx.doi.org/10.4018/IJCAC.2019010104)_ 4. Duan, L.-M.; Lukin, M.; Cirac, J.I.; Zoller, P. Long-distance quantum communication with atomic ensembles [and linear optics. Nature 2001, 414, 413–418. [CrossRef]](http://dx.doi.org/10.1038/35106500) 5. [Qiu, J. Quantum communications leap out of the lab. Nature 2014, 508, 441–442. [CrossRef]](http://dx.doi.org/10.1038/508441a) 6. Lo, H.-K.; Curty, M.; Qi, B. Measurement-device-independent quantum key distribution. Phys. Rev. Lett. **[2012, 108, 130503. [CrossRef]](http://dx.doi.org/10.1103/PhysRevLett.108.130503)** 7. Lucamarini, M.; Yuan, Z.L.; Dynes, J.F.; Shields, A.J. Overcoming the rate–distance limit of quantum key [distribution without quantum repeaters. Nature 2018, 557, 400–403. [CrossRef]](http://dx.doi.org/10.1038/s41586-018-0066-6) 8. Fossier, S.; Diamanti, E.; Debuisschert, T.; Tualle-Brouri, R.; Grangier, P. Improvement of continuous-variable [quantum key distribution systems by using optical preamplifiers. J. Phys. B 2009, 42, 114014. [CrossRef]](http://dx.doi.org/10.1088/0953-4075/42/11/114014) 9. Qu, Z.; Djordjevic, I.B. Four-dimensionally multiplexed eight-state continuous-variable quantum key [distribution over turbulent channels. IEEE Photonics J. 2017, 9, 7600408. [CrossRef]](http://dx.doi.org/10.1109/JPHOT.2017.2777261) 10. [Ralph, T.C. Continuous variable quantum cryptography. Phys. Rev. A 1999, 61, 010303. [CrossRef]](http://dx.doi.org/10.1103/PhysRevA.61.010303) 11. Djordjevic, I.B. On the Discretized Gaussian Modulation (DGM)-based Continuous Variable-QKD. IEEE Access **[2019, 7, 65342–65346. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2917587)** 12. Djordjevic, I.B. Optimized-Eight-State CV-QKD Protocol Outperforming Gaussian Modulation Based [Protocols. IEEE Photonics J. 2019, 11, 4500610. [CrossRef]](http://dx.doi.org/10.1109/JPHOT.2019.2921521) 13. Djordjevic, I.B. Hybrid QKD Protocol Outperforming both DV- and CV-QKD Protocols. IEEE Photonics J. **[2020, 12, 7600108. [CrossRef]](http://dx.doi.org/10.1109/JPHOT.2019.2946910)** 14. Bernstein, D.J.; Buchmann, J.; Dahmen, E. Post-Quantum Cryptography; Springer: Berlin, Germany, 2009. 15. Elliott, C.; Colvin, A.; Pearson, D.; Pikalo, O.; Schlafer, J.; Yeh, H. Current status of the DARPA quantum network (Invited Paper). In Proceedings of the SPIE 5815, Quantum Information and Computation III, Defense and Security, Orlando, FL, USA, 25 May 2005. 16. Sasaki, M.; Fujiwara, M.; Ishizuka, H.; Klaus, W.; Wakui, K.; Takeoka, M.; Miki, S.; Yamashita, T.; Wang, Z.; Tanaka, A.; et al. Field test of quantum key distribution in the Tokyo QKD Network. Opt. Express 2011, 19, [10387–10409. [CrossRef] [PubMed]](http://dx.doi.org/10.1364/OE.19.010387) 17. Alléaume, R.; Branciard, C.; Bouda, J.; Debuisschert, T.; Dianati, M.; Gisin, N.; Godfrey, M.; Grangier, P.; Länger, T.; Lütkenhaus, N.; et al. Using quantum key distribution for cryptographic purposes. J. Theor. _[Comput. Sci. 2014, 560, 62–81. [CrossRef]](http://dx.doi.org/10.1016/j.tcs.2014.09.018)_ 18. [Nagayama, S.; Van Meter, R. Internet-Draft: IKE for IPsec with QKD. 2009. Available online: https:](https://tools.ietf.org/html/draft-nagayama-ipsecme-ipsec-with-qkd-01) [//tools.ietf.org/html/draft-nagayama-ipsecme-ipsec-with-qkd-01 (accessed on 28 July 2020).](https://tools.ietf.org/html/draft-nagayama-ipsecme-ipsec-with-qkd-01) 19. Mink, A.; Frankel, S.; Perlner, R. Quantum Key Distribution (QKD) and Commodity Security Protocols: Introduction and Integration. Intern. J. Netw. Secur. Appl. 2009, 1, 101–112. 20. Yin, J.; Cao, Y.; Li, Y.H.; Liao, S.K.; Zhang, L.; Ren, J.G.; Cai, W.Q.; Liu, W.Y.; Li, B.; Dai, H.; et al. Satellite-based [entanglement distribution over 1200 kilometers. Science 2017, 356, 1140–1144. [CrossRef] [PubMed]](http://dx.doi.org/10.1126/science.aan3211) 21. Dequal, D.; Vallone, G.; Bacco, D.; Gaiarin, S.; Luceri, V.; Bianco, G.; Villoresi, P. Experimental single-photon [exchange along a space link of 7000 km. Phys. Rev. A 2016, 93, 010301. [CrossRef]](http://dx.doi.org/10.1103/PhysRevA.93.010301) ----- _Entropy 2020, 22, 831_ 9 of 9 22. Briegel, H.J. Cluster States. In Compendium of Quantum Physics; Greenberger, D., Hentschel, K., Weinert, F., Eds.; Springer: Berlin/Heidelberg, Germany, 2009; pp. 96–105. 23. Gilbert, G.; Hamrick, M.; Weinstein, Y.S. Efficient construction of photonic quantum-computational clusters. _[Phys. Rev. A 2006, 73, 064303. [CrossRef]](http://dx.doi.org/10.1103/PhysRevA.73.064303)_ 24. Baldi, M.; Bianchi, M.; Chiaraluce, F. Security and complexity of the McEliece cryptosystem based on QC [LDPC codes. IET Inf. Secur. 2013, 7, 212–220. [CrossRef]](http://dx.doi.org/10.1049/iet-ifs.2012.0127) 25. Sun, X.; Zou, D.; Qu, Z.; Djordjevic, I.B. Run-time reconfigurable adaptive LDPC coding for optical channels. _[Opt. Express 2018, 26, 29319–29329. [CrossRef]](http://dx.doi.org/10.1364/OE.26.029319)_ 26. [Ma, X.; Zeng, P.; Zhou, H. Phase-matching quantum key distribution. Phys. Rev. X 2018, 8, 031043. [CrossRef]](http://dx.doi.org/10.1103/PhysRevX.8.031043) 27. Lo, H.-K.; Ma, X.; Chen, K. Decoy state quantum key distribution. Phys. Rev. Lett. 2005, 94, 230504. [[CrossRef] [PubMed]](http://dx.doi.org/10.1103/PhysRevLett.94.230504) 28. Tamura, Y.; Sakuma, H.; Morita, K.; Suzuki, M.; Yamamoto, Y.; Shimada, K.; Honma, Y.; Sohma, K.; Fujii, T.; Hasegawa, T.; et al. The First 0.14-dB/km loss optical fiber and its impact on submarine transmission. J. _[Lightw. Technol. 2018, 36, 44–49. [CrossRef]](http://dx.doi.org/10.1109/JLT.2018.2796647)_ 29. Pan, Z.; Seshadreesan, K.P.; Clark, W.; Adcock, M.R.; Djordjevic, I.B.; Shapiro, J.H.; Guha, S. Secret Key Distillation over a Pure Loss Quantum Wiretap Channel under Restricted Eavesdropping. In Proceedings of the 2019 IEEE International Symposium on Information Theory (ISIT 2019), Paris, France, 7–12 July 2019; pp. 3032–3036. 30. Pan, Z.; Djordjevic, I.B. Security of Satellite-Based CV-QKD under Realistic Assumptions. In Proceedings of the 22nd International Conference on Transparent Optical Networks ICTON 2020, Bari, Italy, 19–23 July 2020. 31. Wang, T.-L.; Gariano, J.; Djordjevic, I.B. Employing Bessel-Gaussian Beams to Improve Physical-Layer [Security in Free-Space Optical Communications. IEEE Photonics J. 2018, 10, 7907113. [CrossRef]](http://dx.doi.org/10.1109/JPHOT.2018.2867173) 32. Wang, T.-L.; Djordjevic, I.B.; Nagel, J. Laser Beam Propagation Effects on Secure Key Rates for [Satellite-to-Ground Discrete Modulation CV-QKD. Appl. Opt. 2019, 58, 8061–8068. [CrossRef]](http://dx.doi.org/10.1364/AO.58.008061) © 2020 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution [(CC BY) license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/.) -----
{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC7517431, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/1099-4300/22/8/831/pdf?version=1596101395" }
2,020
[ "JournalArticle" ]
true
2020-07-29T00:00:00
[ { "paperId": "a35d44be73494a09fee9e465968b9f4b1c5ef876", "title": "Security of Satellite-Based CV-QKD under Realistic Assumptions" }, { "paperId": "da74a376a0e5a8c775d3fcebbe6ffd377a2ce6bf", "title": "Hybrid QKD Protocol Outperforming Both DV- and CV-QKD Protocols" }, { "paperId": "4a51b17ed9d09a4d33b466d5140a2cc387185e7a", "title": "Laser beam propagation effects on secure key rates for satellite-to-ground discrete modulation CV-QKD." }, { "paperId": "747fe4e7823e19ab712cc0b625fd3b4da2fa0d8a", "title": "Physical-Layer Security and Quantum Key Distribution" }, { "paperId": "057a4b21f3b3ce17e1c0071651f44a0efbbeea39", "title": "Secret key distillation over a pure loss quantum wiretap channel under restricted eavesdropping" }, { "paperId": "46b300add8a462e9b5a0826dcbdeed0dacb297b9", "title": "Optimized-Eight-State CV-QKD Protocol Outperforming Gaussian Modulation Based Protocols" }, { "paperId": "2c6f87bd06f8b19e9b65fa98cc7a907d9bbb5201", "title": "On the Discretized Gaussian Modulation (DGM)- Based Continuous Variable-QKD" }, { "paperId": "229a3ae4136eb6f61d80c476eb57f9af5601151c", "title": "Vulnerability of the Synchronization Process in the Quantum Key Distribution System" }, { "paperId": "5f9b0ed05a386a007aa5866388db5f5a2a68a918", "title": "Run-time reconfigurable adaptive LDPC coding for optical channels." }, { "paperId": "c83b9da59c4aeae6eb578a4ad377caa73b9a119e", "title": "Employing Bessel-Gaussian Beams to Improve Physical-Layer Security in Free-Space Optical Communications" }, { "paperId": "85670868b210bb28591ad775d815ef9f961c8566", "title": "Phase-Matching Quantum Key Distribution" }, { "paperId": "0da7787e573e7e91dbbea1cb77c7b8c13e5a775e", "title": "Overcoming the rate–distance limit of quantum key distribution without quantum repeaters" }, { "paperId": "3ef53e090fedfbef898561d18fe07960021ca08c", "title": "Four-Dimensionally Multiplexed Eight-State Continuous-Variable Quantum Key Distribution Over Turbulent Channels" }, { "paperId": "7865f3132d4053c25f04ed148b480799932a825a", "title": "Satellite-based entanglement distribution over 1200 kilometers" }, { "paperId": "21fb93ed12f3bb985eca64296d84b21a806722af", "title": "Features of the Photon Pulse Detection Algorithm in the Quantum Key Distribution System" }, { "paperId": "172835329357c307cb296ce50c622374c0069eb1", "title": "Experimental single photon exchange along a space link of 7000 km" }, { "paperId": "9f595e47ba244c2eda6f47fc98f42e3ae4db61b1", "title": "IKE for IPsec with QKD" }, { "paperId": "5626ae9ea6b010871b0eb4a6237d1dae064f5c14", "title": "Quantum communications leap out of the lab" }, { "paperId": "e0087bee26edc29886beb428767d26d5783097cf", "title": "Security and complexity of the McEliece cryptosystem based on quasi-cyclic low-density parity-check codes" }, { "paperId": "0014862203b30180743559fe378fc4dc007107eb", "title": "Field test of quantum key distribution in the Tokyo QKD Network." }, { "paperId": "46b77116268b9fe992a3d4763a73fbf2d3400157", "title": "Quantum Key Distribution (QKD) and Commodity Security Protocols: Introduction and Integration" }, { "paperId": "bc1bcd12998be8437579cf7a95a8390576f65b5b", "title": "Improvement of continuous-variable quantum key distribution systems by using optical preamplifiers" }, { "paperId": "99291ce0b97a31c786560241fea62604332afbf5", "title": "Post-quantum cryptography" }, { "paperId": "f5480ff6e99948c9d38bafdb77924befe3ad7d87", "title": "Using quantum key distribution for cryptographic purposes: A survey" }, { "paperId": "76682df264c4406b213410837035a2f3696311f1", "title": "Efficient construction of photonic quantum-computational clusters" }, { "paperId": "f3dac0d820a7961310bbdb3966db68a4a3d2706b", "title": "Current status of the DARPA quantum network (Invited Paper)" }, { "paperId": "c45f3b85a2f7efccf834de721e4183992312f859", "title": "Decoy state quantum key distribution." }, { "paperId": "a435c4730520628f0fe48e289782925a4979c251", "title": "Long-distance quantum communication with atomic ensembles and linear optics" }, { "paperId": "7cc963de3b2d36b4f9d5b6de50e6fb5508651556", "title": "Continuous variable quantum cryptography" }, { "paperId": "d356042183a64e44251fe5b624c10800f6984176", "title": "The First 0.14-dB/km Loss Optical Fiber and its Impact on Submarine Transmission" }, { "paperId": "a8d3024031c3248fc812edfbce864af7a11ece12", "title": "Cluster States" }, { "paperId": null, "title": "This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license" } ]
11,683
en
[ { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00d1e468ebce3a3c6197b876c32330c2f54777a3
[]
0.813042
BioShare: An Open Framework for Trusted Biometric Authentication under User Control
00d1e468ebce3a3c6197b876c32330c2f54777a3
Applied Sciences
[ { "authorId": "2027667427", "name": "Quan Sun" }, { "authorId": "2145185836", "name": "Jie Wu" }, { "authorId": "5493242", "name": "Wenhai Yu" } ]
{ "alternate_issns": null, "alternate_names": [ "Appl Sci" ], "alternate_urls": [ "http://www.mathem.pub.ro/apps/", "https://www.mdpi.com/journal/applsci", "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-217814" ], "id": "136edf8d-0f88-4c2c-830f-461c6a9b842e", "issn": "2076-3417", "name": "Applied Sciences", "type": "journal", "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-217814" }
Generally, biometric authentication is conducted either by mobile terminals in local-processing mode or by public servers in centralized-processing mode. In the former mode, each user has full control of his/her biometric data, but the authentication service is restricted to local mobile apps. In the latter mode, the authentication service can be opened up to network applications, but the owners have no control of their private data. It has become a difficult problem for biometric applications to provide open and trusted authentication services under user control. Existing approaches address these concerns in ad-hoc ways. In this work, we propose BioShare, a framework that provides trusted biometric authentication services to network applications while giving users full control of their biometric data. Our framework is designed around three key principles: each user has full control of his/her biometric data; biometric data is stored and processed in trusted environments to prevent privacy leaks; and the open biometric-authentication service is efficiently provided to network applications. We describe our current design and sample implementation, and illustrate how it provides an open face-recognition service with standard interfaces, combines terminal trusted environments with server enclaves, and enables each user to control his/her biometric data efficiently. Finally, we analyze the security of the framework and measure the performance of the implementation.
# applied sciences _Article_ ## BioShare: An Open Framework for Trusted Biometric Authentication under User Control **Quan Sun** **[1,2,]*, Jie Wu** **[1]** **and Wenhai Yu** **[2]** 1 School of Computer Science and Technology, Fudan University, No. 220 Handan Rd., Shanghai 200433, China 2 China UnionPay Co., Ltd., No. 1699 Gutang Rd., Shanghai 201201, China ***** Correspondence: quansun@unionpay.com **Abstract: Generally, biometric authentication is conducted either by mobile terminals in local-** processing mode or by public servers in centralized-processing mode. In the former mode, each user has full control of his/her biometric data, but the authentication service is restricted to local mobile apps. In the latter mode, the authentication service can be opened up to network applications, but the owners have no control of their private data. It has become a difficult problem for biometric applications to provide open and trusted authentication services under user control. Existing approaches address these concerns in ad-hoc ways. In this work, we propose BioShare, a framework that provides trusted biometric authentication services to network applications while giving users full control of their biometric data. Our framework is designed around three key principles: each user has full control of his/her biometric data; biometric data is stored and processed in trusted environments to prevent privacy leaks; and the open biometric-authentication service is efficiently provided to network applications. We describe our current design and sample implementation, and illustrate how it provides an open face-recognition service with standard interfaces, combines terminal trusted environments with server enclaves, and enables each user to control his/her biometric data efficiently. Finally, we analyze the security of the framework and measure the performance of the implementation. **Citation: Sun, Q.; Wu, J.; Yu, W.** BioShare: An Open Framework for Trusted Biometric Authentication under User Control. Appl. Sci. 2022, _[12, 10782. https://doi.org/10.3390/](https://doi.org/10.3390/app122110782)_ [app122110782](https://doi.org/10.3390/app122110782) Academic Editor: Byung-Gyu Kim Received: 12 September 2022 Accepted: 21 October 2022 Published: 25 October 2022 **Publisher’s Note: MDPI stays neutral** with regard to jurisdictional claims in published maps and institutional affil iations. **Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **Keywords: biometric authentication; face recognition; trusted executive environment; enclave** **1. Introduction** With the prevalence of smartphones and e-commerce, biometric identification technologies are widely used by service providers and are favorably accepted by most consumers. Fingerprint identification and face recognition have become standard features on most Android and iOS devices, and payment platforms based on facial recognition are increasingly popular in countries such as China. Meanwhile, privacy protection has become a serious topic and is attracting comprehensive attention from scholars and governments. The European Union released the General Data Protection Regulation (GDPR) [1–3] in 2018, the United States issued the California Consumer Privacy Act (CCPA) [4,5] in 2018, and China published the Personal Information Protection Law (PIPL) [6] in 2021. Google, Facebook, British Airways, and many other companies were fined for data leakage, unauthorized usage, and other data issues. We are currently faced with the dilemma between data usage and privacy protection, and we need a technical solution to achieve both goals within the same framework. Currently, there are two types of solutions available: local processing solutions and centralized-processing solutions. The former method is to process and store personal biometric data in local trusted environments integrated in personal mobile terminals, and the representative implementations include Apple Touch ID/Face ID and Windows Hello. Through local trusted computation, each user has full control of his/her biometric data without the risk of privacy leakage, but authentication services are restricted only ----- _Appl. Sci. 2022, 12, 10782_ 2 of 21 to local mobile apps. As such, the services are secure but not open. The latter method uses a centralized system to collect and process biometric data in centralized-processing mode on public servers, and typical application scenarios include many payment platforms based on facial recognition, as described in Section 2.3. With the help of powerful servers, service providers can provide open biometric authentication services to various network applications efficiently. However, when raw biometric data or extracted feature data are stored by service providers in their database servers, users are put at risk of privacy leakages due to improper maintenance or malicious attacks. While user biometric data are fully controlled by service providers without strict regulation and independent audit, users have no control of their private data and are faced with the abuse of personal data. As a conclusion, the first method is secure but not open, while the second method is open but not secure. We need a framework that provides both secure and open biometric services. In this paper, we propose BioShare, a framework that provides both secure and open biometric services.To achieve this goal, we store and process biomatric data in secure environments (terminal TEEs and server enclaves) to ensure security, and we coordinate the terminals and the servers to provide open services for network applications. Our framework is designed around three key principles: **a. Each user has full control of his/her biometric data. Biometric services are pro-** vided in either forwarding mode or authorized mode. In the forwarding mode, personal biometric data are stored in the user’s terminals. In the authorized mode, biometric data are temporarily stored in the server enclave under user authorization, and users can cancel the authorization at any time. Both modes ensure users have full control of their private data. **b. Open bioauthentication services are efficiently provided to network applica-** **tions. We design standard interfaces for biometric services, develop service request handler** modules running on the server side, and provide open services to network applications. **c. Biometric data are stored and processed in trusted environments to prevent pri-** **vacy leaks. On the terminal side, we use Trusted Executive Environments to protect** biometric data. On the server side, we use enclave technologies to ensure data security. We build secure channels between the terminal side and the server side using cryptographic technologies to ensure data security during transmission. The BioShare framework is designed to provide open and secure biometric services with secure data storage both on terminal TEEs and server enclaves, and through standard interfaces between terminal–server interactions. We implemented BioShare with facial recognition as an example, and measured the performance of the implementation. Based on the test results, the end-to-end response time is 91 ms to 1011 ms, which meets business requirements in most human interactive scenarios, and the security strength is higher than 128 bits, which fulfils the recommended standard for business systems. The rest of the paper is organized as follows: We enumerate related works in Section 2, describe key parts of our current design in Section 3, illustrate key challenges with a sample implementation in Section 4, and then measure and analyze the performance in Section 5. Finally, we discuss several open issues in Section 6 and conclude in Section 7. **2. Related Works** _2.1. Biometric Identification on Personal Mobile Devices_ Currently most biometric services on mobile terminals are implemented as local processing solutions, which is to say that they are secure but restricted to local mobile apps. Since fingerprint identification was introduced into iPhone 5s smartphones by Apple in 2013, biometric recognition technologies have undergone exponential development in the smartphone market. Apple provides biometric services Touch ID [7] and Face ID [8] on iOS platforms, and Microsoft offers Windows Hello [9] on personal Windows devices. Many proven biometric technologies are widely integrated into mobile devices: fingerprint [10,11], face [12–14], iris [15–17], vein [18,19], etc., and many other biometric modalities are under study: gait [20], keystroke [21], handwriting [22], etc. At the same time, auxiliary methods are used to prevent spoofing and attacks, such as liveness ----- _Appl. Sci. 2022, 12, 10782_ 3 of 21 detection [23–25], 3D recognition [26,27], micro-expression recognition [28,29], and movement detection such as head-shaking or blinking [30,31]. In general, biometric services are conducted in terminal trusted environments and are provided for device unlocking, online payment, and bank account verification, etc. However, these services are limited to local apps and are not available for remote network applications. _2.2. Trusted Executive Environment_ In local processing solutions, biometric data are processed and stored in local secured environments to ensure data privacy and service security. Currently, operating systems for both terminals and servers have become huge and complex because they must support multiple devices and rich applications. As a result, they have become more and more vulnerable, because system software vulnerabilities depend on code size and complexity [32]. Therefore, current CPUs offer a special execution environment isolated from the OS, such as ARM TrustZone [33–35], and Intel SGX [36,37] and RISC-V Keystone [38]). On the terminal market, based on hardware capabilities, terminal manufacturers design and implement native TEE OS software, such as Qualcomm QSEE [39], Google Trusty TEE [40], and Huawei iTrustee, etc. GlobalPlatform defines GPTEE specifications [41,42] to resolve compatibility issues between different TEE OSes, and UnionPay releases TEEI infrastructure [43] to support the coexistence and intercommunication of multiple TEEs in the same CPU. On the server market, CPU manufacturers release original Software Development Kits (SDKs) for different CPU models, such as Intel SGX SDK [44] and RISC-V Keystone SDK [45], and some other companies provide unified open-source solutions to resolve the compatibility issues of different SDKs, such as Microsoft Open Enclave SDK [46], Google Asylo [47], and Huawei secGear [48], etc. Currently TEE technology is used on either the terminal side or the server side to provide trusted computation environments for local applications; however, no solution is available to combine the terminal TEEs and server enclaves to form an integrated trusted environment. _2.3. Payment Platforms Based on Facial Recognition_ As a typical implementation of centralized-processing solutions, most facial recognition payment systems are open to network applications, but they process user biometric data in unsecure environments on public servers. Since the first facial recognition payment system was launched by Uniqul in Finland in 2013 [49], the paying-with-your-face service is becoming more and more popular in China. Alipay launched smile-to-pay products named Dragonfly [50], Tencent issued face-payment devices named Frog [51], and UnionPay cooperated with 60+ banks to provide facial payment solutions to merchants [52]. All of these cases have one thing in common: Users must upload private biometric data to service providers, and service providers store and process user biometric data in centralized mode on background servers. However, when biometric data are stored by service providers in their database servers, users are put at risk of privacy leakages due to improper maintenance or malicious attacks. When biometric data are processed by service providers without strict regulation and independent audit, users will lose control of their biometric data and face the abuse of personal data. **3. BioShare Design** The goal of BioShare is to provide trusted biometric authentication services to network applications under users’ full control of personal biometric data. It is designed to achieve: **a. Self-determination: As the owner of personal biometric data, each user is entitled** to grant or cancel the authorization to the server at any time, and can specify or change authorization conditions such as a specified number of times, specified time period, whitelist, or blacklist. ----- _Appl. Sci. 2022, 12, 10782_ 4 of 21 **b. Open Service: Biometric authentication services can be opened up to network appli-** cations rather than being restricted to local applications, and therefore, standard interfaces should be defined for the external invocation and internal processing of service requests. **c. Privacy Protection: All biometric data must be stored and processed in trusted** environments to prevent privacy leaks, and all critical operations such as encryption and decryption must be conducted in trusted environments to ensure security. The following subsections describe how our BioShare design achieves the aforementioned goals. First, we provide an overview of our framework, and then we describe the key components of our framework: the terminal-side modules, including User Application and Trusted Application, the server-side modules, including Service Application and Enclave Module, and the secured channels between both sides. _3.1. Overview_ We designed a unified and scalable infrastructure for BioShare that combines the capabilities of both the terminals and the server to conduct trusted biometric authentication and provide open services to various network applications under user authorization. Figure 1 shows a high-level view of the BioShare framework. It consists of five main components: **a. User Application (UA). BioShare UA is an app deployed on user terminals to** conduct user-interactive actions, such as user signup, service registration, and authorizing the server to perform biometric authentication. **b. Service Application (SA). The SA is designed to handle service requests from** network applications and calls Trusted Application (TA) or Enclave Module (EM) to conduct biometric authentication. **c. Trusted Application (TA). The TA provides trusted biometric services, such as** biometric recognition and comparison, in trusted executive environments (TEE) on the terminal side. **d. Enclave Module (EM). The EM provides trusted biometric services on the server** side under user authorization. **e. Secured Communication Module (SCM). The SCM provides secure channels** between user terminals and the server. **Figure 1. BioShare Architecture.** In the design of BioShare, standard interfaces are defined as API libraries for the external invocation and internal processing of service requests, as shown in Table 1. This facilitates the development and deployment of new biometric authentication services for open access under user control. For example, a new iris authentication service can be implemented by replacing the biometric recognition algorithm and making minor changes to the ----- _Appl. Sci. 2022, 12, 10782_ 5 of 21 User Application. Current terminal manufacturers can easily extend their biometric services to network application scenarios with unified infrastructure and standard interfaces. **Table 1. Standard Interface for BioShare.** **Module** **API Function** **Description** **(Internal Interface) Handle authentication requests from SA by calling** TA function, parameters including biometric type and biometric data **(Internal Interface) Acquire biometric raw data of the current user** with hardware sensors, conduct biometric recognition on raw data, and store feature data into secure storage parameters including biometric type (such as facial, fingerprint, etc.) **(Internal Interface) Conduct biometric comparison between biometric** data input and feature data of the current user, parameters including biometric type, and biometric data **(Internal Interface) export the biometric feature data of the current** user from TEE in an encrypted format, parameters including biometric type **(Open Service Interface) Handle biometric authentication requests** from Network Applications, parameters including user ID, biometric type and biometric data **(Internal Interface) Handle register-service requests from User Appli-** cation, inputs including user ID, terminal ID, and additional parameters. **(Internal Interface) Handle authorize-server requests from User Ap-** plication, parameters including user ID, terminal ID, biometric type, encrypted biometric feature data, and authorization options. **(Internal Interface) Conduct biometric comparison between feature** data input and feature data of user input, parameters including user ID, biometric type, and the biometric data input **(Internal Interface) Import the biometric feature data of a user into** server enclave, parameters including user ID, biometric type and feature data in an encrypted format **(Internal Interface) Encrypt a message with the symmetric encryption** algorithm, parameters including plain text message **(Internal Interface) Decrypt a message with the symmetric encryption** algorithm, parameters including the encrypted message **(Internal Interface) Sign a message, parameters including the message** **User** **Application** **(UA)** **Trusted** **Application** **(TA)** **Service** **Application** **(SA)** **Enclave** **Module** **(EM)** **Secured** **Communication** **Module** **(SCM)** bool authenticate( int biometricType, byte[] biometricData) bool acquireFeatureData( int biometricType) bool authenticate( int biometricType, byte[] biometricData) byte[] exportFeatureData( int biometricType) bool authenticate( String userId, int biometricType, byte[] biometricData) bool registerService( String userId, String terminalId, Map<String,String> params) bool authorizeServer( String userId, String terminalId, int biometricType, byte[] featureData, Map<int,object> options) bool authenticate( string userId, int biometricType, byte[] biometricData) bool importFeatureData( String userId, int biometricType, byte[] encryptedFeatureData) byte[] encryptMessage( byte[] plainMsg) byte[] decryptMessage( byte[] encryptedMsg) byte[] signMessage( byte[] msg) _3.2. User Application (UA)_ A key design principle for BioShare is that the framework should enable each user to control his/her biometric data handily and to authorize the server efficiently on personal mobile terminals. We develop a user application and deploy it on user terminals. The UA handles user commands as illustrated in Algorithm 1. The major functions include: **User Signup: Each user inputs his/her basic information, including user name, mobile** phone number, identity card number, etc., and UA verifies the authenticity of data input and creates an account for the user. Please be reminded that each user can log in on multiple terminals simultaneously with the same account. **Service Registration: Currently, most mobile terminals are capable of providing bio-** metric services to local apps. We extend the service object from local applications to ----- _Appl. Sci. 2022, 12, 10782_ 6 of 21 network applications through a service-registration mechanism: a. The user registers terminal biometric capabilities to the server in UA and the server maintains a user-registration database. b. When the server receives biometric authentication requests for the user from network applications, it will forward the requests to the user’s terminals through standard interfaces of UA, listed in Table 1. c. On the terminal side UA coordinates with Trusted Application (TA) to conduct biometric authentication, and return the result to the server. The service-registration mechanism successfully expands the biometric capabilities of user terminals to empower network applications, but the service may sometimes be unstable due to network issues or terminal power-off. **Authorizing The Server: BioShare is designed as a general framework for various** biometric services, including facial authentication, fingerprint authentication, and other biometric services, so users can choose a specific biometric type and authorize the server to conduct the biometric authentication. To provide stable services, each user can authorize the server to conduct biometric authentication with the following process: a. In UA, the user authorizes the server to conduct specified biometric authentication (such as facial authentication, fingerprint authentication and so on) with specified options, such as a specified number of times, specified time period, whitelist or blacklist, etc., and the server updates authentication information of the user in the user-registration database. b. UA exports the user’s private biometric data from the terminal TEE and imports it into the server enclave, and the data will be stored temporarily in the server enclave according to authorization options. c. When the server receives biometric authentication requests from network applications, it will call the Enclave Module to conduct biometric authentication with stored biometric data and return the result. Each user may cancel his/her authorization anytime and anywhere, and accordingly, his/her private data will be purged from the server enclave. **Algorithm 1 Request-Handling Process in User Application** **Input: The ID Value of the Current User, userId;** Request Action, requestAction; Request Data, requestData **Output: Result (true or false), returnResult** 1: Initialize returnResult with 0 2: if requestAction = UserSignUp then 3: Verify requestData 4: Call TA to acquire biometric data with hardware 5: Create a new account for userId 6: **return true** 7: else if requestAction = ServiceRegistration then 8: Call SA to create new UserRecord into database set UserRecord.UserId=userId and UserRecord.TerminalId=requestData.TerminalId 9: **return true** 10: else if requestAction = AuthorizeServer then 11: Call SA to update database set serRecord.ServerAuthorized=true and serRecord.AuthorizeOptions=requestData.AuthorizeOptions where UserRecord.UserId=userId and UserRecord.TerminalId=requestData.TerminalId 12: Call SCM to build a secure channel between TA and EM 13: Transfer Biometric Data of the current user from TA to EM 14: **return true** 15: else if requestAction = AuthenticateUser then 16: Validate the service request 17: Call TA to conduct biometric authentication 18: **return authentication result from TA** 19: end if 20: return returnResult ----- _Appl. Sci. 2022, 12, 10782_ 7 of 21 _3.3. Service Application (SA)_ BioShare achieves the goal of providing open services via the Service Application running on the server, and Algorithm 2 shows the process flow in detail. SA listens for biometric authentication requests from remote network applications, verifies the request data, and checks the user in the user-registration database. If the user has registered the biometric service, SA will process the request with either of the following modes: **Forwarding Mode:** If the user does not authorize the server to conduct biometric authentication, then SA will forward the request to the user terminals, and on the terminal side, UA coordinates with TA to conduct biometric authentication and returns the result to the server, as described in step “Service registration” of Section 3.2. **Authorized Mode: If the user has authorized the server, SA will call Enclave Module** to conduct biometric authentication and return the result, as described in step “Authorizing the server” of Section 3.2. The BioShare Service Application is also designed to handle command requests from the users, including service-registration requests and authorizing-the-server requests, which are described in Section 3.2. The server maintains a user-registration database that records the registration and authorization information for each user. **Algorithm 2 Process for Handling Service Requests from Network Applications** **Input: The ID of The Target User, userId** Biometric Modality Value, biometricType Biometric Feature/Raw Data, biometricData **Output: Authentication Result, returnResult** (0-false/1-true/-1-failed) 1: Initialize returnResult with -1 2: Receive service requests from Network Applications 3: Verify input data 4: Search database for UserRecord where UserRecord.UserId=userId 5: if UserRecord IS NOT NULL then // user registered 6: **if UserRecord.ServerAuthorized = false then** //forwarding mode 7: Connect to the terminal with UserRecord.TerminalId 8: Call SCM to build a secure channel to TA through UA 9: Call TA to conduct biometric authentication 10: _returnResult = authentication result of TA_ 11: **else// authorized mode** 12: Call EM to conduct biometric authentication 13: _returnResult = authentication result of EM_ 14: **end if** 15: end if 16: return returnResult _3.4. Trusted Application (TA)_ The trusted application is a module running in terminal TEE that receives calls from the UA and conducts biometric operations through the standard interfaces listed in Table 1. Because all actions are performed in the terminal TEE and all private data are stored in tamper-resistant secure storage, biometric services from TA are secure and trusted. **Biometric Data Acquisition: TA captures biometric raw data through biometric hard-** ware integrated into mobile terminals, such as fingerprint devices, cameras, and other devices. TA calls the hardware driver to control the biometric sensor, the sensor captures raw data and writes it to a memory buffer, and finally, the raw data are copied to secure memory. **Biometric Recognition: Biometric recognition transmutes biometric raw data into** feature data in the terminal TEE. A typical process of biometric recognition consists of three steps: a. Data processing. TA processes the raw data with specified algorithms, ----- _Appl. Sci. 2022, 12, 10782_ 8 of 21 such as biometric detection algorithms. b. Liveness detection. As an option, TA performs liveness detection with infrared images, RGB images, or depth images, etc. c. Feature extraction. TA extracts feature data from raw data using a specified model. Taking face recognition as an example, the typical process includes face detection, liveness detection, and facial feature extraction. **Secure Data Storage: Biometric feature data are stored in secure storage in terminal** TEEs and are protected by hardware isolation to ensure privacy protection, and only trusted applications (TAs) can obtain permissions to access biometric data in TEE mode. When biometric data are moved from secure storage to insecure environments, they are encrypted by the SCM module to ensure data security, as in Section 3.6. **Biometric Comparison: TA conducts a biometric comparison by calculating the simi-** larity between source feature values and the target feature values. If the similarity score exceeds the threshold value, TA returns a success, otherwise it returns a failure. To ensure security, private biometric data are transmitted to TA in encrypted formats, and are decrypted in the terminal TEE before biometric comparison. **Constraints: Terminal TEE is a restricted zone in mobile CPU platforms with limited** capabilities, and the limitations include CPU speed, maximum RAM, and maximum secure storage, etc. Table 2 lists the TEE capabilities of typical CPUs. We therefore chose algorithms and models with low computing and memory costs in the implementation of biometric recognition and comparison. **Table 2. TEE Capabilities of Typical CPUs.** **CPU Model** **TEE** **CPU Cores** **Max RAM** **Max Secure Storage** SamSung Exynos 1080 ARM Trustronic 8 100 MB 5 MB Qualcomm Snapdrago 888 QSEE 8 100 MB 16 MB Hisilicon Kirin 9000 iTrustee 8 48 MB 12 MB _3.5. Enclave Module (EM)_ The Enclave Module is a module running on the server side that conducts biometric recognition and comparison using the standard interfaces listed in Table 1. During the user-authorizing process, EM exports private biometric data from Terminal TEEs and stores it temporarily in the server enclave, as described in Section 3.2. Based on the data, EM can conduct biometric authentication without the intervention of user terminals. Because all actions are performed in the server enclave, biometric services provided by EM are secure and trusted. **Biometric recognition:** Biometric recognition transmutes biometric raw data into feature data. EM conducts biometric recognition with the same algorithm and the same steps as TA on the terminal side: data processing, and liveness detection and feature extraction, as described in Section 3.4. **Biometric comparison: EM on the server side conducts a biometric comparison with** the same algorithm as TA on the terminal side, which is described in Section 3.4. **Privacy Protection: Within BioShare, the security mechanisms for server enclaves are** used to protect biometric data during both storage and processing. Specifically, the EPC (Enclave Page Cache) protects all EM code and data in encrypted format during runtime state, the Sealing/Unsealing mechanism secures EM data in encrypted format during persistent storage, and the Attestation technology ensures that genuine code runs in secure enclave environments to avoid data interception and code tampering. **Constraints: Server enclaves are protected execution environments with restricted** capabilities, and the restrictions include available CPU cores, maximum RAM, and memory access latency, etc. Meanwhile, in EM, we have to choose the same algorithm as that of TA on the terminal side due to the incompatibility of different algorithms. ----- _Appl. Sci. 2022, 12, 10782_ 9 of 21 _3.6. Secured Communication Module (SCM)_ The Secured Communication Module ensures secure and undeniable communication between user terminals and the server through cryptographic techniques such as symmetric encryption, asymmetric encryption, and hash and digital signatures. To ensure security, all encryption/decryption actions are conducted in the trusted environment of either the server or the terminal. Secured communication channels combine terminal TEEs and the server enclave, forming an integrated trusted environment for trusted computing. **Secured Channels:** Before the server and the terminal communicate with each other, for example, the UA module on the terminal authorizes the server in Section 3.2 and Algorithm 1 and the SA module on the server forwards service requests to the user terminals in Section 3.3, we use an asymmetric encryption algorithm RSA (with a key length of 3072 bytes) to build secure communication channels between the server and the terminal. RSA algorithms with a key length of 7680 bytes will provide higher security, and alternative algorithms include ECC algorithms with key lengths of 256/384 bytes. First, random numbers are generated in the server enclave and terminal TEEs, then public/private key pairs are generated on both sides, based on the random numbers with an asymmetric encryption algorithm (i.e., the RSA or ECC algorithm). Finally, private keys are stored in trusted environments and public keys are exchanged with each other. Please be reminded that asymmetric encryption applies only to small data sets due to high computing costs. **Data Encryption: In BioShare, user biometric data are only processed in secure en-** vironments including terminal TEEs and server enclaves, and that are transferred in encrypted format outside secure environments and decrypted in secure environments to ensure data security and privacy protection. We use the symmetric encryption algorithm AES-128 to efficiently encrypt/decrypt the biometric data during terminal–server communications. The following are the main steps: a. A random number is generated for each user by the hardware device in the server enclave. b. A working key is generated for the user based on the random numbers using the symmetric encryption algorithm (i.e., AES), and saved in the server enclave. c. The working key is encrypted with the terminal public key and sent to the user’s terminal. d. The working key is decrypted and stored in the terminal TEE, so that both sides share the same working key. **Digital Signatures: To ensure nonrepudiation, we use digital signatures on both the** server and terminal sides. In the terminal TEE, a message can be signed with the following steps: a. Calculate the hash value for the target message with the hash algorithm SHA-256. b. Encrypt the hash value with the user’s private key. c. Send the message along with the ciphertext to the server. On the server enclave, the message can be verified via the following steps: a. Decrypt the ciphertext to a plaintext format. b. Calculate the hash value for the message with the SHA-256 algorithm. c. Check whether the hash value is consistent with the plaintext. Digital signatures are used in the return values of the SA module as a response to service requests from network applications as in Section 3.3 and Algorithm 2. **4. Sample Implementation** BioShare is currently implemented with facial recognition as an example, and deployed on Android mobile terminals with TEE communicating with a server with enclaves. The following sections describe several implementation details and challenges specific to our current implementation. _4.1. General Information_ In our implementation, we chose mainstream CPU architectures with dominant TEE platforms, specifically ARM Cortex-A architecture with ARM TrustZone on the terminal side, and Intel Xeon E3 CPU with Intel SGX support on the server side. Table 3 lists the configurations of the hardware and software on each side. ----- _Appl. Sci. 2022, 12, 10782_ 10 of 21 **Table 3. The Configurations of Mobile Terminals and the Server.** **Terminal** **Server** CPU Mediatek MT8788 (ARM Cortex-A) Intel Xeon E3 RAM 1.3 GB 64 GB TEE ARM TrustZone Intel SGX OS Android 9.0 CentOS Linux 7.4 TEE OS Nebula TEEI 1.1.0 _4.2. User Application (UA)_ The BioShare User Application is implemented as an Android app that is deployed on mobile terminals. The key challenge in the implementation of UA is to maintain a reliable connection between the terminals and the server in a changing environment. **TCP Connections: The user application either creates a short TCP/IP connection to** the server for each call, or keeps a long connection with the server for a certain period. The former requires lower resource costs, while the latter shows better performance. **Notification Mechanism: Because Android or iOS kills the application running in the** background automatically at irregular intervals, the User Application may be unreachable to the server from time to time. We need a notification mechanism through which the service application can notify the terminal to relaunch the user application at any time. Apple Push Notification Service [53], Google Firebase Cloud Messaging [54], and other third-party products provide reliable notification services for iOS and Android devices. _4.3. Service Application (SA)_ The BioShare Service Application is implemented as two system service processes running in the background. One service process is responsible for handling the command requests from user applications on the terminal side, while the other is responsible for handling the service requests from network applications. The two processes coordinate with each other by sharing the same user-registration database. We have developed two different experiments in which network applications call the biometric service in either forwarding mode or authorized mode. Experimenters can currently register terminal biometric capabilities to the server in the User Application on the terminal side, then the biometric service on the server will run in forwarding mode and all service requests from network applications will be forwarded to the user’s terminals registered. Furthermore, experimenters can also authorize the server in the User Application; accordingly, the biometric service on the server will run in authorized mode and all service requests from network applications will be handled locally by the Enclave Module on the server. _4.4. Trusted Application (TA)_ The BioShare Trusted Application is implemented as a trusted module deployed in the Trusted Executive Environment on the terminal side. The key challenge is the choice of the appropriate biometric algorithms and implementations for facial recognition and comparison, due to the limitations of TEE capabilities. We attempt to choose open-source implementations with complete functionality, good performance, low computing and memory overheads, embedded-environment support, and appropriate licensing agreements. **Facial Recognition and Comparison: Currently, many open-source projects for facial** recognition are available from GitHub, and Table 4 lists some popular projects. In our implementation, we chose SeetaFace2 because it is built with C++, is independent of third-party libraries, and is compatible with X86, iOS, and Android systems. We used pretrained models from the SeetaFace2 project. According to the SeetaFace2 document, Cascaded-CNN is used as a face-detection algorithm, achieving 92% in the FDDB public dataset. FEC-CNN is used as a face landmaker algorithm, achieving a 0.069 average ----- _Appl. Sci. 2022, 12, 10782_ 11 of 21 positioning error on the 300-W Challenge public dataset. ResNet50 is used for facial feature extraction/comparison with 25 million parameters, and the model has been pretrained with 33 million photos with an accuracy rate of more than 98% in the general 1:N scenario on a 1000-person dataset when the error acceptance rate is 1%. Figure 2 shows the flow chart of the processes. **Figure 2. The flow chart of face recognition.** **Table 4. Typical Opensource Projects for Facial Recognition.** **Functionality** **Hardware support** **Open-Source Project** **Language** **License** **Recognition** **Comparison** **X86** **Embedded** **SeetaFace Engine** Yes Yes Yes Yes C++ BSD **SeetaFace Engine2** Yes Yes Yes Yes C++ BSD **SeetaFace2** Yes Yes Yes Yes C++ BSD **FaceBoxes** No No Yes Yes C++ BSD **libfacedetection** No No Yes Yes C BSD **OpenCV 4** Yes Yes Yes No C++ BSD **RetinaFace** Yes Yes Yes No Python MIT **deepinsight/insightface** Yes Yes Yes No Python MIT **Liveness Detection: We chose the open-source project FeatherNets for liveness detec-** tion in the implementation, because several models provided by the project achieve the goal of low computing cost (about 80 M FLOPs), small model size (0.35 M parameters), and complete functionality (applicable to infrared images and depth images, etc.). Table 5 shows the test results of the models within the FeatherNets project. **Table 5. Test Results of Liveness Detection Models.** **Model** **ACER** **TPR@FPR = 1%** **TPR@FPR = 0.1%** **FLOPS** FishNet150 0.00144 0.999668 0.998330 6452.72M FishNet150 0.00181 1.0 0.9996 6452.72M FishNet150 0.00496 0.998664 0.990648 6452.72M MobileNet v2 0.00228 0.9996 0.9993 306.17M MobileNet v2 0.00387 0.999433 0.997662 306.17M MobileNet v2 0.00402 0.9996 0.992623 306.17M MobileLiteNet54 0.00242 1.0 0.99846 270.91M MobileLiteNet54-se 0.00242 1.0 0.996994 270.91M FeatherNetA 0.00261 1.00 0.961590 79.99M FeatherNetB 0.00168 1.0 0.997662 83.05M **Ensembled all** **0** **1** **1** **-** _4.5. Enclave Module (EM)_ The BioShare Enclave Module is implemented as trusted code that runs in the Intel SGX enclave on the server side. On other TEEs such as ARM TrustZone and RISC-V keystone, we have to develop new code due to incompatibilities between them. In order to achieve the goal “develop once and deploy anywhere”, we use Microsoft Open Enclave SDK [46] to resolve compatibility issues between different TEEs, and other opensource projects including Google Asylo [47] and Huawei secGear [48] etc., provide similar functionality. ----- _Appl. Sci. 2022, 12, 10782_ 12 of 21 We have developed invocation interfaces to authenticate both facial image data and facial feature data, and experimenters can specify either raw data or feature data in service requests from network applications. We use the open-source code SeetaFace2 to conduct facial recognition and comparison. If the input is facial feature data, the service request will be processed with high performance due to the low computing overhead of the feature comparison algorithm, but this is achieved with low compatibility because the input feature data must be extracted with the same algorithm as that of the Enclave Module. If the input is facial image data, the service request will be processed with higher computing costs but with better compatibility. The experimental results are described in Section 5.2. _4.6. Secured Communication Module (SCM)_ The BioShare Secured Communication Module is implemented as trusted functions running in trusted environments on both the terminal and server sides to provide cryptographic services. In our implementation, we choose Advanced Encryption Standard (AES) as the symmetric encryption algorithm and specify 128 or 192 bytes as the key length to ensure a sufficient encryption intensity. We select RSA or ECC as the asymmetric encryption algorithm, specify 3072 or 7680 bytes as the key length for RSA, and 256 or 384 bytes as the key length for ECC; and we specify the SHA-256 algorithm in the hash value calculation. During the entire lifecycle of each transaction, all biometric data are transmitted in encrypted format in untrusted environments, and decrypted and processed only in trusted environments; and the result is returned with a digital signature to ensure non-repudiation. **5. Evaluation and Analysis** We now evaluate and analyze BioShare in terms of terminal-side performance, serverside performance, communication performance, and holistic security. _5.1. Overview_ BioShare processes service requests from network applications in either forwarding mode or authorized mode, and we measure the time overhead of processing a service request in both modes. In forwarding mode, the server forwards the service request to the user terminal, as shown in Figure 3. The total overhead of processing a service request varies from 91 ms to 1011.5 ms, and the major parts are the request-processing cost and the server-client communication cost, which are evaluated in Sections 5.2 and 5.4. In authorized mode, the server processes service requests in the server enclave under user authorization, as shown in Figure 4. The total overhead of processing a service request varies from 16.1 ms to 168.6 ms, and the major part is the request-processing cost, which is evaluated in Section 5.3. **Figure 3. Overhead Imposed by BioShare in Processing a Service Request with Forwarding Mode.** ----- _Appl. Sci. 2022, 12, 10782_ 13 of 21 **Figure 4. Overhead Imposed by BioShare in Processing a Service Request with Authorized Mode.** _5.2. Performance Evaluation of Terminal TEE_ We measure the performance of the terminal TEE in conducting facial authentication upon different data inputs. The overhead of face recognition is 507.1 ms upon facial raw image input, while the overhead of face comparison reduces to 1.6 ms upon facial feature data input, as shown in Table 6. Figure 5 shows the cumulative distribution of time consumption in face recognition, and we can see that the performance of the terminal TEE is quite stable. We perform further research on each step of face recognition and find that feature extraction and liveness detection are the most time-consuming steps, accounting for 70.9% and 22.9%, respectively, as shown in Figure 6. **Table 6. Overhead of TA in Processing Facial Recognition.** **Input** **Action** **Time (ms)** **Facial Image Data** Face Recognition 507.1 **Facial Feature Data** Face Comparison 1.6 **Figure 5. CDF of Time Consumption of Face Recognition.** We measure the prediction accuracy rate of facial recognition with an actual business dataset of 329 pre-captured photos with a resolution of 640 480 from 110 employees. _×_ In the test, we compare each photo in the dataset with other photos, and the test results are shown in Table 7. The accuracy rate is 99.937%, the false acceptance rate (FAR) is 0.054%, and the false rejection rate (FRR) is 1.524%, which meets business requirements in most scenarios. ----- _Appl. Sci. 2022, 12, 10782_ 14 of 21 **Figure 6. Overhead Proportion of Each Step in Facial Recognition.** **Table 7. Test Results of Facial Recognition and Comparison.** **Prediction Result** **Actual Result** **TRUE** **FALSE** **TRUE** 323 (TP) 5 (FN) **FALSE** 29 (FP) 53599 (TN) We measure the performance difference between the terminal TEE and the server enclave during facial authentication. To ensure the comparability of experimental results, we specify that the same algorithmic code runs in a single thread on a single CPU core in both environments. The evaluation results show that the face-recognition cost of the terminal TEE is more than three times that of the server enclave, the face-comparison overhead on both sides is as low as 1–2 ms, and the face-recognition accuracy is equal on both sides. Please refer to Figure 7 for more details. **Figure 7. Performance Difference between Terminal TEE and Server Enclave.** _5.3. Performance Evaluation of Server Enclave_ The key challenge with server-side facial authentication is the performance of the server enclave, which differs across CPU platforms. Our demo implementation works fine on the Pentium Silver CPU enclave with four physical cores and 128 MB RAM at most for ----- _Appl. Sci. 2022, 12, 10782_ 15 of 21 low or moderate load scenarios, but more powerful enclave environments with more CPU cores and higher RAM support are required in high load scenarios. Table 8 lists some key indicators of enclaves on specific Intel CPU platforms. **Table 8. Key Indicators of Enclaves on Different CPU Platforms.** **Key Indicators** **Pentium Silver** **Xeon E3** **Xeon SP Ice Lake** Maximum Physical Cores 4 6 80 Maximum Encalve RAM 128 MB 128 MB 1 TB Encalve Dynamic Memory Management Yes No Yes Generally, the performance overhead of server enclaves consists of three major parts: enclave memory access latency, enclave switching cost, and enclave dynamic-memorymanagement cost. Figure 8 lists the evaluation results of the three parts in professional tests [55]. Specifically, on the Intel Xeon Icelake platform, the overhead caused by enclave memory access latency is almost negligible, the enclave switching cost is less than 10,000 CPU lifecycles, and the enclave dynamic-memory-management cost is quite high. The performance analysis is instructive for further optimization of the current implementation. (a) Enclave Memory Access Latency (b) Enclave Switching Cost (c) Cost of Allocating Enclave Memory Pages (d) Cost of Deallocating Enclave Memory Pages **Figure 8. Performance Overhead of Enclaves on Intel SGX Platforms.** _5.4. Performance Analysis of Mobile Communication_ Most user terminals communicate with the server through short or long TCP/IP connections over a 4G/5G network, and we measure the end-to-end delays of different communication methods. Based on the experiment, the average delay is 39.7 ms for one-way trips between the terminal and the server, long TCP/IP connections have lower latency than short connections (average 21.1 ms vs. 58.3 ms), and 5G networks show better performance than 4G (average 27.3 ms vs. 52.1 ms). Figure 9 shows the experimental results in detail. ----- _Appl. Sci. 2022, 12, 10782_ 16 of 21 **Figure 9. CDF of End-to-End Communication Delays.** We have performed further research on the performance of the notification mechanism used in BioShare. We kill the user application on the user terminal, relaunch it with the notification services [56], and measure the communication delay from the server to the terminal. Based on the experiment, it takes 415 ms for Aurora Mobile JPush to notify the terminal from the server, which is close to the reported time delay for Apple Push Notification Service and Google Firebase Cloud Messaging, as shown in Table 9. **Table 9. Normal Delay of Notification Services.** **Provider** **Platform** **Notification Service** **Delay (ms)** Apple IOS Apple Push Notification Service 400 Google Android Firebase Cloud Messaging 500 Aurora Mobile Android/iOS Aurora Mobile JPush 415 _5.5. Security Analysis of BioShare_ We evaluate the security level of secure communication channels between the terminal TEE and the server enclave by analyzing the cryptographic algorithms used in BioShare, as shown in Table 10. In the implementation, we specify key lengths of 128/192 for the AES algorithm; thus, the security strength of symmetric encryption is 128/192. We specify key lengths of 3072/7680 for the RSA algorithm and key lengths of 256/384 for the ECC algorithm, and therefore, the security strength of asymmetric encryption is 128/192. We use SHA-256 as the hash algorithm, and the security strength is 256. In conclusion, the security strength of secure communication channels is 128/192, which fulfils and even exceeds the values recommended by the National Institute of Standards and Technology (NIST) [57]. The terminal TEE empowers the Trusted Application with secure hardware, secure OS, secure storage, and secure provisioning, etc. The server enclave empowers the Enclave Module with isolated execution, encrypted RAM, secure storage, sealing and attestation mechanisms, etc. The TEEs on both terminal and server sides provide trusted environments to process and store biometric data at a high security level, and Table 11 lists some security properties of trusted environments. ----- _Appl. Sci. 2022, 12, 10782_ 17 of 21 **Table 10. Security Strength of Cryptographic Algorithms of BioShare.** **NIST Recommendations** **Algorithms of BioShare** **Category** **Algorithms** **(Key Length)** **Algorithm** **Key Length** **Security Strength** AES-128 128 128 Symmetric encryption AES 128+ AES-192 192 192 RSA 3072 128 RSA 2048+ RSA 7680 192 Asymmetric encryption ECC 256 128 ECC 224+ ECC 384 192 Hash SHA 224+ SHA-256 256 256 **Table 11. Security Properties of Terminal TEE and Server Enclave.** **Server Enclave** **(Intel SGX)** **Security Properties** **Terminal TEE** **(ARM Trustzone)** Isolated Execution Yes Yes Encrypted RAM No Yes Secure Storage Yes Yes Remote Attestation Yes Yes Secure provisioning Yes Yes Privileged software attack defense Yes Yes Trusted Path Yes No **6. Discussion and Future Work** This paper makes the case for biometric authentication as a trusted service under users’ full control, and demonstrates how we addressed several challenges in making this service secure, practical, and efficient. Our solution reconciles the demand for open services and the requirement for privacy security, and shows comparative advantages over the currently available solutions. Table 12 lists the details of the comparisons. When compared to terminal local-processing solutions, such as Apple Touch ID and Windows Hello, as shown in Section 2.1, our solution extends open biometric authentication services from local mobile apps to remote network applications with high security levels. When compared to centralized-processing solutions such as facial recognition payment systems, as listed in Section 2.3, BioShare can avoid both unintentional leakages and the intentional abuse of biometric data by keeping private data in an encrypted format within secure storage, and running trusted code in isolated secure environments. Based on our test results, the BioShare implementation meets the performance requirements of most humaninteractive businesses and fulfills the security requirements of recommended standards for business systems. We now discuss important challenges that are outside the focus of this paper. **TEE Fragmentation: BioShare relies on terminal TEEs and server enclaves to handle** service requests; however, the TEE market as a whole is currently fragmented. In both terminal and server markets, different manufacturers offer different TEE solutions that are incompatible with each other, such as QSEE/iTrustee on the terminal side and SGX/TrustZone on the server side, and therefore, we have to develop native code for each manufacturer, and even for each system model. This tremendously increases the development costs and implementation complexity for large-scale applications of BioShare, and we need a universal standard and interface protocol for terminal TEEs as well as server enclaves. ----- _Appl. Sci. 2022, 12, 10782_ 18 of 21 **Table 12. Security Properties of Terminal TEE and Server Enclave.** **Centralized-Processing** **Objectives** **Local-Processing Solutions** **BioShare** **Solutions** Yes (Each user can authorize the server to conduct biometric authentication or cancel the authorization at any time) Each user has full control of his/her biometric data Yes (Biomatric data are stored in the local secure storage of pri- No vate mobile terminals) No (The service is restricted to Yes (The service is open Yes (The service is open to network Open biometric services local mobile apps) to network applications) applications) Nonsecure storage Secure storage (Terminal TEEs and Biometric data storage Secure storage (Terminal TEEs) (Public databases) Server Enclaves) Biometric data processing Trusted Environment (Terminal Nontrusted environ- Trusted Environment (Terminal TEEs environment TEEs) ment (Public servers) and Server Enclaves) Biometric data Not applicable (Local Secure channels, the data are transmitSecure/unsecure transmission processing) ted in decrypted format Prevent data leakage Yes No Yes Resisting data interception Yes (Attestation mechanism of Yes (Attestation mechanism of server No and code tampering attacks terminal TEE) Enclave) Resisting memory dump Yes (Enclave memory protection No No attacks mechanism) Resisting privileged code Yes (Isolated Execution mecha- Yes (Isolated Execution mechanisms No attacks nism of terminal TEEs) of terminal TEEs and server Enclaves) **Network Attacks: BioShare opens up terminal capabilities to network applications,** meaning that the terminals are faced with various network attacks, such as replay attacks and man-in-the-middle attacks. In addition, the server connects to numerous terminals through TCP/IP connections, and therefore is put at risk of cyber-attacks such as impersonation attacks and distributed denial of service attacks. We need to design defense technology and systems against various network attacks on user terminals, as well as the BioShare server. **Business Model: BioShare relies on users to install the User Application on individual** terminals, raising the question of whether there are proper incentives for users to do so. We need a business model for BioShare to ensure that all participants can benefit from service provision. One possible solution is to charge a fee for each service request from network applications, and reward the user with a certain proportion of the income. **7. Conclusions** In this paper, we make the case for a unified framework that provides trusted biometric authentication services to network applications under the user’s full control of personal biometric data. Our demo implementation of BioShare provides: (a) authorized and forward service modes to ensure users’ full control of personal data, (b) standard interfaces to ensure service openness to network applications, (c) trusted computation to prevent privacy leaks on both terminal and server sides, (d) secure channels to ensure data security during transmission, and (e) a unified framework to ensure effective collaboration between user terminals and the server. We showed that our system is efficient, based on standard interfaces, and provides a unified terminal–server collaboration mode to support new biometric services under user control. As part of our future work, we are designing standard TEE solutions to facilitate the development of new biometric service experiments, developing defense technology and systems against various network attacks, and exploring business models to ensure that each user benefits from participation. ----- _Appl. Sci. 2022, 12, 10782_ 19 of 21 **Author Contributions: Conceptualization, Q.S. and J.W.; methodology, Q.S.; investigation, Q.S. and** W.Y.; formal analysis, Q.S. and W.Y.; resources, Q.S. and J.W.; validation, Q.S. and W.Y.; visualization, Q.S.; writing—original draft preparation, Q.S. and W.Y.; writing—review and editing, Q.S. and J.W. All authors have read and agreed to the published version of the manuscript. **Funding: This work was supported in part by the Program of Shanghai Academic/Technology** Research Leader under Grant 19XD1433700. **Institutional Review Board Statement: Not applicable.** **Informed Consent Statement: Not applicable.** **Data Availability Statement: Not applicable.** **Acknowledgments: The National Engineering Laboratory for Electronic Commerce and Electronic** Payment provided the experimental environment for the model validation in this article, and we are thankful for suggestions from academician Chai Hongfeng and manager Chen Chengqian for model improvements. **Conflicts of Interest: The authors declare no conflict of interest.** **References** 1. [General Data Protection Regulation. Available online: https://gdpr-info.eu/ (accessed on 24 February 2022).](https://gdpr-info.eu/) 2. Gobeo, A.; Fowler, C.; Buchanan, W.J. GDPR and Cyber Security for Business Information Systems. In GDPR and Cyber Security _[for Business Information Systems ; River Publishers: Gistrup, Denmark, 2020; pp. i–xix. [CrossRef]](http://doi.org/10.1201/9781003338253)_ 3. Layton, R.; Elaluf-Calderwood, S. A Social Economic Analysis of the Impact of GDPR on Security and Privacy Practices. In Proceedings of the 2019 12th CMI Conference on Cybersecurity and Privacy (CMI), Copenhagen, Denmark, 28–29 November [2019; pp. 1–6. [CrossRef]](http://dx.doi.org/10.1109/CMI48017.2019.8962288) 4. [Bonta, R. California Consumer Privacy Act (CCPA). Available online: https://oag.ca.gov/privacy/ccpa (accessed on 24 February 2022).](https://oag.ca.gov/privacy/ccpa) 5. Stallings, W. Handling of Personal Information and Deidentified, Aggregated, and Pseudonymized Information under the [California Consumer Privacy Act. IEEE Secur. Priv. 2020, 18, 61–64. [CrossRef]](http://dx.doi.org/10.1109/MSEC.2019.2953324) 6. [Personal Information Protection Law. Available online: http://www.npc.gov.cn/npc/c30834/202010/569490b5b76a49c292e64c4](http://www.npc.gov.cn/npc/c30834/202010/569490b5b76a49c292e64c416da8c994.shtml) [16da8c994.shtml (accessed on 24 February 2022).](http://www.npc.gov.cn/npc/c30834/202010/569490b5b76a49c292e64c416da8c994.shtml) 7. [About Touch ID Advanced Security Technology. Available online: https://support.apple.com/en-us/HT204587 (accessed on 24](https://support.apple.com/en-us/HT204587) February 2022). 8. [About Face ID Advanced Technology. Available online: https://support.apple.com/en-us/HT208108 (accessed on 24 February 2022).](https://support.apple.com/en-us/HT208108) 9. [Learn about Windows Hello and Set It up. Available online: https://support.microsoft.com/en-us/windows/learn-about-](https://support.microsoft.com/en-us/windows/learn-about-windows-hello-and-set-it-up-dae28983-8242-bb2a-d3d1-87c9d265a5f0) [windows-hello-and-set-it-up-dae28983-8242-bb2a-d3d1-87c9d265a5f0 (accessed on 24 February 2022).](https://support.microsoft.com/en-us/windows/learn-about-windows-hello-and-set-it-up-dae28983-8242-bb2a-d3d1-87c9d265a5f0) 10. Xu, Y.; Lu, G.; Lu, Y.; Zhang, D. High resolution fingerprint recognition using pore and edge descriptors. Pattern Recognit. Lett. **[2019, 125, 773–779. [CrossRef]](http://dx.doi.org/10.1016/j.patrec.2019.08.006)** 11. Jo, Y.H.; Jeon, S.Y.; Im, J.H.; Lee, M.K. Security analysis and improvement of fingerprint authentication for smartphones. Mob. Inf. _[Syst. 2016, 2016, 8973828. [CrossRef]](http://dx.doi.org/10.1155/2016/8973828)_ 12. Cavazos, J.G.; Phillips, P.J.; Castillo, C.D.; O’Toole, A.J. Accuracy comparison across face recognition algorithms: Where are we on measuring race bias? arXiv 2019. arXiv:1912.07398. 13. Rana, A.; Ciardulli, A. Identity verification through face recognition, Android smartphones and NFC. In Proceedings of the [World Congress on Internet Security (WorldCIS-2013), London, UK, 9–12 December 2013; pp. 162–163. [CrossRef]](http://dx.doi.org/10.1109/WorldCIS.2013.6751039) 14. Baqeel, H.; Saeed, S. Face detection authentication on Smartphones: End Users Usability Assessment Experiences. In Proceedings of the 2019 International Conference on Computer and Information Sciences (ICCIS), Aljouf, Saudi Arabia, 3–4 April 2019; pp. 1–6. [[CrossRef]](http://dx.doi.org/10.1109/ICCISci.2019.8716452) 15. Hongo, K.; Takano, H. Personal Authentication with an Iris Image Captured Under Visible-Light Condition. In Proceedings of the 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC), Miyazaki, Japan, 7–10 October 2018; pp. 2266–2270. [[CrossRef]](http://dx.doi.org/10.1109/SMC.2018.00389) 16. Ali, S.A.; Shah, M.A.; Javed, T.A.; Abdullah, S.M.; Zafar, M. Iris recognition system in smartphones using light version (LV) recognition algorithm. In Proceedings of the 2017 23rd International Conference on Automation and Computing (ICAC), [Huddersfield, UK, 7–8 September 2017; pp. 1–6. [CrossRef]](http://dx.doi.org/10.23919/IConAC.2017.8082011) 17. Raja, K.B.; Raghavendra, R.; Busch, C. Smartphone based robust iris recognition in visible spectrum using clustered K-means features. In Proceedings of the 2014 IEEE Workshop on Biometric Measurements and Systems for Security and Medical [Applications (BIOMS) Proceedings, Rome, Italy, 17 October 2014; pp. 15–21. [CrossRef]](http://dx.doi.org/10.1109/BIOMS.2014.6951530) 18. [Garcia-Martin, R.; Sanchez-Reillo, R. Vein Biometric Recognition on a Smartphone. IEEE Access 2020, 8, 104801–104813. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.3000044) 19. Garcia-Martin, R.; Sanchez-Reillo, R. Deep Learning for Vein Biometric Recognition on a Smartphone. IEEE Access 2021, 9, [98812–98832. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2021.3095666) ----- _Appl. Sci. 2022, 12, 10782_ 20 of 21 20. Zou, Q.; Wang, Y.; Wang, Q.; Zhao, Y.; Li, Q. Deep Learning-Based Gait Recognition Using Smartphones in the Wild. IEEE Trans. _[Inf. Forensics Secur. 2020, 15, 3197–3212. [CrossRef]](http://dx.doi.org/10.1109/TIFS.2020.2985628)_ 21. Coakley, M.J.; Monaco, J.V.; Tappert, C.C. Keystroke biometric studies with short numeric input on smartphones. In Proceedings of the 2016 IEEE 8th International Conference on Biometrics Theory, Applications and Systems (BTAS), Niagara Falls, NY, USA, [6–9 September 2016; pp. 1–6. [CrossRef]](http://dx.doi.org/10.1109/BTAS.2016.7791181) 22. Kutzner, T.; Ye, F.; Bönninger, I.; Travieso, C.; Dutta, M.K.; Singh, A. User verification using safe handwritten passwords on smartphones. In Proceedings of the 2015 Eighth International Conference on Contemporary Computing (IC3), Washington, DC, [USA, 20–22 August 2015; pp. 48–53. [CrossRef]](http://dx.doi.org/10.1109/IC3.2015.7346651) 23. Daniel, N.; Anitha, A. A Study on Recent Trends in Face Spoofing Detection Techniques. In Proceedings of the 2018 3rd International Conference on Inventive Computation Technologies (ICICT), Coimbatore, India, 15–16 November 2018; pp. 583–586. [[CrossRef]](http://dx.doi.org/10.1109/ICICT43934.2018.9034361) 24. Chen, H.; Chen, Y.; Tian, X.; Jiang, R. A Cascade Face Spoofing Detector Based on Face Anti-Spoofing R-CNN and Improved [Retinex LBP. IEEE Access 2019, 7, 170116–170133. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2955383) 25. Simanjuntak, G.D.; Nur Ramadhani, K.; Arifianto, A. Face Spoofing Detection using Color Distortion Features and Principal Component Analysis. In Proceedings of the 2019 7th International Conference on Information and Communication Technology [(ICoICT), Kuala Lumpur, Malaysia, 24–26 July 2019; pp. 1–5. [CrossRef]](http://dx.doi.org/10.1109/ICoICT.2019.8835343) 26. Luo, J.; Hu, F.; Wang, R. 3D Face Recognition Based on Deep Learning. In Proceedings of the 2019 IEEE International Conference [on Mechatronics and Automation (ICMA), Tianjin, China, 4–7 August 2019; pp. 1576–1581. [CrossRef]](http://dx.doi.org/10.1109/ICMA.2019.8816269) 27. Xu, K.; Wang, X.; Hu, Z.; Zhang, Z. 3D Face Recognition Based on Twin Neural Network Combining Deep Map and Texture. In Proceedings of the 2019 IEEE 19th International Conference on Communication Technology (ICCT), Xi’an, China, 16–19 October [2019; pp. 1665–1668. [CrossRef]](http://dx.doi.org/10.1109/ICCT46805.2019.8947113) 28. Li, Y.; Huang, X.; Zhao, G. Joint Local and Global Information Learning With Single Apex Frame Detection for Micro-Expression [Recognition. IEEE Trans. Image Process. 2021, 30, 249–263. [CrossRef] [PubMed]](http://dx.doi.org/10.1109/TIP.2020.3035042) 29. Yao, L.; Xiao, X.; Cao, R.; Chen, F.; Chen, T. Three Stream 3D CNN with SE Block for Micro-Expression Recognition. In Proceedings of the 2020 International Conference on Computer Engineering and Application (ICCEA), Guangzhou, China, 27–29 [March 2020; pp. 439–443. [CrossRef]](http://dx.doi.org/10.1109/ICCEA50009.2020.00101) 30. Suzaki, K.; Shimizu, K.; Oguchi, K. Feasible Personal Identification by Eye Blinking Using Wearable Device. In Proceedings of the 2019 IEEE 15th International Colloquium on Signal Processing Its Applications (CSPA), Batu Feringghi, Malaysia, 8–9 March [2019; pp. 266–269. [CrossRef]](http://dx.doi.org/10.1109/CSPA.2019.8696045) 31. Anjos, A.; Chakka, M.M.; Marcel, S. Motion-based counter-measures to photo attacks in face recognition. IET Biom. 2013, _[3, 147–158. [CrossRef]](http://dx.doi.org/10.1049/iet-bmt.2012.0071)_ 32. Alhazmi, O.; Malaiya, Y. Quantitative vulnerability assessment of systems software. In Proceedings of the Annual Reliability and [Maintainability Symposium, 2005. Proceedings, Alexandria, VA, USA, 24–27 January 2005; pp. 615–620. [CrossRef]](http://dx.doi.org/10.1109/RAMS.2005.1408432) 33. [Arm TrustZone Technology. Available online: https://developer.arm.com/ip-products/security-ip/trustzone (accessed on 24](https://developer.arm.com/ip-products/security-ip/trustzone) February 2022). 34. Cerdeira, D.; Santos, N.; Fonseca, P.; Pinto, S. SoK: Understanding the Prevailing Security Vulnerabilities in TrustZone-assisted TEE Systems. In Proceedings of the 2020 IEEE Symposium on Security and Privacy (SP), Francisco, CA, USA, 17–21 May 2020; [pp. 1416–1432. [CrossRef]](http://dx.doi.org/10.1109/SP40000.2020.00061) 35. [Pinto, S.; Santos, N. Demystifying Arm TrustZone: A Comprehensive Survey. ACM Comput. Surv. 2019, 51, 1–36. [CrossRef]](http://dx.doi.org/10.1145/3291047) 36. Costan, V.; Lebedev, I.; Devadas, S. Secure Processors Part I: Background, Taxonomy for Secure Enclaves and Intel SGX Architecture; Now Publishers Inc.: Delft, The Netherlands, 2017. 37. Costan, V.; Lebedev, I.; Devadas, S. Secure Processors Part II: Intel SGX Security Analysis and MIT Sanctum Architecture; Now Publishers Inc.: Delft, The Netherlands, 2017. 38. Lee, D.; Kohlbrenner, D.; Shinde, S.; Asanovi´c, K.; Song, D. Keystone: An Open Framework for Architecting Trusted Execution Environments. In Proceedings of the Fifteenth European Conference on Computer Systems (EuroSys’20), Heraklion, Greece, [27–30 April 2020; Association for Computing Machinery: New York, NY, USA, 2020. [CrossRef]](http://dx.doi.org/10.1145/3342195.3387532) 39. [Mobile Security Solutions Secure Mobile Technology | Qualcomm. Available online: https://www.qualcomm.com/products/](https://www.qualcomm.com/products/features/mobile-security-solutions) [features/mobile-security-solutions (accessed on 24 February 2022).](https://www.qualcomm.com/products/features/mobile-security-solutions) 40. Trusty TEE Android Open Source Project. [Available online: https://source.android.com/security/trusty (accessed on 24](https://source.android.com/security/trusty) February 2022). 41. [GlobalPlatform Specifications Archive—GlobalPlatform. Available online: https://globalplatform.org/specs-library/ (accessed](https://globalplatform.org/specs-library/) on 24 February 2022). 42. Suzaki, K.; Nakajima, K.; Oi, T.; Tsukamoto, A. Library Implementation and Performance Analysis of GlobalPlatform TEE Internal API for Intel SGX and RISC-V Keystone. In Proceedings of the 2020 IEEE 19th International Conference on Trust, Security and [Privacy in Computing and Communications (TrustCom), Guangzhou, China, 10–13 November 2020; pp. 1200–1208. [CrossRef]](http://dx.doi.org/10.1109/TrustCom50675.2020.00161) 43. Chai, H.; Lu, Z.; Meng, Q.; Wang, J.; Zhang, X.; Zhang, Z. TEEI—A Mobile Security Infrastructure for TEE Integration. In Proceedings of the 2014 IEEE 13th International Conference on Trust, Security and Privacy in Computing and Communications, [Beijing, China, 24–26 September 2014; pp. 914–920. [CrossRef]](http://dx.doi.org/10.1109/TrustCom.2014.121) ----- _Appl. Sci. 2022, 12, 10782_ 21 of 21 44. Intel[®] Software Guard Extensions (Intel[®] [SGX). Available online: https://www.intel.com/content/www/us/en/architecture-](https://www.intel.com/content/www/us/en/architecture-and-technology/software-guard-extensions.html) [and-technology/software-guard-extensions.html (accessed on 24 February 2022).](https://www.intel.com/content/www/us/en/architecture-and-technology/software-guard-extensions.html) 45. [Keystone. An Open Framework for Architecting TEEs. Available online: https://keystone-enclave.org/ (accessed on 24 February 2022).](https://keystone-enclave.org/) 46. [Open Enclave SDK. Available online: https://openenclave.io/sdk/ (accessed on 24 February 2022).](https://openenclave.io/sdk/) 47. [Asylo: An Open, Flexible Framework for Enclave Applications. Available online: https://asylo.dev/ (accessed on 24 February 2022).](https://asylo.dev/) 48. [secGear. Available online: https://gitee.com/src-openeuler/secGear (accessed on 24 February 2022).](https://gitee.com/src-openeuler/secGear) 49. [Vrankulj, A. Uniqul Launches Facial Recognition Payment System. Available online: https://www.biometricupdate.com/201307](https://www.biometricupdate.com/201307/uniqul-launches-facial-recognition-payment-system) [/uniqul-launches-facial-recognition-payment-system (accessed on 24 February 2022).](https://www.biometricupdate.com/201307/uniqul-launches-facial-recognition-payment-system) 50. [About AliPay Facial Payment. Available online: https://opendocs.alipay.com/open/20180402104715814204/intro (accessed on](https://opendocs.alipay.com/open/20180402104715814204/intro) 24 February 2022). 51. [About WeChat Facial Payment. Available online: https://pay.weixin.qq.com/wiki/doc/wxfacepay/ (accessed on 24 February 2022).](https://pay.weixin.qq.com/wiki/doc/wxfacepay/) 52. [About UnionPay Facial Payment. Available online: https://cn.unionpay.com/upowhtml/cn/templates/newInfo-nosub/9ed2b7](https://cn.unionpay.com/upowhtml/cn/templates/newInfo-nosub/9ed2b7ea4873410186ae96112fccfc7d/20191211194221.html) [ea4873410186ae96112fccfc7d/20191211194221.html (accessed on 24 February 2022).](https://cn.unionpay.com/upowhtml/cn/templates/newInfo-nosub/9ed2b7ea4873410186ae96112fccfc7d/20191211194221.html) 53. Notifications Overview—Apple Developer. [Available online: https://developer.apple.com/notifications/ (accessed on 24](https://developer.apple.com/notifications/) February 2022). 54. [Firebase Cloud Messaging. Available online: https://firebase.google.com/docs/cloud-messaging/ (accessed on 24 February 2022).](https://firebase.google.com/docs/cloud-messaging/) 55. Ant-Techfin. [Performance Test Results of Intel SGX. Available online: https://my.oschina.net/u/4587334/blog/5014463](https://my.oschina.net/u/4587334/blog/5014463) (accessed on 24 February 2022). 56. [About JPush. Available online: https://docs.jiguang.cn/jpush/guideline/intro/ (accessed on 24 February 2022).](https://docs.jiguang.cn/jpush/guideline/intro/) 57. [Giry, D. Keylength—NIST Report on Cryptographic Key Length and Cryptoperiod (2020). Available online: https://www.](https://www.keylength.com/en/4/) [keylength.com/en/4/ (accessed on 24 February 2022).](https://www.keylength.com/en/4/) -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/app122110782?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/app122110782, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/2076-3417/12/21/10782/pdf?version=1667369371" }
2,022
[]
true
2022-10-25T00:00:00
[]
18,439
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00d4ca5c46ab55cd82aff4fffd52b182c9a4ca34
[ "Computer Science" ]
0.887592
Trust-based hexagonal clustering for efficient certificate management scheme in mobile ad hoc networks
00d4ca5c46ab55cd82aff4fffd52b182c9a4ca34
Sādhanā
[ { "authorId": "40066145", "name": "V. Janani" }, { "authorId": "143894255", "name": "M. Manikandan" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
DOI 10.1007/s12046 016 0545 0 # Trust-based hexagonal clustering for efficient certificate management scheme in mobile ad hoc networks ## V S JANANI[*] and M S K MANIKANDAN Department of Electronics and Communication Engineering, Thiagarajar College of Engineering, Madurai 625015, India e-mail: jananivs@tce.edu MS received 17 December 2015; revised 18 March 2016; accepted 20 April 2016 Abstract. The wireless and dynamic nature of mobile ad hoc networks (MANET) render them more vulnerable to security attacks. However, providing a security mechanism implicitly has been a major challenge in such an ad-hoc environment. Certificate management plays an important role in securing an ad-hoc network. Certificate assignment, verification, and revocation complexity associated with the Public Key Infrastructure (PKI) framework is significantly large. Smaller the size of the network lesser will be the certificate management complexity. However, smaller the size, large will be the overall infrastructural cost, and also larger will be the overall redundant certificates due to multiple certificate assignment at the boundary regions, that in turn affects the prompt and accurate certificate revocation. By taking these conflicting requirements into consideration, we propose the trust-based hexagonal clustering for an efficient certificate management (THCM) scheme, to bear an absolutely protected MANET Disparate to the existing clustering techniques, we present a hexagonal geographic clustering model with Voronoi technique where trust is accomplished. In particular, to compete against attackers, we initiate a certificate management strategy in which certificate assignment, verification, and revocation are carried out efficiently. The performance of THCM is evaluated by both simulation and empirical analysis in terms of effectiveness of revocation scheme (with respect to revocation rate and time), security, and communication cost. Besides, we conduct a mathematical analysis of measuring the parameters obtained from the two platforms in multiple times. Relevant results demonstrate that our design is efficient to guarantee a secured mobile ad hoc network. Keywords. Clustering; certificate management; MANET; security; trust; Voronoi. ## 1. Introduction Moreover, to manage issues other than secured routing such as authentication, privacy, integrity, and other security Ensuring an efficient security mechanism in a dynamic services, Public Key Infrastructure (PKI) was deduced. communication system is quite a challenging operation. In From past several years the PKI framework has been well MANET, no distinct part is dedicated to support any established that offers securing applications on the MANspecific functionality individually, with reliable routing ETs, due to its effectiveness in providing security in the being an eminent example. For decades, many routing form of digital signatures and certificate management. In protocols have been introduced for mobile ad hoc envi- traditional PKI-based approaches a centralized trusted ronment to achieve efficient secure routing, especially in a certificate authority (CA) provides certificates for the nodes multicast geocast region. These protocols differ in the in a network, by which the nodes are authenticated during approaches for finding routes between nodes in the net- communication, that is, certificates for each node are signed work. A location-based multicast (LBM) protocol for a by CAs and managed by the PKI system by which the secured route discovery was introduced by Ko and Vaidya nodes are authenticated during communication. Research [1]. In LBM protocol, the route has been discovered by ers have identified various security concerns when PKI utilizing location information from the Global Positioning system is used for security applications. These concerns System (GPS). This protocol reduces the overhead of the include: (a) Computational complexity that affects comroute discovery by limiting the search for a new route and putational cost and (b) PKI management, including certhe attackers within the ad hoc network, thus securing the tificate management. However, it is difficult for such a communication. certificate-based PKI strategy in MANETs for its self-or ganizing and infrastructure-less property. Therefore, *For correspondence deploying such a PKI-based communication system where ----- 1136 V S Janani and M S K Manikandan geographical or terrestrial constraints demands (such as battlefields, emergency, and disaster areas) is difficult. An ad hoc network is exposed to many kinds of attacks and so it is difficult to ensure a secure communication. To protect the legitimate nodes from these attacks, a vulnerable system should be considered in ad hoc networks. This can be achieved through the use of an efficient certificate management scheme that conveys trust in PKI. Certificate Management (CM) is considered to be a crucial task that promises trust in PKI. An efficient security solution for CM should confine two main factors: assignment and revocation. An enormous amount of researches has been made in these areas to provide a promising solution for security issues in MANET Certificate Revocation (CR) is an integral mechanism in certificate management, which enlists and removes the node’s certificate that has been identified to launch attack. If a node is found to be compromised or misbehaved, it should be denied from all activities and removed from the network. Certificate Revocation Lists (CRLs) are mechanism through which revocation information is propagated in a PKI framework. A CRL is a signed list by the CA listing all the certificates that are revoked. It is therefore considered as a main challenge for certificate revocation to revoke the certificates of malicious nodes promptly and accurately. In addition, the size of CRLs is important in a PKI system. However, there are several drawbacks in establishing PKI communication system to the ad hoc communications. Some of them are: In a traditional flat PKI system, a CA maintains the certificate authorization and a complete CRL list for the cluster. Single CA issues all of the certificates within a cluster. This list will be passed on to cluster head (CH) to dispatch the certificates to the nodes. Such a structure can be prone to delay, and maintaining such an infrastructure may add up the infrastructural cost to a large extent. Issuing network-wide certificate to the nodes may lead to resource underutilization. We may require to restrict the usage of communication resources by node to certain region only, for example, the region where it has been registered. Revocation checking can be problematic in this structure, since all of the revoked certificates in the network are listed in a single CRL, the number of entries on that CRL can become quite large. Further, there are cases where malicious nodes and their certificates can no longer be revoked in a timely practice. A large CRL takes significant bandwidth to download and consumes significant computational resources on the CA to check the revocation status of a particular node also; the amount of revocation information that can be stored at a CA is limited by the memory available at the CA. Therefore, it is clear that the complexity of the PKI system and also the size of the CRL have to be minimized with prompt and accurate certificate management, in order to make the PKI-based security viable for MANET security deployment. In this pursuit, we make the following contributions in this paper: A certificate assignment strategy is introduced for MANETs in order to reduce the complexity of managing the PKIbased security framework. A cluster region-based certification approach is established, where the entire network is partitioned into several geographical clusters provided by the LBM protocol. (2) To avoid dynamic communication droppings, the nodes in the boundary region of any particular geographical clusters are assigned with multiple certificates corresponding to its current region as well as several other regions in its vicinity, which in turn reduces the size of CRLs. (3) (inspired by Bellur.B’scertificate assignment strategies in [2] To reduce the size of CRLs further, the CA tailors the expiry time of each certificate with the distance of node from the cluster region. (4) Using Voronoi Diagram (VD), the optimal strategy for multiple certificate assignment is resolved. We assume the cluster to be in hexagonal shape for geometrical simplification as presented by Chang and Wang [3] and Zhuang et al [4]. This paper is structured as follows. Section 2 describes the works related to certificate management in MANET are described. Section 3 describes the proposed region-based certificate method. Section 4 presents the operations of nodes with region-specific certificates. An analytic model for the size of CRL and communication cost is derived in section 5, followed by the performance evaluation and simulations in section 6, and empirical analysis in section 7. The concluding remarks appear in section 8. ## 2. Related works In recent years, researchers have focused on MANET security issues as done by Fan et al [5]. It is difficult to provide a complete security solution to mobile networks due to the wireless connectivity, dynamic topology, and infrastructure-less features. To cope with uncertain nodes, clustering techniques have been widely applied in MANET by Cao and Hadjicostic [6], Chau et al [7], Cheng et al [8], Kao et al [9], and Mohamed and Abdelfettah [10]. Many clustering algorithms in the ad hoc network were investigated by Abdelhak et al [11], Khalid et al [12], and Mohd. Junedul Haque [13]. With the objective to reduce the distance calculation complexities of uncertain nodes, an important structure in computational geometry named Voronoi diagrams are applied for wireless application as proposed by Fan et al [14] and Kao et al [9]. Stojmenovic et al [15] introduced a distributed algorithm to compute the Voronoi region of each node. A general algorithm to reduce flooding ratio in routing within a Voronoi network was presented by Kao et al [9]. The topology control and routing in a wireless ad hoc network was done by Ngai et al [16] using Voronoi design and delaunary triangulation. To increase the spatial reuse, the network areas are clustered into congruent polygons with ----- Trust-based hexagonal clustering for efficient certificate 1137 Voronoi geometric features. A hexagonal spatial geometric distribution of nodes was introduced by Zhuang et al [4]. This partitioning technique has shown to increase the network capacity and throughput of the network. It was proven that the regular hexagons have flexibility to be partitioned into smaller hexagonal shapes and grouped together to form larger ones. Most of the previous studies on MANETs have utterly assumed that nodes are cooperative. As an effective mechanism to consider issues in node cooperation, trust has been highly recommended in recent researches as done by Renu et al [17] and Manju and Yudhvir Singh [18]. Jingwei and David [19] quantified trust relationships with the risk in a PKI system. A fully trust-based PKI approach for ad hoc networks was presented by the authors Liu et al [20], Ferdous et al [21], Cho et al [22], and Wei et al [23]. This approach proved to eliminate security vulnerabilities to a large extent with maximized performance characteristics. The performance issues in trust management protocols were addressed by Ing-Ray Chen et al [24] with minimized trust bias and maximized application performance. To provide a trade-off between cryptographic security and vulnerability, the MANET applications necessitated protocols in multicast condition. In wireless networks numerous researches have been done in multicast routing and routing protocols by Deering et al [25] and Mohammad M Qabajeh et al [26]. Kanchan and Asutkar [27] applied clustering, encryption, and cryptography techniques to improve the performance of these routing protocols. The dynamic movement of nodes in the dynamic environment made these existing protocols inept. This drives the need for an improved multicast flooding approach. To reduce route establishing overhead and to improve the performance of routing protocol, a location information-based approach was introduced. Here, to handle the duplication mechanism so that each destination receives atleast a single copy of the original message, flooding algorithm was used. A location-based approach (LBM) proposed by Ko and Vaidya [1] described the flooding algorithm in wireless topology, which used physical location information obtained from the GPS. On the demand for providing security to the legitimate nodes against attackers, many certificate revocation schemes have been proposed in PKI networks and military ad hoc environment. Jormakka and Jormakka [28] presented a certificate revocation scheme designed for a semiad hoc military and civilian network to prevent fake certificate revocations. A survey on the certificate in a distributed system from the year 2000 onward is done by Yki and Mikko [29]. Wei Liu et al in [30] and Mohammad and Javad [31] studied a cluster–based certificate revocation scheme that quickly revoke malicious certificates and retrieve falsely accused certificates in distributed networks. Mohamed M E A Mahmoud et al [32] carried out a study on revoking certificates in a pseudonymous PKI system in which certified key pairs were assigned to maintain privacy in each node. To ensure the validity of certificates in PKI system, a validation technique was proposed by Mohammad Masdari et al [33, 34], in which the trust level of CA on each node was considered. The certificate revocation method, CCRVC presented by Liu et al [35], handles attacker nodes. CCRVC revoked malicious nodes to solve false accusation. URSA proposed by Luo et al [36] implemented a novel ticket certification process that used tickets to recognize and to grant access to well–behaved nodes. This scheme maximized the service availability with a distributed and localized mechanism. Later, Taisuke et al [37] considered the complexities of Certificate Dispersal Problem in a tree structure where the problem was solved in polynomial time. Certificate management with trust in a PKI framework has been used as a security mechanism for attack handling. CR scheme was presented by Park et al [38] and Raya et al [39] to identify and remove certificate of those nodes that were detected as attacker node. This scheme provided security of the network by revoking the compromised or misbehaved nodes. The revocation scheme by Park et al supported a cluster-based network. The cluster head performed the necessary revocation action of removing the nodes in black list in this scheme. Mawloud Omar et al [40] addressed the constraints in node mobility while designing a reliable certificate system. The authors proposed a recovery protocol based on web-of-trust where the nodes themselves issue and manage the public key certificates. A short and safe certificate chain was selected in order to reduce the communication overhead and resist attacks. Nevertheless, there are certain flaws in the existing certificate management mechanism in utilizing PKI-based communication system to a mobile environment. In efficient deployment of revocation scheme add up the resource utilization as well as communication cost. Owing to the absence of topology, providing a promising security to the mobile nodes in MANET is difficult to achieve. We propose an efficient Trust-based Hexagonal cluster for certificate management (THCM) strategy for use in mobile networks, to secure MANET and to reduce the complexities in the PKI-based security system. To partition the uncertain nodes of MANET, a Voronoi-based clustering is performed in hexagon structured polygon to reduce region overlapping drawbacks that occur in traditional clustering shapes. A trust-based hexagonal clustering is incorporated in our scheme, where the CH selection is performed with high trust degree. Considering the communication cost and certificate management complexities, optimal sizes of regions are calculated. ## 3. Proposed system design This section provides a detailed description of our proposed certificate management scheme that significantly reduces the complexity of the PKI system. We begin with ----- 1138 V S Janani and M S K Manikandan partitioning of the network into different geographical regions with a trust-based clustering approach. The proposed certificate assignment and revocation mechanism is implemented in each such geographic cluster that provides secure intra-clustering and inter-clustering communication. ## 3.1 Proposed clustering technique There have been several clustering strategies proposed in literature. In an uncertain clustering (UC) model, it has been assumed that a node or a point ‘ni’ should be located inside a (closed) region with a probability density function (PDF) to describe the distribution of nodes within a region. The uncertain point clustering has been performed with different methods such as K-means, UK-means, pruning, Min-Max BB, partial ED, and so on. To compute the closeness of the node and the cluster representative, different methods based on mean, Euclidean distance, and probability have been in practice. However, these traditional clustering techniques of uncertain nodes increase the computational complexities and communication cost in mobile environment, especially in mobile ad hoc networks. To construct a highly desirable uncertain clustering cell in MANET, we propose to use VD-based clustering in which the clustering issues are managed considering the drawbacks of existing UC methods. In MANET, VD is used to partition network into clusters based on Euclidean distances to nodes in a specific subset of the plane. A Voronoi diagram represents the region of influence around each of a given set of nodes. This geometric structure partitions the entire plane into polygon cells, called Voronoi polygons, formed with respect to n nodes in a plane. It is widely used since it offers an efficient solution for point location. In recent years this structuring concept is widely used for exploring location and routingbased issues. The Voronoi partition or cluster for a given set of nodes is unique and produces polygons that are route connected. A Voronoi polygon, traditionally, constructed as follows: V xð Þi [¼] �yjd xð i; yÞ � d x� j; y�; i 6¼ j� ð1Þ where V xð Þi [: Voronoi polygon of][ x]i xi : Node, y : Set of points closer to xi d xð i; yÞ : Distance from point y and xi and x� j; y� : Distance from point y and xj. Our clustering technique consists of two steps: 1. Cluster construction 2. Cluster head selection. 3.1a Cluster construction: In the first step, Voronoi clusters (VCs) are constructed on a set of nodes N ¼ fn1; n2. . .nkg with a distance function d : S[m] � S[m] ! S (mdimensional space) giving the distance d x; y 0 between ð Þ � any nodes x; y S[m]. The VD partitions the space S[m] in k 2 cells with cluster representatives C ¼ cf 1; c2. . .ckg with the property mentioned by Cao and Hadjicostis [6] as where the neighboring region, Xn mð Þ is the region on one side of the cluster cell edge En mð Þ and Ej j is the empty set. Figure 1. Voronoi-hexagonal clustering in MANET. d xð ; ciÞ\d x� ; cj�8x 2 V cð Þi ; ci 6¼ cj: ð2Þ In the second step, the distance between the nodes and a cluster representative (a node) is calculated. The Voronoi partitioning of a network can be of any polygonal shape and for its beneficial geometrical characteristics, we assume that the uncertainty region of Ni is a regular hexagon with nodes whose center are equidistance to each other with distance d and radius r, where r [ 0. The hexagonal clustering partitions a larger area into adjacent, nonoverlapping areas and can be subdivided into smaller hexagons. Nodes join to form hexagonal clusters and each cluster consists of CH and Cluster Members as shown in figure 1. The distance d a; b between nodes in MANET plays an important ð Þ role in determining the network performance. We shall assume that the nodes of the ad hoc network are independent and randomly distributed in the hexagonal structure. The edges of the hexagonal polygon is perpendicular to the line joining a node with another in N. Considering the radius, for any query point S, (2) can be written as 2 d pð ; ciÞ � d p� ; cj� ¼ ri þ rj: ð3Þ If two nodes overlap, the distance d n� i; nj�\ri rj and þ (3) become unreal, which means the edges cannot be found and we consider the cluster as empty. The hexagonal cluster construction in the MANET is illustrated in Algorithm 1. The expected region of each node ni is initialized as a whole space (step 2). The VC edges and the corresponding neighboring regions of ni are then computed for each node nj (steps 4 and 5). The VD for cluster construction considers an expected region of node ni and the neighboring region of VC edge En mð Þ. The expected region of ni, denoted by Eri is the intersection of all the internal regions; that is, \ Eri ¼ j¼1... Ej j[V]j6¼i Xn mð Þ ð4Þ ----- Trust-based hexagonal clustering for efficient certificate 1139 The clustering polygon can be generated by excluding all the neighboring regions from the domain space. The overlapped regions are reduced to generate the expected region Eri (step 6). For each node nj, we verify the expected region lie inside a Minimum and Maximum Region **A** O2 Bounding (MinMax-RB) of the domain space; MinMax-RB is the minimum or maximum region with sides perpen O1 dicular to the principle axes of S[m] that encloses a finite region. If so, the node nj is then assigned to a cluster. Let us **F** consider six equilateral triangles in a regular hexagon. For calculation we take a single equilateral triangle DOAF. A circle with center cn and radius rn is assumed to intersect the DOAF as in figure 2. On spatial decomposition the **E** region that does not contain the hexagonal region is con sidered as neighboring regions Nn mð Þ and the region where the area of the circle and the neighboring region overlap as overlap region Oi (ie., Oi xð ; yÞ ¼ O1 þ O2 þ O3). The probability of the expected region Eri in a hexagonal cluster with area A and x; y as coordinates of any random node is ð Þ given as Hexagonal Voronoi cluster construction. ----- 1140 V S Janani and M S K Manikandan PEri ¼ A[1][2] " 6 # ZZ X prn[2] [�] Oi xð ; yÞ i¼1 dxdy 5 ð Þ The degree of successive encounter ‘x’ made be trustee on trustor may either be positive (represented as ep xð Þ) or negative (represented as ep xð Þ). Here, to evaluate the trust, we consider three cases of uncertainty degree, i.e., 0, ¼ 0\UD\1 and UD ¼ 1 as shown in figure 3. When the uncertain degree is low (UD 0, the nodes ¼ Þ are highly trustable. This highly certain case shows that the trustor is very much confident with the trustee. If the uncertain degree varies from low to high (0\UD\1Þ, the trustor may not have sufficient confidence with the trustee. On the other hand, a highly uncertain case occurs when the uncertain degree UD 1. At this state the trustor may be ¼ completely unknown about the trustee. The nodes with highest trust degree, that is, UD 0 and ¼ TD ¼ 1, is considered as CH, initially at time T1. As time progresses, the topology changes frequently in a MANET, which varies the cluster nodes and the cluster heads. Hence, the cluster head selection procedure is adaptable for the change in topology. The trust value of each node is recomputed and the CH is selected, comparing the current CH ðCHcurrÞ with the previous CH ðCHpreÞ and location ðLOCpreÞ. The nodes with trust degree between 0 and 1ðthat is; 0\UD\1Þ have undergone distrust test to reduce the rate of risks. On comparison with the trust degree and the distrust degree of such nodes, they are either revoked or considered as cluster members, that is,, the nodes with highest distrust degree DTD ð ¼ 1 or DTD [ TD and UD ¼ 1Þ are revoked and the remaining nodes are assigned as CH. This trust-based cluster head selection eliminates a certain amount of risk in communication within the network. The detailed cluster head selection process is shown in flow chart 1. To perceive the exact location information of any node, each node in the network is enabled with a position identification system. Our proposed scheme makes use of the Figure 3. Trustability. PEri ¼ [p]A[r]n[2] [�] A[6][2] ZZ Oi xð ; yÞdxdy: ð6Þ 3.1b Cluster head selection: In MANET, the nods join or leave the cluster dynamically and thus the CH selection is difficult. We consider a distributed cluster head selection procedure with n nodes, which are of h hops distance within a cluster. It is much easier to select an efficient mechanism to establish security, if trust relationship among the nodes is obtainable for every cooperating node. Hence, to provide a secured communication among cooperative nodes, it is important to calculate the trust and distrust degrees of nodes in the network. The trust of a node can be defined as the probability of belief of a trustor (t) on a trustee (s), varying from 0 (complete distrust) to 1 (complete trust). The probability of trust and distrust of the trustor on information (i) sent by the trustee with context to belief (b) is given in (7) and (8), presented by Jingwei and David [19]. TrustDegree; TD t; s; i; b ð Þ h ^ i ð7Þ P belief t; i madeBy i; s; b beTrue b ¼ ð Þ��� ð Þ ð Þ DistrustDegree; DTD t; s; i; b ð Þ h ^ i P belief t; _i madeBy i; s; b beTrue b : 8 ¼ ð : Þ��� ð Þ ð Þ ð Þ To measure the trust degree explicitly in an ad hoc environment, we present a trust calculation method with uncertainty degree. With this a high level of trust can be achieved for secured communication. The certainty of nodes in MANET is considered as the summation of trust and distrust degrees. Consequently, the uncertainty degree UD by Jingwei and David [19] is defined as ð Þ UD t; s; i; b 1 certainity of nodes: 9 ð Þ ¼ � ð Þ An important factor that affect the trust level of a node is the Encounter History (EH), which specifies the number of successive interactions between the trustor and the trustee in a network. Initially we assume EH as greater than or equal to 0. The trust and the distrust level of any node can be measured with the relation as shown in (10) and (11). TD t; s; i; b ð Þ ¼ DTD t; s; i; b ð Þ ¼ Pn x¼1 [e][p][ x]ð Þ 10 ð Þ EH Pn x¼1 [e][n][ x]ð Þ : 11 ð Þ EH Therefore, (9) UD t; s; i; b 1 ð Þ ¼ � �Pn Pn � x¼1 [e][p][ x]ð Þ x¼1 [e][n][ x]ð Þ : 12 þ ð Þ EH EH ----- Trust-based hexagonal clustering for efficient certificate 1141 **Start** **List the nodes in cluster** **Initialize CHpre, CHcurr,LOCpre and** **present ( ) as 0 and EH ≥1** **co-operativeIs node** **No** **Revoke nodes** **Yes** **Generate probability as trust metric** **Calculate encounter history (EH)** **Compute trust degree (TD) for each** **node** **Yes** **If TD =1 and** **No** **UD=0** **No** **TD ranges [0,1]** **If** **Yes** **≤present( )** **False** **TD(CHpre) =TD(CHcurr)** **False** **&** **Compute distrust** **) ≤1** **EH(CHpre)= EH(CHcur)** **probability** **Select new** **cluster head** **True** **Calculate distrust** **True** **degree (DTD)** **remains** **Assign CHpre and CHcurr** **as cluster head** **If** **DTD >TD/DTD = 1** **Yes** **and** **UD=1** **No** **Broadcast cluster** **Assign nodes as** **head** **cluster members** **Stop** Flow chart 1. Trust-based cluster head selection process. clusters as well as the geographic location information intensively. 3.1c Geocast clusters: We use a geographical positionbased routing scheme of Ko and Vaidya [1] for improving the efficiency of routing in a multicast environment. LBM assumes the availability of GPS for obtaining location information essential for routing in the hexagon clusters. Limiting the search area for finding a path, reduces control overhead and increases bandwidth utilization, in LBM, makes it suitable for mobile networks. LBM uses two approaches for flooding control packets, namely, multicast tree and multicast flooding, in a geographic cluster. ## 3.2 Certificate authority In certificate management the nodes have to obtain valid digital certificates from the CA, before it takes part in the communication. A trusted third party, CA is deployed in cluster-based scheme to enable nodes to preload the certificate. The CA distributes and manages certificates to all nodes within the cluster. The validity of a certificate can be verified by ensuring that the certificate is neither expired nor revoked by the CA. In the proposed cluster-based CM scheme, depending on the location of each node, certificates are assigned. It is constrained that the node uses only the certificate corresponding to their current geographic location and discards those certificates that are not ----- 1142 V S Janani and M S K Manikandan appended with a certificate assigned to that particular node. In addition to this, the nodes in the boundary regions are assigned with multiple certificates corresponding to several clusters in its vicinity, in advance, making flexible roaming between adjacent regions. The multiple certificates assigned to the nodes can be derived from the same key pair (public–private keys) for simplicity. The CRL is chosen to access CA in the mobile environment concatenated with a time stamp as an indication of its updates. This list enumerates the digital certificate’s status of all nodes, that is, date of certificate issued, entity that issued, and the reason for revocation of the certificate. When a node attempts to access the cluster, the CA allows or denies access based on the CRL entry for that particular node 3.2a Region-based CRL concept: To reduce the potential network and computational overhead raised by larger CRLs improve the revocation efficiency, a scheme for partitioning the CRLs into several smaller lists has been in practice. The partitioning of CRL is transparent to all the nodes and for each certificate, available information shall indicate the segment that it should be consulted. In our model CRLthe network is segmented based on the geographic information. The certificate assigned to all the nodes in a particular geographic cluster A are mapped to a CRL Register Head represented by CRLRegNo(A). The proposed system constrained the nodes to append signed message with the certificate corresponding to their present geographically partitioned cluster. All nodes in a given cluster, therefore, append signed messages using the certificate that belong to the CRL Register Head of that particular cluster. For example, the nodes in cluster A appends signed messages with certificate that have same CRL Identity (ID), CRLRegNo(A). During verification of received message, a node in cluster A obtain the CRL corresponding to its current cluster, represented by CRLRegNo(A) by the CA. In addition, the nodes will discard the signed messages that are appended with certificates other than the current location. When a node moves closer to the boundary of a neighboring cluster B, it accepts signed message appended using certificate corresponding to cluster B in addition to its corresponding cluster A. Such nodes receive the CRL Identity of B represented by CRLRegNo(B) issued by the CA. This proposed strategy of CRL partitioning will minimize the CRL size to a great amount and hence the communication costs enormously. To reduce the size of CRLs further, we considered the expiry time of certificates. It is proposed that the CA can alter the expiry time of certificates assigned to nodes of different geographic clusters to be inversely proportional to the distance between the current cluster region and home cluster region of the node., that is,, the certificates get expired when the node moves apart from its home cluster region. This eliminates direct revocation of such expired certificates that reduce the size of CRL. Let the distance between positions p and q be Dist p; q and the boundary of A be Bound A . Then, ð Þ ð Þ Dist Nð i; AÞ ¼ MinpinBound Að Þ½[Dist GPS]ð [ð][N]iÞ; AÞ�: ð13Þ If the node moves closer to the boundary of cluster A then Dist Nð i; AÞ\MaxiRange, where MaxiRange is the maximum range of a cluster. Likewise, if a node is said to be in the center of a cluster, then Dist Nð i; AÞ [ MaxiRange. It is assumed that a geographic cluster B is said to be the neighbor of A, if there exists position p and q and Dist pð ; qÞ\MaxiRange. ## 4. Certificate management strategy The MANET environment can be organized into different clusters with several shapes such as circle, rectangle, and hexagon. To gain advantage in faster searching speed and to have successive search patterns overlapped, we considered the clusters as regular hexagons. We consider the nodes and the CA in the network are knowledgeable about the clustering as well as CRL partitioning. To know the physical location of each node in the cluster, the LBM protocol used in THCM updates its geographic information, whenever required. The LBM protocol in THCM will update geographic information of each node. Moreover, the nodes can be determined even before they are about to migrate from its current cluster location to the neighboring cluster of its vicinity. ## 4.1 Functionalities of certificate management When a hexagonal geographic cluster is organized, the nodes in a particular cluster place request to the CA for assigning the certificate for authenticated participation in the communication. After verification, the CA responds with multiple certificates corresponding to the current location as well as neighboring location of the nodes as shown in figure 4. Our proposed THCM is considered in different phases. Initialization Phase: During this phase, each node will send a request ðCERReqÞ for assigning certificate to the CA. The request sent by the nodes are signed using its private key. Verification Phase: Upon receiving the request, CA first verifies the message using the public key attached with the request. From the GPS information received with the request, CA determines the cluster in which the node is currently located and its neighboring clusters. Assignment Phase: The CA responds to the node with multiple certificates ðCERResÞ corresponding to its current as well as neighboring locations. Besides, the CA responds to the CRLs corresponding to different geographic locations of the nodes. The functionalities in the proposed certificate management scheme in each hexagonal cluster are described as To send a message: To begin a secure communication, each node in a hexagonal geographic cluster should obtain ----- Trust-based hexagonal clustering for efficient certificate 1143 Request for certificate (CER Req) |Nodes (N ) i|(CER Req) Assign certificate ( CER Res)|Verify request Certificate Authority (CA) Determine current Certificate of & neighboring N i regions of Ni| |---|---|---| Figure 4. Certificate request and assignment. signed messages that are appended with corresponding certificates from CRL. A node signs the hash of the message with its private key. This signed message is then appended with the certificate corresponding to the geographic location. To receive a message: This is an important functionality of certificate management where the messages are verified and processed. It includes the following operations (with figure 1). Verification: A node that receives messages verifies three main elements namely; certificate, validity,and signature. Sender’s certificate: The certificate of the sender is verified first to analyze whether it belongs to the current geographic cluster A or its neighboring region B . If the ð Þ ð Þ sender’s certificate belongs to cluster region other than AorB, the message is discarded. Verify validity: If the sender’s certificate corresponds to either AorB, it is further verified to check its validity, that is, it has not expired or has not been revoked. The certificate is discarded if it is expired. Further, the revoked status of the certificate is determined from the appropriate CRL, that is, if the sender’s certificate corresponds to cluster A, it is specified by CRLRegNo(A) and if it corresponds to cluster B, it is specified by CRLRegNo(A). The certificate is discarded, if it is proved as revoked from above verification. Signature: The signature of the message is verified whether it is received from current or neighboring geographic cluster. Accept message: If the messages pass all the above verification procedures, then the messages are accepted. Re-Organize: When a node in cluster A is identified to move closer to the boundary of a neighboring cluster B, the CRL Register Number is reorganized. In addition to the certificate of current cluster location, the node accepts new signed messages that are appended with the certificate corresponding to the new cluster region. It also acquires the CRL of cluster B represented as CRLRegNo(B), issued by the CA. For example, when a node N1 of cluster A moves closer to the boundary of cluster, N1 accepts the CRL of B (CRLRegNo(B)) in addition to the CRLRegNo(A) of cluster A. This re-organize functionality of proposed system maximizes the availability of services to each node and resilience against attacks. Request for new messages: When a node identifies the certificate corresponding to the neighboring cluster that it probably visits in near future is about to expire, the nodes send a request to the CA for new certificates. These rerequests processes for a new set of certificate to the CA and the assignment reply from the CA are performed in three phases; initialization phase, verification phase, and assignment phase as described in section 4. ## 4.2 Certificate assignment In a random network, wireless nodes are distributed randomly over an area. This random distance between nodes in MANET plays an important role in the performance of the system. We propose a random probability distribution of the distance between nodes distributed in a regular hexagon. To reduce the complexities, we use spatial decomposition method of Bettstetter and Wagner [41], for certificate assignment. Figure 5 shows the hexagonal clustering of networks in a MANET environment. For reference we have taken a single hexagon cluster ABCDEF with center [0]O[0] intersected by a circle of center cn and radius rn. The sides of each hexagon is taken as ‘S’. The equilateral triangle DOAF represents one of the six equilateral triangle regions in ABCDEF. When a node sends a request for certificate, the certificate authority will assign one or more certificates depending on the random distance of the CA and the requested node, after verification processes. We consider three different cases of certificate assignment strategy in a mobile ad hoc network. Single certificate assignment: The requestor node R ð Þ and the CA lie completely inside the circle, within the hexagonal region as shown in figure 5(a), that is, the CA and the node R correspond to same geographic cluster BCDEF. During this case the CA assigns a single certificate to R corresponding to its current geographic distance. Multiple certificate assignment: As in figure 5(b), suppose the requestor node R moves closer to the boundary of ð Þ cluster ABCDEF so that the circle cuts the edges of the hexagon cluster ABCDEF. Multiple certificates are assigned to R in this case, corresponding to its current geographic location (hexagon ABCDEF) and neighboring cluster region. Null certificate: Suppose the requestor node R does not belong to the geographic location of CA ABCDEF or at ð Þ ----- 1144 V S Janani and M S K Manikandan **B** **A** **B** **A** R R CA **C** CA **O** **F** **C** **O** **F** D E D E ## (a) (b) **B** **A** R CA **C** **O** **F** D E ## (c) Figure 5. Certificate assignment schemes. (a) Single certificate assignment. (b) Multiple certificate assignment. (c) Null certificate assignment. the boundary of ABCDEF, the request is discarded and no certificate is assigned to R as shown in figure 5(c). ## 4.3 Attack model The proposed certificate management scheme aims to achieve resistance against the following security attacks: Forging Attacks: The revocation information generated in a cluster should be unforgettable, so that any node in the cluster must not be able to generate duplicate revocation information, even though it has the revocation information generated earlier. Collusion Attack: A revoked node should not be able to collude to revoke a trustable node. Revocation Denial Attack: Neither a trustable node nor a distrust node should purposely fail the revocation process of a misbehaved node, internally or externally. ## 5. Analytic model 5.1 Size of CRLs The proposed system benefits the reduction in the size of CRLs. Generally, the size of CRL in a network depends on the number of nodes to which certificates are to be assigned ðNT Þ, rate of revocation (R), validity of node’s certificate (V), and the order of CRL entries (m). The revocation rate R is an integral part for evaluating the CRL size as well as ð Þ the performance of revocation system. It can be stated as the rate of launching attack by an attacker node before its certificates get revoked. Owing to the dynamic movement of nodes in a MANET environment, the order of CRL entries varies frequently, which affects the validity of certificates. The CRL size is given as Size of CRL ¼ NT � R � V � m: ð14Þ ----- Trust-based hexagonal clustering for efficient certificate 1145 When a cluster-based certificate management is applied, the size of CRL also varies with the number of geographically partitioned clusters. It is assumed that the average number of valid certificates assigned to each node in a cluster as V CER. The average number of nodes in a cluster is noted as Navg ¼ [N]RT[T][; where][ R][T][ is the total number of cluster] in the network. Thus, the size of CRL for the proposed system is given by Size of CRL THCM schemeð Þ ¼ Navg � V CER � R � V � m: 15 ð Þ Hence, the size of CRL can be reduced depending on the degree of cluster partitioning. When the size of the cluster is smaller, the cost of certificate gets reduced. Conversely, the complexity of PKI system increases. This is because the cost of certificate especially in the boundaries of the cluster increases. In addition to this, the installation cost of the CA in different clusters adds up the framework cost. It is therefore necessary to determine an optimal size for the cluster to reduce the cost of communication. ## 5.2 Communication cost One of the important issues in MANET communication system is the rise in communication cost due to the certificate assignment and revocation processes within each cluster. In THCM scheme, in order to reduce the cost of communication, the certificates are assigned based on random distance between the CA and any requestor node. The efficient revocation scheme in THCM reduces the communication cost due to revocation. We assume an optimum size for each cluster in CA installation and certificate management. The overall communication cost in a particular geographic cluster includes the cost in sending a request for certificate by any node, cost in issuing certificate by the CA, and the revocation cost. Commc ¼ Creq þ Cassign þ Crevoke: ð16Þ Let the average number of nodes in ABCDEF is (single or multiple certificates), and the length of certificate issues, that is, response length ðReslÞ. The verification process incorporates the revocation of certificates, the expiry check, and the length of revoked message ðRevlÞ, which may change frequently in a dynamic infrastructure like MANET. Cassign ¼ Cverify � Cissuing � Resl: ð19Þ Usually, the CA verifies the request for certificate from any node in a cluster and issues multiple certificates. This increases the communication overhead as well as communication cost to a larger extent. To reduce the cost of communication and overhead, probability density functions (pdf) of the random distance discussed in Section 4 (with reference to figure 5) are carried out. For reference we have taken hexagonal clusters ABCDEF with sides as ‘s’. It is assumed that the nodes within the circle and hexagonal region ABCDEF are assigned with one certificate that correspond to the current home cluster region ABCDEF (figure 5(a)). The nodes at the boundary of the region ABCDEF are assumed to be assigned with the multiple certificates corresponding certificate of home cluster and the neighboring cluster region (figure 5(b)). It is also assumed that the request from the nodes, belonged to other adjacent clusters is discarded (figure 5(c)). The efficient certificate management scheme within the cluster is formulated with the probability of certificate assignment and management. In wireless networks, random distance between nodes is considered as a critical factor that affects the system performance. The closed-form distribution for random distance can be applied to calculate path loss, link capacity, near-far neighbors, transmission power, and other performance metrics in MANET. Here we use a modified random distance calculation concept of Bettstetter and Wagner [41] for probability density function calculations. The random distance formulates the stochastic activities within the mobile network. A random distance of a node is considered as a location-based discrete time process, where each node moves randomly with same length ðDtÞ and duration ðDxÞ. When the node moves to the boundary of a cluster, the probability varies. Let us assume that the coordinates of cn and cm be ðxn; ynÞ and ðxm; ymÞ with fn ¼ [x][n][þ]2[x][m] and fm ¼ [y][n][þ]2[y][m][. Let] cos h ¼ d c[x][m][�]i;c[x]j[n] [and sin][ h][ ¼][ y]d c[m][�]i;c[y]j[n] ð Þ ð Þ[. The probability of the] random distance between nodes in the MANET is calculated with area-ratio approach. Suppose the side of the equilateral triangle a ¼ 1 and the distance be RD, then the probability P Rð D � DÞ is taken as the ratio of the area between the triangle and the circle. In figure 5 the distance is calculated from the center of the hexagon with two different cases, depending on the value of the distribution function D. (i) The circle x[2] y[2] D[2] is completely inside the þ ¼ p 3ffiffi hexagon of area [p][D]6 [2][;][ thatis][;][ 0][ �] [D][ �] 2 [, then random] distribution function is given as NABCDEF ¼ NT p 3 23[ffiffi]s[2] 17 ð Þ A where A is the area of all the clusters. The cost of sending a request to CA by any node depends on the average number of nodes in each region and the length of the request ðReqlÞ, which is the distance of the node and the CA, given by Creq ¼ NT p 3 23[ffiffi]s[2] � Reql ð18Þ A The certificate assignment by CA plays a vital amount in the increment of overall communication cost. It depends on the verification cost, the cost of issuing the certificates ----- 1146 V S Janani and M S K Manikandan FRD Dð Þ ¼ P Rð D � DÞ ¼ [area of hexagon]area of circle ¼ 3[2]p[p] Dffiffiffi3 [2]: ð20Þ (ii) The circle x[2] y[2] D[2] cut-off the edges of the þ ¼ p 3ffiffi hexagon; that is, 2 [�] [D][ �] [1, then random distribu-] tion function is given as On differentiating the cost function with respect to s and equate that to 0, we can minimize the communication cost value in the proposed THCM scheme. For simplification, here we assume D s h; where ¼ � 0 h 1. � � CommC ) s^[4] ¼ 0 B B B @ 25 ð Þ 4A p NT 3Apffiffiffi3 ðReql þ ReslÞ þ �1 � 3[4]p[p][h]ffiffiffi3� 3� 23[ffiffi] � NT 3pA ffiffiffi3 � Resl þ [C]T [�] [3]pA ffiffiffi3 � Revlþ C � � 3pffiffiffi3 T [�] 1 � 3[4]p[p][h]ffiffiffi3 A � Revl 1 C C C A "pD[2] pffiffiffi3 pffiffiffirffiffiffiffiffiffiffiffiffiffiffiffiffi# FRD Dð Þ ¼ p[2]ffiffiffi3 3 [�] [2][D][2][ cos][�][1] 2D [þ] 3 D[2] � [3]4 : 21 ð Þ (iii) The circle x[2] y[2] D[2] completely outside of the þ ¼ hexagon; that is, 1 D 2, then random distribu� � tion function is given as FRD Dð Þ ¼ 0: ð22Þ The probability of the distance between any two nodes in a hexagonal cluster can be written as (from (20), (21), and (22)) 4pD pffiffiffi3 3p ;ffiffiffi3 0 � D � 2 4D pffiffiffi3 pffiffiffi3 p ðffiffiffi3 [p]3 [�] [2 cos][�][1] 2D [;] 2 [�] [D][ �] [1] 0; else ## 6. Performance evaluation and simulation results In this section, we evaluate the performance of the proposed certificate management scheme in terms of effectiveness and reliability of revocation scheme and communication cost. To verify the performance in terms of cost of communication and effectiveness of revocation, we compare THCM with two existing schemes; CCRCV by Liu et al [35] and a voting-based scheme proposed by Luo et al [36]. ## 6.1 Simulation environment The MANET simulation setup is performed in QualNet 4.5 environment with IDE: Visual studio 2013, programming language: VC?? and SDK: NSC_XE-NETSIMCAP (Network Simulation and Capture). The nodes follow a random waypoint approach (RWP) presented by the authors Bettstetter and Wagner [41], Bai and Helmy [42] and Aschenbruck et al [43], where the speed and the direction of each nodes are chosen randomly and independently. When the simulation starts, each node chooses one location randomly as the destination within the simulation field. The nodes then move with constant velocity chosen uniformly and randomly in a range 0½ ; Vm�; where Vm is the maximum range of velocity that a node can travel. When the node reaches its destination, it halts for a time period, referred as halt time ðT haltÞ. If T halt ¼ 0, a continuous mobility can be experienced. However, when the T halt expires, the nodes again move randomly in the simulation field. On the one hand, we evaluate the performance of the proposed THCM by varying the two parameters Vm and T halt for topology alterations, that is, If Vm is less and T halt is high, a relatively stable topology can be achieved. On the other hand a highly dynamic topology can be obtained if Vm is high and T halt is less. PRD Dð Þ ¼ 8 >>>>< >>>>: 23 ð Þ The Cassign and Crevoke also depends on the number of certificate assigned k and the average number of certifið Þ cate revoked per time slot ��CT . Therefore, the overall communication cost is given as, (16) ) p CommC ¼ NT 3 A23[ffiffi]s[2] ðReql þ ReslÞ p þ �p4Dffiffiffi3 �p3 [�] p[2 cos][�][1] p2Dffiffiffi3�� � 2 � NT 3 A23[ffiffi]s[2] � Resl þ [C]T 3 A23[ffiffi]s[2] � Revl þ [C]T [�] p � �p4Dffiffiffi3 �p3 [�] [2 cos][�][1] p2Dffiffiffi3�� � 3 A23[ffiffi]s[2] � Revl 24 ð Þ ----- Trust-based hexagonal clustering for efficient certificate 1147 ## 6.2 Effectiveness of revocation scheme The revocation rate and revocation time are the two core factors that evaluate the effectiveness of any revocation scheme. The potency of the proposed THCM scheme is shown in terms of revocation rate and revocation time as shown in figures 6 and 7. Revocation time is defined as the time period by which an attacker launches an attack before its certificate gets revoked. Whereas, the revocation rate represents the rate of attacker nodes revoked before launching the attacks. To analyze the impact of attacker nodes on revocation, we deploy 100 nodes in the network, whereas the attacker nodes range upto 30% to 35%. As shown in figures 6 and 7, the effectiveness of revocation is performed by comparing the proposed revocation scheme with an existing CCRVC scheme and voting scheme. Figure 6 shows the change in the revocation time with the increase in attacker nodes, between the proposed THCM scheme and existing non-trust-based schemes of Figure 6. Revocation time. Figure 7. Revocation rate. Liu et al [35] and Luo et al [36]. On comparing, the voting scheme takes higher revocation time than the other two schemes. This is due to the waiting period for the votes from different nodes to make revocation decision. THCM maintains a beneficial revocation time even with higher number of attackers. A maximum revocation time of 60 s can be noted in THCM for highest percentage of attacker. The revocation times of the three different schemes for increasing number of attackers are given in table 1. Figure 7 demonstrates the revocation rate for different attacker node levels. It is noted that the revocation rate improves with the increase in attackers for proposed trustbased revocation scheme. The proposed THCM revocation scheme works well on the attacker by calculating the trustability of each node. Even though the rate drops a little in between, it gradually increases for larger number of attackers, that is, a revocation rate of 98% is achieved for 35% of attackers in THCM. The simulation results of the three schemes for various percentage of attackers are given in table 2. ## 6.3 Reliability of revocation The reliability of our scheme can be determined from the proposed algorithm by calculating the probability of success revocations given by Wasef and Shen [44]. 0 B Psuccess ¼ BB1 � @ � p 1 �N � n � p �N n x 1 C C C A : 26 ð Þ Table 1. Revocation time (s) of different key management schemes. Number of attackers Schemes 5 10 15 20 25 30 Voting scheme 110 120 128 130 132 134 CCRCV 20 32 55 68 70 78 THCM 17 23 44 57 59 60 Table 2. Revoked attackers (in %) of different key management schemes. Percentage of attackers Schemes 5 10 15 20 25 30 35 Voting scheme 92 91 90 88 85 82 78 CCRCV 95 96 97 96 95 92 93 THCM 97 98 97 96 96 98 98 ----- 1148 V S Janani and M S K Manikandan Figure 8. Successful revocation probability. Figure 9. Cost of communication. Figure 8 shows the probability of successful revocations ðPsuccessÞ with different values of positive encounters (p), negative encounters (n), and secret key (x), varying the average number of nodes N within the communication ð Þ range of a node in the cluster. It is observed that Psuccess increases with N for constant n and x. It can also be noted that the Psuccess increases with increase in n and decrease in p. This indicates the vulnerability strength of the system against attackers, that is, if the negative encounters n rises, the network is subjected to ð Þ more number of attackers to which a desired security level should be provided. The above discussion proves the reliability of our proposed THCM scheme with desired security level. ## 6.4 Communication cost In the proposed certificate management schemes, the main factors that have high impact on the communication cost are certificate revocation and certificate issuing processes. Figures 9 and 10 represents the efficiency of our scheme on cost factor. The communication cost can be conserved in a successful manner with this scheme. Comparison of two different schemes run in a simulation environment of 100 nodes that follow a random walk mobility model Bettstetter and Wagner [41] (a specific RWP mobility model with T halt ¼ 0), in which each node changes its mobility rate at different time intervals. The proposed THCM scheme is compared in terms of cost of communication, with CCRVC and voting-based scheme for different number of attackers, as in figures. We plotted the cost of certificates of each scheme in figure 9 where we can see that CCRVC is costlier than other two schemes. Although THCM is costlier than votingbased scheme for small numbers of attackers, the cost of the voting scheme increases abruptly with the number of attackers. At the most, the cost range is limited to 128 in THCM, where it is 198 for voting scheme and 235 in CCRVC. Our cost-conservative certificate management scheme is analyzed for different number of certificates revoked, as shown in figure 10. It is noticed that the communication cost increases 180 for the maximum number of attackers, which is lower compared with other two schemes (i.e., voting scheme attained a cost of 240 and CCRVC reaches 212). It is noted that the proposed THCM scheme outperforms the voting–based method in terms of communication cost for different number of attackers as well as certificates revoked. The communication cost in issuing the certificate and communication cost in sending the revocation informations for existing schemes and the proposed THCM scheme are given in tables 3 and 4. Figure 10. Communication cost with revocation. ----- Trust-based hexagonal clustering for efficient certificate 1149 Table 3. Cost of communication (certificates/node) of different key management schemes. Percentage of attackers Schemes 5 10 15 20 25 30 ## 7. Empirical analysis 7.1 Emulator platform A real-world certificate management system is developed with Android Emulator: T-Engine, which is renamed as TRON Forum on April 2015, to analyze the performance of proposed THCM scheme. The emulator introduces the QULANET simulator into a real network. T-Engine enables the users to rapidly build a ubiquitous computing solution utilizing off-the-shelf components with complete mobility permitted as presented by Krikke [45] and Noboru and Ken [46]. The middleware library available for T-Engine supports network protocol, GUI, and specified security tools (as presented by Khan and Sakamura [47] and many other added features in order to emulate real smart mobile nodes. The platform also supports the resource distribution of software and tamper resistant network security. Figure 11 shows the emulator platform run for the proposed THCM scheme with 50 mobile nodes represented as TM i ; 0 i 50. ð Þ � � Our study facilitates to understand the certificate revocation time, the rate of revoked node, and the communication cost. This study also provides solid evidence on the optimal certificate management for the three schemes for different number of attacker nodes. Figure 12(a) and (b) shows the emulator output for the effectiveness of revocation scheme in terms of revocation time and rate of revocation of voting scheme, CCRVC, and THCM schemes. The numbers of attackers vary from 5% to 50% of all the cases. The results in the emulator evidently show there is no significant change in the time and rate of revocation comparing with that of QUALNET. Figure 13(a) and (b) represents the emulator output for the cost factor. The results are plotted for the cost of communication with respect to certificate revocation and certificate issuing processes. Likewise the revocation scheme, the cost-conservative feature of the THCM also Figure 11. Emulator execution. Voting scheme 30 62 100 118 144 198 CCRCV 78 146 170 196 224 230 THCM 56 74 88 102 120 124 Table 4. Communication cost with revocation for different key management schemes. Average number of certificates revoked Schemes 5 10 15 20 25 30 Voting scheme 52 86 112 160 200 236 CCRCV 64 86 124 146 188 212 THCM 50 74 100 128 152 180 ## 6.5 Security analysis This section analyzes the proposed THCM scheme against security attacks discussed in section 4. Resilience against Forging attack: To forge the revocation information, an attacker should determine the trust degree and distrust degree of any node. The attack node should be aware of the positive and negative encounters for calculating the trust or distrust degree. Further, the attacker node should collect the information regarding successive interactions as well as location information of that node to forge the total revocation information. Furthermore, the CA signs the revocation message and sends to all the nodes in the cluster, which cannot be forged. From the above discussion our THCM scheme is resistant enough to forge attack. Resilience against collusion attack: When a node’s certificate that it likely to visit in near future is about to expire, in THCM, request for a fresh set of certificates is sent in advance. Hence, it is assured that the revoked node can never have the entire revocation certificate and so they cannot collude to revoke any other node. Therefore, the proposed THCM is resilient against collusion attack in the network. Resilience against revocation denial attack: THCM conducts the verification phase in section 4 each time, which includes the sender certificate check, validity check, and the signature check. By this the CA detects and discards fallacious process. In addition, since the proposed certificate management scheme adopts a probability certificate assignment technique, same revocation information may be found with more than one node. Consequently, the CA identifies the multiple copies and excludes the duplicate ones. Hence, the THCM scheme exhibit robustness against revocation denial attacks. ----- 1150 V S Janani and M S K Manikandan Figure 12. Emulator output for effectiveness of revocation scheme. (a) Revocation time and (b) Revocation rate. Figure 13. Emulator output for efficiency on cost factor. (a) Cost of communication and (b) communication cost with revocation. shows no much significant variation compared with the simulation results. ## 7.2 Simulation and emulation: a comparison The QUALNET and T-Engine outputs are compared and plotted in figure 14(a), (b), (c), and (d). The graphs shows the performance effectiveness of the proposed THCM scheme compared with the existing certificate management methods. The results show that THCMs have no significant variation in the values while implementing the simulator in a live environment. The values obtained by emulation are very close to that of simulation, which certainly shows the optimal management of THCM scheme. To get an efficient and accurate output, multiple trails were performed with the simulation and emulation parameters using T-Test methodology. T-Test is conducted with 10 numbers of trails in order to prove the accuracy of the output statistically. Various hypotheses were stated to support the T-Testing, which is summarized in table 5. To compare the performance boost of THCM, simulation as well as emulation results was processed through statistical tests and calculations. The mean and standard deviation are calculated from the data acquired through 10 rounds with a limit of significance (LoS) set at 2. The values within the specified LoS are assumed to be acceptable and those above the LoS are assumed as insignificant. The proposed THCM scheme statistically demonstrated that there is no significant difference in the simulation and emulation values. The mean for each parameter is calculated using the following formula: ----- Trust-based hexagonal clustering for efficient certificate 1151 Figure 14. Simulation vs emulation. (a) Revocation time. (b) Revocation rate. (c) Cost of communication. (d) Communication cost with revocation. 1 N N X yi ð27Þ i¼1 where N is the total number of data trials, yi is the observed value and the standard deviation is calculated as � rM�M ¼ sqrt [r]source[2]Na þ [r]source[2]Nb � 28 ð Þ THCM to efficiently partition the network into nonoverlapping clusters and to manage certificates. Our approach enables each node to establish a trust value with other interacting nodes, in each Voronoi hexagonal cluster, with minimal communication cost and maximum utilization of the certificate management scheme. Simulation results show that our scheme achieved a revocation rate of 98% in maximum of 60 s, for a higher percentage of attackers. We seek to lower the cost of certificate assignment and revocation, as shown in the simulation results. We also developed an analytic—statistical approach to study the impact of certificate management on attacker nodes and cost in a real-time MANET emulator. In addition, we provided a simple mathematical analysis to justify the results. We believe that the proposed scheme works efficiently and also has remarkable contributions to the modeling, design, and analysis of an effective certificate management scheme for MANETs. Therefore, our r[2]source [is the variance of source population and] Na and Nb: are the sizes of the two types of samples. ## 8. Conclusion In this paper, we have addressed complete security measure against attackers for mobile ad hoc networks. In contrast to the existing techniques, we have proposed ----- 1152 V S Janani and M S K Manikandan Table 5. Statistical analysis of key management schemes. QUALNET values EMULATOR values Trails x ið Þ Mean ðMean � x ið ÞÞ[2] x ið Þ Mean ðMean � x ið ÞÞ[2] (i) Revocation time 1 0 47.3 2237.29 0 48.8 2381.44 2 19 47.3 800.89 22 48.8 718.24 3 25 47.3 497.29 28 48.8 432.64 4 47 47.3 0.09 45 48.8 14.44 5 57 47.3 94.09 58 48.8 84.64 6 60 47.3 161.29 63 48.8 201.64 7 62 47.3 216.09 64 48.8 231.04 8 66 47.3 349.69 65 48.8 262.44 9 68 47.3 428.49 72 48.8 538.24 10 69 47.3 470.89 71 48.8 492.84 Average mean 47.3 48.8 Standard deviation 22.9261859 23.14649 T-test result 1.605403 T-test hypothesis Since T-Test Result is less than 2, there is no significant difference between simulated and emulated results (ii) Revocation rate 1 98 95.4 6.76 96 93.9 4.41 2 96 95.4 0.36 97 93.9 9.61 3 100 95.4 21.16 96 93.9 4.41 4 96 95.4 0.36 95 93.9 1.21 5 95 95.4 0.16 94 93.9 0.01 6 95 95.4 0.16 92 93.9 3.61 7 93 95.4 5.76 94 93.9 0.01 8 94 95.4 1.96 92 93.9 3.61 9 92 95.4 11.56 92 93.9 3.61 10 95 95.4 0.16 91 93.9 8.41 Average mean 95.4 93.9 Standard deviation 2.2 1.9723083 T-test result 0.1456 T-test hypothesis Since T-Test Result is less than 2, there is no significant difference between simulated and emulated results (iii) Communication cost 1 0 102.3 10465.3 0 103.7 10753.69 2 57 102.3 2052.09 56 103.7 2275.29 3 78 102.3 590.49 77 103.7 712.89 4 92 102.3 106.09 93 103.7 114.49 5 107 102.3 22.09 109 103.7 28.09 6 118 102.3 246.49 120 103.7 265.69 7 130 102.3 767.29 133 103.7 858.49 8 136 102.3 1135.69 137 103.7 1108.89 9 143 102.3 1656.49 147 103.7 1874.89 10 162 102.3 3564.09 165 103.7 3757.69 Average mean 102.3 103.7 Standard deviation 45.39394233 46.637002 T-test result 0.06803 T-test hypothesis Since T-Test Result is less than 2, there is no significant difference between simulated and emulated results (iv) Communication cost per average number of revocation 1 0 128.7 16563.7 0 129.2 16692.64 2 54 128.7 5580.09 53 129.2 5806.44 3 79 128.7 2470.09 84 129.2 2043.04 4 105 128.7 561.69 103 129.2 686.44 5 128 128.7 0.49 128 129.2 1.44 6 155 128.7 691.69 153 129.2 566.44 7 181 128.7 2735.29 185 129.2 3113.64 8 189 128.7 3636.09 187 129.2 3340.84 ----- Table 5 continued Trails Trust-based hexagonal clustering for efficient certificate 1153 QUALNET values EMULATOR values x ið Þ Mean ðMean � x ið ÞÞ[2] x ið Þ Mean ðMean � x ið ÞÞ[2] 9 194 128.7 4264.09 198 129.2 4733.44 10 202 128.7 5372.89 201 129.2 5155.24 Average mean 128.7 129.2 Standard deviation 64.71174546 64.915021 T-test result 0.01725 T-test hypothesis Since T-Test Result is less than 2, there is no significant difference between simulated and emulated results scheme, THCM can be adequately adopted for wireless ad hoc networks. Acknowledgment This research is supported by All India Council for Technical Education (AICTE), Government of India. ## References [1] Ko Y B and Vaidya N H 1999 Geocasting in mobile ad hoc networks: Location-based multicast algorithms. In: Proceedings of IEEE WMCSA, pp. 101–110 [2] Bellur B 2008 Certificate assignment strategies for a pkibased security architecture in a vehicular network. Proceedings IEEE GLOBECOM, pp 1–6 [3] Chang R S and Wang S H 2008 Hexagonal collaboration groups in sensor networks. Proc. IEEE CCNC pp 358–359 [4] Zhuang Y, Gulliver T A and Coady Y 2013 On planar tessellations and interference estimation in wireless ad-hoc networks. IEEE Wireless Commun. Lett. 2(3): 331–334 [5] Fan Y, Yulan Z and Ping X 2015 An overview of ad hoc network security. communications in computer and information science, Springer, vol. 557, pp 129–137 [6] Cao M and Hadjicostis C N 2003 Distributed algorithms for Voronoi diagrams and applications in ad-hoc networks. Technical Report UILUENG-03-2222 [7] Chau M, Cheng R, B Kao B and Ng J 2006 Uncertain data mining: An example in clustering location data. In: Proceedings of PAKDD, pp 199–204 [8] Cheng R, Xie X, Yiu M L, Chen J and Sun L 2010 Uvdiagram: A Voronoi diagram for uncertain data. In: Proceedings of 26th IEEE International Conference on Data Engineering, pp 796–807 [9] Kao B, Lee S D, Lee F, Cheung D and Ho W S 2010 Clustering uncertain data using Voronoi diagrams and R-tree index. IEEE Trans. Knowledge and Data Eng. 22(9): 1219–1233 [10] Mohamed Aissa and Abdelfettah Belghith 2014 A node quality based clustering algorithm in wireless mobile ad hoc networks. In: Proceedings of the 5th International Conference on Ambient Systems, Networks and Technologies, Elsevier, vol. 32, pp. 174–181 [11] Abdelhak B, Abdelhak B and Saad H 2013 Survey of clustering schemes in mobile ad hoc networks, Commun. Netw. pp 8–14 [12] Khalid H, Abdul H Abdullah, Khalid M Awan, Faraz Ahsan, Akhtab Hussain and Johor Bahru 2013 Cluster head election schemes for WSN and MANET: A survey. World Appl. Sci. J. 23(5): 611–620 [13] Mohd Junedul Haque, Mohd Muntjir and Hussain A S 2015 A comparative survey for computation of cluster-head in MANET. Int. J. Comput. Appl. 118 (3): 6–9 [14] Fan P, Li G, Kai Cai and Letaief K B 2007 On the geometrical characteristic of wireless ad-hoc networks and its application in network performance analysis. IEEE Trans. Wireless Commun. 6(4): 1256–1265 [15] Stojmenovic I, Ruhil A P and Lobiyal D K 2006 Voronoi diagram and convex hull based geocasting and routing in wireless networks. Wireless Commun. Mobile Comput. 6: 247–258 [16] Ngai W K, Kao B, Chui C K, Cheng R, Chau M and Yip K Y 2006 Efficient clustering of uncertain data. Proceedings of ICDM pp. 436–445 [17] Renu D, Manju Khari and Yudhvir Singh 2012 Survey of trust schemes on ad-hoc network. Int. J. AdHoc Netw. Syst. springer, 2 pp. 170–180 [18] Manju Khari and Yudhvir Singh 2012 Different ways to achieve trust in MANET. Int. J. AdHoc Netw. Syst. 2(2): 1–10 [19] Jingwei H and David N 2009 A calculus of trust and its application to pki and identity management. In: Proceedings of 8th Symposium on Identity and Trust on the Internet, pp. 23–37 [20] Liu K, Abu-Ghazaleh N and Kang K 2007 Location verification and trust management for resilient geographic routing. J. Parallel Distributed Comput. 67: 215–228 [21] Ferdous R, Muthukkumarasamy V and Sithirasenan E 2011 Trust-based cluster head selection algorithm for mobile ad hoc networks. In: Proceedings of International Joint Conference IEEE TrustCom pp. 589–596 [22] Cho J H, Chan K S, Chen I R 2013 Composite trust-based public key management in mobile ad hoc networks. ACM 28th Symposium on Applied Computing, Coimbra, Portugal, pp 1949–1956 [23] Wei Z, Tang H, Richard Yu, Wang M and Mason P 2014 Security enhancements for mobile ad hoc networks with trust management using uncertain reasoning. IEEE Trans. Vehicular Technol. 63(9): 4647–4658 ----- 1154 V S Janani and M S K Manikandan [24] Ing-Ray Chen, Jia Guo, Fenye Bao and Jin-Hee Cho 2014 Trust management in mobile ad hoc networks for bias minimization and application performance maximization ad hoc networks. Ad hoc networks. Elsevier, vol. 19, pp. 59–74 [25] Deering S, Estrin D, Farinacci D, Jacobson V, Helmy A and Wei L 1997 Protocol independent MulticastVersion 2, dense mode specification. Internet Draft, ftp://ietf.org/internetdrafts/draft-ietf-idmr-pim-dm-spec-05.txt [26] Mohammad M Qabajeh, Aisha H Abdalla, Othman O Khalifa and Liana K Qabajeh2015 A survey on scalable multicasting in mobile ad hoc networks. Wireless Personal Commun. 80(1): 369–393 [27] Kanchan D and Asutkar G M 2016 Enhancement in the performance of routing protocols for wireless communication using clustering, encryption, and cryptography. Artificial intelligence and evolutionary computations in engineering systems, advances in intelligent systems and computing, vol. 394, pp 547–558. [28] Jormakka J and Jormakka H 2014 Revocation of user certificates in a military ad hoc network. Brazilian J. Inform. Security Cryptogr. 1(1): 1–3 [29] Yki K and Mikko S 2014 Survey of certificate usage in distributed access control. Computers & security, Elsevier vol 44, pp 16–32 [30] Wei Liu, Hiroki Nishiyama, Nirwan Ansari and Nei Kato 2011 A study on certificate revocation in mobile ad hoc networks. IEEE International Conference on Communications (ICC), pp 1–5 [31] Mohammad Masdari and Javad P B 2012 Distributed certificate management in mobile ad hoc networks. Int. J. Appl. Inform. Syst. 4(6): 33–40 [32] Mohamed M E A Mahmoud, Jelena Misic, Kemal Akkaya and Xuemin Shen 2015 Investigating public-key certificate revocation in smart grid. IEEE Internet Things J. 2: 490–503 [33] Mohammad Masdari, Sam J and Jamshid B 2015a Improving OCSP-based certificate validations in wireless ad hoc networks. Wireless Personal Commun., 82 (1): 377–400 [34] Mohammad Masdari, Sam J, Jamshid B and Ahmad Khadem-Zadeh 2015b Towards efficient certificate status validations with E-ADOPT in mobile ad hoc networks. Computers & security, Elsevier, vol 49, pp. 17–27 [35] Liu W, Nishiyama H, Ansari N, Yang J and Kato N 2013 Cluster-based certificate revocation with vindication capability for mobile ad hoc networks. IEEE Trans. Parallel Distributed Syst. 24(2): 239–249 [36] Luo H, Kong J, Zerfos P, Lu S and Zhang L 2004 URSA: Ubiquitous and robust access control for mobile ad hoc networks. IEEE/ACM Trans. Netw. 12(6): 1049–1063 [37] Taisuke Izumi, Tomoko Izumi, Hirotaka Ono and Koichi Wada 2015 Approximability of minimum certificate dispersal with tree structures. Theoretical computer science, Elsevier, vol. 591, pp 5–14 [38] Park K, Nishiyama H, Ansari N and Kato N 2010 Certificate revocation to cope with false accusations in mobile ad hoc networks. Proceedings of IEEE 71st Vehicular Technology Conference pp 1–5 [39] Raya M, Manshaei M H, Felegyhazi M and Hubaux J P 2008 Revocation games in ephemeral networks. Proceedings of ACM CCS [40] Mawloud Omar, Hamida B, Lydia Mammeri, Amel Taalba and Abdelkamel T 2016 Secure and reliable certificate chains recovery protocol for mobile ad hoc networks. J. Netw. Comput. Appl., Elsevier, 62: 153–162 [41] Bettstetter C and Wagner C 2002 The spatial node distribution of the random waypoint mobility model. Proceedings German Workshop on Mobile Ad Hoc Networks (WMAN) [42] Bai F and Helmy A 2004 A survey of mobility modeling and analysis in wireless ad hoc networks. Wireless ad hoc and sensor networks. Kluwer academic publishers [43] Aschenbruck N, ErnstR, Gerhards-Padilla E and Schwamborn M 2010 BonnMotion – A mobility scenario generation and analysis tool. In: Proceedings of the 3rd International Conference on Simulation Tools andTechniques [44] Wasef A and Shen X 2009 EDR: Efficient decentralized revocation protocol for vehicular ad hoc networks. IEEE Trans. Veh. Tech. 58(9): 5214–5224 [45] Krikke 2005 T-engine: Japan’s ubiquitous computing architecture is ready for prime time. IEEE Pervasive Comput. 4(2): 4–9 [46] Noboru K and Ken S 2010 Ubiquitous ID: Standards for ubiquitous computing and the internet of things. IEEE Pervasive Comput. 9(4): 98–101 [47] Khan M F F and Sakamura K 2015 Tamper-resistant security for cyber-physical systems with eTRON architecture. IEEE International Conference on Data Science and Data Intensive Systems, Sydney, NSW, pp 196–203 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/s12046-016-0545-0?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/s12046-016-0545-0, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://www.ias.ac.in/article/fulltext/sadh/041/10/1135-1154" }
2,016
[]
true
2016-10-01T00:00:00
[]
19,516
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00d793b0c383f3232ba96b73d6917925ebb75ebe
[ "Computer Science" ]
0.853692
Single-View and Multiview Depth Fusion
00d793b0c383f3232ba96b73d6917925ebb75ebe
IEEE Robotics and Automation Letters
[ { "authorId": "7726696", "name": "José M. Fácil" }, { "authorId": "144583933", "name": "Alejo Concha" }, { "authorId": "2987153", "name": "L. Montesano" }, { "authorId": "143750691", "name": "Javier Civera" } ]
{ "alternate_issns": null, "alternate_names": [ "IEEE Robot Autom Lett" ], "alternate_urls": [ "http://ieeexplore.ieee.org/servlet/opac?punumber=7083369" ], "id": "93c335b7-edf4-45f5-8ddc-7c5835154945", "issn": "2377-3766", "name": "IEEE Robotics and Automation Letters", "type": null, "url": "https://www.ieee.org/membership-catalog/productdetail/showProductDetailPage.html?product=PER481-ELE" }
Dense and accurate 3-D mapping from a monocular sequence is a key technology for several applications and still an open research area. This letter leverages recent results on single-view convolutional network (CNN)-based depth estimation and fuses them with multiview depth estimation. Both approaches present complementary strengths. Multiview depth is highly accurate but only in high-texture areas and high-parallax cases. Single-view depth captures the local structure of midlevel regions, including texture-less areas, but the estimated depth lacks global coherence. The single and multiview fusion we propose is challenging in several aspects. First, both depths are related by a deformation that depends on the image content. Second, the selection of multiview points of high accuracy might be difficult for low-parallax configurations. We present contributions for both problems. Our results in the public datasets of NYUv2 and TUM shows that our algorithm outperforms the individual single and multiview approaches. A video showing the key aspects of mapping in our single and multiview depth proposal is available at https://youtu.be/ipc5HukTb4k.
This paper has been accepted for publication in IEEE Robotics and Automation Letters. DOI:10.1109/LRA.2017.2715400 [IEEE Xplore: http://ieeexplore.ieee.org/document/7949041/](http://ieeexplore.ieee.org/document/7949041/) c 2017 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any _⃝_ current or future media, including reprinting /republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. ----- ## Single-View and Multi-View Depth Fusion ### Jos´e M. F´acil[1], Alejo Concha[1], Luis Montesano[1][,][2] and Javier Civera[1] **_Abstract— Dense and accurate 3D mapping from a monoc-_** **ular sequence is a key technology for several applications and** **still an open research area. This paper leverages recent results** **on single-view CNN-based depth estimation and fuses them** **with multi-view depth estimation. Both approaches present** **complementary strengths. Multi-view depth is highly accurate** **but only in high-texture areas and high-parallax cases. Single-** **view depth captures the local structure of mid-level regions,** **including texture-less areas, but the estimated depth lacks** **global coherence. The single and multi-view fusion we propose is** **challenging in several aspects. First, both depths are related by** **a deformation that depends on the image content. Second, the** **selection of multi-view points of high accuracy might be difficult** **for low-parallax configurations. We present contributions for** **both problems. Our results in the public datasets of NYUv2** **and TUM shows that our algorithm outperforms the individual** **single and multi-view approaches. A video showing the key** **aspects of mapping in our Single and Multi-view depth proposal** **[is available at https://youtu.be/ipc5HukTb4k.](https://youtu.be/ipc5HukTb4k)** **_Index Terms— Deep Learning in Robotics and Automation,_** **Mapping, SLAM** I. INTRODUCTION Estimating an online, accurate and dense 3D scene reconstruction from a general monocular sequence is one of the fundamental research problems in computer vision. The problem has nowadays a high relevance, as it is a key technology in several emerging application markets (augmented and virtual reality, autonomous cars and robotics in general). The state of the art are the so-called direct mapping methods [1], that estimate an image depth by minimizing a regularized cost function based on the photometric error between corresponding pixels in several views. The accuracy of the multi-view depth estimation depends mainly on three factors: 1) The geometric configuration, with lower accuracies for low-parallax configurations; 2) the quality of the correspondences among views, that can only be reliably estimated for high-gradient pixels; and 3) the regularization function, typically the Total Variation norm, that is inaccurate for large texture-less areas. Due to this poor performance on large low-gradient areas, semi-dense maps are sometimes estimated only in high-gradient image pixels for visual direct SLAM (e.g., [2]). Such semi-dense maps are accurate in high-parallax configurations but not a complete model of We gratefully acknowledge the support of NVIDIA Corporation for the donation of a Titan X GPU, the Spanish government (projects DPI201232168 and DPI2015-67275), the Aragon regional government (Grupo DGA T04-FSE) and the University of Zaragoza (JIUZ-2015-TEC-03). 1The authors are with the I3A, University of Zaragoza, Spain _{jmfacil, montesano, jcivera}@unizar.es,_ aconchabelenguer@gmail.com 2Luis Montesano is also with Bit&Brain Technologies SL. RGB sequence Single-view Prediction CNN Depth Fusion Our Fusion Multi-view Estimation Fig. 1: Overview of our proposal. The input is a set of overlapping monocular views. The learning-based singleview and geometry-based multi-view depth are fused, outperforming both of them. All the depth images are colornormalized for better comparison. This figure is best viewed in color. the viewed scene. Low-parallax configurations are mostly ignored in the visual SLAM literature. An alternative method is single-view depth estimation, which has recently experienced a qualitative improvement in its accuracy thanks to the use of deep convolutional networks [3]. Their accuracy is still lower than that of multi-view methods for high-texture and high-parallax points. But, as we will argue in this paper, they improve the accuracy of multi-view methods in low-texture areas due to the highlevel feature extraction done by the deep networks –opposed to the low-level high-gradient pixels used by the multi-view methods. Interestingly, the errors in the estimated depth seem to be locally and not globally correlated since they come from the deep learning features. The main idea of this paper is to exploit the information of single and multi-view depth maps to obtain an improved depth even in low-parallax sequences and in low-gradient areas. Our contribution is an algorithm that fuses these complementary depth estimations. There are two main challenges in this task. First, the error distribution of the single-view estimation has several local modes, as it depends on the image content and not on the geometric configuration. Single and multi-view depth are hence related by a content-dependent deformation. Secondly, modeling the multi-view accuracy is not trivial when addressing general cases, including high and low-parallax configurations. ----- We propose a method based on a weighted interpolation of the single-view local structure based on the quality and influence area of the multi-view semi-dense depth and evaluate its performance in two public datasets –NYU and TUM. The results show that our fusion algorithm improves over both individual single and multi-view approaches. The rest of the paper is organized as follows. Section II describes the most relevant related work. Section III motivates and details the proposed algorithm for single and multi-view fusion. Section IV presents our experimental results and, finally, Section V contains the conclusions of this work. II. RELATED WORK We classify the related work for dense depth estimation into two categories: methods based in multiple views of the scene and those which predict depth from one single image. _A. Multi-View Depth_ In the multi-view depth estimation, [1], [4], [5] are the first works that achieved dense and real-time reconstructions from monocular sequences. Some of the most relevant aspects are the direct minimization of the photometric error –instead of the traditional geometric error of sparse reconstructions– and the regularization of the multi-view estimation by adding the total variation (TV) norm to the cost function. TV regularization has low accuracy for large textureless areas, as shown recently in [6], [7], [8] among others. In order to overcome this [6] proposes a piecewise-planar regularization; the plane parameters coming from multiview superpixel triangulation [9] or layout estimation [10]. [7] proposes higher-order regularization terms that enforce piecewise affine constraints even in separated pixels. [8] selects the best regularization function among a set using sparse laser data. Building on [6], [11] adds the sparse data-driven 3D primitives of [12] as a regularization prior. Compared to these works, our fusion is the first one where the information added to the multi-view depth is fully dense, data-driven and single-view; and hence it does not rely on additional sensors, parallax or Manhattan and piecewiseplanar assumptions. It only relies on the network capabilities for the current domain, assuming that the test data follows the same distribution that the data used for training. Due to the difficulty of estimating an accurate and fully dense map from monocular views there are several approaches that estimate only the depth for the highest-gradient pixels (e.g., [2]). While this approach produces maps of higher density than the more traditional feature-based ones (e.g., [13]), they are still incomplete models of the scene and hence their applicability might be more limited. _B. Single-View Depth_ Depth can be estimated from a single view using different image cues, for example focus (e.g., [14]) or perspective (e.g., [15]). Learning-based approaches, as the one we use, basically discover RGB patterns that are relevant for accurate depth regression. |Col1|High-Gradient|Low-Gradient| |---|---|---| |Multi-View|0.18|1.02| |Single-View|0.36|0.42| TABLE I: Median depth error [m] for single and multiview depth estimation, and high and low-gradient pixels. This evaluation has been done in the sequence living room 0030a from the NYUv2 dataset (one of the sequences with higher parallax). The normalized threshold between high and lowgradient pixels is 0.35 (gray scale). The pioneering work of Saxena et al. [16] trained a MRF to model depth from a set of global and local image features. Before that, [17] presented an early approach to depth prediction from monocular and stereo cues. Eigen _et al. [18] presented a two deep convolutional neural network_ (CNN) stacked, one to predict global depth an the second one that refines it locally. Build upon this method, [3] recently presented a three scale convolutional network to estimate depth, surface normals and semantic labeling. Liu et al. [19] use a unified continuous CRF-and-CNN framework to estimate depth. The CNN is used to learn the unary and pairwise potentials that the CRF uses for depth prediction. Based on [3], [20] incorporates mid-level features in its prediction using skip-layers. It shows competitive results and a small batch-size training strategy that makes their network faster to train. [21] introduces a different method to predict depth from single-view using deep neural networks, showing that training the network with a much richer output improves the accuracy. [22] formulates the depth prediction as a classification problem and the net output is a pixelwise distribution over a discrete depth range. Finally, [23] presents an unsupervised network for depth prediction using stereo images. III. SINGLE AND MULTI-VIEW DEPTH FUSION State-of-the-art multi-view techniques have a strong dependency on high-parallax motion and heterogeneous-texture scenes. Only a reduced set of salient pixels that hold both constraints has a small error, and the error for the majority of the points is large and uncorrelated. In contrast, singleview methods based on CNN networks achieve reasonable errors in all the image but they are locally correlated. Our proposal exploits the best properties of these two methods. Specifically, it uses a deep convolutional network (CNN) to produce rough depth maps and fuses their structure with the results of a semi-dense multi-view depth method (Fig. 1). Before delving into the technical aspects, we will motivate our proposal with some illustrative results. Table I shows the median depth error of the high-gradient and low-gradient pixels for a multi-view and single view reconstruction using a medium/high-parallax sequence of the NYUv2 dataset. For the multi-view reconstruction, the error for the low-gradient pixels increases by a factor of 2. Notice that the opposite happens for the single-view reconstruction: the error of highgradient pixels is the one increasing by a factor of 2. For this experiment, the threshold used to distinguish between ----- 2000 1000 00 1 2 3 4 0 1 2 3 4 0 1 2 3 4 ### Error (m) Fig. 2: Histogram of single-view depth error [m] for three sample sequences. Notice the multiple modes, each one corresponding to a local image structure, this can be seen in the error images in the top row of the figure. high and low-gradient pixels is 0.35 in gray scale (where the maximum gradient would be 1). Furthermore, the single-view depth error usually has a structure that indicates the presence of local correlations. For instance, Fig. 2 shows the histogram of the single-view depth estimation error for three different sequences (two of the NYUv2 dataset and one of the TUM dataset). Notice that the error distribution is grouped in different modes, each one corresponding to an image segment. This effect is caused by the use of the high-level image features of the latest layers of the CNN network, that extend over dozens of pixels in the original image and hence over homogeneous texture areas. The different nature of the errors can be exploited to outperform both individual estimations. This fusion, however, cannot be na¨ıvely implemented with a simple global model as it requires content-based deformations. In the next subsections we detail the specific multi and single-view methods that we use in this work and our fusion algorithm. _A. Multi-view Depth_ For the estimation of the multi-view depth we adopt a direct approach [2], that allows us to estimate a dense or semi-dense map in contrast to the more sparse maps of the feature-based approaches. In order to estimate the depth of a keyframe Ik we first select a set of n overlapping frames {I1, . . ., Io, . . ., In} from the monocular sequence. After that, every pixel x[k]l [of the reference image][ I][k][ is first] backprojected at an inverse depth ρ and projected again in every overlapping image Io. every high-gradient pixel if we want a semi-dense map) x[k]l in the reference image Ik and its corresponding one x[o]l [in] every other overlapping image Io at an hypothesized inverse depth ρl, 1 _C(ρ)_ = _n_ _n_ � _o=1,o≠_ _k_ _t_ � _ϵl(Ik, Io, x[k]l_ _[, ρ][l][)][.]_ (2) _l=1_ � _x[o]l_ [=][ T][ko][(][x]l[k][, ρ][l][) =][ KR]ko[⊤] �� _K[−][1]x[k]l_ _||K[−][1]x[k]l_ _[||]_ _ρl_ � _−_ _tko_ The error ϵl(Ik, Io, x[k]l _[, ρ][l][)][ for each individual pixel][ x]l[k]_ [is] the difference between the photometric values of the pixel and its corresponding one _ϵl(Ik, Io, x[k]l_ _[, ρ][l][)]_ = _Ik(x[k]l_ [)][ −I][o][(][x]l[o][)][.] (3) The estimated depth for every pixel _ρˆ_ = ( ˆρ1 . . . _ρˆl_ _. . ._ _ρˆt )[⊤]_ is obtained by the minimization of the total photometric error C(ρ): _ρˆ_ = arg min _C(ρ)_ (4) _ρ_ _B. Single-view Depth_ For single-view depth estimation we use the Deep Convolutional Neural Network presented by Eigen et al., [3]. This network uses three stacked CNN to process the images in three different scales. The input to the network is the RGB keyframe Ik. As we use the network structure and parameters released by the authors without further training, our input image size is 320 240. The output of the network _×_ is the predicted depth, that we will denote as s. The size of the output is 147 109, that we upsample in our pipeline in _×_ order to fuse it with the multi-view depth. The first scale CNN extract high-level features tuned for depth estimation. This CNN produces 64 feature maps of size 19 14 that are the input, along with the RGB image, _×_ of the second scale CNN. This second stacked CNN refines the output of the first one with mid-level features to produce a first coarse depth map of size 74 55. This depth map is _×_ upsampled and feeds a third stacked CNN that does a local refinement of the depth. This final step is necessary, as the convolution and pooling steps of the previous layers filter out the high-frequency details. The first scale was initialized with two different pre-trained networks: the AlexNet [24] and the Oxford VGG [25]. We use the VGG version, the most accurate one as reported by the authors. This network has been trained in indoor scenes with the NYUDepth v2 dataset [26]. As they used the official train/test splits of the dataset, so do we. We decided to use this neural network because it was the best-performing dense single-view method at the moment we started this work and still it is the one that keeps better trade off between quality and efficiency. We refer the reader to the original work [3] for more details on this part of our pipeline. _C. Depth Fusion_ As we mentioned before, the objective is to fuse the output of each previous method while keeping the best properties of each of them: the single-view reliable local structure and the accurate, but semi-dense multi-view depth estimation. Let _, (1)_ where Tko,Rko and tko are respectively the relative transformation, rotation and translation between the keyframe Ik and every overlapping frame Io. K is the camera internal calibration matrix. We define the total photometric error C(ρ) as the summation of every photometric error ϵl between every pixel (or ----- denote s and m to the single-view depth and the multi-view semi-dense depth estimation, respectively. s is predicted as detailed in section III-B and m = _ρ1_ [is the inverse of the] inverse depth estimated in section III-A. The fused depth estimation fij for each pixel (i, j) of a keyframe Ik is computed as a weighted interpolation of depths over the set of pixels in the multi-view depth image _fij =_ � _Ws[m]ij[uv]_ (muv + (sij − _suv)),_ (5) (u,v)∈Ω where Ω is the semi-dense set of pixels estimated by the multi-view algorithm (e.g. in a high-parallax sequence, they usually correspond with the high-gradient pixels). The interpolation weights Ws[m]ij[uv] model the likelihood for each pixel (u, v) Ω belonging to the same local structure as pixel _∈_ (i, j). The interpolation can be interpreted in two ways. First, the depth gradient (sij − _suv) is added to each multi-view_ depth muv, i.e. we create depth map for each muv with the structure of s and then weigh them with pixel based weights. Second, for each depth sij we modify it according to the weighted discrepancy between (muv − _suv)._ The key ingredient of this interpolation are the weights _Ws[m]ij[uv]_ that model a deformation based on the local image structures. Each weight is computed as the product of four different factors. The first factor RGB image with the point Weigth factors of the point Non-normalized influence Fig. 3: Non-normalized influence of the highlighted red point in the image. First column: RGB input image with a red point over the table, this point represent one pixel estimated by the multi-view algorithm. Second column: each one of the weights calculated separately, the third and fourth weights are shown as a product for a more intuitive view. _Third column: Non-normalized influence of the highlighted_ point in the RGB image. Notice how its influence is cut on the edge of the table. Figure best viewed in electronic format. The product of this four factor makes a non-normalized weight for each pixel in Ω Non-normalized influence Weigth factors of the point _W˜_ _m[s][ij]uv_ [=] Weigth factors of the point 4 � _W˜nsmijuv_ (10) _n=1_ RGB image with the point (i−u)[2] +(j−v)[2] )) _σ1_ _,_ (6) _W˜1msijuv_ = e _−[√]_ simply measures proximity based on the distance of the pixels (i, j) and (u, v). The parameter σ1 controls the radius of proximity for each point. The remainder three factors depend on the structure of the single-view prediction s. The second factor _muv_ 1 _W˜2sij_ = _|∇xsuv −∇xsij| + σ2_ 1 _·_ _|∇ysuv −∇ysij| + σ2_ (7) and represents its area of influence. The parameters σ1, σ2 and σ3 shape the area of influence and have to be selected to balance proximity, gradient and planarity and to avoid discontinuities in the result of the fusion. This was done empirically on a small set of three images. The values of the parameters are 15, 0.1 and 1e 3, respectively, and we kept _−_ them fixed for all our experiments. Fig. 3 shows this area for a point on an image and how it is computed. Notice how the influence expands around the point but is kept inside the same local structure (the table). Once all the factors has been computed, since all the pixels (i, j) are influenced by all the pixels in Ω (see Eq. 5), we normalize the weights for each single-view pixel so all the weights over a pixel (i, j) sum 1. _Ws[m]ij[uv]_ = �(p,kW˜)∈s[m]Ωijuv[W][˜]−[ m]sij[pk]min−(g,hmin)∈(Ωg,hW[˜])∈s[m]Ωij[gh]W[˜] _s[m]ij[gh]_ (11) The normalized weights expand the local influence to the whole image (see Fig. 4 and Fig. 5 for a more detailed view). Notice how the influence expands along planes even if the points in Ω do not reach the end of the plane; and is sharply reduced when the local structure changes. Once these influence weights have been calculated and normalized, the fusion depth estimation, f, for each point (i, j) is a combination of all the selected points in Ω, as presented in Eq. 5. _D. Multi-view Low-Error Point Selection_ Up to now we have assumed that all the points in the multi-view semi-dense depth map Ω have low error. This is easily achievable in high-parallax sequences by using robust estimators –robust cost functions or RANSAC. However, it measures the similarity of depth gradients and assigns larger weights to similar ones. ∇xsij and ∇ysij represent the depth gradient in the x and y direction respectively at the pixel (i, j). σ2 limits the influence of a point to avoid extremely high weights for very similar or identical gradients. We set it to 0.1 in the experiments. Finally, the factors _W˜3msijuv_ and _W˜4msijuv_ strengthen the influence between the points lying in the same plane and are defined as _W˜3msijuv_ = e[−|][(][s][ij] [+][∇][x][s][ij] _[·][(][u][−][i][))))][−][s][uv][|]_ + σ3 (8) and _W˜4msijuv_ = e[−|][(][s][ij] [+][∇][y][s][ij] _[·][(][v][−][j][))))][−][s][uv][|]_ + σ3, (9) where σ3 sets a minimum weight to any point in Ω. This is required to avoid vanishing weights when they are combined with _W[˜]1msijuv_ and _W[˜]2msijuv_ [.] ----- Fig. 4: Normalized influence area of the points. Notice how it expands around local structure areas given a set of points in Ω. First column: RGB image with the points of Ω labeled with different colors. Second column: influence areas computed by our method. Notice how this influence expands in areas with the same local structure but can be misled in areas where there is a lack of points or where the estimation from the neural net is not accurate enough. Figure best viewed in color. Fig. 5: Detail of the influence area. Notice how it expands mainly in the areas with same local structure. Figure best viewed in color. is problematic for the degenerate or quasi-degenerate lowparallax geometries that we also target in this paper. In this case, multi-view depths may contain large errors that will propagate to the fused depth map and it is necessary to filter them out. Unexpectedly, selecting high gradient pixels was not robust enough to remove points with large depth errors and we have developed a two step algorithm that takes into account photometric and geometric information in the first step and the single-view depth map in the second one. The first step selects a fixed percentage of the best correspondence candidates –the best 25% in our experiments– based on the product of a photometric and a geometric scores. On one hand, the photometric criterion focuses on the quality of the correspondences using image information. We apply a modified version of the second best ratio.We first extract the two closest matches for a pixel (smallest photometric errors according to Eq. 3). We then compute the score as a function of the ratio between the distance of the two descriptors (a high ratio suggesting a good match) and the gradient of the distance function along the epipolar line (i.e., the error function presenting a distinct V-shape around this match and suggesting spatial accuracy). On the other hand, the geometric score simply backpropagates the image correspondence error to the depth estimation, resulting in low scores for low-parallax correspondences. In a second stage we also use the structure of the single-view reconstruction and apply RANSAC to estimate a spurious-free linear transformation between the multi and single-view points using only the points pre-filtered in the first stage. We apply this linear model along the entire image, consensus with outliers is found if small patches are used. This reduces further the number of spurious depth values from the multi-view algorithm. The result is a small set of low-error points that we use for the interpolation of the previous section. As mentioned before, in our experiments this algorithm behaves better than a geometric-only compatibility test, especially in the low-parallax sequences of the NYUv2 dataset. IV. EXPERIMENTAL RESULTS In this section we evaluate the algorithm and compare its performance against two state-of-the-art methods: multiview direct mapping using TV regularization (implemented following [1], [28]) and the single-view depth estimation using the network of [3]. We have selected two datasets with different properties. The first one is the NYUv2 Depth Dataset [26], a general dataset aimed at image segmentation evaluation and hence likely to contain low-parallax and lowtexture sequences. We analyze results in six sequences from the test set (i.e. the single-view net had not been trained on these sequences) selected just to include different types of rooms. The second one is the TUM RGB-D SLAM Dataset [27], a dataset oriented to visual SLAM and then likely to present a bias benefiting multi-view depth. In this case, we evaluated two sequences selected randomly. We run our algorithm in a 320 240 subsampled version _×_ of the images, as this is the size of the single-view neural network given by the authors. We also run our multi-view depth estimation at this image size, and upsample the fused depth to 640 480 in order to compare it against the ground _×_ truth D channel from the kinect camera. As our aim is to evaluate the accuracy of the depth estimation, we will assume that camera poses are known for the multi-view estimation. In the TUM RGB-D SLAM Dataset [27] we use the ground truth camera poses. In the NYUv2 Depth Dataset sequences we estimate them using the RGB-D Dense Visual Odometry by Gutierrez-Gomez _et al. [29]. These camera poses will remain fixed and used_ to create the multi-view depth maps. As mentioned before, the parameters of the fusion algorithm were experimentally set prior to the evaluation on a small separate set of images. To evaluate the methods, we computed three different metrics, the RMSE, the Mean Absolute Error in meters and the ----- RMSE SCALE INVARIANT MEAN ERROR (m) MEAN ERROR (m) Sequence TV Eigen[3] Ours(auto) TV Eigen[3] Ours(auto) TV Eigen[3] Ours(auto) Ours(man) bathroom 0018 1.458 0.852 **0.793** 0.405 0.150 **0.145** 1.174 0.692 **0.612** 0.263 bedroom 0013 1.004 0.550 **0.482** 0.212 0.139 **0.136** 0.690 0.441 **0.344** 0.163 dining room 0032 2.212 0.710 **0.694** 0.416 0.209 **0.204** 1.797 0.581 **0.554** 0.318 kitchen 0032 3.599 1.621 **1.572** 0.812 0.592 **0.583** 2.920 1.222 **1.183** 0.805 living room 0025 1.073 0.620 **0.597** 0.289 0.236 **0.219** 0.798 0.471 **0.435** 0.289 living room 0030a 1.031 0.818 **0.792** 0.411 0.228 **0.219** 0.849 0.532 **0.440** 0.329 fr1 desk 1.581 0.433 **0.410** 0.255 0.121 **0.103** 1.211 0.317 **0.294** 0.154 fr1 room 1.467 0.323 **0.301** 0.167 0.092 **0.081** 1.163 0.231 **0.207** 0.102 TABLE II: Left table: Error metrics for the NYUv2 and TUM datasets. For each sequence and metric we compare the TV-regularized multi-view depth, the single-view depth [3] and our fused depth. Right table: Mean error for the fused depth with manual multi-view point selection. (The evaluation has been performed in the first 100 frames of each sequence) |Sequence|Col2|RMSE|SCALE INVARIANT|MEAN ERROR (m)| |---|---|---|---|---| ||Sequence|TV Eigen[3] Ours(auto)|TV Eigen[3] Ours(auto)|TV Eigen[3] Ours(auto)| |NYUDepth v2|bathroom 0018 bedroom 0013 dining room 0032 kitchen 0032 living room 0025 living room 0030a|1.458 0.852 0.793 1.004 0.550 0.482 2.212 0.710 0.694 3.599 1.621 1.572 1.073 0.620 0.597 1.031 0.818 0.792|0.405 0.150 0.145 0.212 0.139 0.136 0.416 0.209 0.204 0.812 0.592 0.583 0.289 0.236 0.219 0.411 0.228 0.219|1.174 0.692 0.612 0.690 0.441 0.344 1.797 0.581 0.554 2.920 1.222 1.183 0.798 0.471 0.435 0.849 0.532 0.440| |TUM|fr1 desk fr1 room|1.581 0.433 0.410 1.467 0.323 0.301|0.255 0.121 0.103 0.167 0.092 0.081|1.211 0.317 0.294 1.163 0.231 0.207| RGB input TV Eigen Ours(auto) Ours(man) Ground Truth Fig. 6: The first six rows are depth images for the NYUDepth v2 dataset [26] and the last two rows are for the TUM Dataset [27]. Color ranges are row-normalized to facilitate the comparison between different methods. First column RGB keyframe, second column TV-regularized multi-view depth, third column single-view depth, fourth column our depth fusion with automatic multi-view point selection, fifth column our depth fusion with manual multi-view point selection, and sixth _column ground truth. Figure best viewed in electronic format._ scale invariant error proposed in [18] _n[1]_ �i _[d]i[2]_ _[−]_ _n1[2][ (][�]i_ _[d][i][)][2]_ where d is (log(y) − log(y[∗])), y and y[∗] are the ground truth depth and the estimated depth respectively. The results are summarized in Table II. Our method outperforms the TV regularization in both datasets obtaining an average improvement over 50% with respect to the mean of the error ----- |NYUv2|SCALE INVARIANT|MEAN ERROR (m)| |---|---|---| ||W1 W1 · W2 Q4 i=1 Wi|W1 W1 · W2 Q4 i=1 Wi| ||0.224 0.216 0.208|0.390 0.376 0.353| |TUM|0.098 0.088 0.064|0.145 0.142 0.128| TABLE III: Mean of error metrics for the NYUv2 and TUM datasets. For each sequence and metric we compare the fusion with the only use of the weight W1, the use of _W1 · W2. and all the weights together._ in meters. As expected, the TV regularization performs better in the TUM sequences and achieves lower errors, but in terms of improvement there seems not to be big differences between both datasets. Our fusion of depths also outperforms the single-view depth reconstruction, the improvement being 10% on average. Both methods perform similarly in both datasets, but except in one sequence, our method is always better or as good as the deep single-view reconstruction. Notice that the improvement does not come exclusively from scale correction; the scale invariant error shows that our method improves the structure estimation in both the single and multi-view cases. The right-most colum of Table II shows the depth errors when the set of multi-view points does not contain outliers. We selected them using the ground-truth data from the D channel, and keeping only those points whose depth error was lower than 10cm. The results are for all sequences better than any method attaining improvements around 70% and 38% with respect to TV and [3], respectively. Although expected, this result highlights the impact of multi-view outliers and the need for good point selection. It also provides an upper bound and shows that there is still room for improvement in this latest part of our algorithm. In Table III we show an experiment to better understand the contribution of each weight of our algorithm. For this evaluation we have considered the spurious-free set of multi-view points in order to avoid the influence of noise. It can be seen that using all the weights has an average of 9.8% improvement in mean absolute error with respect to using just W1 and a 6.5% of improvement with respect to using W1 and W2. Finally, we present the results of some randomly picked images for each sequence of each dataset. Fig. 6 shows the obtained depth images for the NYUDepth v2 and the TUM datasets. The improvement with respect to the regularized multi-view approach is clear visually since the depth structure is much more consistent. Improvements with respect to single-view images are more subtle and are best viewed by looking at the corresponding depth error images of Fig. 7. Usually, the improvement comes from a better relative placement of some local structure. For instance, the walls are darker in the error images (see the bathroom 18, bedroom 13 or fr1 desk in Fig. 7). The effect is more evident when the multi-view points were selected based on the ground truth. This better alignment of local structures reduces the error, as can be seen in the per-sequence error boxplots of Fig. 8. RGB input Eigen Ours(auto) Ours(man) Fig. 7: The first six rows are error images (predicted depth - ground truth) for the NYUDepth v2 dataset [26] and the last two rows are for the TUM Dataset [27]. Color ranges are row-normalized to facilitate the comparison between different methods. Darker blue is better. First column RGB keyframe, second column single-view depth, third column our depth fusion with automatic multi-view point selection, _fourth column our depth fusion with manual multi-view point_ selection. In the third column, in yellow, are highlighted the areas where the improvement of our method can be easily appreciated with respect to single-view’s error. Figure best viewed in electronic format. V. CONCLUSIONS In this paper we have presented an algorithm for dense depth estimation by fusing 1) the multi-view depth estimation from a direct mapping method, and 2) the single-view depth that comes from a deep convolutional network trained on RGB-D images. Our approach selects a set of the most accurate points from the multi-view reconstruction and fuses them with the dense single-view estimation. It is worth remarking that the single-view depth errors do not depend on the geometric configuration but on the image content and hence the transformation is not geometrically rigid and varies locally. The estimation of this alignment is our main contribution and the most challenging aspect of this research. ----- 6 5 4 3 2 1 3.5 3 2.5 2 1.5 1 0.5 0 2.5 2 1.5 1 0.5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 0 0 0 0 Ours(man) Ours(auto) Eigen TV Ours(man) Ours(auto) Eigen TV Ours(man) Ours(auto) Eigen TV Ours(man) Ours(auto) Eigen TV bathroom_0018 bedroom_0013 living_room_0025 living_room_0030a Fig. 8: Box-and-Whiskers plots of the pixel error distribution for four of our test scenes. From left to right: Our method with manual point selection, our method with automatic point selection, single-view depth from Eigen et al. [3] and TV-regularized multi-view depth. Our experiments show that our proposal improves over the state of the art (Eigen et al. [3] for single-view depth and direct mapping plus TV regularization for multi-view depth). Contrary to other approaches, the single-view depth we use is entirely data-driven and hence does not rely on any scene assumption. As mentioned, we take the network of [3] as our single-view baseline, because of its availability and its excellent accuracy-cost ratio. However, our fusion algorithm is independent of the specific network and could be used with any of the single-view approaches mentioned in Section II. Future work will, as suggested by the results, try to improve the multi-view points selection and the fusion of both images using, for instance, iterative procedures or segmentation-based fusion. REFERENCES [1] R. A. Newcombe, S. J. Lovegrove, and A. J. Davison, “DTAM: Dense tracking and mapping in real-time,” in Computer Vision (ICCV), 2011 _IEEE International Conference on, pp. 2320–2327, IEEE, 2011._ [2] J. Engel, T. Sch¨ops, and D. Cremers, “LSD-SLAM: Large-scale direct monocular SLAM,” in European Conference on Computer Vision, pp. 834–849, Springer, 2014. [3] D. Eigen and R. Fergus, “Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture,” in Proceedings of the IEEE International Conference on Computer _Vision, pp. 2650–2658, 2015._ [4] G. Graber, T. Pock, and H. Bischof, “Online 3D reconstruction using convex optimization,” in 2011 IEEE International Conference on _Computer Vision Workshops, pp. 708–711, IEEE, 2011._ [5] J. St¨uhmer, S. Gumhold, and D. Cremers, “Real-time dense geometry from a handheld camera,” in Joint Pattern Recognition Symposium, pp. 11–20, Springer, 2010. [6] A. Concha, W. Hussain, L. Montano, and J. Civera, “Manhattan and Piecewise-Planar Constraints for Dense Monocular Mapping,” in _Robotics: Science and Systems, 2014._ [7] P. Pinies, L. M. Paz, and P. Newman, “Dense mono reconstruction: Living with the pain of the plain plane,” in 2015 IEEE International _Conference on Robotics and Automation, pp. 5226–5231, 2015._ [8] P. Pini´es, L. M. Paz, and P. Newman, “Too much TV is bad: Dense reconstruction from sparse laser with non-convex regularisation,” in _2015 IEEE International Conference on Robotics and Automation_ _(ICRA), pp. 135–142, IEEE, 2015._ [9] A. Concha and J. Civera, “Using superpixels in monocular SLAM,” in _Robotics and Automation (ICRA), 2014 IEEE International Conference_ _on, pp. 365–372, IEEE, 2014._ [10] V. Hedau, D. Hoiem, and D. Forsyth, “Recovering the spatial layout of cluttered rooms,” in 2009 IEEE 12th international conference on _computer vision, pp. 1849–1856, IEEE, 2009._ [11] A. Concha, W. Hussain, L. Montano, and J. Civera, “Incorporating scene priors to dense monocular mapping,” Autonomous Robots, vol. 39, no. 3, pp. 279–292, 2015. [12] D. F. Fouhey, A. Gupta, and M. Hebert, “Data-driven 3d primitives for single image understanding,” in Proceedings of the IEEE International _Conference on Computer Vision, pp. 3392–3399, 2013._ [13] R. Mur-Artal, J. Montiel, and J. D. Tardos, “ORB-SLAM: a versatile and accurate monocular SLAM system,” Robotics, IEEE Transactions _on, vol. 31, no. 5, pp. 1147–1163, 2015._ [14] J. Ens and P. Lawrence, “An investigation of methods for determining depth from focus,” IEEE Transactions on pattern analysis and machine _intelligence, vol. 15, no. 2, pp. 97–108, 1993._ [15] P. Sturm and S. Maybank, “A method for interactive 3d reconstruction of piecewise planar objects from single images,” in The 10th British _machine vision conference (BMVC’99), pp. 265–274, 1999._ [16] A. Saxena, M. Sun, and A. Y. Ng, “Make3D: Learning 3D scene structure from a single still image,” IEEE transactions on pattern _analysis and machine intelligence, vol. 31, no. 5, pp. 824–840, 2009._ [17] A. Saxena, J. Schulte, and A. Y. Ng, “Depth estimation using monocular and stereo cues.,” in IJCAI, vol. 7, 2007. [18] D. Eigen, C. Puhrsch, and R. Fergus, “Depth map prediction from a single image using a multi-scale deep network,” in Advances in Neural _Information Processing Systems, pp. 2366–2374, 2014._ [19] F. Liu, C. Shen, and G. Lin, “Deep convolutional neural fields for depth estimation from a single image,” in IEEE Conference on Computer _Vision and Pattern Recognition, pp. 5162–5170, 2015._ [20] J. Li, R. Klein, and A. Yao, “Learning fine-scaled depth maps from single rgb images,” arXiv preprint arXiv:1607.00730, 2016. [21] A. Chakrabarti, J. Shao, and G. Shakhnarovich, “Depth from a single image by harmonizing overcomplete local network predictions,” in _Advances in Neural Information Processing Systems, pp. 2658–2666,_ 2016. [22] Y. Cao, Z. Wu, and C. Shen, “Estimating depth from monocular images as classification using deep fully convolutional residual networks,” _arXiv preprint arXiv:1605.02305, 2016._ [23] C. Godard, O. Mac Aodha, and G. J. Brostow, “Unsupervised monocular depth estimation with left-right consistency,” arXiv preprint _arXiv:1609.03677, 2016._ [24] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in neural _information processing systems, pp. 1097–1105, 2012._ [25] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014. [26] P. K. Nathan Silberman, Derek Hoiem and R. Fergus, “Indoor segmentation and support inference from RGBD Images,” in ECCV, 2012. [27] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers, “A benchmark for the evaluation of RGB-D SLAM systems,” in Intelligent _Robots and Systems (IROS), 2012 IEEE/RSJ International Conference_ _on, pp. 573–580, IEEE, 2012._ [28] A. Handa, R. A. Newcombe, A. Angeli, and A. J. Davison, “Applications of legendre-fenchel transformation to computer vision problems,” _Department of Computing at Imperial College London. DTR11-7,_ vol. 45, 2011. [29] D. Guti´errez-G´omez, W. Mayol-Cuevas, and J. Guerrero, “Inverse Depth for Accurate Photometric and Geometric Error Minimisation in RGB-D Dense Visual Odometry,” in Robotics and Automation (ICRA), _2015 IEEE International Conference on, pp. 83–89, IEEE, 2015._ -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1611.07245, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://arxiv.org/pdf/1611.07245" }
2,016
[ "JournalArticle" ]
true
2016-11-22T00:00:00
[ { "paperId": "4463dc4a32b948f0230f3b782cbfecaf1c9e5b1d", "title": "Unsupervised Monocular Depth Estimation with Left-Right Consistency" }, { "paperId": "846b35459dd57bc3aee85eb209e97fc180c0cbec", "title": "A Two-Streamed Network for Estimating Fine-Scaled Depth Maps from Single RGB Images" }, { "paperId": "e41812aa74d7d2d8d7089d74900adbda51c6b131", "title": "Learning Fine-Scaled Depth Maps from Single RGB Images" }, { "paperId": "d27c6800894441054685ffdf896fab15cebf5fd6", "title": "Estimating Depth From Monocular Images as Classification Using Deep Fully Convolutional Residual Networks" }, { "paperId": "13f85dae967a9c80f408f599086013ef9dbd1ad8", "title": "Depth from a Single Image by Harmonizing Overcomplete Local Network Predictions" }, { "paperId": "3c52cd2b6f7a375b881f97a6f2c2925c79c9747d", "title": "Incorporating scene priors to dense monocular mapping" }, { "paperId": "4bceb95e022bcea85178c29a8be807042502533c", "title": "Too much TV is bad: Dense reconstruction from sparse laser with non-convex regularisation" }, { "paperId": "057b05c3ac3fc34e11994503cbbfdb01ad16d7f7", "title": "Dense mono reconstruction: Living with the pain of the plain plane" }, { "paperId": "40c9d47bb0fdb39df6afb5fbdfda876f71bb4ca5", "title": "Inverse depth for accurate photometric and geometric error minimisation in RGB-D dense visual odometry" }, { "paperId": "6933c70c747e6a8103f68f1a1db80185401d537b", "title": "ORB-SLAM: A Versatile and Accurate Monocular SLAM System" }, { "paperId": "3f25b3ddef8626ace7aa0865a1a9e3dad1f23fb6", "title": "Deep convolutional neural fields for depth estimation from a single image" }, { "paperId": "cb3a2ddcf305e2ec0f6b94af13d1e631ed261bdc", "title": "Predicting Depth, Surface Normals and Semantic Labels with a Common Multi-scale Convolutional Architecture" }, { "paperId": "c13cb6dfd26a1b545d50d05b52c99eb87b1c82b2", "title": "LSD-SLAM: Large-Scale Direct Monocular SLAM" }, { "paperId": "eb42cf88027de515750f230b23b1a057dc782108", "title": "Very Deep Convolutional Networks for Large-Scale Image Recognition" }, { "paperId": "d7c8ebb5ace5163e303a22e83308846fc167dad9", "title": "Manhattan and Piecewise-Planar Constraints for Dense Monocular Mapping" }, { "paperId": "dd2cf76ae78a3262a094ac865aa9f60c55472c5d", "title": "Depth Map Prediction from a Single Image using a Multi-Scale Deep Network" }, { "paperId": "9fd9cf375033e38262353363b593befbd66426c7", "title": "Using superpixels in monocular SLAM" }, { "paperId": "c855d8b75090e4d4aadd6ce936046327774470c9", "title": "Data-Driven 3D Primitives for Single Image Understanding" }, { "paperId": "5eb4c55740165defacf08329beaae5314d7fbfe6", "title": "A benchmark for the evaluation of RGB-D SLAM systems" }, { "paperId": "abd1c342495432171beb7ca8fd9551ef13cbd0ff", "title": "ImageNet classification with deep convolutional neural networks" }, { "paperId": "c1994ba5946456fc70948c549daf62363f13fa2d", "title": "Indoor Segmentation and Support Inference from RGBD Images" }, { "paperId": "7633c7470819061477433fdae15c64c8b49a758b", "title": "DTAM: Dense tracking and mapping in real-time" }, { "paperId": "88b4037bd43133dcc39329dec1866e5071e6d95b", "title": "Online 3D reconstruction using convex optimization" }, { "paperId": "e577d7f4577e904d41d22bc1edc29d8264b23f66", "title": "Real-Time Dense Geometry from a Handheld Camera" }, { "paperId": "451a06626afe8dd70099c7dfec86de7af909a062", "title": "Recovering the spatial layout of cluttered rooms" }, { "paperId": "41bcea1bec0f0b0e9e2cb4894bf6bfda091a4eae", "title": "Make3D: Learning 3D Scene Structure from a Single Still Image" }, { "paperId": "49531103099c8d17ea34eb09433688e84de4f35f", "title": "Depth Estimation Using Monocular and Stereo Cues" }, { "paperId": "62bdf4743c5ad00a32790a18dd593be0813e571e", "title": "A Method for Interactive 3D Reconstruction of Piecewise Planar Objects from Single Images" }, { "paperId": "7c0f08f0d00d3697f40f461902dff2906fa25e57", "title": "An Investigation of Methods for Determining Depth from Focus" }, { "paperId": null, "title": "Applications of Legendre-Fenchel transformation to computer vision problems" } ]
11,339
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00de0b220e561f58bda5643968bf05822b7f896d
[ "Computer Science" ]
0.914832
Experimental Evaluation of Transmitted Signal Distortion Caused by Power Allocation in Inter-Cell Interference Coordination Techniques for LTE/LTE-A and 5G Systems
00de0b220e561f58bda5643968bf05822b7f896d
IEEE Access
[ { "authorId": "1397187109", "name": "Á. Hernández-Solana" }, { "authorId": "1405613007", "name": "Paloma García-Dúcar" }, { "authorId": "2136979", "name": "A. Valdovinos" }, { "authorId": "2164974275", "name": "Juan Ernesto García" }, { "authorId": "134151207", "name": "J. de Mingo" }, { "authorId": "2231623", "name": "P. L. Carro" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "http://ieeexplore.ieee.org/servlet/opac?punumber=6287639" ], "id": "2633f5b2-c15c-49fe-80f5-07523e770c26", "issn": "2169-3536", "name": "IEEE Access", "type": "journal", "url": "http://www.ieee.org/publications_standards/publications/ieee_access.html" }
Error vector magnitude (EVM) and out-of-band emissions are key metrics for evaluating in-band and out-band distortions introduced by all potential non-idealities in the transmitters of wireless systems. As EVM is a measure of the quality of the modulated signal/symbols, LTE/LTE-A and 5G systems specify mandatory EVM requirements in transmission for each modulation scheme. This paper analyzes the influence of the mandatory satisfaction of EVM requirements on the design of radio resource management strategies (RRM) (link adaptation, inter-cell interference coordination), specifically in the downlink (DL). EVM depends on the non-idealities of the transmitter implementations, on the allocated power variations between the subcarriers and on the selected modulations. In the DL of LTE, link adaptation is usually executed by adaptive modulation and coding (AMC) instead of power control, but some flexibility in power allocation remains being used. LTE specifies some limits in the power dynamic ranges depending on the allocated modulation, which ensures the satisfaction of EVM requirements. However, the required recommendations concerning the allowed power dynamic range when inter-cell interference coordination (ICIC) and enhanced ICIC (eICIC) mechanisms (through power coordination) are out of specification, even though the EVM performance should be known to obtain the maximum benefit of these strategies. We perform an experimental characterization of the EVM in the DL under real and widely known ICIC implementation schemes. These studies demonstrate that an accurate analysis of EVM is required. It allows a better adjustment of the design parameters of these strategies, and also allows the redefinition of the main criteria to be considered in the implementation of the scheduler/link adaptation concerning the allocable modulation coding scheme (MCS) in each resource block.
Received April 9, 2022, accepted April 24, 2022, date of publication April 28, 2022, date of current version May 9, 2022. _Digital Object Identifier 10.1109/ACCESS.2022.3170910_ # Experimental Evaluation of Transmitted Signal Distortion Caused by Power Allocation in Inter-Cell Interference Coordination Techniques for LTE/LTE-A and 5G Systems ÁNGELA HERNÁNDEZ-SOLANA, PALOMA GARCÍA-DÚCAR, ANTONIO VALDOVINOS, JUAN ERNESTO GARCÍA, JESÚS DE MINGO, AND PEDRO LUIS CARRO Aragon Institute for Engineering Research (I3A), University of Zaragoza, 50018 Zaragoza, Spain Corresponding author: Ángela Hernández-Solana (anhersol@unizar.es) This work was supported in part by the Spanish Ministry of Science with European Regional Development Funds (ERDF) under the projects RTI2018-099063-B-I00 and RTI2018-095684-B-I00, and in part by the Government of Aragon (Reference Group T31_20R). **ABSTRACT Error vector magnitude (EVM) and out-of-band emissions are key metrics for evaluating** in-band and out-band distortions introduced by all potential non-idealities in the transmitters of wireless systems. As EVM is a measure of the quality of the modulated signal/symbols, LTE/LTE-A and 5G systems specify mandatory EVM requirements in transmission for each modulation scheme. This paper analyzes the influence of the mandatory satisfaction of EVM requirements on the design of radio resource management strategies (RRM) (link adaptation, inter-cell interference coordination), specifically in the downlink (DL). EVM depends on the non-idealities of the transmitter implementations, on the allocated power variations between the subcarriers and on the selected modulations. In the DL of LTE, link adaptation is usually executed by adaptive modulation and coding (AMC) instead of power control, but some flexibility in power allocation remains being used. LTE specifies some limits in the power dynamic ranges depending on the allocated modulation, which ensures the satisfaction of EVM requirements. However, the required recommendations concerning the allowed power dynamic range when inter-cell interference coordination (ICIC) and enhanced ICIC (eICIC) mechanisms (through power coordination) are out of specification, even though the EVM performance should be known to obtain the maximum benefit of these strategies. We perform an experimental characterization of the EVM in the DL under real and widely known ICIC implementation schemes. These studies demonstrate that an accurate analysis of EVM is required. It allows a better adjustment of the design parameters of these strategies, and also allows the redefinition of the main criteria to be considered in the implementation of the scheduler/link adaptation concerning the allocable modulation coding scheme (MCS) in each resource block. **INDEX TERMS EVM, inter-cell interference coordination, LTE, LTE-A, 5G.** **I. INTRODUCTION** The error vector magnitude (EVM) and out-of-band emissions resulting from the modulation process are the habitual figures of merit adopted by the 4G/5G (i.e., long term evolution –LTE– standards) for evaluating in-band and outband distortions introduced in the transmitter communication system and, thus, the signal accuracy of orthogonal The associate editor coordinating the review of this manuscript and approving it for publication was Tariq Masood . frequency division multiple access (OFDMA) transmissions. These distortions limit the signal-to-noise ratio (SNR) in transmission. EVM is the measure of the difference between the ideal modulated symbols and the measured symbols after the equalization (this difference is called the error vector). In order to exploit the full benefit of the modulation, when base stations (named evolved Node B – eNB– in 4G) perform radio resource management (RRM) strategies (i.e., scheduling, link adaptation, and inter-cell interference management), it is important that eNBs take into account not only the ----- target block error rate (BLER), linked to the expected signalto-interference-plus-noise ratio SINR in reception (derived from channel state information –CSI– reported by user equipment –UE–), but also the influence of EVM in the SNR in transmission in order to guarantee that SNR does not degrade too much at the transmitter. In this way, from release 8 to the last specification of 4G/5G standards, specifications have set mandatory and specific EVM requirements for each modulation scheme (QPSK, 16QAM, etc.). Because EVM depends, in addition to several other factors, on the difference in power allocated per subcarrier, satisfaction or EVM requirements must be considered when selecting the modulation coding scheme (MCS) and the transmission power per subcarrier as part of interference management strategies. In fact, this may severely impact the definition of this type of RRM strategies. Specifications set limits linking power and modulation allocation in order to meet mandatory EVM requirements. These limits are defined as the dynamic power range. However, although the required EVM must be fulfilled for all transmit configurations, it is a key aspect that is absent in almost all studies concerning RRM management, which are generally decoupled from radio frequency (RF) transmission analysis [1]–[14]. There are only a very limited number of contributions in which EVM requirements are considered [15]–[18], showing that the performance of the ideal implementations of RRM schemes are severely reduced. However, they do not explicitly analyze EVM. They assume that the dynamic power range defined in the specifications is a mandatory requirement in any scenario to meet the EVM. Nevertheless, we will see that the dynamic power range defined in LTE unnecessarily limits the flexibility of link adaptation in the inter-cell interference (ICI) coordination (ICIC) design, when these ICIC schemes are based on power coordination. The purpose of this study is to analyze the influence of mandatory satisfaction of EVM requirements on RRM design (related to link adaptation and power allocation constraints), specifically when ICIC mechanisms are applied in the downlink (DL). To the best of our knowledge, this aspect has not been previously analyzed in the literature. QoS in LTE/LTE-A and 5G evolutions depends on RRM strategies, including ICIC and resource and power allocation, operating in an interrelated fashion. By applying rules and restrictions on resource assignments in a coordinated manner between cells concerning the allocable time/frequency resources and the power constraints attached to them, LTE reduces ICI and ensures QoS, particularly at the cell edge. These schemes are required in both the DL and uplink (UL), but they present differences in the rate (MCS selection) and/or power adaptation according to the link channel conditions and data user requirements. Power management takes place in both the DL and UL, although the approaches are clearly differentiated. Both conventional and fractional power controls (FPCs) are applied in the UL [19]. The first case is a subcase of the second. Used to limit ICI and to reduce UE power consumption, the aim of UL power control is to fully or partially compensate (when FPC is applied) the path loss to satisfy the SINR requirements of a selected MCS. As a UE should adopt the same MCS and power at all allocated subcarriers, the satisfaction of EVM requirements in transmission (which are the same as in DL [20]) does not translate into restrictions for RRM implementations. EVM depends only on the nonlinearities of the real transmitter chain. Contrary to the uplink UL, in the downlink DL of LTE/LTE-A systems, to better control interference variations at the UEs in inter-cell scenarios, the first approach and overall goal of downlink power allocation is to budget a constant power spectral density (PSD) for all occupied frequency subcarriers or resource elements (REs) over large time periods. Thus, link adaptation is executed by adaptive modulation and coding (AMC) selection instead of pathloss compensation through power adaptation [19]. A priori, this approach does not preclude the use of ICIC schemes, where frequency coordination aims to reduce ICI by defining different frequency allocation patterns for UEs located in different areas of the cell (for instance, the inner zone and the cell edge) and (in some implementations) by defining different power levels (PSD) on each frequency partition (power coordination). Fig. 1.b1 and Fig. 1.b2 in conjunction with Fig. 1.a illustrate examples of the well-known soft frequency reuse (SFR) and fractional frequency reuse (FFR) schemes, which will be described later. In fact, ICIC derived from FFR and SFR schemes continues to be an important issue to facilitate spatial reuse in both DL [10]–[13] and UL [14] of 5G networks. Nevertheless, the link quality is not limited only by noise and interference at the receiver, which are the effects considered in almost all the RRM studies. Because of the imperfections of the real transmission chains and because the base station (i.e., evolved Node B –eNB- in LTE) transmits simultaneously to several UEs with different MCS and power levels according to the selected frequency partitions, there are distortion effects that limit the SNR in transmission and, as a result, the maximum SINR achievable in reception. The analysis of these effects, characterized by the EVM measure, must be considered. As mentioned above, according to specifications, a maximum EVM per each modulation level must be guaranteed at the transmitter output. With this aim, from release 8, specifications have set and maintained some limits on the difference between the power of an RE and the average RE power for an eNB at the maximum output power (defined as the dynamic power range [21]) to achieve specific EVM requirements for each modulation scheme (QPSK, 16QAM, 64 QAM, 256 QAM) [21]. However, there are two drawbacks to overcome. First, almost all the proposals of RRM in the state-of-theart exclude EVM effects and show the benefits of higher power ranges for the modulation order [15], [6], [7]. Results are obtained under idealized conditions that do not match the actual operation of RF transmitters. Meeting dynamic power range constraints limits the flexibility of using modulations in some power ranges, and drastically degrades the performance ----- **FIGURE 1. Inter-cell interference coordination schemes.** **FIGURE 2. Resource allocation and downlink power allocation in LTE/LTE-A.** of ICIC schemes [15]–[18] compared with ideal implementation. However, in a second place, these dynamic power range restrictions should be interpreted with caution because they had been defined and suggested under specific conditions and simplified assumptions, because ICIC effects were not included in the studies conducted for the specification. The EVM depends on many factors related to the implementation of real transmitters. More flexibility in power allocation is possible while meeting EVM requirements, which are considered mandatory. The actual EVM values may be significantly different from those assumed when the standard limits are stated. Contrary to the simplicity of the assumptions used to define the specification, when different power levels are defined in the transmission spectrum mask, different EVM levels can be obtained depending on the location of the RE and not only on the difference of each power level with respect to the average power. These results are useful for improving the resource allocation. Thus, the objectives of this work are: 1) To emphasize that the proposal and evaluation of RRM strategies, specifically ICIC strategies, need to include mandatory EVM requirements. RRM evaluations that are agnostic of EVM requirements do not properly estimate the actual performance of the proposed schemes. In this context, it is important to note that ICIC and enhanced ICIC (eICIC) based on power coordination remain important ways to facilitate spatial reuse in both downlink and, even, uplink, not only in 4G but also in 5G. 2) To characterize the EVM in the DL transmitter of real RF subsystems, depending on the distribution of modulation and power among subcarriers linked to ICIC and eICIC implementations. The aim is to derive some general performance patterns that allow improving the implementation of these strategies. The objective is to obtain information to be used in the redefinition of the restrictions that must be applied to achieve a better use of resources while meeting the QoS. First, we concisely review the LTE resource allocation basis and constraints in terms of power allocation defined in the specification while reviewing the expected impact of EVM. Then, the motivation for using some well-known ----- ICIC and eICIC mechanisms for homogeneous and heterogeneous (HetNet) deployment scenarios schemes is discussed, and the conditioning factors that arise in terms of satisfaction with the EVM requirements are analyzed. Finally, we evaluate the effect of power allocation on the EVM measured over a standard-compliant LTE downlink signal in a real RF subsystem. **II. RELATED WORK** _A. DL POWER ALLOCATION ACCORDING_ _TO SPECIFICATIONS_ As stated above, conventional power control does not apply to DL, which considers a constant power spectral density for all occupied REs over large time periods and link adaptation through MCS selection. However, in accordance with this goal, owing to their particular requirements, cell-specific reference signals (CRSs), which are embedded into the overall system bandwidth at certain REs, are transmitted with constant power through all DL system bandwidth and across all subframes. CRSs are involved in several of the most important procedures at the air interface: cell search and initial acquisition, downlink channel quality, reference signal received power (RSRP), reference signal received quality (RSRQ) measurements, and cell (re)selection and measurements for handover support. Therefore, their power level must be constant and known by the UEs, being broadcast in mandatory system information block 2 (SIB2). DL power management determines the energy (power) per resource element (EPRE). The reference signal (RS) EPRE (RS-EPRE) is easily obtained by dividing the maximum allowed output power (P[(]max[p][)] [) per antenna port (][p][) in the car-] rier frequency by the number of REs in the entire bandwidth (see Fig. 2). That is, being the ‘‘physical resource block’’ (PRB) the smallest resource unit that can be scheduled for a UE (which is composed by a number NSC[RB] [=][ 12 of subcar-] riers in the frequency domain with �f = 15 kHz subcarrier spacing), the nominal EPRE is obtained by (1): _Pmax_ _EPRE = Emax _[(][p][)]_ _nom_ [=] _NRB[DL]_ [·][ N]SC[ RB] _,_ (1) being NRB[DL] [the number of PRBs in the downlink bandwidth] configuration. When the RS-EPRE is defined, this parameter is used as a reference to determine the DL EPRE of other DL physical signal components or channels (synchronization signals, broadcast channel –PBCH–, DL control channel –PDCCH–, DL shared channel–PDSCH–, control format indicator channel –PCFICH– and physical hybrid automatic repeat-request indicator channel –PHICH–), whose EPRE (i.e., PDSCH EPRE) is set relative to this value. Thus, the specification of LTE allows DL power management to allocate different PDSCH EPRE levels. Nevertheless, the ratio of the PDSCH EPRE to cell-specific RS-EPRE among the PDSCH REs (not applicable to PDSCH REs with zero EPRE) should be maintained for a specific condition. This ratio, which depends on the OFDM symbol, is denoted by either ρB if the PDSCH RE is on the same symbol where there is an RS (symbol index 0 and 4 of each slot) or ρA, otherwise (symbols index 1, 2, 3, 5 and 6) (see Fig. 2a). In our analysis, ρA / ρB is set to 1. In addition, the RE power control dynamic range, which is defined as the difference between the power of an RE and the averaged RE power for an eNB at the maximum output power for a specific reference condition (i.e., the threshold of _ρA and ρB), is limited for each modulation scheme used in the_ PDSCH, according to Table 1 defined in the specification [21] (see Fig. 2.b). In fact, in some specific UE configuration conditions, the allowed ratio ρA in OFDM symbols that do not carry RS is limited to eight values (ρA is equal to the PA parameter [22]), ranging from 6 dB to 3 dB { 6, 4.77, − + − − 3, 1.77, 0, 1, 2, 3}. Note that in all cases, the output − − power per carrier should always be less than or equal to the maximum output power of the eNB. This could be considered an additional limitation, but it refers only to signaling. In fact, in release 10 [19], [23], the relative narrowband TX power (RNTP) indicator was exchanged between the eNB through the X2 interface to support dynamic ICIC. The RNTP bit map provides an accurate indication of the power allocation status of each PRB (RNTP (nRB) with nPRB = 0, . . ., NRB[DL] [−] [1), taking one of the following values: {][−∞][,] 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 1, − − − − − − − − − − − 2, 3}. This power status is defined as the ratio between the maximum intended EPRE of the UE-specific PDSCH REs in OFDM symbols that do not contain an RS and the nominal EPRE. Therefore, greater flexibility is considered. **TABLE 1. E-UTRA BS RE power control dynamic range (Table 6.3.1.1-1** in [21]). **TABLE 2. EVM requirements for E-UTRA carrier (Table 6.5.2-1 in [21]).** In any case, as stated above, power thresholds are set because the different modulations for the DL (QPSK, 16QAM, 64QAM, and 256QAM) require different limits of EVM to exploit the full benefit of the modulation, and the power control range affects EVM. According to the specifications, the EVM for each modulation of the PDSCH is ----- better than the values listed in Table 2. The EVM is defined according to (2) as the square root of the ratio of the mean error vector power (difference between the ideal modulated symbols and the measured symbols after equalization) to the mean reference power expressed as a percentage. The EVM measurement shall be performed over all allocated resource blocks and DL subframes within at least 10ms measurement periods. The basic unit of EVM measurement is defined over one subframe (1ms) in the time domain and NBW[RB] [=] _NSC[RB]_ [=][ 12 subcarriers (180KHz) as defined in (][2][) (annex E] in [21])): these requirements make the application of many ICIC and eICIC mechanisms difficult (for instance, 64QAM is not allowed in reduced power partitions), a more precise analysis of EVM for ICIC (out of specification) is required to specify enhanced requirements for joint power allocation and modulation selection. First, as stated above, EVM depends on a number of factors, including thermal noise in various parts of the transmitter chain, precoding, PA linearity, and predistortion characteristics. They are difficult to quantify theoretically, because they depend on vendor implementation. However, apart from the absolute values, general patterns are identified. 1) PA imperfections are the main contributors to EVM, causing a certain loss of signal orthogonality and, thus, a type of in-band interference. This means that even if the total output power does not significantly change by reducing the power of some PRBs while other PRBs are power boosted (as in ICI schemes described next), more degradation is expected to occur when the power is reduced on the selected PRBs. This is because they are more affected by the in-band interference caused by power-boosted PRBs. Therefore, the interference, and thus, this degradation, could not be uniform in all PRBs. This depends on the distance to the power-boosted PRB and the ratio between their respective power levels. The aim of this study is to analyze the EVM degradation depending on the PRB position to improve radio resource allocation. 2) It is a straightforward conclusion that further power reduction beyond the maximum specified power dynamic range can be considered based on the vendor implementation. The EVM could be better than 7.5% for the working point of PDSCH-EPRE RS-EPRE, = and as a result, a power dynamic range could also be defined as the 64QAM meeting EVM requirements. However, in any case, the EVM impact analysis of setting different PDSCH EPRE levels on several PRB partitions on the same OFDM symbol is still necessary. 3) When eICIC mechanisms are applied (reviewed in Section B), normal and low-power (LP) subframes are distributed within a frame (10ms) as shown in Fig. 3. In this case, when a power reduction is applied to all PRBs in an LP subframe, although the channel powers of the cell RS (CRS) and PBCH are maintained to avoid time-variant CRS transmission power fluctuations, the total output power is reduced, and low EVM degradation is expected in these LP subframes compared with subframes when normal operation occurs. It is expected that the PA operates within the saturation limits for normal subframes and applies a power back-off at the PA. In this case, because EVM measurements should be performed within at least 10ms measurements, EVM variations between normal and low-power subframes must be considered in order to ensure EVM requirements in all subframes. � |Z [′](t, f ) − _I_ (t, f )|[2] _f ∈F(t)_ _,_ (2) � � |I (t, f )|[2] _t∈T_ _f ∈F(t)_ _EVM_ = � � � � � � � _t∈T_ where T is the set of symbols with the considered modulation scheme being active within the subframe, F(t) is the set of subcarriers within the NBW[RB] [subcarriers with the considered] modulation scheme being active in symbol t, I (t, f ) is the ideal signal reconstructed by the measurement equipment in accordance with relevant transmission (TX) models, and _Z_ [′](t, f ) is the modified signal under test. The method for measuring EVM is quite involved (annex E in [21]), but a simple approximation assumes that the error vector resembles white noise. In this case, EVM can be converted to SNR using the following formula: SNR 10 = × log10(1/EVM [2]). Considering this, the limits in Table 1 were set to meet the EVM requirements defined in the specifications (Table 2) [21]. These values were obtained through simulations to ensure that the system performance was not significantly degraded. Specifically, the dynamic power range was defined for these minimum performance requirements. A range of (7.5% - 8%) EVM was proposed as a working assumption for 64QAM modulated PRBs when PDSCH-EPRE RS-EPRE. Thus, a better SNR achiev= able under this condition is 22 dB. To define power range limits, they consider that although there are many causes of EVM, power amplifier (PA) nonlinearities, specifically clipping noise, are the major contributors to EVM. To make the PA implementation efficient, the peak-to-average power ratio (PAPR) of the signal was reduced by clipping the highest peaks. Thus, the signals were slightly modified, indicating this as an additional noise source. The power range was estimated by assuming that if the output power did not change significantly, the clipping noise remained nearly constant and with similar levels for all the PRBs [24] [26]. − Under this assumption, an RE power reduction will lead to a reduced SNR in transmission and a higher EVM that was quantified up to 12.5% (SNR 18 dB) for 16QAM if = the power is reduced by 4 dB and up to 17.5% for QPSK (SNR 15 dB) with 7 dB of power reduction. By applying = a margin, they resulted in 3 dB and 6 dB, as defined − − in Table 2. However, these assumptions are not very close to the performance of actual implementations. In addition, because ----- **FIGURE 3. Enhanced Inter-Cell Interference Coordination (eICIC).** _B. DL POWER MANAGEMENT AND INTER-CELL_ _INTERFERENCE COORDINATION IMPLEMENTATION_ A deeper EVM analysis for ICIC and eICIC may seem unnecessary given the evolution of 4G / 5G systems. However, nothing could be further from the truth. There are diverse mechanisms to combat inter-cell interference in LTE, including ICI cancellation (IC), ICIC, eICIC mechanisms for the HetNet, coordinated multipoint (CoMP), and coordinated beamforming. A good survey of ICIC techniques can be found in [1]–[4], as well as related radio resource strategies [5] in both standard and heterogeneous LTE/LTE-A networks, along with different performance assessments. To illustrate the impact of EVM on ICIC, it is sufficient to consider two main and well-known categories of ICIC: fractional frequency reuse (FFR) and soft frequency reuse (SFR). FFR and SFR are static ICIC techniques that require interventions from mobile network operators to adjust the PRB and power distribution between cell zones according to UE distribution and quality of service demands. Although simple, these approaches are preferred by many operators, including public safety operators, because of their compatibility with the standard, their inherent ease of implementation, and the fact that they require little or no inter-cell communication. More dynamic implementations are possible by applying dynamic coordination between the eNB. For instance, they are possible using narrow-band transmit power (RNTP) indicator exchange through the X2 interface [2], [3]. In any case, from the early stages of the 4G definition to the present 5G context, a large number of works have studied ICI management based on these low-complexity schemes, resulting in the proposal of many derived variants for 4G and 5G [1]–[5], [8], [10]–[14]. First deployments of 5G networks have been made using OFDMA and proposals for new multiple access techniques are also based on OFDMA, thus, ICIC techniques derived from FFR and SFR remain an important issue to facilitate spatial reuse in both DL [10]–[13] and UL [14]. The basic principle behind these schemes is the division of the available PRBs in the carrier spectrum into two partitions/sub-bands: one intended for mobile users (UE) found in the inner part of the cell and the other that is reserved for users found in the outer or cell-edge area (celledge users). Subsequently, several degrees of reuse factors for the inner and outer partitions are applied in a multicell system. In FFR-based approaches, as illustrated in Fig. 1.a with Fig. 1.b2, cell-inner users use the same sub-band in every cell, but the outer sub-band is usually divided into several sub-bands (usually three), and each cell is given a sub-band that is orthogonal to the outer sub-bands in neighboring cells. In this case, the inner region does not share any spectrum with the cell-edge or adjacent cell-edge regions. This strict policy reduces interference on both inner-and cell-edge users but may underutilize the available frequency resources. On the other hand, in SFR-based approaches, the entire bandwidth can be utilized in every cell, and the effective reuse can be adjusted by power coordination between the PRBs used in the inner and outer sub-bands, as illustrated in Fig. 1.a with Fig. 1.b1. Cell-inner users can have access to the cell-edge sub-bands selected by the neighboring cells, but with a lower power level to reduce interference to the neighboring cells. Thus, the SFR achieves higher spectrum efficiency, but there is a higher ICI. Despite these differences, in all ICIC schemes, there is a common set of basic parameters that must be specified, whose adjustment and optimization have a severe impact on the performance of these schemes [1], [2], [6]–[9]: 1) The set and size of the frequency partitions defined in the PDSCH (e.g., we can identify them as NRB[inner] and NRB[edge][).] 2) The power level per RE for each frequency partition (e.g., EPRERB[inner] and EPRERB[outer] ) and, thus, the power range between them (e.g.,� = EPRERB[outer] − _EPRERB[inner][)]. Note that in partial FFR, different power_ levels are not strictly needed, although at the edge, subband resources can be power-boosted to reach the cell coverage limit. 3) The spatial region where the partitions are used (e.g., cell center or cell edge), and thus, the number of user groups or classes. ----- 4) Threshold criterion for classifying users into groups. Note that FFR and SFR are static/semi-static ICIC techniques, however, they always require interventions in the network to adjust the PRB and power distribution between partitions according to the UE distribution and QoS demands. Concerning the power level settings, the power in the cell-edge sub-band(s) should be boosted, while the power in the cell-inner sub-band should accordingly be de-boosted to maintain a constant nominal power and maximum output power. Equation (3) sets the expression to compute the power level settings depending on the � values that will be used in the evaluation. _P[(]max[p][)]_ [[][mW] []][ =][ N]RB[ inner] - 10[(12][·][EPRE]RB[inner] [dBm]�10) +NRB[outer] - 10[(12][·][EPRE]RB[outer] [dBm]�10+�[dB]/10) (3) As anticipated, the problem is that if power control dynamic range constraints are required to be satisfied to meet EVM requirements, it is difficult to use or to obtain the maximum benefit of the power control coordination schemes in practical systems. According to Table 1, for instance, if �> 0 dB, 64QAM cannot be applied to the SFR or FFR when the outer-cell power boost is also considered, which occurs in most cases. However, almost all state-of-the-art studies until the present have evaluated the performance of their SFR-based or FFR-based proposals without considering these constraints [1] [13], being the ratio of the outer and − inner power densities (�) one of the most important parameters in the analysis [1], [2], [6] [9], [11], [12]. In these − studies, 64QAM are frequently used, particularly in the inner sub-bands. As SFR-based approaches improve spectral utilization, this technique could be particularly effective in some cases; for instance, public safety operators are more affected by limitations on available bands and spectrum bandwidth (3–5 MHz). They often provide support to scenarios with few UE but high resource occupancy. However, the SFR causes more interference to all the center and edge users when compared with the FFR and FFR power-boosted cases. By properly adjusting � (for instance, from 0 dB to 10 dB), the operator can control the tradeoff between improving the average cell throughput (low � values) and the cell edge throughput (high _�_ values). Recently, in [10], the authors propose a flexible soft frequency reuse (F-SFR) that enables a self-organization of a common SFR in the networks with an unpredictable and dynamic topology of flying base stations. Authors propose a graph theory-based algorithm for bandwidth allocation and for a transmission power setting in the context of SFR. They use a deep neural network (DNN) to significantly reduce the computation complexity. However, same as in [8], where a multi-layer SFR is proposed, or [6], [7], [9], where the ratio between the power density in the outer cell region and in the inner cell region is evaluated (from 0 dB to 10 dB or 12 dB), the performance is obtained without considering the effects linked to a real transmitter implementation. Under this assumption, a priori, the dynamic power range is not limited, and the selection of the optimal values, in order to lead to high performance, only depends on the interference caused by co-channel neighbor cells, in addition to RRM (link planning and adaptation) implementations. The same assumptions are applied in the studies conducted in [11], where FFR and SFR with K edge sub-bands are considered and the power ratio ranges from 0 to 20 dB. The same occurs in [12], where authors propose a generalized model of FFR for ultra-dense networks. Knowing that, according to [19], the transmission power in DL should not dynamically change, an FFR scheme extended to N (from 2 to 4) power/frequency subbands/groups is proposed. Power levels of each frequency group are appropriately selected to optimize the system operation while the total power consumption remains unchanged. Power ratio between groups varies from 3 to 13 dB, but the optimization does not consider the mandatory requirements of the specifications in order to limit the dynamic power range (linked to power allocation and link adaptation) to meet the error vector magnitude (EVM) requirements at the transmitters. In fact, this is a key aspect that is absent in all the referred studies and in almost all the studies carried out concerning ICI management in DL. All of these studies, and many others available in the literature, are of great interest (i.e., some interesting reviews are available in [1] [5]). How− ever, a practical limitation of all the proposals is that they do not satisfy the EVM requirements, which has a significant impact on the optimal power allocation, dynamic scheduling, and link adaptation. Considering the interest in ICIC based on power coordination, it is clear that the power control dynamic range must be re-evaluated for ICIC, considering the effects of real transmitter implementations for several values of � and MCS distributions among the bandwidth partitions. To our knowledge, there is only a very limited number of contributions in which satisfaction of EVM requirements is considered (not explicitly but in some way) [15] [18]. − In these contributions, ICIC schemes are proposed and studied in HetNet scenarios, but the problem is similar. Power control dynamic range defined in Table 1 is imposed in eICIC for HetNet deployments, as illustrated in Fig. 3, where low-power nodes (LPN) are deployed under macrocell coverage. In this case, cell range expansion (CRE) is used to extend the coverage of the LPN, whereas the low power almost blank subframe (LP-ABS) technique is used to decrease the interference caused by the macrocell to the LPN in the extension area (that is, to the cell-edge user in the LPN). LP-ABS is a time-domain ICIC. Contrary to the traditional ABS mechanism, where the macrocell stops its PDSCH transmissions in predefined black subframes intended only for LPN transmissions, in LP-ABS, the macro eNB maintains its data transmissions on the ABS subframes, but the PDSCH EPRE is reduced, with α (where 0 ≤ _α ≤_ 1) the reducing factor. Fig. 3 illustrates the LP-ABS concept with _α (e.g. −3 dB). Similar to �_ in ICIC schemes, α is the key design parameter. The studies in [15] [18] are conducted − ----- from the point of view of RRM design. Thus, the effects of RF implementation are not explicitly considered, but the authors emphasize that the maximum value of allowed power boosting relative to the nominal value must be properly designed to limit the dynamic range and the EVM requirements [21]. In [15], the authors perform a good analysis of the impact of 6 dB, 9 dB, and 12 dB power reductions, − − − concluding that although small values are sufficient to reach the maximum performance in macrocells applying eICIC, larger reductions (i.e., 12 dB) make it possible the application of larger cell range expansion offsets and the consequent improvement of macrocell performance owing to higher picocell offloading ratios. Under ideal conditions (without considering EVM requirements), the same study shows that although it is normal that the modulation order decreases when the transmission power is reduced, a large percentage of 16QAM and 64QAM transmissions are often used in low power subframes. This is because many of these transmissions are directed to the inner UEs, are little affected by interference, and have good channel conditions. However, in [15], when LTE specification constraints (Table 1) are considered, the authors remark that LP-ABS subframes could only be deboosted to -6 dB from RS-EPRE without significant specification changes, and only if the modulation is constrained to QPSK during these de-boosted subframes. If modulations are limited to QPSK in low-power subframes, as specified, perceptible degradation in the macrocell performance occurs. Increasing the dynamic range (i.e. up to 9 dB) for all the modulations yields in a degradation in the EVM. The same consideration is applied in [16], [17], limiting the power de-boost to 6 dB. They conclude that the support of high power reductions will only be possible at the expense of better EVM requirements for 0 dB to meet the EVM requirements for large power reductions. In a similar way, in [18], where a coordinated multi-point transmission (CoMP) scheduling is applied in combination with ICIC techniques with different power reduction levels, the authors compare the achieved user data rate and system throughput performance without (ideal case) and with LTE constraints (in this case, a dynamic power range is applied and a lower modulation order should be used to conserve modulation accuracy). They show that when the LTE constraints are employed (modulation order is constrained based on the used power offset level), the obtained user data and system throughput performance under the ideal case are drastically degraded regardless of the ICIC technique used. This shows that the LTE constraints should be explicitly considered in any practical RRM proposal. The limitation of all these studies is that they are based on the theoretical power control dynamic range defined in the specifications without an explicit EVM evaluation. However, EVM depends on the vendor implementation and ICIC was out of the studies conducted by the specifications in order to set the dynamic power range. The only actual limitation is that the EVM for the different modulation schemes in the PDSCH should be better than the limits defined in Table 2. Some power back-offs can be applied to the PA in LP-ABS subframes compared with normal subframes (which impacts the CRS TX power) and need to be evaluated [27]. Thus, the requirements listed in Table 1 must be applied with caution in the LP-ABS. Further power reductions compared with those considered in the dynamic power range can be applied. However, as defined above, the EVM measurements were performed for each PRB within at least 10 measurement periods. This implies that the EVMs from the normal and LP-ABS are averaged. Thus, to ensure a good system performance, the differences between EVMs in PRBs that are not affected by power reductions must be explicitly considered. Concerning the SFR and FFR-based schemes, the most relevant issue is to evaluate the distribution of the EVM along the PRBs in the entire carrier bandwidth according to �. Owing to the loss of orthogonality in the transmitted signal caused by many imperfections in the transmitter chain, EVM is expected to vary on the PRBs of the same sub-band depending on their position relative to the boundary between the inner and outer sub-bands. This information can be used in resource allocation, allowing a more precise MCS selection according to the expected EVM on the PRB. In summary, RRM and RF transmission studies have generally been decoupled in the literature. However, power coordination (linked to ICIC and eICIC), and link adaptation cannot be agnostic to the RF implementation. Being EVM an essential indicator to quantify the transmission performance of a wireless communication, the aim of this work is to quantify the effects in terms of EVM degradation in a real RF subsystem depending on the modulation when power allocation schemes linked to ICIC and eICIC are considered. The kernel of this contribution is that there are no similar studies in the literature. The goal is to avoid unnecessary performance limitations when applying ICIC and eICIC in LTE/LTE-A and 5G cells by restricting the variation range of � and α and allocable MCSs. Absolute values depend on vendor implementation, but some generalizable results can be obtained from a detailed study of power coordination linked to ICIC variants derived from the SFR and/or FFR schemes. **III. EXPERIMENTAL RESULTS** In this work, we have carried out an experimental characterization of EVM degradation in a real RF subsystem for several MCS allocations when different power levels are applied as part of the ICIC and eICIC schemes proposed in 4G/5G networks. To evaluate the effect of power allocation on the EVM measured over the transmitted signal, we have generated a standard-compliant LTE downlink signal (OFDM modulation) with QPSK, 16QAM, and 64QAM modulated subcarriers and a bandwidth of BW 5 MHz. Thus, a total = of 25 PRBs are available. The test signal, generated with MATLAB, which is used in the experiments, follows the LTE frame structure, consisting of different physical signals and channels, including PDSCH, PDCCH, RS, and synchronization data. However, ICIC only applies to the PDSCH; thus, the ----- power and modulation variation in each PRB is only carried out in the PDSCH. The power level and MCS can be independently selected for each PRB with the EVM obtained for each PRB. We evaluated different distributions for the inner and outer sub-bands according to the patterns defined for the SFR scheme in sectors A, B, and C (see Fig. 1.a with Fig. 1.b1). The conclusions derived from the results obtained for all the patterns are similar; therefore, without loss of generality, we will include those obtained for pattern C. That is, the outer sub-band is allocated to the first PRB. The two most relevant parameters that affect the EVM are the power ratio between the outer and inner power densities (�), defined in Fig. 1, and the distance from the PRB, where the EVM is evaluated as the jump point between the inner and outer sub-bands. Taking this into account and without loss of generality, the results shown here correspond to a scenario in which the sizes of the outer and inner sub-bands are adjusted to be almost equal. That is, NRB[inner] = 12 and NRB[outer] = 13. Power levels are set for different � rates according to (2). For example, Fig. 4 shows an LTE frame with 25 PRBs (5 MHz bandwidth), where the first 13 PRBs have a power level 9 dB higher than that of the last 12 PRBs, and a 64-QAM modulation scheme is used for all PRBs. **FIGURE 4. Test signal: LTE frame with 25 PRBs where the first 13 PRBs** have a power level 9dB higher than the last 12 PRBS (64QAM modulation). _A. EXPERIMENTAL SETUP_ The complete experimental test bench is shown in Fig. 5 using an equivalent block diagram. The experimental setup used in this study is shown in Fig. 6. The digital development platform used for the implementation of digital signal processes and the digital I/Q modulator and demodulator consists of an FPGA Zynq-7000 AP SoC connected to a PC that controls a high-speed analog module with an integrated RF agile transceiver, the Analog Devices AD9361 software defined radio (SDR). It comprises an RF 2 2 transceiver with integrated 12-bit digital analog con× verters (DACs) and analog to digital converter (ADCs), and **FIGURE 5. Block diagram scheme stands for the experimental setup.** **FIGURE 6. Laboratory experimental test setup.** has a tunable channel bandwidth (from 200 kHz to 56 MHz) and receiver (RX) gain control. It is used as a generator and receptor for the LTE signal, as described above. The RF carrier frequency is set at 1.815 GHz within band 3 of the LTE standard [21], which is called DCS. Because the signal power at the output of the board is low, it is amplified using a low-noise amplifier (LNA) (Minicircuits ZX60-P33ULN ). The signal is then amplified with a + PA (Minicircuits ZHL-4240), which has a 1-dB compression point of 26 dBm and an approximate gain of 41.7 dB at the test frequency. As previously stated, the most important cause of the increased EVM level in the transmitted signal is the nonlinear distortion caused mainly by the RF power amplifier (PA), which depends on the operating point of the RF PA. For this reason, in this work, several tests are conducted by varying the operating point of the RF power amplifier, and consequently its RF output power, to evaluate the impact of the nonlinearities of the PA on the EVM level. Fig. 7 shows the amplitude-to-amplitude modulation (AM/AM) characteristics of the RF PA used in the experimental setup at a linear (red dots) and nonlinear (blue dots) operating points, corresponding to an averaged RF output power of 18.4 dBm and 22 dBm, respectively. The output signal is shown on an oscilloscope (Agilent Infiniium DSO90804A), which measures the signal power. A splitter (Minicircuits ZAPD-2-21-3W-S ) has been added + to the setup to measure the signal power before amplification. A second splitter (Minicircuits ZN2PD2-50-S ) is used to + capture the amplified output signal and send it to the feedback ----- The predistorted output signal, u(n), is obtained from the baseband input x(n) using (5): _M_ � _apmx (n −_ _m) |x (n −_ _m)|[p][−][1],_ (5) _m=0_ _u (n) =_ _N_ � _p=1_ **FIGURE 7. AM/AM characteristics of the PA measured in the** experimental test setup in a linear and nonlinear scenario. loop to be demodulated on the board. The demodulation process is carried out on a digital platform and analyzed on a PC with MATLAB. An attenuator of 30 dB is used at the output of the PA to avoid damaging the oscilloscope. Starting from this testbed, it is known, as we refer above, that the nonlinear distortion caused mainly by the RF power amplifier is the most important cause of increased EVM level in the transmitted signal. Therefore, depending on the operating point of the RF PA, some type of linearization technique may be necessary in a real implementation to reduce the nonlinear distortion produced by the RF power amplifier and thus decrease the EVM level. Therefore, we performed an analysis using a digital predistorter (DPD) included in the system when the PA works in a nonlinear region. DPD processing is performed in the FPGA Zynq-7000 AP SoC, as explained above. In this study, a classical polynomial model based on a truncated Volterra series is chosen for the amplifier model and is defined as (4): _M_ � _bkmx (n −_ _m) |x (n −_ _m)|[k][−][1],_ (4) _m=0_ where M is the memory depth, N is the nonlinear order, m is the memory tap delay, and apm are the predistorter model coefficients. They are calculated in the first stage of the feedback path (post-distorter), whose input is v(n) and is defined as (6): _v (n) = y (n) /Gnorm, where Gnorm = GlinRF = βGRF_ _, (6)_ where GlinRF is the linearized RF complex gain, β is the gain factor, and GRF is the complex gain without linearization defined as GRF = max |y(n)| / max |x(n)|. This factor β is used to compensate for the gain reduction owing to the linearization process. DPD performance can be improved by carefully adjusting this factor as long as the DPD model remains stable [29]. A more detailed description of this well-known method and how to obtain the input signal matrix expression as well as the coefficient vector can be found in [30]. This model can fit the nonlinearity and memory effects of a power amplifier. In this study, the DPD parameters, nonlinearity order, and memory depth are fixed (N 7 and M 0) for all the = = downlink RF input signal powers. This corresponds to a basic model without memory, but the aim of this work is not to optimize the DPD but to evaluate the improvement of the EVM with the use of a DPD. _B. RESULTS_ A set of experiments has been conducted to evaluate the real effects in terms of EVM degradation in a real RF subsystem considering the MCS allocation, the operating point of the RF PA output power, and the power ratio between the outer and inner power densities (�). The final objective is to obtain information to set real restrictions concerning EVM requirements that affect ICIC and eICIC implementations in order to improve resource allocation strategies, including scheduling and link adaptation through MCS selection. As mentioned above, the results presented here correspond to a pattern where the first 13 PRBs correspond to the outer sub-band of an SFR scheme and the last 12 PRBs correspond to the inner sub-band. Similar analyses and equivalent conclusions have been obtained for other patterns of the inner and outer sub-bands and sizes of the sub-bands. First, Fig. 9 shows the EVM measured in each PRB in various situations depending on the operating point of the RF PA, each of which corresponds to the respective average output power of the PA, with no difference in the power level (� = 0 dB) between the inner and outer sub-bands and considering 64 QAM. This allows us to observe the influence of the operating point of the PA on the measured EVM. As expected, the higher the RF output power, the more _y (n) =_ _N_ � _k=1_ where N is the nonlinear order, M is the memory depth, _x(n) and y(n) are the baseband input and output signals,_ respectively, and bkm are the model coefficients. This model allows us to obtain the DPD characteristics using an indirect learning structure, as explained in [28] and shown in Fig. 8. **FIGURE 8. Predistorter Scheme using an indirect learning structure.** ----- **FIGURE 9. EVM measured in each PRB varying the operating point of the** RF power amplifier. (Modulation 64QAM and no difference in power level between PRBs � **= 0 dB).** **FIGURE 10. EVM measured at an intermediate operating point of the RF** PA, corresponding to Pout = 20.3 dBm, varying the power level � between the first 13 and the last 12 PRBs from 0 dB to 9 dB. (Modulation 64QAM). nonlinearities in the transmitter, and the higher the EVM value along the whole carrier band. In addition, Fig. 9 evidences how all the possible nonlinearities in the transmitter lead to the loss of orthogonality between the subcarriers, which generates inter-subcarrier interference (in-band interference), affecting PRBs differently across the band. Because of the decreasing spectral power of individual subcarriers in the side lobes, PRBs located at the edges of the carrier spectrum are affected by a fewer number of interfering subcarriers able to add a significant interference power. Fig. 9 shows this effect: PRBs in the middle of the system band (i.e., PRB#8 to PRB#16) are more affected by the in-band interference and present higher EVM values. Then, EVM decreases slightly at the ends of the band (i.e., PRB#0, PRB#1 or PRB#23, PRB#24). This effect is more significant as the RF PA works in a nonlinear operating point. It should be noted that the EVM requirements of the standard (see Table 2) are not met for higher output power levels. In these cases, it is necessary to include a DPD in the transmitter to meet the specifications for 64QAM (8%). **FIGURE 11. EVM measured at a nonlinear operating point of the RF PA,** corresponding to Pout = 22 dBm, varying the power level � between the first 13 and the last 12 PRBs from 0 dB to 9 dB. (Modulation 64QAM). It is expected that the effect of in-band interference will be more evident when power coordination, linked to SFR, is applied. That is, a power ratio � is applied between the outer sub-band (used by the users located in the cell edge) and the inner sub-band (used by the users located in the cellcenter). After seeing in Fig. 9 the performance for different points of operation, in Fig. 10 and Fig. 11 we analyze the impact of � values in two different points of operation, always considering the more demanding modulation (64QAM) from the EVM point of view. Fig. 10 shows the EVM measured at an intermediate operating point of the RF PA corresponding to Pout 20.3 dBm, and Fig. 11 shows the = EVM measured at a nonlinear operating point corresponding to Pout 22 dBm. The difference between the power levels = of the outer (first 13 PRBs) and inner (last 12 PRBs) subbands (�) varied from 0 dB to 9 dB. As expected, the PRBs of the inner sub-band (powered down) will lead to a reduced SNR in transmission and a higher EVM than the PRBs of the outer sub-band. This is why the RE power control dynamic ranges (dB) suggested in the specification depend on the MCS. In addition, we observe that the EVM strongly depends on the distance from the PRB (where the EVM is evaluated) to the jump point between the inner and outer sub-bands. It should be noted that the PRB in the transition zone between the two sub-bands is considerably affected. However, EVM degradation diminished when we moved away from the transition zone. The effect is more noticeable with a larger value of �, and for the nonlinear operation point of the RF PA. This is because the in-band interference will affect to a greater extent the subcarriers that are transmitted with less power and are close to others that are transmitted with greater power, because the side lobes of the latter will have a higher relative power with respect to the main lobe of the subcarriers transmitted with less power. This occurs in the transition zone between the inner and outer sub-bands. In this area, the PRBs of the inner sub-band (i.e., PRB#13) suffer more in-band interference level coming from the nearest PRBs of the outer sub-band ----- (which are power boosted with respect to those of the inner sub-band) than coming from PRBs of the own inner sub-band. This results in an increase of EVM in PRBs of the transition zone (i.e., PRB#13), which decreases in the PRBs as long as they are farther to it. For this reason, as PRB move away from the transition zone in the inner sub-band, the high power subcarriers are further away and affect less, so the PRBs at the band edge (i.e., PRB#23 and PRB#24) will present lower EVM values. On the contrary, in the outer sub-band, PRBs are affected by the subcarriers of the inner sub-band, which have less power, and by the subcarriers of their own sub-band with similar power. This results in a lower EVM that derives in the corresponding jump between inner and outer sub-bands. Compared with the inner sub-band, EVM appears to remain almost unchanged. However, we see that in the outer subband, the EVM increases as we move away from the edge of the carrier band (i.e., PRB#0) to the center because PRBs are affected by more subcarriers adding significant interference on each side. This increase stabilizes when we approach the transition zone (i.e., PRB#12) because subcarriers of the inner sub-band become part of the group of most significant interfering ones and they have less power. A good characterization of the EVM performance in the inner sub-band will allow us to make suitable decisions at the scheduler concerning the allowed allocable MCS in each PRB. For instance, in Fig. 10, the EVM requirements (8%) are satisfied for � = 3 dB and � = 6 dB in all PRBs. However, when � = 9 dB is budgeted, 64QAM selection is also allowed for PRB#18to PRB#24. In fact, a general indication from the transmitter SNR point of view is to allocate the highest MCS as far as possible from the jump point. As anticipated in Section II, when ICIC strategies are applied, detailed and individualized analyses are required, which are not considered in the standard. Concerning EVM degradation in the low-power sub-band (in this case, the inner-sub-band), it is clear that it depends on the � factor. However, the specific relationship between EVM and � must be analyzed by considering the transmitter implementation, particularly the actual PA, its operating point, and the use of any linearization technique. The type of analysis that makes it possible to obtain the dynamic power ranges defined in Table 2 cannot be ignored. However, the specific values should not be misunderstood: �> 0 does not prevent the use of 64QAM when ICIC and eICIC are applied. Comparing Fig. 10 and Fig. 11, we can see that the greater the nonlinearity in the transmitter, the greater the difference between the EVM values for PRB#13 and PRB#24. In Fig. 12, the effect of nonlinearity can be clearly observed. In Fig.12, EVM is measured by varying the operating point of the RF PA when the difference in power level � between both sub-bands is set to 6 dB. As shown in Fig. 10 and Fig. 11, there is an increase in the EVM level in the transition zone between the two sub-bands, which decreases as we move away from the transition zone. In this case, it is observed that this effect is more significant when more **FIGURE 12. EVM measured varying the operating point of the RF power** amplifier with a difference in power level � **= 6dB between the first** 13 and the last 12 PRBs. (Modulation 64QAM). nonlinearities exist in the transmitter, as already observed when we compare Fig. 10 and Fig. 11. In Fig. 11, because the baseline EVM is approximately 8% for � = 0 dB, the standard requirements are not satisfied when �> 0. In this case, a digital predistorter (DPD) must be included in the transmitter to reduce the measured EVM. When the transmitter works in a more linear zone (i.e., Pout 16.4 dBm), the = non-idealities persist but are less significant, which allows a better preservation of orthogonality. Because of the outer sub-band is power de-boosted, a lower SNR is achieved in the inner sub-band compared to the outer sub-band, resulting in a higher EVM. However, the decreasing effect that occurs when we move away from the transition zone to the edge of the carrier band is almost negligible. The results of Figs. 9–12 also allow us to infer some relevant conclusions regarding the management of eICIC strategies to combat the interference in HeNet, regardless of the ICIC scheme used in the macrocell to combat inter-cell interference from other macrocells. As mentioned in Section II, it is expected that the PA operates within saturation limits for normal subframes and applies a power back-off at the PA. However, the total output power is also reduced, and a low EVM degradation is expected in these LP subframes compared with subframes where normal operation occurs. If the EVM measurements are performed over a period of at least 10ms, the EVM values are averaged. This means that the EVM requirements can be satisfied in the LP subframes, whereas in normal subframes, it cannot be guaranteed. Thus, the measurements must consider both types of subframes separately. The EVM variations between the normal and lowpower subframes should be considered to satisfy the EVM requirements in all subframes. Taking the working point of the PA that corresponds to a nonlinear zone (Pout 22 dBm), we want to evaluate the = influence of changing the modulation scheme between the inner and outer sub-bands. In Fig. 12, the tests have been performed with 64QAM modulation in all PRBs, whereas ----- **FIGURE 13. EVM measured with different modulations between inner** and outer sub-bands (difference in power level � **= 6dB, Pout = 22 dBm).** **FIGURE 14. EVM measured in a linear (Pout = 18.4 dBm) and nonlinear** operating point (Pout = 22 dBm) of the RF PA with and without DPD and with different modulations between inner and outer sub-bands (difference in power level � **= 6dB).** Fig. 13 shows the results with different modulation schemes. In all cases, an increase in the EVM level appears in the transition zone due to the jump in the power level (� = 6 dB), but it is more significant as long as the order of the modulation used in the inner sub-band decreases. For instance, in this specific implementation, when 64QAM is considered in the outer sub-band, in the inner sub-band the EVM increases from 12.9% when 64QAM is used to 15.2 and 16.2% if 16QAM and QPSK are selected, respectively. For a given SNR, the EVM is lower as the modulation order increases. In addition, as SNR increases, the slope of the EVM improvement is larger for the lower modulation orders. In Fig. 13 we see how EVM decreases faster for QPSK as we move away from the transition zone. Concerning the impact of the MCS used on the outer subband, the results are not conclusive; however, in general, the EVM in the inner sub-band decreases as long as a lower modulation is used in the outer sub-band. Related to the EVM in the outer sub-band, slightly lower EVM values are obtained as the EVM grows in the inner band. Finally, as shown in Fig. 11, in a nonlinear operating point situation of the RF PA, the EVM requirements are **FIGURE 15. EVM measured in a linear (Pout = 18.4 dBm) and nonlinear** operating point (Pout = 22 dBm) of the RF PA with and without DPD and with different modulations between inner and outer sub-bands (difference in power level � **= 3dB).** not satisfied; therefore, a DPD must be included in the transmitter to reduce the measured EVM. To evaluate the effects of the inclusion of a digital predistorter (DPD) at the transmitter, Fig. 14 and 15 show the EVM measured in each PRB at two different operating points of the RF PA: one linear (Pout 18.4 dBm) and the other nonlinear = (Pout 22 dBm). In both cases, the results have been = obtained in two situations: when a QPSK modulation scheme was used in the first 13 PRBs (outer sub-band) and 64QAM in the last 12 PRBs (inner sub-band), and with the same 64QAM modulation scheme in all PRBs. To observe the influence of � on the measured EVM, Fig. 14 shows that the first 13 PRBs have a power level � = 6 dB higher than the last 12 PRBs, while in Fig. 15, � = 3 dB. As in the previous figures, Fig. 14 and Fig. 15 show how a significant increase in the EVM level appears in PRBs 12 and 13 (transition zone), and a decreasing effect as moving away from the transition zone to the edge of the carrier band. This effect is relevant in the nonlinear scenario and is almost negligible in the linear PA and when DPD is applied. In fact, when a DPD is applied, the EVM decreases in all PRBs. For instance, in Fig. 14, EVM reaches values higher than 10% in all the last 12 PRBs when Pout 22 dBm. When a DPD is applied, EVM = decreases in all PRBs below 7%, reaching a value of 3% in the PRBs with a higher power level. The most relevant aspect is that the differences among the PRBs are negligible. Similar conclusions can be obtained from Fig. 15, which shows that the power level difference (�) between the inner and outer sub-bands only affects the specific expected EVM values. It can also be observed that in these situations, the influence of the modulation scheme in the EVM is not significant. **IV. CONCLUSION** In this study, we analyzed the influence of mandatory satisfaction of EVM requirements at the transmitter in the design of radio resource management strategies (RRMs) for DL in 4G/5G mobile systems. Specifically, we experimentally analyzed the real effects of the power allocation schemes linked ----- to ICIC and eICIC in terms of EVM degradation in transmissions. This aspect has not been addressed in studies on ICIC or eICIC, which usually overlook these EVM requirements, resulting in ideal evaluations of RRM proposals and overestimations of the user data transmission and system throughput performance. Only a few works have considered LTE constraints related to the dynamic power range for each modulation order to ensure EVM requirements. However, constraints for ICIC was out of the studies conducted by the specifications. Therefore, the analysis of this work avoids the unnecessary performance limitations that can be achieved when applying ICIC and eICIC in LTE/LTE-A and 5G cells by unnecessarily restricting the range of variation of the allocable power masks and MCSs. As it is known, the particular numerical results obtained depend on the specific transmitter implementation. Thus, the contribution does not lie in providing a precise numerical quantification of the effects, but in the analysis and verification of some EVM behavior patterns that should be considered to maximize the performance of the ICIC and eICIC schemes while ensuring QoS. We can conclude that the two most relevant parameters that affect the EVM are the power ratio between the outer and inner sub-bands (�), PA operation points and the distance from the PRB where the EVM is evaluated to the jump point between the inner and outer subbands. It has been shown that the PRB of low-power subbands in the transition zone between the two sub-bands is considerably affected by the power jump. However, EVM degradation diminished when we moved away from the transition zone. Future research work could address the design of RRM strategies based on the type of analysis performed in this work. This will allow us to make more suitable decisions at the scheduler concerning the allowed allocable MCS in each PRB. **REFERENCES** [1] A. S. Hamza, S. S. Khalifa, H. S. Hamza, and K. Elsayed, ‘‘A survey on inter-cell interference coordination techniques in OFDMA-based cellular networks,’’ IEEE Commun. Surveys Tuts., vol. 15, no. 4, pp. 1642–1670, 4th Quart., 2013. [2] D. G. González, M. García-Lozano, S. Ruiz, and J. Olmos, ‘‘On the need for dynamic downlink intercell interference coordination for realistic long term evolution deployments,’’ Wireless Commun. Mobile Comput., vol. 14, no. 4, pp. 409–434, Mar. 2014. [3] C. Kosta, B. Hunt, A. U. Quddus, and R. Tafazolli, ‘‘On interference avoidance through inter-cell interference coordination (ICIC) based on OFDMA mobile systems,’’ IEEE Commun. Surveys Tuts., vol. 15, no. 3, pp. 973–995, 3rd Quart., 2013. [4] E. Pateromichelakis, M. Shariat, A. U. Quddus, and R. Tafazolli, ‘‘On the evolution of multi-cell scheduling in 3GPP LTE/LTE-A,’’ IEEE Commun. _Surveys Tuts., vol. 15, no. 2, pp. 701–717, 2nd Quart., 2013._ [5] Y. L. Lee, T. C. Chuah, J. Loo, and A. Vinel, ‘‘Recent advances in radio resource management for heterogeneous LTE/LTE-A networks,’’ IEEE _Commun. Surveys Tuts., vol. 16, no. 4, pp. 2142–2180, 4th Quart., 2014._ [6] T. Novlan, R. Ganti, A. Ghosh, and J. Andrews, ‘‘Analytical evaluation of fractional frequency reuse for OFDMA cellular networks,’’ _IEEE Trans. Wireless Commun., vol. 10, no. 12, pp. 4294–4305,_ Dec. 2011. [7] B. M. Hambebo, M. M. Carvalho, and F. M. Ham, ‘‘Performance evaluation of static frequency reuse techniques for OFDMA cellular networks,’’ in Proc. 11th IEEE Int. Conf. Netw., Sens. Control, Apr. 2014, pp. 355–360. [8] M. S. Hossain, F. Tariq, G. A. Safdar, N. H. Mahmood, and M. R. A. Khandaker, ‘‘Multi-layer soft frequency reuse scheme for 5G heterogeneous cellular networks,’’ in Proc. IEEE Globecom Workshops _(GC Wkshps), Dec. 2017, pp. 1–6._ [9] F. Hamdani, A. Maurizka, M. M. Ulfah, and Iskandar, ‘‘Power ratio evaluation for soft frequency reuse technique in LTE-A heterogeneous networks,’’ in Proc. 11th Int. Conf. Telecommun. Syst. Services Appl. _(TSSA), Oct. 2017, pp. 1–5._ [10] M. S. Hossain and Z. Becvar, ‘‘Soft frequency reuse with allocation of resource plans based on machine learning in the networks with flying base stations,’’ IEEE Access, vol. 9, pp. 104887–104903, 2021. [11] A. D. Firouzabadi, A. M. Rabiei, and M. Vehkaperä, ‘‘Fractional frequency reuse in random hybrid FD/HD small cell networks with fractional power control,’’ IEEE Trans. Wireless Commun., vol. 20, no. 10, pp. 6691–6705, Oct. 2021. [12] S. C. Lam and X. N. Tran, ‘‘Fractional frequency reuse in ultra dense networks,’’ Phys. Commun., vol. 48, Oct. 2021, Art. no. 101433. [13] Z. H. Abbas, M. S. Haroon, F. Muhammad, G. Abbas, and F. Y. Li, ‘‘Enabling soft frequency reuse and Stienen’s cell partition in two-tier heterogeneous networks: Cell deployment and coverage analysis,’’ IEEE Trans. Veh. Technol., vol. 70, no. 1, pp. 613–626, Jan. 2021. [14] H. Carvajal, N. Orozco, D. Altamirano, and C. De Almeida, ‘‘Performance analysis of non-ideal sectorized SFR cellular systems in rician fading channels with unbalanced diversity,’’ IEEE Access, vol. 8, pp. 133654–133672, 2020. [15] B. Soret and K. I. Pedersen, ‘‘Macro transmission power reduction for HetNet co-channel deployments,’’ in Proc. IEEE Global Commun. Conf. _(GLOBECOM), Dec. 2012, pp. 4126–4130._ [16] B. Soret, A. D. Domenico, S. Bazzi, N. H. Mahmood, and K. I. Pedersen, ‘‘Interference coordination for 5G new radio,’’ IEEE Wireless Commun., vol. 25, no. 3, pp. 131–137, Jun. 2018. [17] B. Soret and K. I. Pedersen, ‘‘On-demand power boost and cell muting for high reliability and low latency in 5G,’’ in Proc. IEEE 85th Veh. Technol. _Conf. (VTC Spring), Jun. 2017, pp. 1–5._ [18] T. Cogalan, S. Videv, and H. Haas, ‘‘Coordinated scheduling for aircraft in-cabin LTE deployment under practical constraints,’’ in Proc. IEEE 87th _Veh. Technol. Conf. (VTC Spring), Jun. 2018, pp. 1–6._ [19] Evolved Universal Terrestrial Radio Access (E-UTRA); Physical Layer _Procedures, document TS-36.213 V14.16.0, 3GPP, Release 14, Sep. 2020._ [20] Evolved Universal Terrestrial Radio Access (E-UTRA); User Equipment _(UE) Radio Transmission and Reception, document TS 36.101 V14.22.0,_ Release 14, 3GPP, Mar. 2022. [21] Evolved Universal Terrestrial Radio Access (E-UTRA); Base Station (BS) _Radio Transmission and Reception, document TS 36.104 V14.10.10,_ Release 14, 3GPP, Mar. 2021. [22] Evolved Universal Terrestrial Radio Access (E-UTRA); Radio Resource _Control (RRC) Protocol Specification, document TS 36.331 V14.16.0,_ Release 14, 3GPP, Jan. 2021. [23] Y. Wang, W. Zhang, F. Peng, and Y. Yuan, ‘‘RNTP-based resource block allocation in LTE downlink indoor scenarios,’’ in Proc. IEEE Wireless _Commun. Netw. Conf. (WCNC), Apr. 2013, pp. 334–3341._ [24] Proposal for eNB TX Dynamic Range Requirements, document R4080113, Nokia Siemens, 3GPP TSG-RAN Working Group 4 Meeting #46, Sorrento, Italy, Feb. 2008. [25] BS TX Dynamic Range, document R4-080038, TP for 36.104, NTT DoCoMo, NXP, 3GPP TSG-RAN Working Group 4 Meeting #46, Sorrento, Italy, Feb. 2008. [26] BS TX Dynamic Range, Panasonic, document R4-080084, 3GPP TSG RAN WG4 (Radio) Meeting #46, Sorrento, Italy, Feb. 2008. [27] LS on BS Implications Due to LP-ABS for feICIC, document R4122088, Huawei, HiSilicon, 3GPP TSG-RAN WG4 Meeting #62, Jeju, South Korea, Mar. 2012. [28] C. Eun and E. J. Powers, ‘‘A new Volterra predistorter based on the indirect learning architecture,’’ IEEE Trans. Signal Process., vol. 45, no. 1, pp. 223–227, Jan. 1997. [29] A. Zhu, P. J. Draxler, J. J. Yan, T. J. Brazil, D. F. Kimball, and P. M. Asbeck, ‘‘Open-loop digital predistorter for RF power amplifiers using dynamic deviation reduction-based Volterra series,’’ IEEE Trans. Microw. Theory _Techn., vol. 56, no. 7, pp. 1524–1534, Jul. 2008._ [30] L. Ding, G. T. Zhou, D. T. Morgan, Z. Ma, J. S. Kenney, J. Kim, and C. R. Giardina, ‘‘A robust digital baseband predistorter constructed using memory polynomials,’’ IEEE Trans. Commun., vol. 52, no. 1, pp. 159–165, Jan. 2004. ----- ÁNGELA HERNÁNDEZ-SOLANA received the degree in telecommunications engineering and the Ph.D. degree from the Universitat Politècnica de Catalunya (UPC), Spain, in 1997 and 2005, respectively. She has been working at UPC and the University of Zaragoza, where she has been an Associate Professor, since 2010. She is a member of the Aragón Institute of Engineering Research (I3A). Her research interests include 5G/4G technologies, heterogeneous communication networks and mission-critical communication networks, with emphasis on transmission techniques, radio resource management and quality of service, mobility management and planning, and dimensioning of mobile networks. PALOMA GARCÍA-DÚCAR was born in Zaragoza, Spain, in 1972. She received the degree in telecommunications engineering and the Ph.D. degree from the University of Zaragoza, in 1996 and 2005, respectively. In 1995, she was employed at Teltronic, S.A.U., where she worked with the Research and Development Department, involved in the design of radio communication systems (mobile equipment and base station), until 2002. From 1997 to 2001, she has collaborated in several projects with the Communication Technologies Group, Electronics Engineering and Communications Department, University of Zaragoza. In 2002, she joined the Centro Politécnico Superior, University of Zaragoza, where she is currently an Assistant Professor. She is also involved as a Researcher with the Aragon Institute of Engineering Research (I3A). Her research interests include the area of linearization techniques of power amplifiers and signal processing techniques for radio communication systems. ANTONIO VALDOVINOS received the degree in telecommunications engineering and the Ph.D. degree from the Universitat Politècnica de Catalunya (UPC), Spain, in 1990 and 1994, respectively. He was with UPC and the University of Zaragoza, where he has been a Full Professor, since 2003. He is a member of the Aragón Institute of Engineering Research (I3A). His research interests include 5G/4G technologies, heterogeneous communication networks and mission-critical communication networks, with emphasis on transmission techniques, radio resource management and quality of service, mobility management, and planning and dimensioning of mobile networks. JUAN ERNESTO GARCÍA was born in Zaragoza, Spain, in 1997. He received the bachelor’s and master’s degrees in telecommunications engineering from the University of Zaragoza, in 2019 and 2021, respectively. In 2020, he was employed with the Communication Technologies Group, Department of Electronics Engineering and Communications, University of Zaragoza, after collaborating with them during the final bachelor’s degree thesis, where he worked in the research of several lineralization techniques for critical mobile communication systems. In 2021, he joined Indra Sistemas S. A., where he is currently working in the Solution and Product Area as a System Engineer. He is still collaborating as a Researcher with the Aragon Institute of Engineering Research (I3A). His research interests include the area of radio-frequency design and signal processing techniques for critical radio communication systems. JESÚS DE MINGO was born in Barcelona, Spain, in 1965. He received the Ingeniero de Telecomunicación degree from the Universidad Politécnica de Cataluña (UPC), Barcelona, in 1991, and the Doctor Ingeniero de Telecomunicación degree from the Universidad de Zaragoza, in 1997. In 1991, he joined the Antenas Microondas y Radar Group, Departamento de Teoría de la Señal y Comunicaciones, until 1992. In 1992, he was employed at Mier Comunicaciones, S.A., where he worked in the solid state power amplifier design, until 1993. Since 1993, he has been an Assistant Professor, since 2001, an Associate Professor, and since 2017, a full Professor with the Departamento de Ingeniería Electrónica y Comunicaciones, Universidad de Zaragoza. He is a member of the Aragon Institute of Engineering Research (I3A). His research interests include the area of linearization techniques of power amplifiers, power amplifier design, and mobile antenna systems. PEDRO LUIS CARRO was born in Zaragoza, Spain, in 1979. He received the M.S. degree in telecommunication engineering and the Ph.D. degree from the University of Zaragoza, in 2003 and 2009, respectively. In 2002, he carried out his master thesis on antennas for mobile communications at Ericsson Microwave Systems, A.B., Göteborg, Sweden, with the Department of GSM and Antenna Products. From 2002 to 2004, he was employed at RYMSA S.A., where he worked with the Space and Defense Department as an Electrical Engineer, involved in the design of antennas and passive microwave devices for satellite communication systems. From 2004 to 2005, he worked with the Research and Development Department, Telnet Redes Inteligentes, as a RF Engineer, involved in radio over fiber systems. In 2005, he joined the University of Zaragoza, as an Assistant Professor with the Electronics Engineering and Communications Department. His research interests include the area of mobile antenna systems, passive microwave devices, and power amplifiers. in -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/ACCESS.2022.3170910?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ACCESS.2022.3170910, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://ieeexplore.ieee.org/ielx7/6287639/9668973/09764687.pdf" }
2,022
[ "JournalArticle" ]
true
null
[ { "paperId": "39816e34282494bed4308a78808f0f3a7bcdaaf4", "title": "Fractional Frequency Reuse in Ultra Dense Networks" }, { "paperId": "1b337a801602f8de8ee90cb1219bea384c945994", "title": "Fractional Frequency Reuse in Random Hybrid FD/HD Small Cell Networks With Fractional Power Control" }, { "paperId": "a3414c45c1d82767ee6826c251a18cef99690007", "title": "Enabling Soft Frequency Reuse and Stienen's Cell Partition in Two-Tier Heterogeneous Networks: Cell Deployment and Coverage Analysis" }, { "paperId": "650823d4268242cfc2bc3d79351afd5bf0f61acf", "title": "Performance Analysis of Non-Ideal Sectorized SFR Cellular Systems in Rician Fading Channels With Unbalanced Diversity" }, { "paperId": "3467505bdd45a9f5d851703cbfda3dad8b5f0faf", "title": "Coordinated Scheduling for Aircraft In-Cabin LTE Deployment under Practical Constraints" }, { "paperId": "c4887b774c799427762bc87b26a9bc48020d98d2", "title": "Interference Coordination for 5G New Radio" }, { "paperId": "e32c9efcd9b54c5198e1f2f5315fc04552fae7c3", "title": "Multi-Layer Soft Frequency Reuse Scheme for 5G Heterogeneous Cellular Networks" }, { "paperId": "03ee57135a64bdbe29a8330217d91982d29f3033", "title": "Power ratio evaluation for soft frequency reuse technique in LTE-A heterogeneous networks" }, { "paperId": "b6cb5df07c118a909341c26710ad0572711c8efc", "title": "On-Demand Power Boost and Cell Muting for High Reliability and Low Latency in 5G" }, { "paperId": "f2d8f65b392b857751d5d13a325cf759fb96f4f6", "title": "Recent Advances in Radio Resource Management for Heterogeneous LTE/LTE-A Networks" }, { "paperId": "a7e7034df2b0e9bc81ed7b91eb599245aa29abaf", "title": "Performance evaluation of static frequency reuse techniques for OFDMA cellular networks" }, { "paperId": "0a45f58f55b09a32034cc11014bff1074059e8cb", "title": "On the need for dynamic downlink intercell interference coordination for realistic Long Term Evolution deployments" }, { "paperId": "8937613dda07ee1df0291a43dd46d88252aabda1", "title": "RNTP-based resource block allocation in LTE downlink indoor scenarios" }, { "paperId": "a80354929ab76744a4844912febf22655909d705", "title": "A Survey on Inter-Cell Interference Coordination Techniques in OFDMA-Based Cellular Networks" }, { "paperId": "c4cd5b75c9453f2e4e29d3508728fd7b373ec490", "title": "On Interference Avoidance Through Inter-Cell Interference Coordination (ICIC) Based on OFDMA Mobile Systems" }, { "paperId": "1d07843cf816e58338e741ddf24575576875ddb6", "title": "On the Evolution of Multi-Cell Scheduling in 3GPP LTE / LTE-A" }, { "paperId": "ad1de27490d83edceac2843a6d3b488b78afabbe", "title": "Macro transmission power reduction for HetNet co-channel deployments" }, { "paperId": "36e6a7f5dbe5e969dc0e41aa1b703e3b5a3d48ab", "title": "Analytical Evaluation of Fractional Frequency Reuse for OFDMA Cellular Networks" }, { "paperId": "c46d2fd1d644b66cb75784026f2241e0e8994bb0", "title": "Open-Loop Digital Predistorter for RF Power Amplifiers Using Dynamic Deviation Reduction-Based Volterra Series" }, { "paperId": "be250d0548418ee7f8519261920d10b8c26a6a3a", "title": "A robust digital baseband predistorter constructed using memory polynomials" }, { "paperId": "a3b7c381b5cbd276a324a8105b0fb0cf2bc6449f", "title": "Soft Frequency Reuse With Allocation of Resource Plans Based on Machine Learning in the Networks With Flying Base Stations" }, { "paperId": null, "title": "Evolved Universal Terrestrial Radio Access (E-UTRA); Radio Resource Control (RRC) Protocol Specification , document TS 36.331 V14.16.0, Release 14" }, { "paperId": null, "title": "LS on BS Implications Due to LP-ABS for feICIC" }, { "paperId": null, "title": "Dynamic Range, Panasonic, document R4-080084, 3GPP TSG RAN WG4 (Radio) Meeting #46" }, { "paperId": null, "title": "Dynamic Range, document R4-080038, TP for 36.104, NTT DoCoMo, NXP, 3GPP TSG-RAN Working Group 4 Meeting #46" }, { "paperId": null, "title": "BS TX Dynamic Range, Panasonic , document R4-080084" }, { "paperId": "e418432b5c4d0d53ad60f1926fbeb9767e841bea", "title": "A new Volterra predistorter based on the indirect learning architecture" }, { "paperId": null, "title": "The spatial region where the partitions are used (e.g., cell center or cell edge), and thus, the number of user groups or classes" }, { "paperId": null, "title": "The power level per RE for each frequency partition (e.g., EPRE innerRB and EPRE outerRB ) and, thus, the power range between them (e.g., (cid:49) = EPRE outerRB − EPRE inner ) RB" }, { "paperId": null, "title": "User Equipment (UE) Radio Transmission and Reception" } ]
19,266
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00e55a9d00b3beab2f3a952e50d8a7359ae8f1c8
[ "Computer Science" ]
0.930337
Consortium Blockchain-Based Decentralized Stock Exchange Platform
00e55a9d00b3beab2f3a952e50d8a7359ae8f1c8
IEEE Access
[ { "authorId": "1825767965", "name": "Hamed Al-Shaibani" }, { "authorId": "2515757", "name": "Noureddine Lasla" }, { "authorId": "2140299480", "name": "Mohamed M. Abdallah" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "http://ieeexplore.ieee.org/servlet/opac?punumber=6287639" ], "id": "2633f5b2-c15c-49fe-80f5-07523e770c26", "issn": "2169-3536", "name": "IEEE Access", "type": "journal", "url": "http://www.ieee.org/publications_standards/publications/ieee_access.html" }
The global implementation architecture of the traditional stock market distributes responsibilities and data across different intermediaries, including financial and governmental organizations. Each organization manages its system and collaborates with the others to facilitate trading on the stock exchange platform, and typically buy-sell orders go through different parties before settlement. This design architecture that involves a complex chain of intermediaries has several limitations and shortcomings, such as a single point of failure, a longer time for financial settlements, and weak transparency. Blockchain technology consists of a network of computer nodes that securely share a common ledger without the need of having any kind of intermediaries. In this paper, we present a novel blockchain-based architecture for a fully decentralized stock market. Our architecture is based on a private Ethereum blockchain to create a consortium network leveraging organizations that are already involved in the traditional stock exchange to act as validating nodes. In our architecture, the stock exchange trading logic is completely implemented on a smart contract, while considering the existing governmental market regulations. Since the new platform does not introduce significant changes to the stock exchange trading logic and does not eliminate any of the traditional parties from the system, our proposal promotes efficient adoption and deployment of decentralized stock exchange platforms. In addition, we present a proof of concept implementation of the new architecture, including the smart contract for trade exchange, as well as a virtualization-based test network to assess the platform performance. The test network consists of virtual nodes that run the developed stock exchange smart contract where we measure the buy-sell orders throughput and latency under different network sizes and trading workload scenarios. The obtained results have shown that the proposed trading platform can reach a throughput of 311.8 tx/sec, which is equivalent to 89% of the optimal throughput when the sending rate is 350 tx/sec. This throughput is largely sufficient to meet the requirement of major stock exchanges, such as Singapore stock market.
Received May 29, 2020, accepted June 15, 2020, date of publication June 29, 2020, date of current version July 20, 2020. _Digital Object Identifier 10.1109/ACCESS.2020.3005663_ # Consortium Blockchain-Based Decentralized Stock Exchange Platform HAMED AL-SHAIBANI, NOUREDDINE LASLA, (Member, IEEE), AND MOHAMED ABDALLAH, (Senior Member, IEEE) Division of Information and Computing Technology, College of Science and Engineering, Hamad Bin Khalifa University (HBKU), Doha, Qatar Corresponding author: Hamed Al-Shaibani (halshaibani@mail.hbku.edu.qa) **ABSTRACT The global implementation architecture of the traditional stock market distributes responsi-** bilities and data across different intermediaries, including financial and governmental organizations. Each organization manages its system and collaborates with the others to facilitate trading on the stock exchange platform, and typically buy-sell orders go through different parties before settlement. This design architecture that involves a complex chain of intermediaries has several limitations and shortcomings, such as a single point of failure, a longer time for financial settlements, and weak transparency. Blockchain technology consists of a network of computer nodes that securely share a common ledger without the need of having any kind of intermediaries. In this paper, we present a novel blockchain-based architecture for a fully decentralized stock market. Our architecture is based on a private Ethereum blockchain to create a consortium network leveraging organizations that are already involved in the traditional stock exchange to act as validating nodes. In our architecture, the stock exchange trading logic is completely implemented on a smart contract, while considering the existing governmental market regulations. Since the new platform does not introduce significant changes to the stock exchange trading logic and does not eliminate any of the traditional parties from the system, our proposal promotes efficient adoption and deployment of decentralized stock exchange platforms. In addition, we present a proof of concept implementation of the new architecture, including the smart contract for trade exchange, as well as a virtualization-based test network to assess the platform performance. The test network consists of virtual nodes that run the developed stock exchange smart contract where we measure the buy-sell orders throughput and latency under different network sizes and trading workload scenarios. The obtained results have shown that the proposed trading platform can reach a throughput of 311.8 tx/sec, which is equivalent to 89% of the optimal throughput when the sending rate is 350 tx/sec. This throughput is largely sufficient to meet the requirement of major stock exchanges, such as Singapore stock market. **INDEX TERMS Blockchain, smart contract, stock exchange, trading.** **I. INTRODUCTION** The stock market is a platform composed of financial and governmental organizations that participate in exchanging shares, bonds, or other securities in a transaction known as a trade. The growth of this market has a positive and direct impact on the financial growth of a country’s economy since it offers opportunities that attract investors to trade and exchange shares. Studies conducted in USA [1] and Pakistan [2] examine this relationship by comparing the stock market performance with the Gross Domestic Product The associate editor coordinating the review of this manuscript and approving it for publication was Christian Esposito . (GDP). The authors conclude that the performance of the stock exchange is directly proportional to the performance of a country’s economy. For instance, Pakistan achieved growth in both the GDP and stock market of about 30% and 6.08%, respectively, between 2003 and 2008. Therefore, the stability and security of the stock platform are vital in increasing the confidence to invest and trade and eventually results in better economic growth for the country. Despite the wide popularity and adoption of the conventional stock exchange platform architecture, it suffers from different limitations and shortcomings as follows [3], 1) due to the centralization architecture of the stock exchange platform, each participating system such as brokers and the stock ----- exchange is considered as a single point of failure, 2) there is inconsistency in the data managed by each system resulting in errors and extended recovery time when failures happen, 3) the availability of the system on a daily basis is limited which can impact exercises such as auditing the data as well as providing transparent access to the users throughout the day, and 4) the long time the system takes to perform financial settlements; typically it takes three days after trading happens to achieve the settlement. There have been many attempts by major vendors who build trade execution and matching engines in the financial sector to address the above limitations. For instance, Nasdaq, which is one of the leading tradings and matching technology vendors, is offering different services and products such as hosting, managing and providing services to the complete end to end trading processes. They also provide surveillance systems integrated with other systems to allow their clients to monitor and regulate the process of trading and settlement. However, this approach still suffers from limitations such as the long settlement time, the limited level of data transparency, and the defined trading hours. Most importantly, clients might require hardware specifications or the need to follow an enforced architecture that involves the distribution of the systems involved in the trading process across multiple organizations where each manages its system separately. For instance, some stock market regulators enforce specific cash settlement and clearance solutions, which is a different system than what Nasdaq provides [4]. This results in maintaining multiple systems, which increases the chance of having a single point of failure among the participating systems as well as increasing the complexity of the overall trading platform system architecture. According to [5], blockchain can solve many of the identified limitations affecting the traditional stock exchange platform, such as the lack of transparency, the long settlement time between brokers and the central bank, and the high transaction fees paid to brokers for each generated trade. In [3], the centralized architecture of Bucharest stock exchange market has been analyzed to address its limitations with the main objective of addressing the issue of the high fees the investor pays to the broker for each successfully executed trade. The authors define the new stock market in a smart contract and deploy it into the Ethereum public network. This implementation requires a form of payment, in Ethereum cryptocurrency (Ether), for each performed transaction. Their conclusion shows that decentralizing the Bucharest stock exchange platform can help in reducing the total transaction fees. The research objective and implementation approach to decentralize the Bucharest stock exchange platform varies with our objective and the approach we took in several key areas. First, we are using a consortium blockchain network in which all participants are known and trusted, and there is no form of cryptocurrency fees that will be used to pay the miners in the network. Second, our main objective is to optimize the performance of the decentralized system rather than reducing the fees, as we measure the throughput and **TABLE 1. List of acronyms.** latency to ensure our implementation meets the required level of the stock market platform. Also, our consensus algorithm is based on Proof of Authority (PoA). It provides better performance in terms of execution time and power efficiency in comparison with the public network consensus algorithms such as Proof of Work (PoW) used by the decentralized Bucharest stock exchange. The consensus algorithms will be discussed in more detail later in this paper. TABLE 1 provides definitions of the acronyms used in the paper. In this paper, we propose a consortium blockchain-based stock exchange platform that meets the performance requirement of the stock exchange platform while also addressing the limitations of the traditional stock exchange. Our proposal is based on Ethereum blockchain technology in which all necessary business regulations and rules defined in a smart contract shared across a permissioned blockchain network with the participating financial and governmental institutions. We perform experimental tests by deploying the smart contract on virtual nodes and measuring the network performance under different workloads by increasing the number of generated trades and the number of validating nodes. Our results show that this architecture does meet the required performance of the stock exchange platform in terms of latency and throughput under different test scenarios. The remaining of this paper is organized as follows: Section II provides an overview of the traditional stock exchange platform. Section III presents an overview of blockchain technology and discusses the different implementations and consensus algorithms of blockchain for public and private implementations. Section IV discusses the related work focusing on the implementation design based on blockchain for stock market and an E-auction system. Our proposed blockchain-based stock exchange framework is presented in Section V, where we discuss the system architecture and the smart contract defining the functionalities and business logic of stock exchange. In section VI, ----- **TABLE 2. Major entities of a traditional stock market.** we evaluate the performance of the proposed architecture in terms of transaction throughput and latency. Finally, Section VII concludes the paper. **II. TRADITIONAL STOCK EXCHANGE OVERVIEW** The stock market can be defined as ‘‘an aggregation of buying and selling offers corresponding to an asset’’ [6]. The asset has a form of bonds, shares, or other securities that the market offers. The person who trades in the stock market is called investor or trader and needs first to open a trading account with the Central Securities Depository (CSD), which takes the responsibility of managing investors’ trading accounts and personal data. Due to market regulation, investors cannot directly place an order into the system and need to go through a third party, namely brokers. In the case of international traders, a special financial entity, named Custodian, is employed to place orders in the local market. The matching engine of the entered buy and sell orders is hosted by the Stock Exchange (SE) entity. The Central Bank (CB) manages the financial settlement between brokers and custodians. All the participating governmental and financial organizations need to follow the rules and regulations defined by the Financial Market Authority (FMA). FMA is also responsible for continuously monitoring the stock exchange platform and reviewing the data. A summary of the different entities involved in the traditional stock market with their respective description is given in TABLE 2. _A. TRADING OVERVIEW_ The traditional stock market is a centralized platform as shown in FIGURE 1, which presents this architecture and the flow of events that take place when a new investor participates in the platform. First, the investor needs to open an investor account from CSD and obtains his/her National Investor Number (NIN). The investor then needs to open a trading account with a broker by providing the mandatory NIN account. Once the investor information is validated, he/she can place orders of buying or selling shares through the associated broker services such as the website or the mobile application. The broker takes the responsibility of using its Order Management System (OMT), which acts as an interface with SE to submit the investor’s order. Once a successful trade is generated for that order, SE sends the **FIGURE 1. Centralized stock exchange platform flow of events.** acknowledgment message to the broker, who then notifies the user via the different services provided by the broker. The shares owned by the investor are updated in the broker account as well as the investor account held in CSD system. The market regulator FMA has access to both SE and CSD systems to monitor the market and validate the trades during and after trading hours. _B. TRADING HOURS AND PHASES_ There are usually four different phases that a stock market goes into in most implementations, as shown in FIGURE 2. The market starts with a 30 minutes pre-open phase in which investors can enter their orders, but no trades are generated. Based on bids and offers entered, opening prices for listed securities are calculated, so when the market opens, those calculated prices will be the buy/sell prices used by the investors. The next phase is market opening for trading in which the listed securities can be traded, and orders entered in the pre-open phase gets executed. This is the main phase of the stock exchange, where orders keep entering, and trades get generated. The duration of this phase is approximately 3 hours and 30 minutes, and this time can vary from one stock exchange to another as per the regulations of the country hosting it. The market then prepares for closing and enters the pre-close phase, which is estimated to last for 10 minutes. ----- **FIGURE 2. Stock exchange market trading hours and market phases.** In this phase, algorithms run to generate the closing prices for listed securities, and investors can still enter their orders, but no trade is generated. Finally, the market enters the closing phase, which is estimated to last for 5 minutes in which entered orders get executed. The market then closes for the day. _C. ORDER TYPES_ Investors can place different types of orders in terms of buying or selling shares [3]. These types are listed below to explain how a successful match between buy and sell transactions happen, as some orders allow the investors to specify conditions on which if triggered, the order gets executed and hence, results in generating trades: - Market Order: The buyer would like to buy the shares with the current price of the share in the market. The same applies to the seller. This order type does not guarantee the highest financial gain from the transaction but ensures the order is immediately executed when entered - Limit Order: The buyer sets a maximum limit he/she is willing to pay to buy a particular share. The order gets executed for any sell offer that is equal to or less than this limit. In the case of the seller, the limit will be the minimum price he/she is willing to sell with. The order gets executed for a buy order with a price higher than or equals to the minimum limit. - Validity Defined Order: This order can be associated with either a limit or market order types. The entered order will remain valid for a single day (Day Order) or until a certain date (Good Till Date Order). In most stock exchange markets, the system cancels the order in approximately 62 days if no validity date is provided (Open Order). - Fill or Kill (FOK) Order: Either execute the full order (Sell or buy all indicated shares) or cancel the order - Immediate or Cancel (IOC) Order: The order is immediately executed, and the remaining quantity that has not been fulfilled will be canceled. **III. BLOCKCHAIN OVERVIEW** To address the limitations and shortcomings of the traditional stock exchange platform, we opt for Blockchain technology. Blockchain can be defined as a ‘‘network of computers, all of which must approve a transaction has taken place before it is recorded, in a ‘chain’ of computer code. The details of the transfer are recorded on a public ledger where anyone on the network can see the information [7]. It consists of blocks with each containing a pointer in the form of a hash of the previous block and verified transaction data protected with hash signatures [8]–[10]. Transactions in blockchain are broadcasted in the network and are validated by a process known as mining that is performed by special nodes in the network known as miner nodes [7]. Miner nodes are specific nodes that append a new block to the chain once the block becomes full. It is extremely difficult to change a block in the chain as it requires to have subsequent blocks to be recreated, and hence this mechanism prevents modification and maintains a high level of security. FIGURE 3 shows the content of the first three blocks. As shown, each block consists of the list of transactions, the hash of the previous block (except for the first block), a nonce value, which is a number that can be used only once. In some blockchain implementations such as Bitcoin the nonce is altered by the miner such that the hash of the block is equal to or less than a certain numerical target value provided by the network as a challenge. The block also contains the hash of the block itself. ----- **FIGURE 3. The first three blocks in a Blockchain.** In order to keep track of all transactions, blockchain ledger is used in a network where participants have access to the same ledger replicating the transactions among all peer nodes in that network. This replication ensures that the overall system built on blockchain can resume if multiple participating nodes failed to connect to the network. The nodes in the network use addresses or identifiers known as public keys to be distinguished by, and hence, defined roles, privacy, and anonymity can be efficiently maintained [11]. Miner nodes rely on the fact that all transactions in the network are duplicated across all nodes involved. Therefore, ‘‘Distributed Consensus’’ needs to be achieved, meaning that an agreement on the validity of the blockchain is achieved by all nodes involved and they all share the same version of the Blockchain [12]. Some implementations of Blockchain such as Ethereum, uses protocols to present the logic that need to be followed and this is known as a smart contract. According to [13] a smart contract can be defined as ‘‘the computer protocols that digitally facilitate, verify, and enforce the contracts made between two or more parties on blockchain’’. The smart contract ensures that the defined logic is validated and needs to be followed. It does not need a centralized entity to validate the defined conditions in the smart contract since once it is deployed, all participating nodes in the network need to follow the defined logic in it. _A. BLOCKCHAIN IMPLEMENTATION TYPES_ In Blockchain, all nodes need to reach a state of agreement or consensus on the next block that needs to be added to the chain, especially that in a peer to peer network such as blockchain these nodes don’t trust each other. However, there are many consensus mechanisms that can be used depending on the implementation type and the blockchain technology used. There are mainly two types of implementation for Blockchain network: permissionless and permissioned. 1) PERMISSIONLESS BLOCKCHAIN (PUBLIC) Permissionless or public implementation of blockchain allows any user to become a node and connect to the network through the internet. The implementation utilizes the concept of Peer-to-Peer network (P2P) that uses distributed architecture in which no client takes the form of an administrator. All clients on the network are connected by flat topology where each peer shares the same rights and privileges as other peers and have access to the same resources as other peers [7], [14], [15]. 2) PERMISSIONED BLOCKCHAIN (PRIVATE) Unlike the permissonless blockchain, nodes in permissioned blockchain are identified and authenticated. In some implementations, an entity takes the responsibility of managing the roles and responsibilities of the nodes and granting them permission to the data accordingly. _B. CONSENSUS ALGORITHMS_ There are several consensus algorithms for the implementation of public blockchain network such as Proof of Work (PoW), Proof of Stake (PoS), Proof of Burn (PoB), Delegated Proof of Stake (DPoS), and Proof of Importance (PoI). PoW and PoS are the two most famous and commonly used consensus algorithms [16]. PoW is an intensive hashing mechanism that provides a difficult mathematical challenge for the block miners to solve, and whoever manages to solve the challenge first will become the block miner [17]. This protocol ensures integrity among all nodes but suffers greatly when it comes to its performance time, processing, and energy power required [17]. On the other hand, PoS consensus protocol is more power-efficient and reduces mining costs. This protocol takes less time than PoW to validate a transaction as it relies on validators taking part in voting for the next block, and the weight of each validator’s vote is dependent on how much it deposited in that system. Nodes that are allowed to create a block act as validators who need to deposit some cryptocurrency as a stake in the network that will be locked to have the chance of being selected as the next block miners. The more stake a validator has in the network, the higher the chance of it being selected to validate the new block. Such protocol ensures acting correctly since any validator who violates the network rules or acts maliciously will lose its stake deposited in the network [18]. PoS has several advantages such as consuming less power and energy, better performance time, and the mechanism of having a stake that can be lost for any malicious behavior is expected to pressure validators to act genuinely more than in PoW. In the case of permissioned blockchain networks, PoA and RAFT are popular consensus algorithms where the participants are known and trusted in a private network. According to [17], PoA is an algorithm that attracted a lot of attention due to its offered performance resulting from lighter exchanged messages. It operates in rounds where several nodes are elected, with one of them acting as a mining leader charged with the task of proposing the new block and eventually reaching consensus. These elected nodes are called ‘‘authorities,’’ and each has its unique ID in which if we have N authorities, at least N/2 1 are assumed to be honest. This + algorithm follows a ‘‘mining rotation schema’’ to distribute the block creation among the authorities in a fair manner, and for each round step, a mining leader from the authority is elected to mine the new block [17]. ----- **TABLE 3. Comparison between the consensus algorithms.** In [19], the author argues that RAFT consensus is easy to understand and implement, which makes it efficient to use when building applications and systems. It works by having a set timer for all authorized nodes, which can validate new blocks in ‘‘terms’’ that can be seen as rounds that get repeated over time. For a given term, as the timer runs out for the first authority, it enters what is called ‘‘candidate state’’, in which it votes for itself to become a leader and broadcasts requests to other authorities to vote for it. If the majority positively voted for the candidate node, then it becomes the leader of that term. Once a leader is elected, its role is to replicate the transaction logs across all other nodes. The logs reach finality and get committed by the leader if and only if it reached the majority of the nodes, once this happens, the leader will commit the log and asks the rest to do the same via a broadcast message. In case the majority of the nodes are offline, the leader will not be able to commit the logs, and there is a high risk of losing the log if the leader and the remaining nodes went offline [19]. TABLE 3 provides a comparison between the permissionless and permissoined consensus algorithms presented in this paper. For the proposed solution, all network participants should be known and trusted. The selected consensus algorithm should allow an authorized participant to act as an administrator for the overall platform since FMA regulates the stock market, and its role is required to be perceived. Moreover, the network should be Byzantine fault tolerance in case some of the network validators act maliciously. PoA consensus algorithm satisfies these requirements. Moreover, for our proof-of-concept implementation, the Geth implementation of PoA, named Clique, is adopted. Clique has a rotation schema for leader election, such that in each round, the leader of the round announces the block and it gets added to the blockchain by the receiving nodes [17]. **IV. RELATED WORK** According to [20], implementations of blockchain in the financial sector focus on four main areas, which are improving the transaction processing time, having sustainability for banking and financial transactions, improving financial data privacy and security, and automating financial contracts. For transaction time improvement, the authors highlight that the current banking systems rely on centralized databases that require several days to achieve financial settlements for the executed transactions [21]. The solution that blockchain offers to solve this problem, according to the authors, is to automate financial transaction settlement by setting up a single account structure that will be used by financial institutions, as well as speeding up international fund transfers [21]. Sustainability is another problem that banks and financial institutions suffer from, especially when a bankruptcy of one bank can have a strong impact on the overall financial sectors. The authors in [24] argue that implementing blockchain can lead the financial sector to achieve stability, especially when the decentralized ledger of money is independent of financial regulations of countries and regions. Financial data security and privacy currently face many challenges due to the nature of the centralized data storage that banking and financial instructions rely on [22]. This can lead to data breaches that does reveal not only financial data, but also personal and demographic data that were also stored in the centralized storage. In addition, banking transactions do not provide sufficient anonymity or extending the freedom of privacy that clients would like to have. Blockchain addresses these two issues by decentralizing the data and ensuring they are securely stored in the participating nodes, which add high complexity to unauthorized attempts to alter or access the stored data. Each participant is authorized to perform changes according to the role assigned while maintaining anonymity on transactions performed [23]. Finally, authors in [20] highlight that blockchain automates financial contracts in terms of execution by eliminating the need of a third party in the middle and allowing a financial transaction to be triggered between the two involved parties. To demonstrate such a feature, money transfer usually takes a couple of days, especially in developing countries, as some controls and regulations need to be verified. When such a transaction is implemented using a financial contract in blockchain, it will no longer require a third party intervention as long as both parties perform their roles as defined in the contract. The financial transaction will be securely executed, and the money will be transferred within minutes [24]. We have analyzed two particular implementations that resemble a close similarity to our idea. The first paper discusses the concept of decentralizing the stock market platform by using Blockchain technology while the second paper utilizes the concept of smart contract in blockchain to build a bidding platform. _A. DECENTRALIZING BUCHAREST STOCK MARKET_ _PLATFORM_ In [3], the authors discuss the limitation of the traditional stock market and propose a solution to implement the trading platform on Blockchain. Their research objective is to showcase how transaction fees can be reduced if blockchain ----- **FIGURE 4. Bucharest centralized stock exchange platform [3].** technology is used as a trading platform instead of the traditional stock exchange platform the Bucharest stock market uses. To test their experiment, the authors have implemented two systems with the first one being modeled according to Bucharest centralized stock exchange platform as per FIGURE 4 in which all orders entered by different brokers are gathered in a single system. The second system implemented is a decentralized blockchain based solution that uses a smart contract to simulate the stock exchange platform. This design does not require to have brokers to enter orders, and instead, investors can interact directly with the system and enter the order themselves. By doing so, the fees paid to brokers are eliminated, and the fees the investors pay per transaction in the proposed blockchain based trading system overall is less than the fees paid in the traditional stock exchange platform. The authors conclude that the fees in the decentralized system will increase as the number of orders in the order book increases since the transaction complexity will become higher. Therefore, the decentralized system will be giving a better transaction fee than the centralized system when the order book is partially full. _B. BIDDING SYSTEM BASED ON BLOCKCHAIN SMART_ _CONTRACT_ An e-auction system has several elements that are in common with the stock exchange platform. It consists of bidders, auctioneers, and third-party intermediaries who provide the platform that connects bidders to auctioneers and allows posting products, checking the highest bidding price, and declaring the winner with the highest bidding price. The authors in [10] suggest building an e-auction system without having intermediaries between the sellers and buyers by using Ethereum based smart contract. Their objective is to solve two main problems the current e-auction systems have, which are the limited level of security offered by the online platform and the high transaction fees users have to pay. The authors claim that their blockchain based solution addresses the first problem by ensuring security related to data shared among the different users of the system is appropriately managed and perceived. The second problem is addressed by reducing the transaction cost by removing any intermediary in the system. FIGURE 5 shows a flowchart representing the bidding process taking **FIGURE 5.** Flowchart showcasing the bidding process [10]. place from start to finish. First, the seller posts the bidding information and the starting price. The bidders bid the price in the sealed envelope, and when it is received by the auctioneer, the sealed envelope price that is the highest is announced as the current highest price. If no price received higher than the current bidder’s highest price or the ending time is due, it is announced as the winning price, and the auctioneer can send the product and receive the money from the winning bidder [7]. By applying the proposed blockchain based e-auction platform as an experiment, the authors conclude that the smart contract can enforce confidentiality, non-repudiation, and prevention of unauthorized alteration of entered bidding orders. TABLE 4 showcases the main differences between the proposed platform and the already discussed related work. For instance, our main research objective is to improve the performance of the system in terms of availability, security, and transparency by adopting a consortium blockchain based on Ethereum with PoA consensus algorithm. We are maintaining all key participants in the traditional stock exchange platform to be part of the proposed platform. We are not introducing major changes that conflict with the roles and regulations imposed by the government. For the case of the decentralized Bucharest stock exchange, the research objective is to reduce the transaction fees paid to the brokers by making significant changes to the existing architecture and eliminating the broker completely from the platform. The new proposed architecture is based on permissionless Ethereum network that uses PoW consensus algorithm. This new platform introduces new fees that are less than the fees paid to the brokers in the traditional stock exchange for cases with partially full order book. The objective of the second related work is to build a secure e-auction system without having intermediaries. The authors used a permisionless Etherem network with a PoW consensus algorithm and made changes to the ----- **TABLE 4. Comparing our paper with related works.** traditional bidding platform by removing intermediaries managing it. Both related works rely on using a public blockchain network which has poor performance and cannot handle the required throughput and latency of the current stock exchange. **V. PROPOSED BLOCKCHAIN-BASED STOCK EXCHANGE** **FRAMEWORK** In this section, we describe our proposed decentralized stock exchange platform that is based on a consortium blockchain between financial and organizational entities that are already part of the traditional stock market. We first give an overview of the system architecture, define the roles and responsibilities of the participating entities, and finally present the smart contract holding and managing the stock exchange trading logic. _A. SYSTEM ARCHITECTURE_ As shown in FIGURE 6, our system is composed of a consortium blockchain network, a smart contract, and financial and organizational entities. The consortium blockchain facilitates transactions between the different participating entities and manages the stockeExchange smart contract that handles the stock trading logic. We select a permissioned blockchain as the entities are all known and also because private version of blockchain is more effective in terms of transaction throughput and latency. The consortium network is composed of a set of authorized participants (validators) which are the CSD, FMA, Broker, Government, and SE. Each of them has specific roles and responsibilities as per the traditional stock exchange platform. The StockExchange smart contract defines all the trading logic as well as the different functions that can be performed by the participating entities, such as, create broker, create new investor, assign share to investor, etc. Each participating entity has a private key along with the associated address and public key that are used for authentication. Therefore, the smart contract ensures that each entity is only allowed to trigger functions according to its associated privileges. TABLE 5 summaries the StockExchange smart contract functionalities and the entities authorized to execute each of them. The detail description of the role of each of the participating entities is given in the following : - FMA: it is responsible for creating and maintaining the smart contract as well as defining all the trading logic and functionalities. It also monitors the trading process and ensures that all defined rules and regulations are properly maintained. It interacts with the smart contract to create and maintain companies with shares and to create and maintain brokers. **FIGURE 6.** System architecture. **TABLE 5. StockExchange smart contract authorizations.** - CSD: is responsible for creating and maintaining investor accounts. It interacts with the smart contract to create investor accounts and assign shares to them. - Broker: it takes the role of trading on behalf of investors. It interacts with the smart contract by associating investors to it and entering buy and sell orders for the associated investors. Brokers are also authorized to assign shares from CSD investor accounts to the investor trading account managed by the broker. Each investor can have multiple trading accounts managed by a different broker for each, while each investor must have a single unique investor account (NIN). - Government: it validates the investor data sent by CSD - SE: it is responsible for matching orders queued in the order book, and generating trades. ----- _B. StockExchange SMART CONTRACT_ We define a smart contract called ‘‘StockExchange’’ to include the business logic and the authorization roles of each participating entity. The smart contract manages the buy and sell orders and generates respective trades whenever a buy order offers a price that is equal to or more than the sell order’s price. The different steps of the trading, implemented in the smart contract, are detailed below: 1) Let QB be a queue of all buy orders sorted in ascending order such that i represents the index of the maximum element in the queue denoted as B[P]. Let B[Q] denotes the quantity of the shares in B[P]. 2) Let QA be a queue of all sell orders sorted in descending order such that y represents the index of the minimum element in the queue denoted as A[P]. Let A[Q] denotes the quantity of the shares in A[P]. 3) We assume that all orders entered are of the types limit order or market order and that partially matched orders are possible in cases where B[Q] _A[Q]_ ̸= 4) If B[P] _A[P]_ and B[Q] _A[Q], both orders are fully_ ≥ = matched, and a trade is generated. Both i and y indices are decremented by 1. 5) If B[P] _A[P]_ and B[Q] _A[Q], B[P]_ is partially matched with ≥ ≥ _A[P], and a trade is generated. The value of B[Q]_ is updated such that B[Q] _B[Q]_ _A[Q]_ and y index is decremented = − by 1. 6) If B[P] ≥ _A[P]_ and B[Q] _< A[Q], A[P]_ is partially matched with _B[P], and a trade is generated. The value of A[Q]_ is updated such that A[Q] _A[Q]_ _B[Q]_ and i index is decremented = − by 1. FIGURE 7 shows the sequence diagram between the participating entities and the smart contact, including all the steps required before generating trades and matching buy/sell orders. The detail description of each of the diagram steps are given in the following: 1) FMA defines the list of all brokers that the stock market consists of by calling the ‘‘addBroker’’ function. The system replies with a message showing the successful creation of the broker. 2) FMA defines the companies that are listed in the stock market along with their details such as number of shares they consist of and their prices. The function ‘‘addCompany’’ is used for this purpose. 3) CSD Validates the investor’s data integrity by sending it to the government. The government replies to the smart contract to update the investor validation status. 4) CSD assigns to each validated investor a new investor account number ‘‘NIN’’ by using the function ‘‘addNin’’. 5) The broker associates an investor to its account using the function ‘‘AssociateBrokerToInvestor’’. The smart contract then validates by checking the NIN Account subsystem to ensure that the NIN exists. If it does, the NIN gets associated to the broker account successfully. 6) The broker assigns shares that are stored in the investor’s NIN account maintained by CSD to the trading account maintained by the broker, by calling the function ‘‘AssignShareToNin’’. 7) Buy orders are entered by the broker into the smart contract. Once these orders are entered, the ‘‘StockExchange‘‘ subsystem logs and stores the order in a sorted queue and tries to match these orders with existing sell orders pending in the sell queue list. If a successful match is generated, the system replies back to the broker that successful trades have been generated for the entered orders. If no match could be generated, the broker will be informed that the orders have been successfully entered the system. 8) This step is similar to step 7 as brokers enter sell orders into the smart contract. If a successful match is generated, the broker is informed about it or else; the broker will be informed that the orders have successfully entered the system. _C. SECURITY AND SYSTEM EFFICIENCY ANALYSIS_ The proposed blockchain-based stock exchange architecture ensures the following security and system efficiency : 1) Transparency: the level of transparency provided by using blockchain guarantees that all transactions and data maintained by the system are visible to the authorized participants and cannot be manipulated. However, any change requires consensus and commitment from all network participants before it gets validated. In contrast, the traditional stock exchange suffers from insufficient transparency level as each party has its system and can hide or manipulate the data before sharing it with other participants. 2) High availability: the proposed architecture addresses the single point of failure by ensuring high availability through decentralizing the data across multiple participants. The smart contract can still be executed even if some nodes were disconnected from the network. Contrary to the traditional stock market, if any of the system participants is unavailable, the whole market is affected. 3) Network efficiency: in the stock exchange, the quality of network connectivity has a critical impact on investors’ profits. For instance, an order sent by an investor through his/her associated broker can be delayed by the network if the broker has connectivity issues, or it is physically located far from the SE. Orders that were entered later by other brokers, with better network connectivity or located physically closer to the SE, will be executed first. This results in a financial loss to the investor despite entering the order first and can cause a lack of fairness and trust in the overall platform. The blockchain network provides better connection utilization between the different participants since nodes are distributed in different physical locations. The node ----- **FIGURE 7.** Sequence diagram between participating entities and the StockExchange smart contract. physically located closest to the users interacting with the smart contract will receive the transactions and broadcast them to the remaining nodes in the network. 4) Consistency: since the ledger is shared across different participants, they all have the same version of the data, and any change that happens in one node will be immediately reflected in the ledgers of the other nodes. This solves the issue of having conflicting data that are not synchronized across the participating systems as it is in the traditional stock exchange platform. For instance, if an investor updates his/her personal data directly with CSD without updating it in the broker system too, a delay in authenticating transactions happens thus, impacting the investor’s profit. 5) Cost efficiency: in our proposed architecture, contrary to the traditional stock exchange, all the participating entities use the same common software and platform, which consists of an Ethereum smart contract. This solution architecture is much simpler and cost-effective as it considerably decreases the overall system complexity and cost for maintenance and technical support. In addition, the proposed architecture is highly available and does not require a separate disaster recovery environment. This saves a high cost compared to the traditional stock market, where each participating entity needs to have a specific disaster recovery site. 6) Flexible configuration: the proposed architecture provides more flexibility and scalability in comparison with the traditional stock exchange platform when it comes to adjusting the functionalities and introducing new changes to the trading logic. Since the proposed design architecture consists of the StockExchange smart contract, new and existing functionalities, as well as authorization, can all be managed in one place. The smart contract can then be shared in the network without requiring participants to make changes in the hardware and storage, which makes it much easier to adopt. 7) Smart contract security: In order to design secure smart contracts, authors in [24] and [25] recommend a set of analysis tools to identify security issues and vulnerabilities in the smart contract code. Among the most famous analysis tools, we selected SmartCheck [24] to assess the proposed ‘‘StockExchange’’ smart contract. SmartCheck allowed us to identify multiple security-related issues and optimize some functions in ----- our initial design. Such issues include the extra gas consumption due to the use of multiple loops and bad array manipulation, which, if not appropriately addressed, can lead to a storage overlap attack where it collides with other data in the storage. Moreover, the tool provided multiple recommendations, such as upgrading the solidity code to the latest version as well as emphasizing on the declarations of public and private modifiers. **VI. PERFORMANCE EVALUATION** In this section, we evaluate the performance of our proposed blockchain-based stock exchange platform, in terms of transaction throughput and latency, to showcase its capability in handling the transaction load of the current stock market. We validate our results with Singapore Exchange, which is one of the emerging markets offering a diversity of listed securities for tading. For this purpose, we developed a testing framework that consists of three main modules: network module, transaction generation and listening module, and performance evaluation module. The description of each module is given below: 1) Network module: This module is used to create the consortium blockchain test network that hosts our defined ‘‘StockExchange’’ smart contract. It consists of the entities that resemble the stock exchange participants, which are SE, FMA, Brokers, government, and CSD. These entities are represented as Ethereum nodes by using Docker container technology, where each node runs the Geth Ethereum client. The selected consensus algorithm is PoA ( see Section III for more details). 2) Transaction generation and listening module: This module is implemented using a JavaScript API that serves as a generator of transaction workload and listens to the blockchain for block confirmation events. By continuously listening to the network, this module records information such as block number, validation time, and the number of transactions per block. The transaction workload consists of buy and sell orders, and the total number of generated transactions at each round of testing is configurable. To ensure that each pair of buy and sell orders generates a trade, we generate for each buy order a corresponding sell order. The workload generator and data listener module interact with a special gateway node in the network that receives the transactions and broadcasts them to the network. 3) Performance evaluation module: This module is used to analyze the information stored in the data listener module and measure the performance of each experiment by calculating the throughput and latency for entered orders and generated trades. The throughput or number of transactions per second (TPS) is calculated as the total number of transactions (N ) divided by the time it takes to validate them, which is the time difference between the block with the first transaction and the block with the last transaction : _TPS = N_ _/Btime,_ where Btime is the difference in validation time between the last and the first blocks. The latency is a measurement that shows the difference in time between the time a transaction is sent and the time it gets validated in a block. It is calculated as the total time it takes to process X number of transactions divided by X . _A. EXPERIMENT_ We have implemented our proposed stock exchange platform, which has been built on top of a consortium blockchain network, using Solidity, the de-facto scripting language to write smart contracts in Ethereum. The created smart contract consists of the following main functions: 1) addBroker: adds a new broker to the system by entering its name, its symbol, and the maximum amount of money it is allowed to spend buying shares in a single trading session. 2) addCompany: a new company is added after entering its name, symbol, its total number of shares, and the price per share. 3) ValidateNIN: investor’s data received by CSD is sent to the government for validation. This data consists of the investor name, age, nationality, and ID number. The government responds in the form of true or false value, which CSD uses as a condition to either proceed or cancel the creation of the new NIN account. 4) addNin: for each validated investor, a unique investor number is assigned. This investor number is associated with the investor’s personal data, including the total number of shares owned by the investor. 5) AssociateBrokerToInevestor: it assigns a broker to an investor by entering the broker’s name, symbol, investor name, and NIN. 6) AssignShareToNin: shares are assigned to a given NIN and the total number of shares in the NIN is updated. 7) buyShares: a buy order that has the company’s symbol, number of shares, price, and NIN enters a queue of buy orders. For each new buy order, the queue is sorted such that the order with the highest price is placed first, followed by the rest in descending order. 8) sellShares: a sell order that has the company’s symbol, number of shares, price, and NIN enters a queue of sell orders. For each sell order, the queue gets sorted such that the order with the lowest price is placed the first, followed by the rest in acceding order. 9) DoMatch: this function is called as part of each ‘‘buyShares’’ and ‘‘sellShares’’ functions. It takes the first item in the buy orders queue and compares it with the first item in the sell orders queue. If the price of the buy order is more or equal to the price of the sell ----- **FIGURE 8. Throughput vs sending rate (tx/sec).** order, then a trade is generated. The matched orders are removed from the queues and the queues again get sorted. The NIN accounts of the buyer and seller will be updated accordingly. Several experiments have been conducted to measure the performance of our Stock exchange trading platform in terms of throughput and latency in which we adjusted our workload and network size for every round of testing. Six different workloads have been used in the form of the sending rate of transactions per second, which are: 100, 200, 300, 350, 400, and 450 tx/sec. The network size has also been adjusted such that the blockchain consisted of 1, 5, 10, and 20 validators for each test scenario. The time to construct two consecutive blocks has been fixed to 2 seconds, and the total number of transactions has also been fixed to 10,000 transactions where 5,000 represent buy orders, and the remaining 5,000 transactions are sell orders. We categorized our test cases into the following workload scenarios : 1) ‘‘With Trades’’: in this scenario, orders are entered such that each pair of buy and sell orders generates a trade. It requires high computational power as the entered orders trigger the doMatch function that requires removing the matched orders from the queue and sorting the queues again as well as updating the buyer and seller NIN accounts accordingly. 2) ‘‘Without Trade’’: In this scenario, the buy and sell orders are not matched, and hence, no trade is generated. In terms of computational needs, this scenario yields the best throughput as it skips the doMatch function, which has to sort and to update investor accounts. The experiments are conducted on a workstation machine with Intel(R) Xeon(R) Gold 6130 CPU, 2.10 GHz, 64 core CPU, 256GB RAM, and running Ubuntu 18.04.2. FIGURE 8 illustrates the measured throughput under different sending rates and number of validators. In the case of a single validator node shown in FIGURE 8a, the throughput is very close to the sending rate up to 350 tx/sec. This is also valid in scenarios with 5 and 10 validators as shown in FIGURE 8b and FIGURE 8c, respectively. However, when the number of validators increases, the throughput gets considerably affected, as shown in FIGURE 8d with 20 validating nodes. This is due to the limited available computation power, as all the nodes in different scenarios share the same workstation machine. To emphasis the effect of computation power on the throughput, TABLE 6 shows the average throughput for transactions with and without trades ----- **TABLE 6. Average throughput (with and without trades) for different network sizes.** **TABLE 7. Trading data for Singapore exchange for the month of** April 2020. **FIGURE 9. Throughput at 350 tx/sec Vs. number of validators.** **FIGURE 10. Latency Vs number of validators.** for different network sizes. Our finding shows that our system can support networks up to 10 validators and transaction rates up to 350 tx/sec. We plot in FIGURE 9 the result of the experiment when considering the two previously defined workload scenarios and their average value at a transaction rate of 350 tx/sec. The worst throughput is noticed for the case of the first workload scenario (with trade) as it requires more computation resources to complete the trade. For a network with up to 10 validating nodes, the average throughput is about 311.8 tx/sec. It is equivalent to 89% of the optimal throughput, which is the ratio of the average throughput value (311.8 tx/sec) to the optimal throughput value (350 tx/sec). FIGURE 10 illustrates the effect of the different sending rates and the number of validators on the average transaction latency. The results show that the latency is inversely proportional to the throughput. For a network with up to 10 validating nodes, and a sending rate up to 350 tx/sec, the average latency is about 5.5 seconds. Considering the block time, which is 2 seconds and the time needed to propagate the block in the network, this can be considered as a reasonable delay. However, the latency significantly increases for large network sizes and high sending rates, where it can reach 40 seconds. It can be related to the following two reasons : 1) The higher the sending rate, the larger the block size, and hence, it will require more time to propagate the block to all the nodes in the network. 2) The computational resources play a major role in the network’s ability to handle high sending rates. This is due to the fact that each transaction needs to go into several steps such as validation, propagation to the network, execution, and its inclusion in a new block that will then be propagated again to the network to be executed by the other nodes. These steps require sufficient computational power to be able to handle high sending rates. The obtained results have shown that our proposed trading platform can reach a transaction throughput of about 311.8 tx/sec. By analyzing the trading data obtained from Singapore Stock Exchange during the month of April 2020 [26], TABLE 7 shows that the total number of performed trades is 10,285,596 for a period of 21 trading days with 7 trading hours each day, which results in 489790.2857 trades per day. If all these trades are to be processed by the platform within the same two hours of a certain day, we will have 244895.1429 trades per hour, which results in 68.02642857 trade per second. Since each generated trade consists of buy and sell orders matched together, the estimated total number of generated transactions per second is 3 times the total number of trades/sec, which is equivalent to 204.0792 tx/sec. It is clear that our proposed platform can easily meet the requirement of this market by only considering the available computation resources used during the experiment. We believe that increased performance could be achieved if more computation resources can be used during the experimental evaluation. ----- **VII. CONCLUSION** In this paper, we have presented a new blockchain-based architecture for a fully decentralized stock market platform. Our architecture is based on Ethereum smart contract that is implemented on a consortium and permissioned network. To be aligned with the regulations of the stock market, we chose the validating nodes to be the financial and governmental organizations that are already involved in the traditional stock exchange platform. This new architecture addresses the limitations of the traditional stock exchange platform such as the single point of failure in the participating systems by replicating the data and smart contract across all participating nodes, the complexity and inefficiency of the data management which our solution solves by providing a shared ledger that can be easily updated and maintained, the limited level of transparency since now all transactions can be seen, the limited daily time to access the platform’s data as now it is easier to monitor the blockchain and access it throughout the day, and offering a faster financial and cash settlement time instead of the three days needed after the trading session. In order to evaluate the performance of our system, several experiments were conducted where the throughput and latency were evaluated. We have used different workloads and network sizes to evaluate the performance and found that the achieved performance can meet the requirement of the stock market platform for network sizes up to 10 validators and up to a sending rate of 350 tx/sec. However, we found that for larger workloads or network sizes, the performance significantly declines due to the limited computational resources used in the experiment. However, since the proposed solution will run on a consortium permissioned network, we believe that the participating entities will be capable of accommodating the necessary computation resources in order to meet the latency and throughput levels of the stock exchange. We plan to conduct further study to address privacy-related concerns and include cryptography encryption in the same ledger such that only allowed participants can see their relevant transaction data. Our future work will also cover further enhancements in the proposed smart contract. For instance, we will cover the possibility of introducing new changes to an already deployed smart contract without causing disturbance to the overall stock exchange platform. **REFERENCES** [1] B. Comincioli, ‘‘The stock market as a leading indicator: An application of Granger causality,’’ Univ. Avenue Undergraduate J. Econ., vol. 1, no. 1, pp. 1–14, 1996. [2] M. S. Nazir, M. Nawaz, and U. Gilani, ‘‘Relationship between economic growth and stock market development,’’ Afr. J. Bus. Manage., vol. 4, pp. 3473–3479, Dec. 2010. [3] C. Pop, C. Pop, A. Marcel, A. Vesa, T. Petrican, T. Cioara, I. Anghel, and I. Salomie, ‘‘Decentralizing the stock exchange using blockchain an ethereum-based implementation of the bucharest stock exchange,’’ in Proc. _IEEE 14th Int. Conf. Intell. Comput. Commun. Process. (ICCP), Sep. 2018,_ pp. 459–466. [4] N Inc. (2019). Trading and Matching Technology Provides Flexible, _Multi-Asset_ _Trading_ _Capabilities_ _for_ _Marketplaces_ _of_ _all_ _Sizes._ [Online]. Available: https://www.nasdaq.com/solutions/tradingand-matching-technology [5] L. Lee, ‘‘New kids on the blockchain: How Bitcoin’s technology could reinvent the stock market,’’ SSRN Electron. J., vol. 12, no. 2, pp. 81–132, 2016. [6] V. V. Bhandarkar, A. A. Bhandarkar, and A. Shiva, ‘‘Digital stocks using blockchain technology the possible future of stocks?’’ Int. J. Manage., vol. 10, no. 3, pp. 44–49, Jun. 2019. [7] T. Ahram, A. Sargolzaei, S. Sargolzaei, J. Daniels, and B. Amaba, ‘‘Blockchain technology innovations,’’ in Proc. IEEE Technol. Eng. Man_age. Conf. (TEMSCON), Jun. 2017, pp. 137–141._ [8] M. Samaniego, U. Jamsrandorj, and R. Deters, ‘‘Blockchain as a service for IoT,’’ in Proc. IEEE Int. Conf. Internet Things (iThings) _IEEE_ _Green_ _Comput._ _Commun._ _(GreenCom)_ _IEEE_ _Cyber,_ _Phys._ _Social Comput. (CPSCom) IEEE Smart Data (SmartData), Dec. 2016,_ pp. 433–436. [9] T. Lundqvist, A. de Blanche, and H. R. H. Andersson, ‘‘Thing-to-thing electricity micro payments using blockchain technology,’’ in Proc. Global _Internet Things Summit (GIoTS), Jun. 2017, pp. 1–6._ [10] Y.-H. Chen, S.-H. Chen, and I.-C. Lin, ‘‘Blockchain based smart contract for bidding system,’’ in Proc. IEEE Int. Conf. Appl. Syst. Invention (ICASI), Apr. 2018, pp. 208–211. [11] A. Dorri, S. S. Kanhere, and R. Jurdak, ‘‘Towards an optimized BlockChain for IoT,’’ in Proc. 2nd Int. Conf. Internet-of-Things Design Implement., Apr. 2017, pp. 173–178. [12] M. Conoscenti, A. Vetro, and J. C. De Martin, ‘‘Blockchain for the Internet of Things: A systematic literature review,’’ in Proc. _IEEE/ACS 13th Int. Conf. Comput. Syst. Appl. (AICCSA), Nov. 2016,_ pp. 1–6. [13] S. Wang, L. Ouyang, Y. Yuan, X. Ni, X. Han, and F.-Y. Wang, ‘‘Blockchainenabled smart contracts: Architecture, applications, and future trends,’’ _IEEE Trans. Syst., Man, Cybern. Syst., vol. 49, no. 11, pp. 2266–2277,_ Nov. 2019. [14] A. Kaushik, A. Choudhary, C. Ektare, D. Thomas, and S. Akram, ‘‘Blockchain—Literature survey,’’ in _Proc._ _2nd_ _IEEE_ _Int._ _Conf._ _Recent Trends Electron., Inf. Commun. Technol. (RTEICT), May 2017,_ pp. 2145–2148. [15] H. Kuzuno and C. Karam, ‘‘Blockchain explorer: An analytical process and investigation environment for bitcoin,’’ in Proc. APWG Symp. Electron. _Crime Res. (eCrime), Apr. 2017, pp. 9–16._ [16] M. Salimitari and M. Chatterjee, ‘‘A survey on consensus protocols in blockchain for IoT networks,’’ Sep. 2018, arXiv:1809.05613. [Online]. Available: https://arxiv.org/abs/1809.05613 [17] S. D. Angelis, L. Aniello, R. Baldoni, F. Lombardi, A. Margheri, and V. Sassone, ‘‘Pbft vs proof-of-authority: Applying the cap theorem to permissioned blockchain,’’ in Proc. Italian Conf. Cyber Secur., Jan. 2018, p. 11. [Online]. Available: https://eprints.soton.ac.uk/415083/ [18] T. T. A. Dinh, R. Liu, M. Zhang, G. Chen, B. C. Ooi, and J. Wang, ‘‘Untangling blockchain: A data processing view of blockchain systems,’’ IEEE Trans. Knowl. Data Eng., vol. 30, no. 7, pp. 1366–1385, Jul. 2018. [19] D. Ongaro and J. Ousterhout, ‘‘In search of an understandable consensus algorithm,’’ in Proc. USENIX Conf. USENIX Annu. Tech. Conf. (USENIX _ATC). Berkeley, CA, USA: USENIX Association, 2014, pp. 305–320._ [Online]. Available: http://dl.acm.org/citation.cfm?id=2643634.2643666 [20] J. Jaoude and R. Saade, ‘‘Blockchain applications—Usage in different domains,’’ IEEE Access, vol. 7, pp. 45372–45373, 2019, doi: [10.1109/ACCESS.2019.2902501.](http://dx.doi.org/10.1109/ACCESS.2019.2902501) [21] G. William Peters and E. Panayi, ‘‘Understanding modern banking ledgers through blockchain technologies: Future of transaction processing and smart contracts on the Internet of money,’’ 2015, arXiv:1511.05740. [Online]. Available: http://arxiv.org/abs/1511.05740 [22] Q. K. Nguyen, ‘‘Blockchain–A financial technology for future sustainable development,’’ in Proc. 3rd Int. Conf. Green Technol. Sustain. Develop. _(GTSD), Nov. 2016, pp. 51–54._ [23] S. Singh and N. Singh, ‘‘Blockchain: Future of financial and cyber security,’’ in Proc. 2nd Int. Conf. Contemp. Comput. Informat. (IC3I), Dec. 2016, pp. 463–467. [24] S. Rouhani and R. Deters, ‘‘Security, performance, and applications of smart contracts: A systematic survey,’’ IEEE Access, vol. 7, pp. 50759–50779, 2019. [25] M. Demir, M. Alalfi, O. Turetken, and A. Ferworn, ‘‘Security smells in smart contracts,’’ in Proc. IEEE 19th Int. Conf. Softw. Qual., Rel. Secur. _Companion (QRS-C), Jul. 2019, pp. 442–449._ [26] SGX. (Apr. 2020). Market Statistics Report. [Online]. Available: https://www2.sgx.com/research-education/historical-data/marketstatistics ----- HAMED AL-SHAIBANI received the B.Sc. degree (Hons.) in computer science from Qatar University, Doha, Qatar, in 2010, and the M.Sc. degree in strategic business unit management from HEC Paris, Doha, in 2016. He is currently pursuing the Ph.D. degree in computer science and engineering with Hamad Bin Khalifa University, Doha. His main research interests include blockchain, cybersecurity, and networking. NOUREDDINE LASLA (Member, IEEE) received the B.Sc. degree from the University of Science and Technology Houari Boumediene (USTHB), in 2005, the M.Sc. degree from the Superior Computing National School (ESI), in 2008, and the Ph.D. degree from USTHB, in 2015, all in computer science. He is currently a Postdoctoral Research Fellow with the Division of Information and Computing Technology, Hamad Bin Khalifa Univeristy, Qatar, with expertise in distributed systems, network communication, and cyber security. MOHAMED ABDALLAH (Senior Member, IEEE) received the B.Sc. degree from Cairo University, in 1996, and the M.Sc. and Ph.D. degrees from the University of Maryland at College Park, in 2001 and 2006, respectively. From 2006 to 2016, he held academic and research positions at Cairo University and Texas A&M University at Qatar. He is currently a Founding Faculty Member with the rank of Associate Professor with the College of Science and Engineering, Hamad Bin Khalifa University (HBKU). His current research interests include wireless networks, wireless security, smart grids, optical wireless communication, and blockchain applications for emerging networks. He has published more than 150 journals and conferences and four book chapters, and co-invented four patents. He was a recipient of the Research Fellow Excellence Award at Texas A&M University at Qatar, in 2016, the Best Paper Award in multiple IEEE conferences including the IEEE BlackSeaCom 2019, the IEEE First Workshop on Smart Grid and Renewable Energym in 2015, and the Nortel Networks Industrial Fellowship for five consecutive years, from 1999 to 2003. His professional activities include an Associate Editor of the IEEE TRANSACTIONS ON COMMUNICATIONS and the IEEE OPEN ACCESS JOURNAL OF COMMUNICATIONS, a Track Co-Chair of the IEEE VTC Fall 2019 conference, a Technical Program Chair of the 10th International Conference on Cognitive Radio Oriented Wireless Networks, and a Technical Program Committee Member of several major IEEE conferences. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/ACCESS.2020.3005663?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ACCESS.2020.3005663, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://ieeexplore.ieee.org/ielx7/6287639/8948470/09127940.pdf" }
2,020
[ "JournalArticle" ]
true
2020-06-29T00:00:00
[ { "paperId": "23a1cec389cf82955a8d3046c5aa1bf25f6f6584", "title": "Blockchain in Internet of Things: a Systematic Literature Review" }, { "paperId": "089f26359d7de562b9564e94e1c26b3bf2661c2b", "title": "Security Smells in Smart Contracts" }, { "paperId": "58067fa88a88b60130c1a432808c14c1a975ab21", "title": "Digital Stocks Using Blockchain Technology the Possible Future of Stocks?" }, { "paperId": "12c37ed1451eb3279bac8cf83c7adaf7a1032b43", "title": "Security, Performance, and Applications of Smart Contracts: A Systematic Survey" }, { "paperId": "659dd200ab51398295b55c7ea1779684d21d28c7", "title": "Blockchain Applications – Usage in Different Domains" }, { "paperId": "5d6891f8efc2f5607ae69fefd343239f447fe602", "title": "A Survey on Consensus Protocols in Blockchain for IoT Networks" }, { "paperId": "8bef3ecaaaadacf6316ce30275b98b40b6800bd5", "title": "Decentralizing the Stock Exchange using Blockchain An Ethereum-based implementation of the Bucharest Stock Exchange" }, { "paperId": "e476438e555b24130e70c73f4529c26c45a2049c", "title": "Blockchain based smart contract for bidding system" }, { "paperId": "82d2b9d09cc339fdeac05abfb8a31f9c6eace948", "title": "Untangling Blockchain: A Data Processing View of Blockchain Systems" }, { "paperId": "9092a7802f6e56dd5b6d1be30c8b5588a22e53fe", "title": "Blockchain technology innovations" }, { "paperId": "5eda01b02535084687d52e03a013af294ffdf1b5", "title": "Thing-to-thing electricity micro payments using blockchain technology" }, { "paperId": "8124e1c470957979625142b876acad033bdbe69f", "title": "Blockchain — Literature survey" }, { "paperId": "59230229b0a6eb52c272d92d2392f9735997a350", "title": "Blockchain explorer: An analytical process and investigation environment for bitcoin" }, { "paperId": "61d69925287bd0e3adea7d6fe9ccaffaf29207cd", "title": "Towards an Optimized BlockChain for IoT" }, { "paperId": "fa5e83b4830e11c612623c0a623b94cdaacc4619", "title": "Blockchain: Future of financial and cyber security" }, { "paperId": "631cc57858eb1a94522e0090c6640f6f39ab7e18", "title": "Blockchain as a Service for IoT" }, { "paperId": "bfbce9ae7fd2828c7ca6ecbbe6c46ddc7d5e3e75", "title": "Blockchain - A Financial Technology for Future Sustainable Development" }, { "paperId": "0a84a077ada2acb6918e7764fafcd28f667dae28", "title": "Understanding Modern Banking Ledgers Through Blockchain Technologies: Future of Transaction Processing and Smart Contracts on the Internet of Money" }, { "paperId": "9979809e4106b29d920094be265b33524cde8a40", "title": "In Search of an Understandable Consensus Algorithm" }, { "paperId": "32bf65155f54914b6f42673ef732970ffcb88b33", "title": "Relationship between economic growth and stock market development" }, { "paperId": "5e2bdde4f4cdeec63594c644d63cda52838f08c2", "title": "Who Trades?" }, { "paperId": null, "title": "‘‘Blockchain-enabledsmartcontracts:Architecture,applications,andfuturetrends,’’" }, { "paperId": null, "title": "Trading and Matching Technology Provides Flexible, Multi-Asset Trading Capabilities for Marketplaces of all Sizes" }, { "paperId": "4052d479c69c9b424df4f352d61cb5b4df85f352", "title": "PBFT vs Proof-of-Authority: Applying the CAP Theorem to Permissioned Blockchain" }, { "paperId": "b3578884b80376bcf792b993c9769761f32c41bf", "title": "New Kids on the Blockchain: How Bitcoin's Technology Could Reinvent the Stock Market" }, { "paperId": "742020bb58065c8712d5f9236cefadecfe8b9c1a", "title": "The Statistics Report" }, { "paperId": "e24148556696b1366472879301a549ebb29a25d9", "title": "The Stock Market as a Leading Indicator: An Application of Granger Causality" }, { "paperId": null, "title": "We assume that all orders entered are of the types limit order or market order and that partially matched orders are possible in cases where B Q (cid:54)= A Q" }, { "paperId": null, "title": "Buy orders are entered by the broker into the smart contract" }, { "paperId": null, "title": "CSD assigns to each validated investor a new investor account number ‘‘NIN’’ by using the function ‘‘addNin’’" }, { "paperId": null, "title": "Let QA be a queue of all sell orders sorted in descending order such that y represents the index of the minimum element in the queue denoted as A P" }, { "paperId": null, "title": "CSD Validates the investor’s data integrity by sending it to the government" }, { "paperId": null, "title": "AssociateBrokerToInevestor" }, { "paperId": null, "title": "FMA defines the companies that are listed in the stock market along with their details such as number of shares they consist of and their prices" }, { "paperId": null, "title": "a buy order that has the company’s symbol, number of shares, price, and NIN enters a queue of buy orders" }, { "paperId": null, "title": "AssignShareToNin" }, { "paperId": null, "title": "If B P ≥ A P and B Q ≥ A Q , B P is partially matched with A P , and a trade is generated" } ]
15,005
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00e65f396b4febdc799e03c997608fbeb1035905
[ "Computer Science" ]
0.878881
Achlys: Towards a Framework for Distributed Storage and Generic Computing Applications for Wireless IoT Edge Networks with Lasp on GRiSP
00e65f396b4febdc799e03c997608fbeb1035905
2019 IEEE International Conference on Pervasive Computing and Communications Workshops (PerCom Workshops)
[ { "authorId": "66031965", "name": "Igor Kopestenski" }, { "authorId": "39874026", "name": "P. V. Roy" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
Internet of Things (IoT) continues to grow exponentially, in number of devices and the amount of data they generate. Processing this data requires an exponential increase in computing power. For example, aggregation can be done directly at the edge. However, aggregation is very limited; ideally we would like to do more general computations at the edge. In this paper we propose a framework for doing general-purpose edge computing directly on sensor networks themselves, without requiring external connections to gateways or cloud. This is challenging because sensor networks have unreliable communication, unreliable nodes, and limited (if any) computing power and storage. How can we implement production-quality components directly on these networks? We need to bridge the gap between the unreliable, limited infrastructure and the stringent requirements of the components. To solve this problem we present Achlys, an edge computing framework that provides reliable storage, computation, and communication capabilities directly on wireless networks of IoT sensor nodes. Using Achlys, the sensor network is able to configure and manage itself directly, without external connectivity. Achlys combines the Lasp key/value store and the Partisan communication library. Lasp provides efficient decentralized storage based on the properties of CRDTs (Conftict-Free Replicated Data Types). Partisan provides efficient connectivity and broadcast based on hybrid gossip. Both Lasp and Partisan are specifically designed to be extremely resilient. They are able to continue working despite high node churn, frequent network partitions, and unreliable communication. Our first implementation of Achlys is on a network of GRiSP embedded system boards. We choose GRiSP as our first implementation platform because it implements high-level functionality, namely Erlang, directly on the bare hardware and because it directly supports Pmod sensors and wireless connectivity. We give some first results on using Achlys for building edge systems and we explain how we plan to evolve Achlys in the future. Achlys is a work in progress that is being done in the context of the LightKone European H2020 research project, and we are in the process of implementing and evaluating a proof-of-concept application in the area of precision agriculture.
# Achlys : Towards a framework for distributed storage and generic computing applications for wireless IoT edge networks with Lasp on GRiSP ### Igor Kopestenski ICTEAM Institute Universit´e catholique de Louvain igor.kopestenski@uclouvain.be **_Abstract—Internet of Things (IoT) continues to grow expo-_** **nentially, in number of devices and the amount of data they** **generate. Processing this data requires an exponential increase** **in computing power. For example, aggregation can be done** **directly at the edge. However, aggregation is very limited; ideally** **we would like to do more general computations at the edge.** **In this paper we propose a framework for doing general-** **purpose edge computing directly on sensor networks themselves,** **without requiring external connections to gateways or cloud.** **This is challenging because sensor networks have unreliable** **communication, unreliable nodes, and limited (if any) computing** **power and storage. How can we implement production-quality** **components directly on these networks? We need to bridge** **the gap between the unreliable, limited infrastructure and the** **stringent requirements of the components. To solve this problem** **we present Achlys, an edge computing framework that provides** **reliable storage, computation, and communication capabilities** **directly on wireless networks of IoT sensor nodes. Using Achlys,** **the sensor network is able to configure and manage itself** **directly, without external connectivity. Achlys combines the Lasp** **key/value store and the Partisan communication library. Lasp** **provides efficient decentralized storage based on the proper-** **ties of CRDTs (Conflict-Free Replicated Data Types). Partisan** **provides efficient connectivity and broadcast based on hybrid** **gossip. Both Lasp and Partisan are specifically designed to be** **extremely resilient. They are able to continue working despite** **high node churn, frequent network partitions, and unreliable** **communication. Our first implementation of Achlys is on a** **network of GRiSP embedded system boards. We choose GRiSP** **as our first implementation platform because it implements high-** **level functionality, namely Erlang, directly on the bare hardware** **and because it directly supports Pmod sensors and wireless** **connectivity. We give some first results on using Achlys for** **building edge systems and we explain how we plan to evolve** **Achlys in the future. Achlys is a work in progress that is being** **done in the context of the LightKone European H2020 research** **project, and we are in the process of implementing and evaluating** **a proof-of-concept application in the area of precision agriculture.** I. INTRODUCTION The edge computing paradigm has been widely accepted as an important concept for sustainability of future Cloud Service Providers (CSPs) and Mobile Network Operators (MNOs) [1], [2]. It is well acknowledged by both enterprise and academia as a valid approach and is actively under research [3], [4], [5]. Newer and more performant infrastructures are continuously ### Peter Van Roy ICTEAM Institute Universit´e catholique de Louvain peter.vanroy@uclouvain.be elaborated both by CSPs and MNOs [6]. Concurrently, IoT devices are getting closer to being actually ubiquitous, i.e., closer to Mark Weiser’s idea of hundreds of wireless comput_ing devices per person. This is already true in some scenarios,_ e.g., airplanes generate around 10 TB of data every 30 minutes. Such cases require very responsive and robust systems for sensor data processing, and could not rely on remote hosts for it, even if these are close to the edge. Since the Internet of Things is rapidly expanding and devices are becoming more powerful, IoT applications are putting severe strain on cloud providers. The edge computing paradigm is one way to solve this problem by distributing the workload in a more sustainable way across the whole network. With this paradigm, computational and storage resources move closer to the edge and IoT applications are able to preserve their QoS. Tasks that were previously done in the cloud are now be delegated to intermediates between datacenters and IoT edge networks. Existing designs are enabling this by bridging edge networks with intermediate gateways[7]. However, it is generally considered that the sensor and actuator networks themselves, such as traditional Wireless Sensor Networks (WSNs), are too limited and unreliable to do their own management. Thus even in recent distributed sensor and IoT networks, a gateway node or a cloud connection is necessary, which adds a single point of failure and increases infrastructure complexity. If such a point is unable to provide its service, network management becomes impossible and sensor data cannot be retrieved anymore. And if gateways need to be permanently available, even short intervals of downtime can disrupt the entire system. A recent survey has portrayed the full landscape of emerging paradigms that strive towards the edge and fog principles[8]. It describes a few designs that share common goals with Achlys, such as mist computing and suggests that such architectures are a best fit for systems that require autonomous behavior, ability for distributed processing directly on IoT devices, little or no Internet connectivity and privacy preservation. _A. The Achlys framework_ This paper presents Achlys, a framework that directly addresses the problem of general-purpose edge computing. ----- Achlys increases the resilience of sensor/actuator edge networks so that they are able to reliably execute application tasks directly on the edge nodes themselves. Achlys provides reliable decentralized communication, storage, and computation abilities, by leveraging CRDTs (Conflict-Free Replicated Data Types) and hybrid gossip algorithms. This lowers cost, reduces dependencies, and simplifies maintenance. Our system has no single point of failure. Achlys consists of three parts: GRiSP embedded system boards[1], a Lasp CRDT-based key/value store, and the Partisan hybrid gossip-based communication component. Experimental releases and source code are available at ikopest.me, and can be deployed on GRiSP boards or run in an Erlang test shell. Achlys adds a task model to Lasp, which allows applications to be written by storing both their code and results directly inside Lasp. In this way, applications are as resilient as Lasp itself. The task model was first developed as part of a master’s thesis [9]. Application code is replicated automatically by Lasp on all nodes. From the developer’s viewpoint, Achlys applications are written in a similar way as applications written for transactional databases. The developer mindset is that the Lasp database always contains correct data. Application tasks can be executed at any time on any node by the task model. On every node, the task model periodically reads tasks from Lasp and executes them. An executing task reads data from Lasp, computes the updated data using node-specific sensor information, actuates node-specific actuators, and stores both the updated data and an updated task in Lasp. Because of the convergence properties of CRDTs, the same task can be executed more than once on different nodes without affecting correctness. This is a necessary condition for resilience, to keep running despite node failures and node turnover. We choose GRiSP because it directly implements Erlang on the bare hardware, which simplifies system development, and because it directly supports Pmod sensors and actuators and has built-in wireless connectivity. Computation and storage abilities of GRiSP are limited, but adequate for many management tasks (in fact, with the current GRiSP infrastructure, our system is comparable to state-of-the-art laptops released just before the year 2000). Our current system is a prototype that is able to run applications on networks of GRiSP boards. With this system we are in the process of implementing and evaluating a proof-of-concept application of Achlys to precision agriculture, in collaboration with Gluk Advice BV, as part of the LightKone European H2020 research project. _B. Motivating example_ We motivate this work with an example related to the precision agriculture use case we are developing. We envision a scenario where a farmer has decided to equip his farms with a Subsurface Drip Irrigation system. Despite that it is one of the most efficient precision irrigation technologies, it remains difficult to have sufficiently precise information about moisture levels, so as to make optimal use of the very 1grisp.org expensive irrigation infrastructure. Inadequate irrigation can be extremely detrimental for production levels, as well as water supplies. We propose a Self-Sufficient Precision Agriculture Management System for Irrigation in order to allow the farmer to efficiently irrigate his farms. In addition to productivity gains, we intend to offer a solution that is near zeroconfig, i.e., it is able to configure itself with no intervention needed by the farmer. Finally, we want to provide a system that manages itself and requires little to no maintenance. It is possible to connect to the system, but this is only used for setting policy and is not needed for day-to-day running. The system’s management mechanisms are autonomous, independent of any third-party providers. Currently, the farmer’s irrigation system distributes water across the entire zone that is equipped with tubes delivering water to plantations when the farmer activates a pump. Since moisture levels can vary significantly inside a single farm, uniform irrigation can be detrimental as some parts will not receive enough water while in other parts water is wasted and irrigation is above the optimal levels. Our system is made of a set of distributed edge nodes that sense moisture levels and actuate on activation or deactivation of underground irrigation tubes. The farm is divided into sectors whose moisture levels are measured by an edge node, and irrigation is adjusted accordingly. Our solution activates the main irrigation pump and valves when necessary, and controls the water flow such that once sufficient moisture is achieved, actuators shut down the water flow of that sector while irrigation continues in other sectors. Irrigation decisions are made by an online optimization algorithm that runs continuously on the edge nodes themselves. The system thus provides completely autonomous basic management, without the need for any kind of Internet connection or computer. The system continuously optimizes its operation to provide adequate irrigation with minimal cost, and reconfigures itself whenever it detects a change in configuration. It is possible to change the irrigation policy by connecting a PC node to the edge network. This way, we extend the basic autonomous system with additional features that allow the farmer to use cloud infrastructures such as storage or high computational power when it is desired. For example, the edge nodes could be asked to measure how much water their sector has consumed based on temperatures during that period, and compute some metrics locally that could be extracted from the edge cluster and stored in the cloud at the end of each month. Learning processes applied to that data could again allow farmers to adjust the system behavior with their computer to gain in efficiency. _C. Common challenges_ In contrast to core cloud datacenters, edge networks are composed of large numbers of heterogeneous devices. Due to the highly dynamic and unpredictable nature of edge network topologies, nodes can also be temporarily isolated from the network. For these reasons, implementing desirable features ----- such as reliable computation and storage directly on edge IoT networks is particularly complex. Possible solutions are proposed by industry actors such as 5G operators and CSPs, and are generally based on Points of Presence (POPs) located near client nodes and available through gateways. In addition, edge applications must be implemented to manage the limited resources of IoT nodes. Therefore efficient deployments at the edge require an adequate load balancing mechanism that ensures that there is no overload on any node. An important goal of edge computing is to offload efficiently the core of the network. Since there are several intermediary entities between cloud datacenters and IoT devices such as servers or smaller datacenters, optimal offloading would be achieved if components of each layer are able to process some requests autonomously and only rely on higher level nodes when necessary. Therefore, edge nodes should also strive for maximum independence, and take advantage of computational and storage resources of IoT devices to complement the edge POP solutions. Moreover, if edge computing extends the traditional cloud computing paradigm only up to peripheral POPs, it makes IoT nodes highly dependent on connectivity and exposes single points of failure. We suggest that the edge paradigm can be implemented in a way that allows offloading of the core at any level in the global network, even in the most peripheral parts. If IoT devices are able to provide some basic functionality, then higher level devices can rely on them to reduce their own workload. And the edge paradigm could therefore maximize the global offloading since it would distribute the load over all the parts of the edge that are able to perform tasks such as computations or data storage in a reliable way accordingly to their hardware resources. However, despite being standardized to some extent[10], [11], a global production ready end-to-end solution has not yet been deployed at scales coming close to those of traditional cloud architectures. There are still many engineering and practical considerations that must be addressed. In this regard, the LightKone H2020 European Project aims at providing a novel approach for general purpose computations at the edge. LightKone directly addresses the added complexity due to heterogeneity of IoT devices, which makes a general purpose computation model at the edge very attractive. _D. Structure of the article_ The remainder of this article is structured as follows. Section II gives a brief overview of current edge computing state of the art and some key enabling technologies for Achlys. Section III gives a structural overview of Achlys followed by use case examples in Section IV. Finally, Section V gives conclusions about the current state and future evolution of Achlys. II. CONTRIBUTIONS In this section, we briefly discuss the contribution of the Achlys application framework in relationship with the global edge computing paradigm. _A. Fault tolerance_ Ensuring fault tolerance is an essential part for generic edge computing[12]. In order to fit the vision of the LightKone project, Achlys strives to guarantee this property. This implies that Achlys must be able to continue functioning even in case of system failure. These failures can be, but are not limited to : _• Network partition or intermittent communication : a_ node or a set of nodes that are isolated from the rest of the network must be able to run and to preserve interoperability with other nodes when the network is repaired. _• Hardware failure or offline operation : if a hardware_ component becomes dysfunctional or goes offline (to save power), it should be contained so that the application preserve a maximum amount of features. _B. Task model_ Achlys provides a general purpose task model solution using Erlang higher-order functions. Since Erlang functions are just values (i.e., constants), they can be copied over a network like any other constant. Using this ability, Achlys is able to provide programmers with an API that allows them to easily disseminate generic tasks in a cluster and be able to retrieve the results if desired. As handling heterogeneity is a highly complex task for smart services at the edge[13], [14], [15], the Achlys prototype can also use this ability to make the task model homogeneous, despite heterogeneity of the infrastructure. This is compatible with a larger vision of future Internet, in which physical components will be virtualized[16], [17], [18] _C. Data consistency at the edge_ Conflict resolution is one of the central problems of distributed and decentralized applications, and is the subject of extensive research[19], [20], [21], [22]. For example, when multiple actors modify the same data entity across a network partition, what should be done when the partition is repaired? CRDTs (Conflict-Free Replicated Data Types) provide a solution to this problem. They are mathematically designed to provide consistent replication with very weak synchronization between replicas: only eventual replica-to-replica communication is needed. The Lasp library uses a wide range of CRDT types for its data storage. To the developer, Lasp looks like a replicated distributed key/value store that runs directly on the IoT network. Section III-C gives more information on Lasp and CRDTs. III. OVERVIEW OF THE SYSTEM DESIGN We now present a more detailed description of the Achlys system, an Erlang[2] implementation of a framework that combines the power of δ-CRDTs[22] in the Lasp store[23], [24], the Partisan communication component[25], and the GRiSP Runtime software. It provides application developers a way to 2erlang.org ----- build resilient distributed edge IoT applications. We leverage hybrid gossiping and the use of CRDTs in order to propose a platform that is able to provide reliable services directly on edge nodes, which are able to function autonomously even when no gateway or Internet access is available. _A. GRiSP base_ The GRiSP base board is the embedded system used to deploy Achlys networks in the current experimental phase. Its main advantage over other hardware is that it has sufficient resources[26] to run relevant Erlang applications, that is[3] : _• Microcontroller : Atmel SAM V71, including :_ **– ARM Cortex M7 CPU clocked at 300MHz** **– 64 MBytes of SDRAM** **– A MicroSD socket** _• 802.11b/g/n wireless antenna_ _• SPI, GPIO, 1-Wire and UART interfaces_ _B. GRiSP_ Figure 1 depicts how the GRiSP architecture is designed. The RTEMS[4] (RTOS-like set of libraries) component is embedded inside the Erlang VM and makes it truly run on bare _metal. Achlys greatly benefits from this unique design since_ it allows a much more direct interaction with the GRiSP base hardware. The GRiSP board directly supports Digilent Pmod[5] modules. The latter offer a very wide range of sensing and actuating features that can be accessed at application level in Erlang. It is not necessary to write drivers in C in order to add new hardware features to extend the range of functionalities. Fig. 1. The GRiSP software stack compared to traditional designs. The hardware layer for GRiSP refers to the GRiSP base board that is currently available. Reprinted from GRiSP presentation by Adam Lindberg. 3 For full specifications please refer to grisp.org. 4rtems.org 5digilentinc.com _C. Lasp_ Lasp is a key part of the Achlys framework for both storage and computation. It provides both replicated data and computation, and guarantees that values will eventually converge on all nodes[23], [24]. Since our GRiSP boards run Erlang directly on bare metal, Lasp is a suitable option for consistency as it runs directly in Erlang[24]. The Lasp port to GRiSP was initiated in a master’s thesis at UCL[9]. Lasp supports various CRDTs including sets, counters and registers, all of which do consistent conflict resolution. For instance, a GCounter is a counter type that can only be incremented, and when all the operations performed on individual nodes converge, the entire cluster will see the same value, i.e., the sum of all increments of the counter across all nodes. This is a very simple example; more complex examples such as sets and dictionaries are also part of Lasp, which allow both adding and removing of elements with the same convergence property. Achlys is thus able to handle concurrent modifications and guarantee that all nodes are eventually consistent, just as on cloud storage services. This works even on nodes with limited resources and intermittent connectivity. The only effect of node failures and intermittent connectivity is to slow down the convergence. CRDTs in Lasp are implemented using additional metadata that allows each operation at each node to be taken into consideration. In fact, the Lasp library uses an efficient implementation of CRDTs called delta-based dissemination mode, which propagates only delta-mutators[27], [22], i.e., update operations, instead of the full state, to achieve consistency. This uses significantly less traffic between nodes than a naive implementation that propagates the full state. _D. Partisan_ Partisan[25] is the communication component used by Lasp to disseminate information between nodes. It provides a highly resilient alternative communication layer used instead of the default distributed Erlang communication. This layer combines the HyParView[28] membership and Plumtree[29] broadcast algorithms, to ensure both connectivity and communication, even in extremely dynamic and unreliable environments. Both algorithms use the hybrid gossip approach. Hybrid gossip is a sweet spot that combines the efficiency of standard distributed algorithms (e.g., spanning tree broadcast in Plumtree) with the resilience of gossip. For example, in the case of Plumtree, the gossip algorithm is used to repair the spanning tree. Partisan comes with a set of configurable peer service modules that are each suited for different types of networks. Since the HyParView manager ensures reliable communication in networks with high attrition rates such as edge clusters, it is used in our configuration by default. But it can easily be adjusted to match other types of topologies, and also enables hosts that can be members of multiple clusters to use the optimal peer service. This makes it suitable for clusters of IoT nodes that are able to communicate with each other despite unreliable networks, but also to be able to communicate with other cluster types such as star or mesh topologies where reliable communication does not require the same amount of ----- DEPLOYED ACHLYS CLUSTER PHYSICAL OVERLAY Fig. 2. An example of an Achlys network measuring temperature and pressure. effort. Partisan is therefore flexible as well as resilient, and we are still able to configure edge nodes to communicate with servers, reliable clients or gateways without overhead that would be generated using HyParView in stable networks. _E. Example Achlys network_ Figure 2 shows a conceptual overview of a WSN setup of Achlys nodes that highlights several elements : _• The_ bottom layer consists of functions _temp1,2(∆, m),p(∆, m) that represent input streams of_ data based on environment variables measured by the nodes, in this example temperatures and pressure. _• The physical topology reflects the real world configura-_ tion of the nodes where each edge implies that the two vertices are able to establish radio communication. _• The virtual overlay that we are able to build using_ Partisan. Achlys provides functions to clusterize GRiSP nodes through Partisan and therefore partitions such as shown by dotted blue lines are abstracted away by the eventual consistency and partition tolerance properties. Physically isolated parts of the network keep functioning, and seamlessly recover once the links are reestablished. _F. Local aggregation_ The vast majority of raw IoT sensor data is usually very short-lived inside systems and ultimately leads to unnecessary storage. Hence in Achlys we introduce configurable parameters for aggregation of sensor data. This way programmers can still benefit from distributed storage but also take advantage of local memory or MicroSD cards to aggregate raw measurements and propagate mean values. The network loads and global storage volume are thus decreased and overall scalability is improved. _G. Generic task model_ Achlys provides developers the ability to embed Erlang higher-order functions through a simple API as shown in Table I. This allows building applications using replicated higherorder functions. Each node receives tasks and can locally decide based on load-balancing information and destination targeting information if it needs to execute it. We used this model to generate replicated meteorological sensor data aggregations via generic functions supplied with specific tasks. A live dashboard of the currently converged view of the data was built and run on a laptop connected to the GRiSP sensor network. As long as that web client host was able to reach any node in the network, it could output its live view of the distributed storage. IV. ADDITIONAL USE CASES _A. Live IoT sensor dashboard_ Since our framework is implemented in Erlang, it is also possible to integrate it in Elixir[6] applications. Elixir is a programming language that is built on top of the Erlang runtime system and adds several very popular web development features. This makes it possible to implement a web server that runs only at the edge and that can interact with an entire Achlys cluster as soon as a single node is reachable. Our previous work has already allowed us to display a minimal version of this use case implementation in the context of the LightKone project. Figure 3 gives a screenshot of a live display of recorded magnetic field data recorded with Digilent Pmod NAV modules attached to GRiSP boards in the cluster. Achlys allows the live monitoring to be accessed from any edge node and it is guaranteed that the distributed database will be consistent regardless of the physical location as long as there is eventually a connection. This use case is a good display of the modular designs that can be implemented with Achlys. We can run isolated applications on IoT networks, add or remove nodes, and easily implement other types of custom nodes that will automatically work the existing cluster. For live monitoring of environmental variables with IoT actuators, the self-configuring and autonomous execution can be a starting point, and users are free to perform predictive analysis and machine learning y bridging the cluster with the cloud computing services. Once the learning process yields a new set of rules, the users can propagate them to all the network nodes and it will become autonomous again. _B. Internet of Battlefield Things_ The U.S. Army Research Laboratory has announced an entire new research program dedicated uniquely to IoT devices[30]. It is focused on deriving the new theoretical models and systems that will bring the key advantages in military conflicts of the following decades. The IoBT researchers display a very strong interest in some properties that are not considered in edge computing IoT architectures designed for commercial use. The authors describe the IoBT network as dynamic and ubiquitous with a high degree of pervasiveness, and self-awareness and self-configuration of networks are explicitly stated as requirements. Based on the 6elixir-lang.org TOPOLOGY ----- |Function|Arguments|Description| |---|---|---| |add_task|{Name, Targets, Fun}|Adds the task tuple {Name, Targets, Fun} to the tasks CRDT| |remove_task|Name|Removes the task named Name from the tasks CRDT| |start_task|Name|Starts the task named Name| |find_and_start_task|nil|Fetches any available task from the tasks CRDT and executes it| |start_all_tasks|nil|Starts all tasks in the tasks CRDT| TABLE I GENERIC TASK MODEL API FUNCTIONS. Fig. 3. A web client running on the edge and monitoring magnetic field sensor data from the distributed database. @ Periodical calls to find @ available tasks on each node erlang:send_after(Cycle, self(), trig) @ Handle the periodical message handle_info(trigger) -> find_and_start_task() ... @ From single node @ Create and propagate new task F = Sense(DeltaInterval, Threshold) add_task({senseTask, Destinations, F}) Fig. 4. Example usage of the task model API for listening on the tasks CRDT and running available functions. described areas of interest in IoBT, we imagine the use case of Achlys nodes in forward deployments of ground infrantry squads. These deployments can lead groups to be isolated in remote areas where no means of communication with remote operators is possible. During nights, the surroundings must be constantly watched and soldiers need to spread across areas they are exposed to higher risks once alone. If soldiers were to be equipped with a set of Achlys nodes, they would be able to maintain their cluster membership even with all the crew members on the terrain, and since the sensors can immediately produce data that can indicate distress such as heat or trembling, the real-time threat analysis can be immediately be propagated to all the members in the area. In urban combat scenarios, this is particularly desirable as individuals are often in buildings or confined spaces where they cannot see each other while Achlys would be able to propagate alarms to all the reachable members. And during all missions Achlys nodes would record success metrics, environmental variables, and network topology changes that would be stored during entire operations, and once groups return to bases they all provide the data to nodes with high computing capacities that can make predictive analysis that would be used again in future deployments and could easily be passed between units if they are in communication range. We have observed that the key features of Achlys that provide reliable communication and storage correspond to numerous requirements of IoBT since they must remain operational regardless of their environment. V. CONCLUSION We introduce Achlys, a framework for general purpose edge computing that runs directly on sensor/actuator networks with unreliable nodes, intermittent communication, and limited computation and storage resources. Our current Achlys prototype is written in Erlang and runs on a wireless ad hoc network of GRiSP sensor/actuator boards. Achlys uses the Lasp and Partisan libraries to provide reliable storage, computation, and communication. Lasp is a distributed key/value store based on CRDTs, which is used both for storage (with efficient δ-CRDTs) and as dynamic management tool to allow dissemination of general-purpose computing functions inside the network. Partisan is a communication component that provides highly resilient broadcast and connectivity for dynamic networks with intermittent connectivity, by using the Plumtree and HyParView hybrid gossip algorithms. Because it is written in Erlang, Achlys can also be used on any infrastructure based on the Erlang runtime. For example, it can be used on scalable web servers written in Elixir, because Elixir interoperates seamlessly with Erlang. Our experiments show encouraging results and validate the feasibility of our IoT edge computing model. This allows us to focus on improving efficiency and usability aspects of Achlys. In particular, further engineering work will be dedicated to minimize the resource usage of the framework and its dependencies. We intend to provide a framework that supports applications on embedded systems in actual deployments, and thus storage, computation and memory requirements of Achlys need to be carefully managed. Techniques for compact storage will be investigated such that we increase the amount of information that is passed through CRDTs while keeping the size identical. Furthermore, ----- other optimizations in terms of networking and self-adaptation will be done in order to elaborate more intelligent clustering mechanisms. This will be reported in further work and measurements of efficiency and resilience with fine-grained adjustments of Partisan’s parameters will help us implement a context-aware networking behavior for Achlys, reducing unnecessary bandwidth usage and connections. Finally, in order to reduce application size, we will study if unused modules can be excluded from the releases deployed on the embedded systems, and if compression features that are available through Erlang compiler flags can preserve the features of Achlys while leaving more space for applications developed on top. While we will keep improving Achlys, it will also be used for proof-of-concept implementations of use case scenarios in edge computing, and in particular for precision agriculture. ACKNOWLEDGMENT This work is partially funded by the LightKone European H2020 project under Grant Agreement 732505. The authors would like to thank Giorgos Kostopoulos of Gluk Advice BV for information on precision agriculture. REFERENCES [1] I.-P. Belikaidis, A. Georgakopoulos, P. Demestichas, U. Herzog, K. Moessner, S. Vahid, M. Fitch, K. Briggs, B. Miscopein, B. Okyere, and V. Frascolla, “Trends and challenges for autonomic RRM and MAC functionality for QoS provision and capacity expansions in the context of 5G beyond 6GHz,” in 2017 European Conference on Networks _and Communications, EuCNC 2017, Oulu, Finland, June 12-15, 2017._ IEEE, pp. 1–5. [2] V. Frascolla, F. Miatton, G. K. Tran, K. Takinami, A. D. Domenico, E. C. Strinati, K. Koslowski, T. Haustein, K. Sakaguchi, S. Barbarossa, and S. Barberis, “5G-MiEdge: Design, standardization and deployment of 5G phase II technologies: MEC and mmWaves joint development for Tokyo 2020 Olympic games,” in IEEE Conference on Standards _for Communications and Networking, CSCN 2017, Helsinki, Finland,_ _September 18-20, 2017._ IEEE, pp. 54–59. [3] Smart World of IoT– The Edge is Getting Smarter, Smaller, and Moving Further Out! - Part 1. [Online]. Available: [http:](http://www.embedded-computing.com/iot) [//www.embedded-computing.com/iot](http://www.embedded-computing.com/iot) [4] A. Kumar, M. Zhao, K. Wong, Y. L. Guan, and P. H. J. Chong, “A Comprehensive Study of IoT and WSN MAC Protocols: Research Issues, Challenges and Opportunities,” pp. 1–1. [5] J. Lin, W. Yu, N. Zhang, X. Yang, H. Zhang, and W. Zhao, “A Survey on Internet of Things: Architecture, Enabling Technologies, Security and Privacy, and Applications,” vol. 4, no. 5, pp. 1125–1142. [6] S. Ziegler and A. Br´ekine, “D6.3 – Second year report on standardization, dissemination and exploitation achievements,” p. 59. [7] R. Morabito, R. Petrolo, V. Loscr`ı, and N. Mitton, “LEGIoT: A Lightweight Edge Gateway for the Internet of Things,” vol. 81, pp. [1–15. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/](https://linkinghub.elsevier.com/retrieve/pii/S0167739X17306593) [S0167739X17306593](https://linkinghub.elsevier.com/retrieve/pii/S0167739X17306593) [8] A. Yousefpour, C. Fung, T. Nguyen, K. Kadiyala, F. Jalali, A. Niakanlahiji, J. Kong, and J. P. Jue, “All One Needs to Know about Fog Computing and Related Edge Computing Paradigms: A Complete [Survey.” [Online]. Available: http://arxiv.org/abs/1808.05283](http://arxiv.org/abs/1808.05283) [9] A. Carlier, I. Kopestenski, and D. Martens, “Lasp on Grisp : Implementation and evaluation of a general purpose edge computing system for Internet of Things.” [[10] I. ETSI, “ETSI MEC.” [Online]. Available: https://www.etsi.org/images/](https://www.etsi.org/images/files/ETSIWhitePapers/etsi_wp30_MEC_Enterprise_FINAL.pdf) [files/ETSIWhitePapers/etsi wp30 MEC Enterprise FINAL.pdf](https://www.etsi.org/images/files/ETSIWhitePapers/etsi_wp30_MEC_Enterprise_FINAL.pdf) [11] N. ETSI, “ETSI NGP.” [12] K. Karlsson, W. Jiang, S. Wicker, D. Adams, E. Ma, R. van Renesse, and H. Weatherspoon, “Vegvisir: A Partition-Tolerant Blockchain for the Internet-of-Things,” in 2018 IEEE 38th International Conference _on Distributed Computing Systems (ICDCS)._ IEEE, pp. 1150–1158. [[Online]. Available: https://ieeexplore.ieee.org/document/8416377/](https://ieeexplore.ieee.org/document/8416377/) [13] M. Du, K. Wang, Y. Chen, X. Wang, and Y. Sun, “Big Data Privacy Preserving in Multi-Access Edge Computing for Heterogeneous Internet of Things,” vol. 56, no. 8, pp. 62–67. [14] J. Jin, J. Gubbi, S. Marusic, and M. Palaniswami, “An Information Framework for Creating a Smart City Through Internet of Things,” vol. 1, no. 2, pp. 112–121. [15] A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi, “Internet of Things for Smart Cities,” vol. 1, no. 1, pp. 22–32. [Online]. [Available: http://ieeexplore.ieee.org/document/6740844/](http://ieeexplore.ieee.org/document/6740844/) [16] F. Esposito, A. Cvetkovski, T. Dargahi, and J. Pan, “Complete edge function onloading for effective backend-driven cyber foraging,” in 2017 _IEEE 13th International Conference on Wireless and Mobile Computing,_ _Networking and Communications (WiMob), pp. 1–8._ [17] V. Ivanov, Y. Jin, S. Choi, G. Destino, M. Mueck, and V. Frascolla, “Highly efficient representation of reconfigurable code based on a radio virtual machine: Optimization to any target platform,” in 25th European _Signal Processing Conference, EUSIPCO 2017, Kos, Greece, August 28_ _- September 2, 2017._ IEEE, pp. 893–897. [18] C. S. Meiklejohn and P. Van Roy, “Loquat: A framework for large-scale actor communication on edge networks.” IEEE, pp. 563–568. [Online]. [Available: http://ieeexplore.ieee.org/document/7917624/](http://ieeexplore.ieee.org/document/7917624/) [19] C. Baquero, P. S. Almeida, and A. Shoker, “Making Operation-based CRDTs Operation-based,” p. 15. [20] A. Barker. Implementing a Garbage-Collected Graph CRDT (Part 1 of [2). [Online]. Available: http://composition.al/CMPS290S-2018-09/2018/](http://composition.al/CMPS290S-2018-09/2018/11/12/implementing-a-garbage-collected-graph-crdt-part-1-of-2.html) [11/12/implementing-a-garbage-collected-graph-crdt-part-1-of-2.html](http://composition.al/CMPS290S-2018-09/2018/11/12/implementing-a-garbage-collected-graph-crdt-part-1-of-2.html) [21] R. Brown, Z. Lakhani, and P. Place, “Big(ger) Sets: Decomposed [delta CRDT Sets in Riak,” pp. 1–5. [Online]. Available: http:](http://arxiv.org/abs/1605.06424) [//arxiv.org/abs/1605.06424](http://arxiv.org/abs/1605.06424) [22] P. S. Almeida, A. Shoker, and C. Baquero, “Delta State Replicated Data Types,” vol. 111, pp. 162–173. [Online]. Available: [http:](http://arxiv.org/abs/1603.01529) [//arxiv.org/abs/1603.01529](http://arxiv.org/abs/1603.01529) [23] C. S. Meiklejohn, V. Enes, J. Yoo, C. Baquero, P. Van Roy, and A. Bieniusa, “Practical Evaluation of the Lasp Programming Model at Large Scale - An Experience Report,” pp. 109–114. [Online]. Available: [http://arxiv.org/abs/1708.06423](http://arxiv.org/abs/1708.06423) [24] C. Meiklejohn and P. Van Roy, “Lasp: A Language for Distributed, Coordination-free Programming,” in _Proceedings_ _of_ _the_ _17th_ _International Symposium on Principles and Practice of Declarative_ _Programming, ser. PPDP ’15._ ACM, pp. 184–195. [Online]. Available: [http://doi.acm.org/10.1145/2790449.2790525](http://doi.acm.org/10.1145/2790449.2790525) [25] C. Meiklejohn and H. Miller, “Partisan: Enabling Cloud-Scale Erlang [Applications.” [Online]. Available: http://arxiv.org/abs/1802.02652](http://arxiv.org/abs/1802.02652) [26] L. Adam. Wireless Small Embedded Erlang Applications with Grisp [Hardware Boards. [Online]. Available: http://www.erlang-factory.com/](http://www.erlang-factory.com/berlin2016/adam-lindberg) [berlin2016/adam-lindberg](http://www.erlang-factory.com/berlin2016/adam-lindberg) [27] P. S. Almeida, A. Shoker, and C. Baquero, “Efficient State-based CRDTs [by Delta-Mutation.” [Online]. Available: http://arxiv.org/abs/1410.2803](http://arxiv.org/abs/1410.2803) [28] J. Leitao, J. Pereira, and L. Rodrigues, “HyParView: A Membership Protocol for Reliable Gossip-Based Broadcast,” in Proceedings of _the 37th Annual IEEE/IFIP International Conference on Dependable_ _Systems and Networks, ser. DSN ’07._ IEEE Computer Society, pp. [419–429. [Online]. Available: https://doi.org/10.1109/DSN.2007.56](https://doi.org/10.1109/DSN.2007.56) [29] ——, “Epidemic Broadcast Trees,” in Proceedings of the 26th IEEE _International Symposium on Reliable Distributed Systems, ser. SRDS_ ’07. IEEE Computer Society, pp. 301–310. [Online]. Available: [http://dl.acm.org/citation.cfm?id=1308172.1308243](http://dl.acm.org/citation.cfm?id=1308172.1308243) [30] Internet of Battlefield Things (IOBT) — U.S. Army Research [Laboratory. [Online]. Available: https://www.arl.army.mil/www/default.](https://www.arl.army.mil/www/default.cfm?page=3050) [cfm?page=3050](https://www.arl.army.mil/www/default.cfm?page=3050) -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1901.05030, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://arxiv.org/pdf/1901.05030" }
2,019
[ "JournalArticle", "Conference" ]
true
2019-01-15T00:00:00
[ { "paperId": "54f2aef120d2e79b15bd296cc72fa38dc810c485", "title": "SmartEdge'19 - The Third International Workshop on Smart Edge Computing and Networking - Welcome and Committeees" }, { "paperId": "974a421f5883779b305db6f90a23be2fa2875e27", "title": "A Comprehensive Study of IoT and WSN MAC Protocols: Research Issues, Challenges and Opportunities" }, { "paperId": "3e4388586eb2538657a24ca1ace3c6877ec2235b", "title": "All One Needs to Know about Fog Computing and Related Edge Computing Paradigms: A Complete Survey" }, { "paperId": "fb11bc82b66e6980d8611b7f73cc2700e3c117d5", "title": "Big Data Privacy Preserving in Multi-Access Edge Computing for Heterogeneous Internet of Things" }, { "paperId": "3cb4ae87b7ed895a6563ad459329c7f8b8a19d47", "title": "Vegvisir: A Partition-Tolerant Blockchain for the Internet-of-Things" }, { "paperId": "d0895a0a9a6793ee2c477309160b6c7817466e8f", "title": "LEGIoT: A Lightweight Edge Gateway for the Internet of Things" }, { "paperId": "2f6996ddf2f23a9b708416014211f8f3f629f33d", "title": "Partisan: Enabling Cloud-Scale Erlang Applications" }, { "paperId": "0f56a4dcb0a20cfd6b57a2300deb36330982b1c0", "title": "Complete edge function onloading for effective backend-driven cyber foraging" }, { "paperId": "2d0887ae74333a09466c8c9d2a00ac77569412f6", "title": "5G-MiEdge: Design, standardization and deployment of 5G phase II technologies: MEC and mmWaves joint development for Tokyo 2020 Olympic games" }, { "paperId": "b811fc7b786b82ac05cde9d87679d368d5a86e69", "title": "Practical evaluation of the Lasp programming model at large scale: an experience report" }, { "paperId": "50ce4eee3e22f9b432792042ffc6136503fed37e", "title": "Highly efficient representation of reconfigurable code based on a radio virtual machine: Optimization to any target platform" }, { "paperId": "0e7f869040c80bec4ecfee61f081b22ac3babb13", "title": "Trends and challenges for autonomic RRM and MAC functionality for QoS provision and capacity expansions in the context of 5G beyond 6GHz" }, { "paperId": "cd3d4459ff0ff590a3e3258b0f774d6963cd4c90", "title": "A Survey on Internet of Things: Architecture, Enabling Technologies, Security and Privacy, and Applications" }, { "paperId": "1edf49dfd5f09aea94e753e8594e95666ebe131a", "title": "Loquat: A framework for large-scale actor communication on edge networks" }, { "paperId": "57adad122502face49b8dea49246b512539c1452", "title": "Big(ger) sets: decomposed delta CRDT sets in Riak" }, { "paperId": "05009141302f022fcf7387b37c8d9a2c415ead7b", "title": "Delta state replicated data types" }, { "paperId": "bd068826b712bc5a320af8ebef71b1cc9c88a836", "title": "Lasp: a language for distributed, coordination-free programming" }, { "paperId": "9a17a57fdd7bbeebdd0c5fe4a50d786034d49182", "title": "Efficient State-Based CRDTs by Delta-Mutation" }, { "paperId": "3ae7d736576b01665ba7f8a5f60430973e3f7b6b", "title": "Making operation-based CRDTs operation-based" }, { "paperId": "52f168c6c4f42294c4c9f9305bc88b6d25ffec9a", "title": "Internet of Things for Smart Cities" }, { "paperId": "967a2b6638fc0a0374e76ecc1768ef10b7c216cd", "title": "An Information Framework for Creating a Smart City Through Internet of Things" }, { "paperId": "e2d0cb2ecbf930048ae3329aa47a814d8ebace5d", "title": "Delta" }, { "paperId": "528a0945acb242d4f36c361e10f9e612c0b631b9", "title": "Epidemic Broadcast Trees" }, { "paperId": "06e04a7a24100dbc0f72f22bc8e6dde4b2a27d8b", "title": "HyParView: A Membership Protocol for Reliable Gossip-Based Broadcast" }, { "paperId": null, "title": "Implementing a Garbage-Collected Graph CRDT (Part 1 of 2)" }, { "paperId": null, "title": "Wireless Small Embedded Erlang Applications with Grisp Hardware Boards" }, { "paperId": null, "title": "Internet of Battlefield Things (IOBT) — U.S. Army Research Laboratory" }, { "paperId": null, "title": "D6.3 – Second year report on standardization, dissemination and exploitation achievements" }, { "paperId": null, "title": "ETSI MEC" }, { "paperId": null, "title": "Lasp on Grisp : Implementation and evaluation of a general purpose edge computing system for Internet of Things . ” [ 10 ] I . ETSI , “ ETSI MEC . ” [ Online ]" }, { "paperId": null, "title": "Smart World of IoT – The Edge is Getting Smarter , Smaller , and Moving Further Out ! - Part 1 . [ Online ]" } ]
10,173
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00e730f1a001c200cd93883e1cbeb0337c11faee
[ "Computer Science", "Medicine" ]
0.82044
Multiagent Continual Coordination via Progressive Task Contextualization
00e730f1a001c200cd93883e1cbeb0337c11faee
IEEE Transactions on Neural Networks and Learning Systems
[ { "authorId": "49785134", "name": "Lei Yuan" }, { "authorId": "2216719801", "name": "Lihe Li" }, { "authorId": "2188107173", "name": "Ziqian Zhang" }, { "authorId": "2166590799", "name": "Fuxiang Zhang" }, { "authorId": "2174870366", "name": "Cong Guan" }, { "authorId": "2152850415", "name": "Yang Yu" } ]
{ "alternate_issns": null, "alternate_names": [ "IEEE Trans Neural Netw Learn Syst" ], "alternate_urls": [ "https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=5962385" ], "id": "79c5a18d-0295-432c-aaa5-961d73de6d88", "issn": "2162-237X", "name": "IEEE Transactions on Neural Networks and Learning Systems", "type": null, "url": "http://ieeexplore.ieee.org/servlet/opac?punumber=5962385" }
Cooperative multiagent reinforcement learning (MARL) has attracted significant attention and has the potential for many real-world applications. Previous arts mainly focus on facilitating the coordination ability from different aspects (e.g., nonstationarity and credit assignment) in single-task or multitask scenarios, ignoring the stream of tasks that appear in a continual manner. This ignorance makes the continual coordination an unexplored territory, neither in problem formulation nor efficient algorithms designed. Toward tackling the mentioned issue, this article proposes an approach, multiagent continual coordination via progressive task contextualization (MACPro). The key point lies in obtaining a factorized policy, using shared feature extraction layers but separated independent task heads, each specializing in a specific class of tasks. The task heads can be progressively expanded based on the learned task contextualization. Moreover, to cater to the popular centralized training with decentralized execution (CTDE) paradigm in MARL, each agent learns to predict and adopt the most relevant policy head based on local information in a decentralized manner. We show in multiple multiagent benchmarks that existing continual learning methods fail, while MACPro is able to achieve close-to-optimal performance. More results also disclose the effectiveness of MACPro from multiple aspects, such as high generalization ability.
## MULTI-AGENT CONTINUAL COORDINATION VIA PROGRESSIVE TASK CONTEXTUALIZATION A PREPRINT **Lei Yuan[1,2], Lihe Li[1], Ziqian Zhang[1], Fuxiang Zhang[1,2], Cong Guan[1], Yang Yu[1,2,][∗]** 1 National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China 2 Polixir.ai yuanl, lilh, zhangzq, zhangfx, guanc @lamda.nju.edu.cn, yuy@nju.edu.cn _{_ _}_ #### ABSTRACT Cooperative Multi-agent Reinforcement Learning (MARL) has attracted significant attention and played the potential for many real-world applications. Previous arts mainly focus on facilitating the coordination ability from different aspects (e.g., non-stationarity, credit assignment) in singletask or multi-task scenarios, ignoring the stream of tasks that appear in a continual manner. This ignorance makes the continual coordination an unexplored territory, neither in problem formulation nor efficient algorithms designed. Towards tackling the mentioned issue, this paper proposes an approach Multi-Agent Continual Coordination via Progressive Task Contextualization, dubbed **MACPro. The key point lies in obtaining a factorized policy, using shared feature extraction layers** but separated independent task heads, each specializing in a specific class of tasks. The task heads can be progressively expanded based on the learned task contextualization. Moreover, to cater to the popular CTDE paradigm in MARL, each agent learns to predict and adopt the most relevant policy head based on local information in a decentralized manner. We show in multiple multi-agent benchmarks that existing continual learning methods fail, while MACPro is able to achieve close-tooptimal performance. More results also disclose the effectiveness of MACPro from multiple aspects like high generalization ability. #### 1 Introduction Cooperative Multi-agent Reinforcement Learning (MARL) has attracted prominent attention in recent years [1], and achieved great progress in multiple aspects, like path finding [2], active voltage control [3], and dynamic algorithm configuration [4]. Among the multitudinous methods, researchers, on the one hand, focus on facilitating coordination ability via solving specific challenges, including non-stationarity [5], credit assignment [6], and scalability [7]. Other works, on the other hand, investigate the cooperative MARL from multiple aspects, like efficient communication [8], zero-shot coordination (ZSC) [9], policy robustness [10], etc. A lot of methods emerge as promising solutions for different scenarios, including policy-based ones [11,12], value-based series [13,14], and many other variants, showing remarkable coordination ability in a wide range of tasks like SMAC [15]. Despite the great success, the mainstream cooperative MARL methods are still restricted to being trained in one single task or multiple tasks simultaneously, assuming that the agents have access to data from all tasks at all times, which is unrealistic for physical agents in the real world that can only attend to one task at a time. Continual Reinforcement Learning plays a promising role in the mentioned problem [16], where the agent aims to avoid catastrophic forgetting, as well as enable knowledge transfer to new tasks (a.k.a. stability-plasticity dilemma [17]), while maintaining scalable to a large number of tasks. Multiple approaches have been proposed to address one or more of these challenges, including regularization-based ones [18–20], experience maintaining techniques [21, 22], and task structure sharing categories [23–25], etc. However, the multi-agent setting is much more complex than the single-agent one, as the interaction among agents might cause additional considerations [26]. Also, coordinating with multiple teammates is proved intrinsically tough [27]. Previous works model this problem as multi _∗Corresponding Author_ ----- task [9] or just uni-modal coordination among teammates [28]. In light of the significance and ubiquity of cooperative MARL, it is thus imperative to consider the continual coordination in both the problem formulation and the algorithm design to tackle this issue. In this work, we develop such a continual coordination framework in cooperative MARL where tasks appear sequentially. Concretely, we first develop a multi-agent task context extraction module, where information of each state in a specific task is extracted and integrated by a product-of-expert (POE) mechanism into a latent space to capture the task dynamic information, and a contrastive regularizer is further applied to optimize the learned representation, with which similar tasks representation are pulled together while dissimilar ones are pushed apart. Afterward, an expandable multi-head policy architecture whose separate independent heads are synchronously expanded with the newly instantiated context, along with a carefully designed shared feature extraction module. Finally, considering the popular CTDE (Centralized Training with Decentralized Execution) paradigm in mainstream cooperative MARL, we leverage the local information of each agent to approximate the policy head selection process via policy distillation in the centralized training process, with which agents can select the most optimal ones to coordinate with other teammates in a decentralized manner. For the evaluation of the proposed approach, MACPro, we conduct extensive experiments on various cooperative multi-agent benchmarks in the continual setting, including level-based foraging (LBF) [29], predator-prey (PP) [11], and the StarCraft Multi-Agent Challenge benchmark (SMAC) [30], and compare MACPro against previous approaches, strong baselines, and ablations. Experimental results show that MACPro considerably improves upon existing methods. More results demonstrate its high generalization ability and its potential to be integrated with different value-based methods to enhance their continual learning ability. Visualization experiments provide additional insight into how MACPro works. #### 2 Related Work **Cooperative Multi-agent Reinforcement Learning Many real-world problems are made up of multiple interactive** agents, which could usually be modeled as a Multi-Agent Reinforcement Learning (MARL) problem [26,31]. Further, when the agents hold a shared goal, this problem refers to cooperative MARL [32], showing great progress in diverse domains like path finding [2], active voltage control [3], and dynamic algorithm configuration [4], etc. Many methods are proposed to facilitate coordination among agents, including policy-based ones (e.g., MADDPG [11], MAPPO [12], FD-MARL [33]), value-based series like VDN [13], QMIX [14], Linda [34], or other techniques like transformer [35], these approaches, have demonstrated remarkable coordination ability in a wide range of tasks (e.g., SMAC [30], Hanabi [12], GRF [35]). Besides the mentioned approaches and the corresponding variants, many other methods are also proposed to investigate the cooperative MARL, including efficient communication [8] to relieve the partial observability caused by decentralized policy execution, policy deployment in an offline manner [36], model learning in MARL [37], policy robustness when some perturbations exist [10], and training paradigm like CTDE (centralized training with decentralized execution) [38], ad hoc teamwork [27], etc. Despite the mentioned progress, the vast majority of current approaches either focus on training the MARL policy on a single task, or the multi-task setting where all tasks appear simultaneously, lacking the attention to the continual coordination problem. In these methods, MRA [39] focuses on creating agents that generalize across populationvarying Markov games, proposing meta representations for agents that explicitly model the game-common and gamespecific strategic knowledge. MATTAR [40] assumes there are some basic tasks, training with which can accelerate the training process in other similar tasks, and develops a multi-agent multi-task training framework. TrajeDi [9] and some variants (or improved versions) like MAZE [41], concentrate on coordinating with different teammates or even unseen ones like a human, these methods are also under the assumption that we can access all the training tasks all the time. [28] introduces a multi-agent learning testbed that supports both zero-shot and few-shot settings based on Hanabi, but it only considers the uni-modal coordination among tasks, and the experimental results demonstrate methods like VDN [13] trained in the proposed testbed can coordinate well with unseen agents, without any additional assumptions made by previous works. **Continual Reinforcement Learning Continual Learning is conceptually related to incremental Learning and Life-** long Learning as they all assume that tasks or samples are presented in a sequential manner [17,42,43]. For continual reinforcement learning [16], EWC [18] learns new Q-functions by regularizing the l2 distance between the optimal weights of the new task and previous ones. It requires additional supervision information like task changes to update its objective, and then selects a specific Q-function head and a task-specific exploration schedule for different tasks. CLEAR [44] is a task-agnostic method that does not require task information during the continual learning process, and leverages big experience replay buffers to prevent forgetting. Coreset [45] prevents catastrophic forgetting by choosing and storing a significantly smaller subset of the previous task’s data, which is used to rehearse the model ----- ### Testing task 1 task 1,2 task 1, …, 𝑚𝑚 Figure 1: An example of multi-agent continual coordination, where tasks (e.g., the position of food changes in the Level Based Foraging (LBF) [29]) change along with the timeline. We thus need to train a policy πππm to solve the concurrent task as well as maintain the knowledge of previous tasks (i.e., avoid catastrophic forgetting) . during or after finetuning. Some other works like HyperCRL [46], and [47] utilize a learned world model to promote continual learning efficiency. Considering the scalability issue along with the task number, CN-DPM [48] and LLIRL [49] decompose the whole task space into several subsets of the data (tasks), and then utilize techniques like Dirichlet Process Mixture or Chinese Restaurant Process to expand the neural network for efficient continual supervised Learning and reinforcement learning tasks, respectively. OWL [24] is a recently proposed approach that learns a multi-head architecture and achieves high learning efficiency, and CSP [25] incrementally builds a subspace of policies for training a reinforcement learning agent on a sequence of tasks. Other researchers also design benchmarks like Continual world [50], or baselines [51] to verify the effectiveness of different methods in single-agent reinforcement learning. [28] investigate whether agents can coordinate with unseen agents by introducing a multi-agent learning testbed based on Hanabi. Still, it only considers the uni-modal coordination among tasks. Our work takes a further step in this direction for problem formulation and algorithm design. #### 3 Problem Formulation This work considers a cooperative multi-agent reinforcement learning problem under partial observation, which can be formalized as a Dec-POMDP [52], with tuple = _N,_ _,_ _, Ω, P, O, R, γ_, where N = 1, _, n_,, = _M_ _⟨_ _S_ _A_ _⟩_ _{_ _· · ·_ _}_ _S_ _A_ , Ω are the set of agents, states, joint actions, and local observation, respectively. P : ∆( ) _A[1]_ _× · · · × A[n]_ _S × A →_ _S_ stands for the transition probability function, O : S × N → Ω and R : S × A → R are the corresponding observation function and reward function, and γ [0, 1) is the discounted factor. Multiple interactive agents in a Dec-POMDP _∈_ coordinate with teammates to complete a task under a share reward R, at each time step, agent i receives the local observation o[i] = O(s, i) and outputs the action a[i] . The formal objective of the agents is to maximize the _∈A[i]_ expected cumulative discounted reward E[[�][∞]t=0 _[γ][t][R][(][s][t][,aaa][t][)]][ by learning an optimal joint policy.]_ In this work, we focus on a continual coordination problem where agents in a team are exposed to a sequence of (infinite) tasks Y = (M1, · · ·, Mm, · · · ). Each task involves a sequential decision making problem and can be formulated as a Dec-POMDP Mm = ⟨Nm, Sm, Am, Ωm, Pm, Om, Rm, γ⟩, as shown in Fig. 1. These agents are continually evaluated on all previous tasks (but cannot be trained with these tasks) and the present task. Therefore, the agent’s policy needs to transfer to new tasks while maintaining the ability to perform previous tasks. Concretely, agents that have learned M tasks are expected to maximize the MARL objective for each task in YM = {M1, · · ·, MM _}. We_ consider the setting where task boundaries are known during the centralized training phase. During the decentralized execution phase, agents cannot access global but only local information to finish the tasks sampled from YM . #### 4 Method In this section, we will describe the detailed design of our proposed method, MACPro. First, we propose a novel training paradigm, including a shared feature extraction part and an adaptive policy heads expansion module based ----- (b) Dynamic Network Expansion (a) Task Contextualization Learning (c) Decentralized Task Approximation & Execution Figure 2: The overall framework of MACPro. (a) We design an efficient multi-agent task contextualization learning module to capture the uniqueness of each emerging task. (b) The training paradigm, including a shared feature extraction part and an adaptive policy heads expansion module based on the learned contexts. (c) Each agent utilizes its local information to approximate the actual task head in a decentralized way. on the learned contexts (Fig. 2(a)). Next, we design an efficient multi-agent task contextualization learning module to capture the uniqueness of each emerging task (Fig. 2(b)). Finally, considering the CTDE property in mainstream cooperative MARL, we train each agent to utilize its local information to approximate the actual task head (Fig. 2(c)). **4.1** **Multi-agent Task Contextualization Learning** In continual reinforcement learning where tasks keep altering sequentially, it is crucial to capture the unique context of each emerging new task. However, the behavioral descriptor of the multi-agent task is much more complex than the single-agent setting due to the interactions among agents [1]. Thus this subsection aims to tackle this issue by developing an efficient multi-agent task contextualization learning module. Specifically, consider a trajectory τ = (s0, · · ·, sT ) with horizon T roll-out by any policies, we utilize a global trajectory encoder gθ parameterized by θ to encode τ into a latent space. Concretely, the trajectory representation is represented by a multivariate Gaussian distribution N (µθ(τ ), σθ[2][(][τ] [))][ whose parameters are computed by][ g][θ][(][τ] [)][. As the] trajectory horizon T may alter for different tasks (e.g., 3m and 5m in SMAC [30]), we here apply a transformer [53] architecture (see App. H) to extract feature from each trajectory, thus the latent context of a whole trajectory can be represented as T Gaussian distributions N (µ0, σ0[2][)][,][ · · ·][,][ N] [(][µ][T] _[, σ]T[2]_ [)][, where][ N] [(][µ][i][, σ]i[2][)][ stands for the][ i][th][ essential] parts of the trajectory. Next, considering the importance of different states in a trajectory, we apply the product-ofexperts (POE) technique [54] to acquire the joint representation of a trajectory, which is also a Gaussian distribution _N_ (µθ(τ ), σθ[2][(][τ] [))][, where:] (1) _µθ(τ_ ) = _σθ[2][(][τ]_ [) =] � _T_ �� _T_ �−1 � � _µt(σt[2][)][−][1]_ (σt[2][)][−][1] _t=0_ _t=0_ � _T_ �−1 � (σt[2][)][−][1] _t=0_ _,_ _._ The detailed derivative process between the joint distribution and each single one can be seen in App. J. The previous part can obtain representation for each trajectory. Nevertheless, the learned representation lacks any dynamic information about a specific multi-agent task. As the difference between any dynamic model lies in transition and reward functions [55], we here apply a loss function to force the learned trajectory representation to capture the dynamic information of each task. Specifically, we learn a context-aware forward model h including three predictors: _hs, ho, hr which are responsible to predict the next state, local observations, and reward given the current state, local_ ----- observations, actions, and task contextualization, respectively: _T_ � � _Lmodel = Eτ_ _∈D′_ _||hs[st,ooot,aaat, z] −_ _st+1||2[2][+]_ _t=0_ _||ho[st,ooot,aaat, z] −_ _ooot+1||2[2][+]_ (hr[st,ooot,aaat, z] − _rt)[2][�],_ (2) where z is the task contextualization sampled from the joint task distribution, is the replay buffer for task con_D[′]_ textualization learning, which stores a small amount of trajectories for each task. However, as there are tasks with different correlations, the mentioned optimization object Lmodel might be insufficient for differentiable context acquisition. Therefore, we apply another auxiliary contrastive loss [56] by pulling together semantically similar data points (positive data pairs) while pushing apart the dissimilar ones (negative data pairs): � _Lcontg =Eτj_ _,τk∈D′_ 111{yj = yk}DJ (gθ(τj)||gθ(τk))+ (3) 1 � 111{yj ̸= yk} _DJ_ (gθ(τj)||gθ(τk)) + ε _,_ where 111{·} is the indicator function, yj and yk are the label(s) of the task(s) from which τj and τk are sampled, respectively, ε is a small positive constant added to avoid division by zero. DJ (P _||Q) = DKL(P_ _||Q) + DKL(Q||P_ ) is the Jeffrey’s divergence [57] used to measure the distance between two distributions, and DKL denotes the KullbackLeibler divergence. Thus the overall loss term is: _Lcontext = Lmodel + αcontg_ _Lcontg_ _,_ (4) where αcontg is the coefficient balancing the loss terms. **4.2** **Adaptive Dynamic Network Expansion** With the previously learned global trajectory encoder gθ, we can obtain a unique contextualization for each task. Now, this subsection comes to the design of a context-based continual learning mechanism, which incrementally clusters a stream of stationary tasks in the dynamic environment into a series of contexts and opts for the optimal policy head from the expandable multi-head neural network. Formally, for multiple tasks that appear sequentially, we design a policy network consisting of a shared feature extractor φ with multiple layers of neural network (the index of agent is omitted in this part for simplicity), which can promote knowledge sharing among different tasks. Furthermore, as there may be some multimodal tasks, a single head for all tasks could make the policy overfit to some specific tasks. One way to solve this problem is to learn a customized head for each task like OWL [24]. However, this solution is of poor scalability as the number of heads increases linearly over the number of tasks that could be infinitely many. Thus, we develop an adaptive network expansion paradigm based on the similarity between task contextualizations. Specifically, we assume that the agents have already experienced M tasks and have K policy heads {ψ[k]}k[K]=1 [so far (][K][ ≤] _[M]_ [). For each head, we store][ bs][ trajecto-] ries in buffer D[′], and we use gθ to obtain the corresponding task contextualizations with mean values {{µ[j]k[}]j[bs]=1[}][K]k=1[.] When encountering a new task (M + 1), we first utilize the feature extractors φ and all the existing heads {ψ[k]}k[K]=1 [to] derive a set of behavior policies {πππk}k[K]=1 [to collect][ bs][ trajectories each on task][ (][M][ + 1)][, denoted as][ {{][τ][ j]k _[}]j[bs]=1[}][K]k=1[.]_ Next, we use gθ to derive the mean values {{µ[′]k[j][}]j[bs]=1[}][K]k=1 [of their contextualizations and calculate the similarities] between the existing mean values {{µ[j]k[}]j[bs]=1[}][K]k=1 [as follows:] _l = (l1, · · ·, lK), l[′]_ = (l1[′] _[,][ · · ·][, l]K[′]_ [)][,] _bs_ � _µ[i]k[||][2][,]_ _i=1_ _bs_ � _µ[i]k[||][2][, k][ = 1][,][ · · ·][, K.]_ _i=1_ where lk = [1] _bs_ _lk[′]_ [= 1] _bs_ _bs_ � _||µ[j]k_ _[−]_ _bs[1]_ _j=1_ _bs_ � _||µ[′]k[j]_ _[−]_ _bs[1]_ _j=1_ (5) Here l is the vector describing the dispersion of the K existing contextualizations, and l[′] is the vector describing the distance between the K new contextualizations and the existing ones. Let k∗ = arg min1≤k≤K lk[′] [, such that the] ----- _k[th]_ _∗_ [pair of existing and new contextualizations are closest among all][ K][ pairs. With an adjustable threshold][ λ][new][, if] _lk[′]_ _∗_ _[≤]_ _[λ][new][l][k]∗_ [, indicating task][ (][M][ +1)][ is similar to the task(s) that head][ ψ][k][∗] [takes charge, we thus merge it to this/these] learned task(s) and use the unified head ψ[k][∗] for them. Otherwise, none of the learned tasks are similar with the new one, a new head ψ[K][+1] is created. This phase processes along with the task sequence, enjoying high scalability and learning efficiency. The previous part solves the head expansion issue, while a single shared feature extractor may inevitably cause forgetting. We here apply an l2-regularizer to relieve this issue by constraining the parameters of the shared part don’t change too drastically when learning task (M + 1): _Lreg =_ _n_ � _||φi −_ _φ[M]i_ _[||][2][,]_ (6) _i=1_ where φ[M]i is the saved snapshot of agent i’s feature extractor φi after training on task M . As we can apply MACPro to any value-based methods, we thus obtain the temporal difference error LTD as [r + γ maxa′ Q[tot](s[′], a[′]; θ[−]) − _Q[tot](s, a; θ)][2], where θ[−]_ are parameters of a periodically updated target network. The overall loss term of training agents’ policies is defined as follows: _LRL = LTD + αregLreg,_ (7) where αreg is the coefficient balancing the two loss terms. 1 Run 5m_vs_6m 2s1z_vs_3z away Keep still 13m 2s3z 2 4m 3s5z Run Run 8m_vs_9m 2s2z_vs_4s 5 4 3 towards randomly (a) LBF (b) PP (c) Marines (d) SZ Figure 3: Experimental environments used in this paper. (a) Level-based foraging (LBF), where the position of the food changes in different tasks as indicated by the number on the food. (b) Predator prey (PP), where in different tasks, the position of landmarks, the agents’ acceleration, maximum speed, positions, and the fixed heuristic policies the prey uses are different. (c) & (d) Where Marines and SZ from StarCraft Multi-Agent Challenge (SMAC) involves various numbers and types of battle agents. **4.3** **Decentralized Task Approximation** Although we have obtained an efficient continual learning approach for any tasks that appear in a sequential way, it is still far away from the MARL setting, as it requires the trajectory of global states to obtain the task representation, while agents in a MARL system can only acquire its local information. Towards tackling the mentioned issue, we here develop a distillation solution. Concretely, for agent i with its local trajectory history τ _[i]_ = (o[i]0[,][ · · ·][, o][i]T [)][, we design a local trajectory encoder][ f][θ]i[′] [that is similar to the global trajectory] encoder gθ. fθi[′] [takes][ τ][ i][ as input and outputs][ N] [(][µ][θ]i[′] [(][τ][ i][)][, σ]θ[2]i[′] [(][τ][ i][))][. We thus optimize][ f][θ]i[′] [by minimizing the Jeffrey’s] divergence between the distributions: |Col1|Col2|Col3|Col4|1| |---|---|---|---|---| |||||| |||||2| |||||| |5||4||3| |Run away|Keep still| |---|---| |Run towards|Run randomly| 3s5z 8m_vs_9m 4m 2s2z_vs_4s � � _Loracle = E(τ,τ i)∈D′_ _DJ_ 2s1z_vs_3z 13m �� _gθ(τ_ )||fθi[′] [(][τ][ i][)] _,_ (8) 5m_vs_6m where denotes gradient stop, τ, τ _[i]_ stand for the global and local trajectory of a same task, respectively. To accelerate _·_ this learning process and make it consistent with task contextualization learning, we design a local auxiliary contrastive loss: � _Lcontl =Eτj_ _,τk∈D′_ 111{yj = yk}DJ (fθi[′] [(][τ][ i]j [)][||][f][θ]i[′] [(][τ][ i]k[))+] 1 � (9) 111{yj ̸= yk} _DJ_ (fθi[′] [(][τ][ i]j [)][||][f][θ]i[′] [(][τ][ i]k[)) +][ ϵ] _._ 2s3z ----- MACPro Oracle Random OWL Coreset EWC Finetuning (a) LBF (b) PP (c) Marines (d) SZ Figure 4: Performance comparison with baselines. Where each task is trained for 400k steps in LBF, 500k steps in other benchmarks, and each plot indicates the average performance across all tasks seen so far. The overall loss term of this part is: _Lapprox = Loracle + αcontl_ _Lcontl_ _,_ (10) where αcontl is the coefficient balancing the loss terms. During the decentralized execution phase, agents firstly roll-out P episodes to probe the environment. Concretely, for each probing episode p(p = 1, _, P_ ), agents randomly choose one policy head to interact with the evaluating task to _· · ·_ collect trajectory τp[i][, and calculate the mean value][ µ][θ]i[′] [(][τ][ i]p[)][ of the trajectory representation][ f][θ]i[′] [(][τ][ i]p[)][. Finally, each agent] _i selects the most optimal task head via comparing the distance with the K existing task contextualization as follows:_ |Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8| |---|---|---|---|---|---|---|---| ||||(a)|LBF|||| ||||||||| |re 4: Perfor r benchmarks||(c) M mance compa, and each pl||arines rison with bas ot indicates th||elines. Wher e average per|| |Col1|Col2|Col3|Col4| |---|---|---|---| ||(b)|PP|| ||||| |task is trai ce across al|(d) ned for 400k st l tasks seen so f|SZ eps in LBF, 5 ar.|00k steps in| _k[⋆i]_ = argmin1≤k≤K 1≤minp≤P _[||][µ][θ]i[′]_ [(][τ][ i]p[)][ −] _bs[1]_ _bs_ � _µ[j]k[||][2][,]_ (11) _j=1_ and use head ψ[k][⋆i] with feature extractor φ for testing. #### 5 Experimental Evaluation In this section, we design extensive experiments for the following questions: 1) Can our approach MACPro achieve high continual ability compared to other baselines in different scenarios, and how each component influences its performance? (Sec. 5.2) ? 2) What task representation is learned by our approach, and how does it influence the continual learning ability (Sec. 5.3) ? 3) Can MACPro be integrated into multiple cooperative MARL methods, and how does each hyperparameter influence its performance (Sec. 5.4) ? **5.1** **Environments and Baselines** For the evaluation benchmarks, we select four multi-agent environments (see Fig. 3). Where Level Based Foraging (LBF) [29] is a cooperative grid world game with agents that are rewarded if they concurrently navigate to the food and collect it. The position of the food changes in different tasks as indicated by the number of the food. Predator Prey (PP) [11] is another popular benchmark where agents (predators) need to chase the adversary agent (prey) and encounter it to win the game. In different tasks, the position of landmarks, the agents’ acceleration, maximum speed, positions, and the fixed heuristic policies the prey uses are different. And Marines and SZ from StarCraft Multi-Agent Challenge (SMAC) [30], involving various numbers of the agent. To evaluate if MACPro can achieve good performance on these benchmarks when different tasks appear continually, we apply it to a popular valued-based method QMIX [14]. Compared baselines include Finetuning, which directly tunes the learned policy on the current task; EWC [18], a regularization-based method that constrains the whole agent network from dramatic change; Coreset [45], which uses a shared replay buffer over all the tasks so that data on old tasks can rehearse the agents during finetuning on the new task. Also, OWL [24] is included as it is similar to our work but applies the bandit algorithm for head selection. To further study the head selection process, we design Random, ----- |Col1|Col2| |---|---| ||| |earning res ng that the ) because t|ults on PP. agents are tr he task has| |Method|Task 1 Task 2 Task 3 Task 4 Task 5 Average| |---|---| |Ours W/o model W/o cont g W/o POE W/o oracle W/o cont l W/o cont g,l|111...000000 ±±± 000...000000 111...000000 ±±± 000...000000 0.80 ± 0.40 0.71 ± 0.38 111...000000 ±±± 000...000000 000...999000 ±±± 000...000999 0.93 0.10 0.68 0.46 0.67 0.47 0.85 0.21 0.93 0.10 0.81 0.18 ± ± ± ± ± ± 111...000000 ±±± 000...000000 0.67 ± 0.47 111...000000 ±±± 000...000000 000...999777 ±±± 000...000333 0.55 ± 0.41 0.84 ± 0.18 0.99 0.01 0.86 0.19 0.97 0.03 0.74 0.17 0.69 0.44 0.85 0.05 ± ± ± ± ± ± 0.21 0.39 0.60 0.49 0.06 0.12 0.24 0.28 0.24 0.38 0.27 0.07 ± ± ± ± ± ± 0.98 0.03 0.86 0.19 0.76 0.17 0.75 0.20 0.97 0.04 0.86 0.06 ± ± ± ± ± ± 0.79 ± 0.40 0.99 ± 0.02 111...000000 ±±± 000...000000 0.34 ± 0.42 0.54 ± 0.38 0.73 ± 0.12| |---|---| with MACPro selecting a head randomly during testing, and Oracle, where MACPro’s head selection is based on the ground-truth heads information. More details about benchmarks and baselines can be seen in App. G. **5.2** **Competitive Results and Ablations** **Continual Learning Ability Comparison** At first glance, we compare MACPro against the mentioned baselines to investigate the continual learning ability as shown in Fig. 5. We can find that Finetuning achieves the most inferior performance in different benchmarks, showing that a conventional reinforcement learning training paradigm is improper for continual learning scenarios. Other successful approaches for single agent continual learning, like Coreset, EWC, and OWL, also suffer from performance degradation in the involved benchmarks, demonstrating the necessity of specific consideration for MARL settings. The Oracle baseline, where we give all the ground-truth task identification when testing, can be seen as an upper bound of performance on the related benchmarks, acquiring superiority over all baselines in all benchmarks, demonstrating a multi-head architecture can solve the multi-modal tasks while conventional approaches fail. Our approach MACPro, obtains comparable performance to Oracle, indicating the efficiency of all the designed modules. Random, which selects a head randomly when testing, suffers from terrible performance degradation compared with MACPro and Oracle, showing that the success of MACPro is owing to the appropriate head selection mechanism but not a larger network with multiple heads. Furthermore, we display the performance on every single task in PP in Fig .5, We can find that baselines Fientuning, EWC, and Coreset all suffer from performance degradation on one task after training on it, i.e., catastrophic forgetting, demonstrating the necessity of specific consideration for MARL continual learning. other baselines, OWL and Random, fail to choose the appropriate head for testing and does not perform well on all tasks. Learning the new task as quickly as Finetuning without forgetting the old ones, our method MACPro obtains excellent performance. The comparable average performance to Oracle also indicates that MACPro can accurately choose the optimal head for testing. More results can be seen in App. I. ----- **Ablation Studies** As MACPro is composed of multiple components, we here design ablation studies on benchmark LBF to investigate their impacts. First, for task contextualization learning, we derive W/o model by removing the forward model h and its corresponding loss term Lmodel, and using the contrastive loss only to optimize the global trajectory encoder gθ. Next, instead of extracting the representation of trajectories with POE, we use the average of the Gaussian distributions generated by the transformer network as the representation, and we call it W/o POE. Further, we also introduce W/o oracle, which has a similar number of parameters as MACPro, to investigate whether the superiority of MACPro over QMIX is due to the increase in the number of parameters. Finally, we remove both global contrastive loss Lcontg and local contrastive loss Lcontl to derive W/o contg,l. As shown in Tab. 1, we can find that when the model loss is removed, W/o model suffers from performance degradation in most tasks, indicating the necessity for task representation learning. Furthermore, the POE mechanism also slightly influences the learning performance, demonstrating the special integration of multiple representations of trajectories can facilitate representation learning. Consequently, when removing the oracle loss function, W/o oracle sustains great performance degradation, and even fails in task 3, indicating a simply larger network cannot fundamentally improve the performance. We also find contrastive learning loss has a positive effect on performance. We further design W/o contg and W/o contl by setting _αcontg = 0 and αcontl = 0 to study the impact of contrastive loss. Both variants suffer from performance degradation,_ indicating the necessity of contrastive learning. 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 MACPro Random OWL EWC Finetuning Coreset MACPro Random OWL EWC Finetuning Coreset (a) LBF (b) PP 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 MACPro Random OWL EWC Finetuning Coreset MACPro Random OWL EWC Finetuning Coreset (c) Marines (d) SZ Figure 6: Generalization results. **The Generalization results** As we focus on training each emerging task sequentially, it produces a significant risk of overfitting. What’s more, the ultimate goal of continual learning agents is to not only perform well on seen tasks, but also utilize the learned experiences to complete future unseen tasks. Here, we design experiments to test the generalization ability of MACPro compared with multiple baselines. Concretely, we design 20 additional tasks (details can be seen in App. G) for each benchmark that agents have not encountered before to conduct zero-shot experiments. As shown in Fig. 6, MACPro demonstrates the most superior performance compared to the multiple compared baselines, indicating that it has strong generalization ability due to the multi-agent task contextualization |Col1|co|nt|g|Col5| |---|---|---|---|---| |n|d|ic|a|ti| |||||| |||||| |||||| |contrastive learning loss has a positive effect on performance. α = 0 and α = 0 to study the impact of contrastive lo contg contl indicating the necessity of contrastive learning. 0.8 Mean 0.6 Return 0.4 Test 0.2 0.0 MACPro Random OWL EWC Finetuning Coreset (a) LBF|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16|Col17| |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| |||||||||||||||||| |||||||||||||||||| |||||||||||||||||| |||||||||||||||||| |||||||||||||||||| |||||||||||||||||| |||||||||||||||||| |||||||||||||||||| |||||||||||||||||| |||||||g||C|or|es|e|t||||| |||||||||||||||||| |||||||||||||||||| |||||||||||||||||| |Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11| |---|---|---|---|---|---|---|---|---|---|---| |||||||||||| |||||||||||| |||||||||||| |||||||||||| |||||||||||| |||||||||||| |MACPro Random OWL EWC Finetuning (d) SZ||||||||||| |||||||Co|r|es|et|| |||||||||||| |||||||||||| |||||||||||| ----- |MACPro (ours) MACPro w/o adaptive expansion Finetuning # Heads of MACPro (ours) # Heads of MACPro w/o adaptive expansion # Heads of Finetuning 10 4 1 Merge task 3 to task 1 Expand a new head for task 6 Projection at the end task 2 task 1,3 Run task 1,3 task 2,4,5 toR wu an rds task 2,4,5 task 1,3 task 9,10 away Similar Dissimilar aR wu an y task 6 Run Keep task 6,7,8 away still|Col2|Col3|MACPr # Head|Col5|o (ours) s of MACPro (ours)|Col7|Col8|Col9|Col10|Col11|MACPro # Heads|w/o adaptive e of MACPro w/o|xpansion adaptive expans|Fi ion #|Col16|netuning Heads of Finetuning|Col18|Col19|10 4 1| |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| ||||||||||||||||||||| ||||||||||||||||||||| ||||||||||||||||||||| ||||||||||||||||||||| ||||||||||||||||||||| ||||||Merge task 3|||to task 1|||E|xpand a ne|w head for ta|sk 6||Projection at||the end|| ||||||||||||||||||||| ||task 2 task 1,3 Run away Similar Run away||task|1,3|Run away||||||task 1,3|task 2,|4,5 Run towards||tas|k 2,4,5 tas|k 1,3 t||| |||||||Run away|||||||||||||| |||||||||||Dissimilar task 6 Run Keep away still|||||||||| Figure 7: Task contextualization analysis. Where similar tasks have the same background color, e.g., task 1 and task 3 correspond to the green background. When encountering a new task, we sample latent variables generated by gθ(τ ) and apply dimensionality reduction to them by principal component analysis (PCA) [58], denoted as . _△_ learning module and the decentralized task approximation procedure. Note that the baseline Oracle is not tested here because there is no ground-truth head selection on unseen tasks. **5.3** **Task Contextualization Analysis** Then, we visualize the development of continual learning performance, along with changes in task representation and factored heads, to demonstrate how our method works. Concretely, we build a task sequence with 10 tasks of benchmark PP. As shown in Fig. 7, when t = 1.0M, the incoming task 3 is similar to task 1, and their latent variables are distributed in the same area (the green ellipse). Task 3 shares the same head as task 1, leading to an unchanged number of task heads. When a dissimilar task is encountered at t = 2.5M, none of the learned tasks are similar to the incoming task 6. The latent variables of task 6 are distributed in a new area (the red ellipse), and MACPro does expand a new head accordingly. This process continuously proceeds, until the learning procedure ends at t = 5.0M, when the latent variables of all ten tasks are distributed in four separate clusters, and MACPro has four heads, respectively. The latent variables of the representations fθi[′] [(][τ][ i][)][ encoded by individual trajectory encoders, denoted as][ ◦][, are also] displayed (we omit them in the first two 3D figures for simplicity). It shows that the representations learned by fθi[′] [(][τ][ i][)] is close to gθ(τ ), enabling accurate decentralized task approximation and good performance. Consequently, we can further find the learning curve in the top row of Fig. 7, along with the number of separate heads that changes according to the corresponding task representations. We also compare two extreme-case methods, where Finetuning holds a single head for all tasks, enjoying high scalability but strong catastrophic forgetting. On the contrary, MACPro w/o adaptive expansion maintains one head for each task and can achieve high learning efficiency, but the heads’ storage cost may impede it when facing a large number of tasks. Our method MACPro, achieves comparable or even better learning ability but consumes fewer heads, showing high learning efficiency and scalability. **5.4** **Integrative Abilities and Sensitive Studies** MACPro is agnostic to specific value-based cooperative MARL methods. Thus we can use it as a plug-in module and integrate it with existing MARL methods like VDN [13], QMIX [14], and QPLEX [59]. As shown in Tab. 2, when integrating with MACPro, the performance of the baselines vastly improves, indicating that MACPro has high generality ability for different methods to facilitate continual learning ability. As MACPro includes multiple hyperparameters, here we conduct experiments on benchmark PP to investigate how each one influences the continual learning ability. First, αreg controls the extent of restriction on changing the parameters of the shared feature extractor φi. If it is too small, the dramatic change of φi’s parameters may induce severe forgetting. On the other hand, if it is too large, agents remember the old task at the expense of not learning the new task. We thus find each hyperparameter via grid-search. As shown in Fig. 8 (a), we can find that αreg = 500 is the best choice in this benchmark. Furthermore, another adjustable hyperparameter αcontg influences the training of global trajectory encoder gθ in multi-agent task contextualization learning. Fig. 8 (b) shows that αcontg = 0.1 performs the best. In decentralized task approximation, αcontl balances the learning of local trajectory encoder fθi[′] [. We find that] in Fig. 8 (c) αcontl = 0.1 performs the best. During decentralized execution, agents first probe P episodes before ----- Table 2: Integrative Abilities. Envs Method VDN QMIX QPLEX W/ MACPro 000...929292 ± ± ± 0 0 0...020202 000...909090 ± ± ± 0 0 0...090909 000...979797 ± ± ± 0 0 0...030303 LBF W/o MACPro 0.21 ± 0.01 0.20 ± 0.00 0.21 ± 0.01 W/ MACPro 000...626262 ± ± ± 0 0 0...060606 000...808080 ± ± ± 0 0 0...020202 000...636363 ± ± ± 0 0 0...050505 PP W/o MACPro 0.29 ± 0.05 0.27 ± 0.04 0.30 ± 0.03 evaluation to derive the task contextualization and select the optimal head. The more episodes agents can probe, the more information about the evaluating task agents can gain. However, setting P to a very large value is not practical. We find in Fig. 8 (d) P = 20 is enough for accurate task approximation. (a) Sensitivity of 𝛼𝛼reg (b) Sensitivity of 𝛼𝛼cont𝑔𝑔 (c) Sensitivity of 𝛼𝛼cont𝑙𝑙 (d) # probe episodes 𝑃𝑃 Figure 8: Test results of parameter sensitivity studies. #### 6 Final Remarks Observing the great significance and practicability of continual learning, this work takes a further step towards continual coordination in cooperative MARL. We first formulate this problem, where agents are centralized trained with access to global information, then an efficient task contextualization learning module is designed to obtain efficient task representation, and an adaptive dynamic network expansion technique is applied, we finally design a local continual coordination mechanism to approximate the global optimal task head selection. Extensive experiments demonstrate the effectiveness of our approach. To the best of our knowledge, the proposed MACPro is the first multi-agent contin |Envs|Method|VDN QMIX QPLEX| |---|---|---| |LBF|W/ MACPro W/o MACPro|000...999222 ±±± 000...000222 000...999000 ±±± 000...000999 000...999777 ±±± 000...000333 0.21 ± 0.01 0.20 ± 0.00 0.21 ± 0.01| |---|---|---| |PP|W/ MACPro W/o MACPro|000...666222 ±±± 000...000666 000...888000 ±±± 000...000222 000...666333 ±±± 000...000555 0.29 ± 0.05 0.27 ± 0.04 0.30 ± 0.03| |---|---|---| ----- ual algorithm capable of multi-agent scenarios, which needs a heuristic-designed environment process. Future work on more reasonable and efficient ways, such as environment automatic generation or applying it to real-world applications would be of great value. #### References [1] A. Oroojlooy and D. Hajinezhad, “A review of cooperative multi-agent deep reinforcement learning,” Applied _Intelligence, pp. 1–46, 2022._ [2] G. Sartoretti, J. Kerr, Y. Shi, G. Wagner, T. S. Kumar, S. Koenig, and H. Choset, “Primal: Pathfinding via reinforcement and imitation multi-agent learning,” IEEE Robotics and Automation Letters, vol. 4, no. 3, pp. 2378–2385, 2019. [3] J. Wang, W. Xu, Y. Gu, W. Song, and T. C. Green, “Multi-agent reinforcement learning for active voltage control on power distribution networks,” in NeurIPS, 2021, pp. 3271–3284. [4] K. Xue, J. Xu, L. Yuan, M. Li, C. Qian, Z. Zhang, and Y. Yu, “Multi-agent dynamic algorithm configuration,” in _NeurIPS, 2022._ [5] G. Papoudakis, F. Christianos, A. Rahman, and S. V. Albrecht, “Dealing with non-stationarity in multi-agent deep reinforcement learning,” preprint arXiv:1906.04737, 2019. [6] J. Wang, Z. Ren, B. Han, J. Ye, and C. Zhang, “Towards understanding cooperative multi-agent q-learning with value factorization,” in NeurIPS, 2021, pp. 29 142–29 155. [7] F. Christianos, G. Papoudakis, M. A. Rahman, and S. V. Albrecht, “Scaling multi-agent reinforcement learning with selective parameter sharing,” in ICML, 2021, pp. 1989–1998. [8] C. Zhu, M. Dastani, and S. Wang, “A survey of multi-agent reinforcement learning with communication,” _preprint arXiv:2203.08975, 2022._ [9] H. Hu, A. Lerer, A. Peysakhovich, and J. N. Foerster, “”other-play” for zero-shot coordination,” in ICML, 2020, pp. 4399–4410. [10] J. Guo, Y. Chen, Y. Hao, Z. Yin, Y. Yu, and S. Li, “Towards comprehensive testing on the robustness of cooperative multi-agent reinforcement learning,” preprint arXiv:2204.07932, 2022. [11] R. Lowe, Y. Wu, A. Tamar, J. H. P. Abbeel, and I. Mordatch, “Multi-agent actor-critic for mixed cooperativecompetitive environments,” in NIPS, 2017, pp. 6379–6390. [12] C. Yu, A. Velu, E. Vinitsky, J. Gao, Y. Wang, A. Bayen, and Y. Wu, “The surprising effectiveness of PPO in cooperative multi-agent games,” in NeurIPS, 2022. [13] P. Sunehag, G. Lever, A. Gruslys, W. M. Czarnecki, V. F. Zambaldi, M. Jaderberg, M. Lanctot, N. Sonnerat, J. Z. Leibo, K. Tuyls, and T. Graepel, “Value-decomposition networks for cooperative multi-agent learning based on team reward,” in AAMAS, 2018, pp. 2085–2087. [14] T. Rashid, M. Samvelyan, C. Schroeder, G. Farquhar, J. Foerster, and S. Whiteson, “Qmix: Monotonic value function factorisation for deep multi-agent reinforcement learning,” in ICML, 2018, pp. 4295–4304. [15] R. Gorsane, O. Mahjoub, R. J. de Kock, R. Dubb, S. Singh, and A. Pretorius, “Towards a standardised performance evaluation protocol for cooperative MARL,” in NeurIPS, 2022. [16] K. Khetarpal, M. Riemer, I. Rish, and D. Precup, “Towards continual reinforcement learning: A review and perspectives,” Journal of Artificial Intelligence Research, vol. 75, pp. 1401–1476, 2022. [17] G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,” Neural Networks, vol. 113, pp. 54–71, 2019. [18] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska et al., “Overcoming catastrophic forgetting in neural networks,” Proceedings of _the national academy of sciences, vol. 114, no. 13, pp. 3521–3526, 2017._ [19] C. Kaplanis, M. Shanahan, and C. Clopath, “Policy consolidation for continual reinforcement learning,” in ICML, 2019, pp. 3242–3251. [20] E. Lecarpentier, D. Abel, K. Asadi, Y. Jinnai, E. Rachelson, and M. L. Littman, “Lipschitz lifelong reinforcement learning,” in AAAI, 2021, pp. 8270–8278. [21] D. Lopez-Paz and M. Ranzato, “Gradient episodic memory for continual learning,” in NIPS, 2017, pp. 6467– 6476. ----- [22] L. Caccia, E. Belilovsky, M. Caccia, and J. Pineau, “Online learned continual compression with adaptive quantization modules,” in ICML, 2020, pp. 1240–1250. [23] S. Sodhani, F. Meier, J. Pineau, and A. Zhang, “Block contextual mdps for continual learning,” in L4DC, 2022, pp. 608–623. [24] S. Kessler, J. Parker-Holder, P. J. Ball, S. Zohren, and S. J. Roberts, “Same state, different task: Continual reinforcement learning without interference,” in AAAI, 2022, pp. 7143–7151. [25] J.-B. Gaya, T. Doan, L. Caccia, L. Soulier, L. Denoyer, and R. Raileanu, “Building a subspace of policies for scalable continual learning,” preprint arXiv:2211.10445, 2022. [26] K. Zhang, Z. Yang, and T. Bas¸ar, “Multi-agent reinforcement learning: A selective overview of theories and algorithms,” Handbook of Reinforcement Learning and Control, pp. 321–384, 2021. [27] R. Mirsky, I. Carlucho, A. Rahman, E. Fosong, W. Macke, M. Sridharan, P. Stone, and S. V. Albrecht, “A survey of ad hoc teamwork: Definitions, methods, and open problems,” preprint arXiv:2202.10450, 2022. [28] H. Nekoei, A. Badrinaaraayanan, A. C. Courville, and S. Chandar, “Continuous coordination as a realistic scenario for lifelong learning,” in ICML, 2021, pp. 8016–8024. [29] G. Papoudakis, F. Christianos, L. Sch¨afer, and S. V. Albrecht, “Benchmarking multi-agent deep reinforcement learning algorithms in cooperative tasks,” in NeurIPS, 2021. [30] M. Samvelyan, T. Rashid, C. S. de Witt, G. Farquhar, N. Nardelli, T. G. J. Rudner, C. Hung, P. H. S. Torr, J. N. Foerster, and S. Whiteson, “The Starcraft multi-agent challenge,” in AAMAS, 2019, pp. 2186–2188. [31] L. Busoniu, R. Babuska, and B. De Schutter, “A comprehensive survey of multiagent reinforcement learnin,” _IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), vol. 38, no. 2, pp._ 156–172, 2008. [32] A. OroojlooyJadid and D. Hajinezhad, “A review of cooperative multi-agent deep reinforcement learning,” _preprint arXiv:1908.03963, 2019._ [33] H. Wang, Y. Yu, and Y. Jiang, “Fully decentralized multiagent communication via causal inference,” IEEE Trans_actions on Neural Networks and Learning Systems, 2022._ [34] J. Cao, L. Yuan, J. Wang, S. Zhang, C. Zhang, Y. Yu, and D.-C. Zhan, “Linda: Multi-agent local information decomposition for awareness of teammates,” preprint arXiv:2109.12508, 2021. [35] M. Wen, J. G. Kuba, R. Lin, W. Zhang, Y. Wen, J. Wang, and Y. Yang, “Multi-agent reinforcement learning is a sequence modeling problem,” in NeurIPS, 2022. [36] F. Zhang, C. Jia, Y.-C. Li, L. Yuan, Y. Yu, and Z. Zhang, “Discovering generalizable multi-agent coordination skills from multi-task offline data,” in ICLR, 2023. [37] X. Wang, Z. Zhang, and W. Zhang, “Model-based multi-agent reinforcement learning: Recent progress and prospects,” preprint arXiv:2203.10603, 2022. [38] X. Lyu, Y. Xiao, B. Daley, and C. Amato, “Contrasting centralized and decentralized critics in multi-agent reinforcement learning,” in AAMAS, 2021, pp. 844–852. [39] S. Zhang, L. Shen, and L. Han, “Learning meta representations for agents in multi-agent reinforcement learning,” _preprint arXiv:2108.12988, 2021._ [40] R. Qin, F. Chen, T. Wang, L. Yuan, X. Wu, Z. Zhang, C. Zhang, and Y. Yu, “Multi-agent policy transfer via task relationship modeling,” preprint arXiv:2203.04482, 2022. [41] K. Xue, Y. Wang, L. Yuan, C. Guan, C. Qian, and Y. Yu, “Heterogeneous multi-agent zero-shot coordination by coevolution,” preprint arXiv:2208.04957, 2022. [42] M. Masana, X. Liu, B. Twardowski, M. Menta, A. D. Bagdanov, and J. van de Weijer, “Class-incremental learning: survey and performance evaluation on image classification,” preprint arXiv:2010.15277, 2020. [43] D. Kudithipudi, M. Aguilar-Simon, J. Babb, M. Bazhenov, D. Blackiston, J. Bongard, A. P. Brna, S. Chakravarthi Raja, N. Cheney, J. Clune et al., “Biological underpinnings for lifelong learning machines,” _Nature Machine Intelligence, vol. 4, no. 3, pp. 196–210, 2022._ [44] D. Rolnick, A. Ahuja, J. Schwarz, T. P. Lillicrap, and G. Wayne, “Experience replay for continual learning,” in _NeurIPS, 2019, pp. 348–358._ [45] A. Chaudhry, M. Rohrbach, M. Elhoseiny, T. Ajanthan, P. K. Dokania, P. H. Torr, and M. Ranzato, “On tiny episodic memories in continual learning,” preprint arXiv:1902.10486, 2019. ----- [46] Y. Huang, K. Xie, H. Bharadhwaj, and F. Shkurti, “Continual model-based reinforcement learning with hypernetworks,” in ICRA, 2021, pp. 799–805. [47] S. Kessler, P. Miło´s, J. Parker-Holder, and S. J. Roberts, “The surprising effectiveness of latent world models for continual reinforcement learning,” preprint arXiv:2211.15944, 2022. [48] S. Lee, J. Ha, D. Zhang, and G. Kim, “A neural dirichlet process mixture model for task-free continual learning,” in ICLR, 2020. [49] Z. Wang, C. Chen, and D. Dong, “Lifelong incremental reinforcement learning with online bayesian inference,” _IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 8, pp. 4003–4016, 2022._ [50] M. Wolczyk, M. Zajac, R. Pascanu, L. Kucinski, and P. Milos, “Continual world: A robotic benchmark for continual reinforcement learning,” in NeurIPS, 2021, pp. 28 496–28 510. [51] S. Powers, E. Xing, E. Kolve, R. Mottaghi, and A. Gupta, “Cora: Benchmarks, baselines, and metrics as a platform for continual reinforcement learning agents,” in CoLLAs, 2022, pp. 705–743. [52] F. A. Oliehoek and C. Amato, A Concise Introduction to Decentralized POMDPs. Springer, 2016. [53] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in NIPS, 2017, pp. 5998–6008. [54] G. E. Hinton, “Training products of experts by minimizing contrastive divergence,” Neural computation, vol. 14, no. 8, pp. 1771–1800, 2002. [55] F.-M. Luo, T. Xu, H. Lai, X.-H. Chen, W. Zhang, and Y. Yu, “A survey on model-based reinforcement learning,” _preprint arXiv:2206.09328, 2022._ [56] S. Chopra, R. Hadsell, and Y. LeCun, “Learning a similarity metric discriminatively, with application to face verification,” in CVPR, 2005, pp. 539–546. [57] H. Jeffreys, The Theory of Probability. OUP Oxford, 1998. [58] S. Wold, K. Esbensen, and P. Geladi, “Principal component analysis,” Chemometrics and Intelligent Laboratory _Systems, vol. 2, no. 1-3, pp. 37–52, 1987._ [59] J. Wang, Z. Ren, T. Liu, Y. Yu, and C. Zhang, “QPLEX: duplex dueling multi-agent q-learning,” in ICLR, 2021. [60] K. Son, D. Kim, W. J. Kang, D. E. Hostallero, and Y. Yi, “QTRAN: Learning to factorize with transformation for cooperative multi-agent reinforcement learning,” in ICML, 2019, pp. 5887–5896. [61] K. Cho, B. Van Merri¨enboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio, “Learning phrase representations using rnn encoder-decoder for statistical machine translation,” preprint arXiv:1406.1078, 2014. [62] S. Hu, F. Zhu, X. Chang, and X. Liang, “Updet: Universal multi-agent reinforcement learning via policy decoupling with transformers,” in ICLR, 2021. [63] S. Iqbal, C. A. S. De Witt, B. Peng, W. B¨ohmer, S. Whiteson, and F. Sha, “Randomized entity-wise factorization for multi-agent reinforcement learning,” in ICML, 2021, pp. 4596–4606. ### Appendix #### G Details About Baselines and Benchmarks This part gives a detailed description about the relevant baselines, benchmarks, and the three related value-based cooperative MARL methods. **G.1** **Baselines** **Finetuning is a simple method based on a single feature extraction model and policy head to learn a sequence of tasks,** ignoring the changes in tasks and directly tuning the learned policy on the current task. However, if the current task is different from the previous ones, the parameters of the policy network would change dramatically to acquire good performance on the current task, thus inducing the phenomenon of catastrophic forgetting. **EWC [18] is one of the regularization-based approaches to address the catastrophic forgetting problem. Concretely,** it tries to maintain expertise on old tasks by selectively slowing down learning on the weights that are important for ----- (a) 5x5 grid (b) 6x6 grid Figure 9: Benchmark LBF used in this paper. Figure 10: Benchmark PP used in this paper, where the prey’s policy is to run in the opposite direction of the nearest predator. them. Specifically, the loss function for learning the current task M is _L(θ) = LM_ (θ) + _[λ]_ 2 � _Fj(θj −_ _θM_ _−1,j)[2],_ (12) _j_ where LM (θ) is the loss for task M only, and Fi is the i[th] diagonal element of the Fisher information matrix F . θM _−1_ is the saved snapshot of θ after training task M 1, and j labels each parameter. λ is a adjustable coefficient to control _−_ the trade-off between the current task and previous ones. In this paper, we set the parameters of the agent’s Q network as θ, and calculate the Fisher information matrix F with temporal difference error. Unlike EWC that constrains the change of network parameters when learning a new task, Coreset [45], one of the replay-based methods, prevents catastrophic forgetting by choosing and storing a significantly smaller subset of data of the previous tasks. When learning the current task, the stored data is also utilized for training the policy, which is expected to remember the previous tasks. In this paper, we set the replay buffer to uniformly store trajectories of all seen tasks, including the current one. A small batch of trajectories of one randomly chosen task is sampled from the buffer to train the agents’ network. **OWL [24] is a recent approach that learns a multi-head architecture and achieves high learning efficiency when the** tasks in a sequence have conflicting goals. Specifically, it learns a factorized policy with a shared feature extractor but separate heads, each specializing in only one task. With a similar architecture to our method MACPro, we can apply it to learn task sequences in a continual manner. During testing, OWL uses bandit algorithms to find the policy that achieves the highest test task reward. However, this strategy could bring performance degradation since agents choose action uniformly at the beginning of the episodes. **G.2** **Benchmarks** We select four multi-agent environments for the evaluation benchmarks. Level Based Foraging (LBF) [29] is a cooperative grid world game (see Fig. 9). Where the positions of two agents and one food are represented by discrete states, and agents are randomly spawned at cells (0, 0), (0, 1), (1, 0), (1, 1). Each agent observes the relative position ----- (a) task 5m_vs_6m in Marines series (b) task 2s3z in SZ series Figure 11: Benchmark Marines and SZ used in this paper. of other agents and the food, moves a single cell in one of the four directions (up, left, down, right), and gains reward 1 if and only if both agents navigate to the food and be at a distance of one cell from the food. In continual learning ability comparison, we design 5 tasks in a 5x5 grid, with the food at cell (0, 4), (2, 4), (4, 4), (4, 2), (4, 0) (green food in Fig. 9 (a)), respectively. To test the generalization ability of different methods, we further design 20 tasks in both 5x5 and 6x6 grid, and the food position is changed as well (red food in Fig. 9 (a)(b)). **Predator Prey (PP) [11] is another popular benchmark where three agents (predators) need to chase the adversary** agent (prey) and collide it to win the game (see Fig. 10). Here agents and landmarks are represented by circles with different sizes and colliding means circles’ intersection. Positions of the two fixed landmarks, positions and speed of the predators and the prey are encoded into continuous states. The predators and the prey can accelerate in one of the four directions (up, left, down, right). In different tasks, the position of landmarks, the predators and the prey’s acceleration, maximum speed, and spawn areas, and the fixed heuristic policies the prey uses are different. Specifically, the prey (1) runs in the opposite direction of the nearest predator, (2) stays still at a position far away from the predators, (3) runs towards the nearest predator, and (4) runs in a random direction with great speed. Predators gains reward 1 if n of them collide with the prey at the same time (n = 1 in case (1)(2)(4) and n = 2 in case (3)). In the generalization test, for one original task, we create one corresponding additional task by adding one constant ξ to the original value of different task parameters, including landmark’s size, x-coordinate, y-coordinate, predator’s and prey’s size, acceleration, and maximum speed. We set ξ = 0.01, 0.02, 0.03 on the four original tasks to derive 20 _±_ _±_ addition tasks to test the generalization ability. The other benchmarks are two task series, named Marines and SZ (Fig. 11), from StarCraft Multi-Agent Challenge (SMAC) [30], involving various numbers of Marines, Stalkers/Zealots in two camps, respectively. The goal of the multi-agent algorithm is to control one of the camps to defeat the other. Agents receive a positive reward signal by causing damage to enemies, killing enemies, and winning the battle. On the contrary, agents receive a negative reward signal when they receive damage from enemies, get killed, and lose the battle. Each agent observes information about the map within a circular area around it, and takes actions, including moving and firing when it is alive. In continual learning ability comparison, the Marines series consists of (1) 5m vs 6m, (2) 13m, (3) 4m, and (4) 8m vs 9m, and the SZ series consists of (1) 2s1z vs 3z, (2) 2s3z, (3) 3s5z, and (4) 2s2z vs 4s, where m stands for marine, which can attack an enemy unit from a long distance at a time, s stands for the stalker, which attacks like a marine and has a self-regenerate shield, and z stands for zealot, which also has a self-regenerate shield but can only attack an enemy unit from a short distance. For the generalization test, we first decrease the default sight range and shoot range by 1 to create four additional tasks for both Marines and SZ. Then, we design scenarios 3m, 5m, 6m, 7m, 8m, 9m, 10m, _{_ 11m, 12m, 4m vs 5m, 6m vs 7m, 7m vs 8m, 9m vs 10m, 10m vs 11m, 11m vs 12m, 12m vs 13m for Marines, _}_ and scenarios 1s1z, 1s2z, 1s3z, 2s1z, 2s2z, 2s4z, 3s1z, 3s2z, 3s3z, 3s4z, 4s2z, 4s3z, 4s4z, 2s2z vs 4z, 3s3z vs 6s, _{_ 4s4z vs 8z for SZ. If vs in the task name, it indicates that the two camps are asymmetric, e.g., in 5m vs 6m, there are _}_ 5 marines in our camp and 6 enemy marines. Otherwise, it indicates that the two camps are symmetric, e.g., in 2s3z, there are 2 stalkers and 3 zealots in both camps. **G.3** **Value Function factorization MARL methods** As we investigate the integrative abilities of MACPro in the manuscript, here we introduce the value-based methods used in this paper, including VDN [13], QMIX [14], and QPLEX [59]. The difference among the three methods lies in the mixing networks, with increasing representational complexity. Our proposed framework MACPro follows the _Centralized Training with Decentralized Execution (CTDE) paradigm used in value-based MARL methods._ These three methods all follow the Individual-Global-Max (IGM) [60] principle, which asserts the consistency between joint and local greedy action selections by the joint value function Qtot(τ _, a) and individual value functions_ ----- MLP | �| 𝑊𝑊1 ��� 𝑠𝑠𝑡𝑡 𝑄𝑄1(𝜏𝜏𝑡𝑡1, 𝑎𝑎1𝑡𝑡 )���𝑄𝑄𝑛𝑛(𝜏𝜏𝑡𝑡𝑛𝑛, 𝑎𝑎𝑡𝑡𝑛𝑛) |𝑄𝑄𝑡𝑡𝑡𝑡𝑡𝑡(𝝉𝝉𝑡𝑡, 𝒂𝒂𝑡𝑡) 𝑄𝑄𝑖𝑖(𝜏𝜏𝑡𝑡 𝑖𝑖, 𝑎 𝑄𝑄𝑡𝑡𝑡𝑡𝑡𝑡(𝝉𝝉𝑡𝑡, 𝒂𝒂𝑡𝑡)|Col2|𝑎𝑎𝑡𝑡 𝑖𝑖)| |---|---|---| |MLP 𝑏2𝑏 𝑠𝑡𝑠𝑡 Mixing Network 𝜋𝜋 | ȉ | MLP 𝑊2𝑊 MLP MLP 𝑏1𝑏 𝑄1𝑄(𝜏𝑡𝜏1 𝑡, 𝑎𝑎𝑡1 𝑡) ȉȉȉ 𝑄𝑛𝑄𝑛(𝜏𝑡𝜏𝑛 𝑡𝑛, 𝑎𝑎𝑡𝑛 𝑡𝑛) ℎ𝑡𝑖𝑖 𝑡−1 GRU | ȉ |||𝜀𝜀 ℎ𝑡𝑖𝑖 𝑡| ||| ȉ | MLP 𝑊2𝑊 MLP 𝑏1𝑏 | ȉ ||| Agent 1 ��� Agent 𝑛𝑛 (𝑜𝑜𝑡𝑡1, 𝑎𝑎1𝑡𝑡−1) ��� (𝑜𝑜𝑡𝑡𝑛𝑛, 𝑎𝑎𝑡𝑡−1𝑛𝑛 ) 𝑄𝑄𝑖𝑖(𝜏𝜏𝑡𝑡𝑖𝑖, 𝑎𝑎𝑡𝑡𝑖𝑖) 𝜋𝜋 MLP GRU MLP (𝑜𝑜𝑡𝑡𝑖𝑖, 𝑎𝑎𝑡𝑡−1𝑖𝑖 ) (a) Mixing Network (b) Overall Structure (c) Agent Network Figure 12: The overall structure of QMIX. (a) The detailed structure of the mixing network, whose weights and biases are generated from a hyper-net (red) which takes the global state as the input. (b) QMIX is composed of a mixing network and several agent networks. (c) The detailed structure of the individual agent network. �Qi(τ _[i], a[i])�ni=1[:]_ _∀τ ∈_ **_T, arg maxQtot(τ_** _, a) =_ **_a∈A_** � � (13) arg maxQ1 �τ [1], a[1][�] _, . . ., arg maxQn (τ_ _[n], a[n])_ _._ _a[1]∈A_ _a[n]∈A_ **VDN [13] factorizes the global value function Q[VDN]tot** [(][τ] _[,][ a][)][ as the sum of all the agents’ local value functions]_ �Qi(τ _[i], a[i])�ni=1[:]_ _n_ _Q[VDN]tot_ (τ _, a) =_ � _Qi_ �τ _[i], a[i][�]_ _._ (14) _i=1_ **QMIX [14] extends VDN by factorizing the global value function Q[QMIX]tot** (τ _, a) as a monotonic combination of the_ agents’ local value functions �Qi(τ _[i], a[i])�ni=1[:]_ _i_ _, [∂Q]tot[QMIX](τ_ _, a)_ _> 0._ (15) _∀_ _∈N_ _∂Qi (τ_ _[i], a[i])_ We mainly implement MACPro on QMIX for its proven performance in various papers and its overall structure is 𝑄𝑄[𝑡𝑡𝑡𝑡𝑡𝑡](𝝉𝝉, 𝒂𝒂) 𝑄𝑄[1](𝜏𝜏[1], 𝑎𝑎[1]) 𝑄𝑄[𝑛𝑛](𝜏𝜏[𝑛𝑛], 𝑎𝑎[𝑛𝑛]) 𝑠𝑠𝑖𝑖𝑎𝑎𝑎𝑎 𝑠𝑠𝑗𝑗𝑒𝑒𝑒𝑒 𝑜𝑜𝑖𝑖 𝑜𝑜𝑖𝑖 … (a) Mixing Network (b) Individual Q Network Figure 13: Network architecture used in Marines and SZ. shown in Fig. 12. QMIX uses a hyper-net conditioned on the global state to generate the weights and biases of the local Q-values and uses the absolute value operation to keep the weights positive to guarantee monotonicity. The above two structures propose two sufficient conditions of the IGM principle to factorize the global value function but these conditions they propose are not necessary. To achieve a complete IGM function class, QPLEX [59] uses a duplex dueling network architecture by decomposing the global value function as: _Q[QPLEX]tot_ (τ _, a) = Vtot(τ_ ) + Atot(τ _, a) =_ _n_ _n_ � _Qi_ �τ _, a[i][�]_ + � _i=1_ _i=1_ (16) �λ[i](τ _, a) −_ 1� _Ai_ �τ _, a[i][�]_ _,_ ----- where λ[i](τ _, a) is the weight depending on the joint history and action, Ai_ �τ _, a[i][�]_ is the advantage function conditioning on the history information of each agent. QPLEX aims to find the monotonic property between individual Q function and individual advantage function. #### H The Architecture, Infrastructure, and Hyperparameters Choices of MACPro We give detailed description of the network architecture, the overall flow, and the parameters of MACPro here. **H.1** **Network Architecture** We here give details about multiple neural networks in (1) agent networks, (2) task contextualization learning, and (3) decentralized task approximation. In benchmark LBF and PP, the number of agents, the dimension of state, observation, and action remains unchanged in different tasks. Specifically, for (1) agent networks, we apply the technique of parameter sharing and design the feature extractor φ as a 5-layer MLP and a GRU [61]. The hidden dimension is 128 for the MLP and 64 for the GRU. Then, each separated head is a linear layer which takes the output of the feature extractor as input and outputs the Q-value of all actions. For (2) task contextualization learning, we design a global trajectory encoder gθ and a contextaware forward model h. gθ consists of a transformer encoder, a MLP, and a POE module. The 6-layer transformer encoder takes trajectory τ = (s1, · · ·, sT ) as input and outputs T 32-dimensional embeddings. Then, the 3-layer MLP transforms these embeddings into means and standard deviations of T Gaussians. Finally, the POE module acquires the joint representation of the trajectory, which is also a Gaussian distribution N (µθ(τ ), σθ[2][(][τ] [))][. The context-aware] forward model h is a 3-layer MLP that takes as input the concatenation of current state, local observations, actions, and task contextualization sampled from the joint task distribution, and outputs the next state, next local observations, and reward. The hidden dimension is 64 and the reconstruction loss is calculated by mean squared error. For (3) decentralized task approximation, the local trajectory encoders fθi[′] [(][i][ = 1][,][ · · ·][, n][)][ have the same structure as the] global trajectory encoder gθ. In benchmark Marines and SZ, a new difficulty arises since the number of agents, the dimension of state, observation, and action could vary from task to task, making the networks used in LBF and PP fail to work. Inspired by the popularly used population-invariant network (PIN) technique in MARL [40, 62, 63], we design a different feature extractor, head and a monotonic mixing network [14] that learns the global Q-value as a combination of local Qvalues. For the feature extractor (see Fig. 13), we decompose the observation oi into different parts, including agent _i’s own information o[own]i_, ally information o[al]i [, and enemy information][ o]i[en][. Then we feed them into attention networks] to derive a fixed-dimension embedding e: _q = MLPq(o[own]i_ ), **Kal = MLPKal** ([o[al]i [1] _[, . . ., o]i[al][j]_ _[, . . .][ ])][,]_ **Val = MLPVal** ([o[al]i [1] _[, . . ., o]i[al][j]_ _[, . . .][ ])][,]_ T � _eal = softmax(qKal_ _/_ _dk)Val,_ (17) **Ken = MLPKen** ([o[en]i [1] _, . . ., o[en]i_ _[j]_ _, . . . ]),_ **Ven = MLPVen** ([o[en]i [1] _, . . ., o[en]i_ _[j]_ _, . . . ]),_ T � _een = softmax(qKen_ _/_ _dk)Ven,_ _e = [MLP(o[own]i_ ), eal, een], where [·, ·] is the vector concatenation operation, dk is the dimension of the query vector, and bold symbols are matrices. Embedding e is then fed into a MLP and a GRU to derive the output of the feature extractor φi. Finally, the output is fed into the policy head, a 3-layer MLP, to derive the Q-value. Furthermore, the dimension of states could also vary in Marines and SZ. Like the way we deal with observations, state s is decomposed into ally information _s[al]i_ [, and enemy information][ s]j[en][. Then their embeddings are fed into an attention network to derive a][ fixed-dimension] embedding es. Finally, we feed es into the original mixing network whose structure is used in benchmarks LBF and PP. Besides the networks mentioned above, the global trajectory encoder gθ, forward model h, local trajectory encoders _fθi[′]_ [are also involved with this issue. For][ g][θ][ and][ f][θ]i[′] [, we first apply the same technique to derive the fixed-dimension] embeddings of states and observations, then feed them into the transformer encoders. For the forward model h, we treat each agent’s action as a part of its own observation and feed their concatenation into the attention network to derive ----- an embedding, which will be feed into h with the embedding of state and task contextualization. Then, h outputs a fixed-dimension embedding. We decode it into the next state, local observations, and reward with task-specific MLP decoders to calculate the reconstruction loss Lmodel. **Algorithm 1 MACPro: Training** **Input: Task sequence Y = {task 1, · · ·, task M** _}_ **Initialize: trajectory encoder gθ, forward model h, individual trajectory encoders fθi[′]:n** [, agents’ feature extractors] _φ1:n_ 1: for m = 1, · · ·, M do 2: Set up task m 3: **if m = 1 then** 4: _ψ1:[1]_ _n_ _[←]_ [Initialized new head] 5: **else** 6: // Dynamic Network Expansion 7: Calculate l, l[′] according to Equation 5 8: Find k∗ = arg min1≤k≤K lk[′] 9: **if lk[′]** _∗_ _[≤]_ _[λ][new][l][k]∗_ **[then]** 10: Merge task m to the task(s) that ψ1:[k][∗]n [tasks charge,][ ψ]1:[m]n 1:n _[←]_ _[ψ][k][∗]_ 11: **else** 12: _ψ1:[m]n_ _[←]_ [Initialized new head] 13: **end if** 14: **end if** 15: (Optional) Reset the ϵ-greedy schedule 16: **for t = 1, · · ·, Ttask m do** 17: Collect trajectories with {φ1:n, ψ1:[m]n[}][, store in buffers][ D][,][ D][′] 18: Update {φ1:n, ψ1:[m]n[}][ according to][ L][RL] 19: **if t mod κ1 = 0 then** 20: // Task Contextualization Learning 21: Train gθ, h according to Lcontext 22: Train fθ1:[′] _n_ [according to][ L][approx] 23: **end if** 24: **if t mod κ2 = 0 then** 25: Save head ψ1:[m]n 26: **end if** 27: **end for** 28: Evaluate task 1, _, m according to Alg. 2_ _· · ·_ 29: Empty the experience replay buffer = _D_ _∅_ 30: end for **H.2** **The Overall Flow of MACPro** To illustrate the process of training, the overall training flow of MACPro is shown in Alg. 1. Where Lines 3 14 _∼_ express the process of dynamic agent network expansion, where we use the task contextualization to decide whether we should initialize a new head for the current task. To initialize a new head (Line 12), we propose two strategies. One strategy is to copy the parameters of the head learned from the last task. The other is to construct a entirely new head by resetting the parameters randomly. We use the first strategy in LBF, PP, Marines, and the second one in SZ. Both strategies work well in the experiments. Then, we started training on the new task. In Line 15 we can choose to reset the ϵ-greedy schedule to enhance exploration (adopted in SZ) or not (adopted in LBF, PP, Marines), where the schedule is to decay ϵ from 1 to 0.05 in 50K timesteps. Next, we iteratively update the parameters of each component in Lines 16 27, where we also save the current head. Finally, we test all seen tasks, empty the replay buffer for the _∼_ current task, and switch to the next task. Besides, the execution flow of MACPro is shown in Alg. 2. In the execution phase, for each testing task, agents first toll-out P episodes to probe the environment and derive the context (Line 3 7). With the gathered local information, _∼_ each agent independently selects an optimal head to perform on this task (Line 8, 9). ----- **Algorithm 2 MACPro: Execution** **Input: Task sequence {task 1, · · ·, task M** _}, feature extractors φ1:n, heads {ψ1:[k]_ _n[}][K]k=1_ **Parameter: Number of probing episodes P** 1: for m = 1, · · ·, M do 2: Set up task m 3: **for p = 1, · · ·, P do** 4: Randomly choose an integer k from 1, _, K_ _{_ _· · ·_ _}_ 5: Agents collect one trajectory τττ p with {φ1:n, ψ1:[k] _n[}]_ 6: Each agent i calculates the mean value µθi[′] [(][τ][ i]p[)][ of the trajectory representation][ f][θ]i[′] [(][τ][ i]p[)] 7: **end for** _bs_ 8: Each agent i selects the optimal head ψi[k][⋆i], where k[⋆i] = argmin1≤k≤K 1≤minp≤P _[||][µ][θ]i[′]_ [(][τ][ i]p[)][ −] _bs[1]_ _j�=1_ _µ[j]k[||][2][.]_ 9: Agents test with {φi, ψi[k][⋆i] _}i[n]=1_ 10: end for Table 3: Hyperparameters in experiment. Hyperparameter Value Number of testing episodes 32 _P (Number of probing episodes)_ 20 Learning rate for updating networks 0.0005 Number of heads in transformer encoders 3 Number of layers in transformer encoders 6 _bs (Batch size of the sampled trajectories)_ 32 _λnew (Threshold for merging similar tasks)_ 1.5 Hidden Dim (Dimension of hidden layers) 64 Attn Dim (Dimension of Key in attention) 8 Entity Dim (Dimension of Value in attention) 64 Z Dim (Dimension of the encoded Gaussians) 32 _κ1 (Interval (steps) of updating both encoders)_ 1000 _κ2 (Interval (steps) of saving the learning head)_ 10000 _αcontl (Coefficient of local contrastive loss Lcontl_ ) 0.1 _αcontg (Coefficient of global contrastive loss Lcontg_ ) 0.1 _αreg (Coefficient of the l2- regularization Lreg on φ)_ 500 Buffer Size of D (Maximum number of trajectories in D) 5000 Buffer Size of D[′] (Maximum number of trajectories in D[′]) 5000 **H.3** **Hyperparameters Choices** Our implementation of MACPro is based on the PyMARL[2] [30] codebase with StarCraft 2.4.6.2.69232 and uses its default hyper-parameter settings. For example, the discounted factor used to calculate the temporal difference error is set to the default value 0.99. The selection of the additional hyperparameters introduced in our approach, e.g., time interval of saving the heads, is listed in Tab. 3. We use this set of parameters in all experiments shown in this paper except for the ablations. #### I The Complete Continual Learning Results In this part, we compare MACPro against the multiple mentioned baselines and ablations to investigate the continual learning ability, and display the performance on every single task seen so far in Fig. 14 16. _∼_ [2https://github.com/oxwhirl/pymarl](https://github.com/oxwhirl/pymarl) |Hyperparameter|Value| |---|---| |Number of testing episodes P (Number of probing episodes) Learning rate for updating networks Number of heads in transformer encoders Number of layers in transformer encoders bs (Batch size of the sampled trajectories) λ (Threshold for merging similar tasks) new Hidden Dim (Dimension of hidden layers) Attn Dim (Dimension of Key in attention) Entity Dim (Dimension of Value in attention) Z Dim (Dimension of the encoded Gaussians) κ (Interval (steps) of updating both encoders) 1 κ (Interval (steps) of saving the learning head) 2 α (Coefficient of local contrastive loss Lcontl) contl α (Coefficient of global contrastive loss Lcontg) contg α reg (Coefficient of the l 2- regularization Lreg on φ) Buffer Size of D (Maximum number of trajectories in D) Buffer Size of D′(Maximum number of trajectories in D′)|32 20 0.0005 3 6 32 1.5 64 8 64 32 1000 10000 0.1 0.1 500 5000 5000| |---|---| ----- |MACPro Oracle Figure 15: Th|MACP|ro Oracle| |---|---|---| |MACPro Oracle Figure 16:|MACP|ro Oracle| |---|---|---| ----- #### J Product of a Finite Number of Gaussians Suppose we have N Gaussian experts with means µi1, µi2, · · ·, µiN and variances σi[2]1[, σ]i[2]2[,][ · · ·][, σ]iN[2] [, respectively.] Thus the product distribution is still Gaussian with mean µi and variance σi[2][:] � _µi1_ � _µi =_ + _[µ][i][2]_ + · · · + _[µ][iN]_ _σi[2][,]_ _σi[2]1_ _σi[2]2_ _σiN[2]_ (18) 1 1 = [1] + [1] + + _._ _· · ·_ _σi[2]_ _σi[2]1_ _σi[2]2_ _σiN[2]_ It can be proved by induction. _Proof. We want to prove Eqn. 18 is true for all N_ 2. _≥_ - Base case: Suppose N = 2 and p1(x) = N (x|µ1, σ1), p2(x) = N (x|µ2, σ2), then _p1(x)p2(x) =_ _−_ [(][x][ −] _[µ][1][)][2]_ 2σ1[2] _−_ [(][x][ −] _[µ][2][)][2]_ 2σ2[2] � 1 _·_ _√_ exp 2πσ2 � � 1 _√_ exp 2πσ1 � � (x − _µ1)[2]_ + [(][x][ −] _[µ][2][)][2]_ _−_ 2σ1[2] 2σ2[2] �� 1 = exp 2πσ1σ2 1 = exp 2πσ1σ2 � 1 = exp 2πσ1σ2  2 [+][µ][2][σ]1[2] 1[σ]2[2][+][µ][2]2[σ]1[2]  _x[2]_ _−_ 2 _[µ][1]σ[σ]1[2][2][+][σ]2[2]_ _x +_ _[µ][2]σ1[2][+][σ]2[2]_ − 1 _[σ]2[2]_  2 _σ[σ]1[2][2][+][σ]2[2]_  �x − _[µ][1]σ[σ]21[2][2][+][+][µ][σ]2[2][2][σ]1[2]_ �2  − 2 _σ[σ]1[2]1[2][+][σ][σ]2[2]2[2]_ _−_ [(][µ]2[1][ −]σ1[2][σ][µ]2[2][2][)][2]  (19) � � � 1 = exp � 2π (σ1[2] [+][ σ]2[2][)]  1 _·_ _√2π_ _√σσ11[2]σ[+]2[σ]2[2]_ _· exp_ − _−_ [(][µ][1][ −] _[µ][2][)][2]_ 2 (σ1[2] [+][ σ]2[2][)] �2    �x − _[µ][1][σ]σ21[2][2][+][+][µ][σ]2[2][2][σ]1[2]_ 1 _[σ]2[2]_ 2 _σ[σ]1[2][2][+][σ]2[2]_ �x − _[µ][1]σ[σ]12[2][2][+][+][σ][µ]2[2][2][σ]1[2]_ 1 _[σ]2[2]_ 2 _σ[σ]1[2][2][+][σ]2[2]_   _._  �2 1 = A · _√2π_ _√σσ11[2]σ[+]2[σ]2[2]_  exp  _−_  Eqn. 19 can be seen as PDF of N (µ, σ) times A where µ = ( _σ[µ]1[1][2]_ [+][ µ]σ2[2][2] [)][σ][2][,][ 1]σ[2][ =] _σ11[2]_ [+] _σ12[2]_ _[.]_ - Induction step: Suppose it is true when N = n, and the product distribution of n Gaussian experts has mean _µ˜ = (_ _σ[µ]1[1][2]_ [+][ · · ·][ +][ µ]σn[n][2] [)˜][σ][2][ and variance] _σ˜1[2][ =]_ _σ11[2]_ [+][ · · ·][ +] _σ1n[2]_ [, then for][ n][ + 1][ Gaussian experts:] 1 1 1 = [1] + + [1] + _,_ _· · ·_ _σ[2][ = 1]σ˜[2][ +]_ _σn[2]+1_ _σ1[2]_ _σn[2]_ _σn[2]+1_ � _µ˜_ � � _µ1_ _µ =_ _σ[2]_ = + + _[µ][n]_ + _[µ][n][+1]_ _· · ·_ _σ˜[2][ +][ µ]σn[n][2]+1[+1]_ _σ1[2]_ _σn[2]_ _σn[2]+1_ (20) � _σ[2]._ - Eqn. 18 has been proved by the above derivation. ----- If we write Tij = (σij[2] [)][−][1][, Eqn. 18 can be written as:] � �[N] �� �[N] �−1 _µi =_ _µijTij_ _Tij_ _,_ _j=1_ _j=1_ (21) � �[N] �−1 _σi[2]_ [=] _Tij_ _,_ _j=1_ and is exactly what we’re trying to prove. -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2305.13937, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": null, "url": "" }
2,023
[ "JournalArticle" ]
false
2023-05-07T00:00:00
[ { "paperId": "3bfff39e9db142b56de1ab87b051fa68d04881e0", "title": "A communication-based identification of critical drones in malicious drone swarm networks" }, { "paperId": "24d67d0f938b74393b05315be2f63422a64a5773", "title": "Robust Multi-agent Communication via Multi-view Message Certification" }, { "paperId": "e1f2f782b85af692db2cd6121cc4277c65843563", "title": "Multilayer Network Community Detection: A Novel Multi-Objective Evolutionary Algorithm Based on Consensus Prior Information [Feature]" }, { "paperId": "4251052867fb5a14b616a3ed0856a968bf5d5d4e", "title": "Expert System-Based Multiagent Deep Deterministic Policy Gradient for Swarm Robot Decision Making" }, { "paperId": "6895c7e5143d64d23cd1a1bd721edc07be8b6f02", "title": "The Effectiveness of World Models for Continual Reinforcement Learning" }, { "paperId": "bdcd8dd1c2051f063b651e3e94b47596c9827a3b", "title": "Building a Subspace of Policies for Scalable Continual Learning" }, { "paperId": "497f580ebfe562149b4aa886c48c466aa1313d40", "title": "Multiagent Reinforcement Learning With Heterogeneous Graph Attention Network" }, { "paperId": "85b2dd15c3f1024c3651eb258eafa4d1358c187e", "title": "Multi-agent Dynamic Algorithm Configuration" }, { "paperId": "94310aa68c694b8b87d59f3d3f8b1b3fd5a4c901", "title": "Towards a Standardised Performance Evaluation Protocol for Cooperative MARL" }, { "paperId": "3fe351b19fb9d35f9d947cd93b5a378631ca3388", "title": "Dynamics-Adaptive Continual Reinforcement Learning via Progressive Contextualization" }, { "paperId": "59714a5c3af2778fdf888760a20ca87b5e7f6358", "title": "Heterogeneous Multi-agent Zero-Shot Coordination by Coevolution" }, { "paperId": "b6b6bc529e665ebf97326d084a71159634ae10a7", "title": "A Survey on Model-based Reinforcement Learning" }, { "paperId": "43ca5ca1f840dd7eaade725789868e8998aacfc8", "title": "Game-Based Backstepping Design for Strict-Feedback Nonlinear Multi-Agent Systems Based on Reinforcement Learning" }, { "paperId": "fe26607ca95c0d1d9005810b4ad12845ee69e9cf", "title": "Multi-Agent Reinforcement Learning is a Sequence Modeling Problem" }, { "paperId": "bd01cc1a04a6fc8e9e5948ded5a1ad05b8dde3d7", "title": "Fully Decentralized Multiagent Communication via Causal Inference" }, { "paperId": "3992a87e7c53faaf1ac357006edac9b833ffb0bb", "title": "Towards Comprehensive Testing on the Robustness of Cooperative Multi-agent Reinforcement Learning" }, { "paperId": "fc9e25150df546bf7a6745e3a8b547ad96b3ded1", "title": "A Novel Representation Learning for Dynamic Graphs Based on Graph Convolutional Networks" }, { "paperId": "24ed28d44c33de145b05217814d678e3486acbc5", "title": "Model-based Multi-agent Reinforcement Learning: Recent Progress and Prospects" }, { "paperId": "d60742f6c1df48bc820a26564ae51cf527732ae2", "title": "A survey of multi-agent deep reinforcement learning with communication" }, { "paperId": "f4ae07a57e1cfe39d45f982ce32056b5ebe53ce8", "title": "Multi-Agent Policy Transfer via Task Relationship Modeling" }, { "paperId": "f295b3e251f4cc2bb5e866330ba74b582b978aa5", "title": "Biological underpinnings for lifelong learning machines" }, { "paperId": "b73e775c8d4da08bc33deee07e0f9759a9e22dfa", "title": "Game of Drones: Multi-UAV Pursuit-Evasion Game With Online Motion Planning by Deep Reinforcement Learning" }, { "paperId": "e14be0c78e9576ac11d09c55189770f203872248", "title": "Event-Triggered Communication Network With Limited-Bandwidth Constraint for Multi-Agent Reinforcement Learning" }, { "paperId": "22108c16ca1941c3cdf3a62b232269516fa5f237", "title": "Multi-Agent Reinforcement Learning for Active Voltage Control on Power Distribution Networks" }, { "paperId": "7a9846fbb9a580f522ff93f201a6bf15f80d112b", "title": "CORA: Benchmarks, Baselines, and Metrics as a Platform for Continual Reinforcement Learning Agents" }, { "paperId": "99b9595ac6a13bf986db40a301e3f11442eec36b", "title": "Block Contextual MDPs for Continual Learning" }, { "paperId": "38ff8a98fce54c61362eec2f0362774f647e4bf7", "title": "LINDA: multi-agent local information decomposition for awareness of teammates" }, { "paperId": "61e6674bcdf297dda6744c8fe69cbc0d8ec6006c", "title": "Learning Meta Representations for Agents in Multi-Agent Reinforcement Learning" }, { "paperId": "fa583b02b83e81982da3bea64fcb6899d9c5cd20", "title": "UNMAS: Multiagent Reinforcement Learning for Unshaped Cooperative Scenarios" }, { "paperId": "b5cf7d647d0910180a543cf6ff80e0ee08d6ba46", "title": "Hierarchical and Stable Multiagent Reinforcement Learning for Cooperative Navigation Control" }, { "paperId": "94cefa04e0f834272d85cc425e0adfb27fd17e08", "title": "Same State, Different Task: Continual Reinforcement Learning without Interference" }, { "paperId": "090273ad6b3720027e34f9183576dd2812bb4454", "title": "Continual World: A Robotic Benchmark For Continual Reinforcement Learning" }, { "paperId": "beb09ed0e6272d2ff96f1d5966a8bfea8d5fb7e8", "title": "Multiagent Meta-Reinforcement Learning for Adaptive Multipath Routing Optimization" }, { "paperId": "16bc9c08dbbca2f4f0567b2f5e1de8ee2fc1b72a", "title": "Continuous Coordination As a Realistic Scenario for Lifelong Learning" }, { "paperId": "3a315c81a98851f0614c09fef6a14c30d6a1e63c", "title": "The Surprising Effectiveness of PPO in Cooperative Multi-Agent Games" }, { "paperId": "742e45f9ccf43fdfb3bc565e3bd0f7252902c0d3", "title": "Scaling Multi-Agent Reinforcement Learning with Selective Parameter Sharing" }, { "paperId": "58e4959fbaf98547507755c3de68ebd2e117ac89", "title": "Contrasting Centralized and Decentralized Critics in Multi-Agent Reinforcement Learning" }, { "paperId": "d5250c59351e5cc7ef842a3c4c89c1a62bd45180", "title": "UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers" }, { "paperId": "9faecf3e18a833f2d49b030d591cc2ded0b54336", "title": "Towards Continual Reinforcement Learning: A Review and Perspectives" }, { "paperId": "0d67345f2f682e6d8149c66b35a4713b8cfec775", "title": "Class-Incremental Learning: Survey and Performance Evaluation on Image Classification" }, { "paperId": "e10682f8999bd5f05d995732cb418df7ae45d8a7", "title": "Continual Model-Based Reinforcement Learning with Hypernetworks" }, { "paperId": "052c100d45f949c06e8419b504e319b442cb3f0a", "title": "QPLEX: Duplex Dueling Multi-Agent Q-Learning" }, { "paperId": "a365d7819697984d8e7e2ae65ec5a1f15f4e3e25", "title": "Lifelong Incremental Reinforcement Learning With Online Bayesian Inference" }, { "paperId": "4ec9c22bb689dd136726cfe4159344066f681cfb", "title": "Benchmarking Multi-Agent Deep Reinforcement Learning Algorithms in Cooperative Tasks" }, { "paperId": "08f1140524e075c3de22c508277df2a7e598d5f4", "title": "Randomized Entity-wise Factorization for Multi-Agent Reinforcement Learning" }, { "paperId": "0c0df99ac6c0091172ba424cc024da417408fbaa", "title": "Towards Understanding Cooperative Multi-Agent Q-Learning with Value Factorization" }, { "paperId": "1b6ed62002979ae366854deb475b3413667e3f2b", "title": "\"Other-Play\" for Zero-Shot Coordination" }, { "paperId": "207c7b8ea8f94463383a089e4f7f24b64503f9c0", "title": "Lipschitz Lifelong Reinforcement Learning" }, { "paperId": "8904e9986ea256f37da535f282c7cd727736db4a", "title": "A Neural Dirichlet Process Mixture Model for Task-Free Continual Learning" }, { "paperId": "54d4a221db5a91a2487b1610374843fafff5a23d", "title": "Multi-Agent Reinforcement Learning: A Selective Overview of Theories and Algorithms" }, { "paperId": "386620b2572266baad688b288807f61e5a84719e", "title": "Online Learned Continual Compression with Adaptive Quantization Modules" }, { "paperId": "060571527cdf3036adb78911c5b1c065b92c4714", "title": "SMIX(λ): Enhancing Centralized Value Functions for Cooperative Multiagent Reinforcement Learning" }, { "paperId": "f7c15e9ac6653330b7dd18a89301a3b333927db3", "title": "A review of cooperative multi-agent deep reinforcement learning" }, { "paperId": "615e443f15778e9fdde27fecebd5c6d028816e27", "title": "Dealing with Non-Stationarity in Multi-Agent Deep Reinforcement Learning" }, { "paperId": "56ca9b304e0476feaa6dd4fd1cccc8c0a1a9d8eb", "title": "QTRAN: Learning to Factorize with Transformation for Cooperative Multi-Agent Reinforcement Learning" }, { "paperId": "9c54962b0fd011d5fe3f5b5275cc6ba091a2c7ae", "title": "On Tiny Episodic Memories in Continual Learning" }, { "paperId": "82055eed2ba0d7156a54c586249742c848e5d565", "title": "The StarCraft Multi-Agent Challenge" }, { "paperId": "c5765a6c1844743337cee4fec985c748af3b1070", "title": "Policy Consolidation for Continual Reinforcement Learning" }, { "paperId": "d9ff7a9344dd5d6653bd7a02bfd704422bb29951", "title": "Experience Replay for Continual Learning" }, { "paperId": "4fe9142a47b35638edcc59013ad5e9e87bd93ea7", "title": "PRIMAL: Pathfinding via Reinforcement and Imitation Multi-Agent Learning" }, { "paperId": "c4e824a574d396803cf4677b7d0ad4e28ad54804", "title": "Value-Decomposition Networks For Cooperative Multi-Agent Learning Based On Team Reward" }, { "paperId": "ffc211476f2e40e79466ffc198c919a97da3bb76", "title": "QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning" }, { "paperId": "9ea50b3408f993853f1c5e374690e5fbe73c2a3c", "title": "Continual Lifelong Learning with Neural Networks: A Review" }, { "paperId": "204e3073870fae3d05bcbc2f6a8e263d9b72e776", "title": "Attention is All you Need" }, { "paperId": "7c3ece1ba41c415d7e81cfa5ca33a8de66efd434", "title": "Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments" }, { "paperId": "118fae4b4d07453561f1eded88654f812c7c61ec", "title": "Gradient Episodic Memory for Continual Learning" }, { "paperId": "2e55ba6c97ce5eb55abd959909403fe8da7e9fe9", "title": "Overcoming catastrophic forgetting in neural networks" }, { "paperId": "836ad0c693bc5ef171ee2b07b3f4d1bd2a0ae24c", "title": "A Concise Introduction to Decentralized POMDPs" }, { "paperId": "0b544dfe355a5070b60986319a3f51fb45d1348e", "title": "Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation" }, { "paperId": "f61efe204165b40b5fd942f110cf3656e09aa1b0", "title": "Principal Component Analysis (PCA)" }, { "paperId": "4aece8df7bd59e2fbfedbf5729bba41abc56d870", "title": "A Comprehensive Survey of Multiagent Reinforcement Learning" }, { "paperId": "cfaae9b6857b834043606df3342d8dc97524aa9d", "title": "Learning a similarity metric discriminatively, with application to face verification" }, { "paperId": "52070af952474cf13ecd015d42979373ff7c1c00", "title": "Training Products of Experts by Minimizing Contrastive Divergence" }, { "paperId": "cc95b268a5f0ac69af25194dc0c7b052ec8d2303", "title": "Discovering Generalizable Multi-agent Coordination Skills from Multi-task Offline Data" }, { "paperId": "e4faee8a8b0db2ce29d5eb7290e63980d1336149", "title": "A Survey of Ad Hoc Teamwork: Definitions, Methods, and Open Problems" }, { "paperId": "2e1ca97d8f7604e3b334ff4903bfa67267379317", "title": "The Surprising Effectiveness of Latent World Models for Continual Reinforcement Learning" }, { "paperId": "c3e151f71168a5f348bdebfde11752ca603fa6d0", "title": "Theory of Probability" }, { "paperId": null, "title": "received the B.Sc. and M.Sc. degrees from the School of Mechanical Engineering and Automation, Northeastern University, Shenyang, China, in 2012 and 2015, respectively. He is" }, { "paperId": null, "title": "His research interests include multiagent reinforcement learning and multiagent systems" }, { "paperId": null, "title": "COORDINATION VIA PROGRESSIVE TASK" }, { "paperId": null, "title": "received the B.Sc. degree in from the School of Artificial Intelligence, University, Nanjing, China" } ]
24,650
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00ea729f4029076d418123a919825ae23de9932a
[]
0.854739
EVOLUTION AND ANALYSIS OF SECURED HASH ALGORITHM (SHA) FAMILY
00ea729f4029076d418123a919825ae23de9932a
Malaysian Journal of Computer Science
[ { "authorId": "144611259", "name": "B. Khan" }, { "authorId": "9255543", "name": "R. F. Olanrewaju" }, { "authorId": "2139481280", "name": "Malik Arman Morshidi" }, { "authorId": "2833169", "name": "R. N. Mir" }, { "authorId": "122232289", "name": "M. L. Mat Kiah" }, { "authorId": "2179526146", "name": "Abdul Mobeen Khan" } ]
{ "alternate_issns": null, "alternate_names": [ "Malays J Comput Sci" ], "alternate_urls": null, "id": "7ed515a5-673b-43d5-8f58-9f0e1363926e", "issn": "0127-9084", "name": "Malaysian Journal of Computer Science", "type": "journal", "url": "https://ejournal.um.edu.my/index.php/MJCS/" }
With the rapid advancement of technologies and proliferation of intelligent devices, connecting to the internet challenges have grown manifold, such as ensuring communication security and keeping user credentials secret. Data integrity and user privacy have become crucial concerns in any ecosystem of advanced and interconnected communications. Cryptographic hash functions have been extensively employed to ensure data integrity in insecure environments. Hash functions are also combined with digital signatures to offer identity verification mechanisms and non-repudiation services. The federal organization National Institute of Standards and Technology (NIST) established the SHA to provide security and optimal performance over some time. The most well-known hashing standards are SHA-1, SHA-2, and SHA-3. This paper discusses the background of hashing, followed by elaborating on the evolution of the SHA family. The main goal is to present a comparative analysis of these hashing standards and focus on their security strength, performance and limitations against common attacks. The complete assessment was carried out using statistical analysis, performance analysis and extensive fault analysis over a defined test environment. The study outcome showcases the issues of SHA-1 besides exploring the security benefits of all the dominant variants of SHA-2 and SHA-3. The study also concludes that SHA-3 is the best option to mitigate novice intruders while allowing better performance cost-effectively.
**EVOLUTION AND ANALYSIS OF SECURE HASH ALGORITHM (SHA) FAMILY** **_Burhan Ul Islam Khan[1*], Rashidah Funke Olanrewaju[2], Malik Arman Morshidi[3], Roohie Naaz Mir[4], Miss Laiha_** **_Binti Mat Kiah[5] and Abdul Mobeen Khan[6 ]_** 1,2,3Department of ECE, KOE, International Islamic University Malaysia (IIUM), Kuala Lumpur, 50728, Malaysia 4Department of Computer Science and Engineering, National Institute of Technology (NIT), Srinagar, 190006, India 5Department of Computer System & Technology, Universiti Malaya (UM), Kuala Lumpur, 50603, Malaysia 6 ALEM Solutions and Technologies, Aspen Commercial Tower, Dubai, 11562, United Arab Emirates Email: burhan.iium@gmail.com[1*] (corresponding author), frashidah@iium.edu.my[2], mmalik@iium.edu.my[3], naaz310@nitsri.net[4], misslaiha@um.edu.my[5], consultmobeen@gmail.com[6] DOI: https://doi.org/10.22452/ mjcs.vol35no3.1 **ABSTRACT** _With the rapid advancement of technologies and proliferation of intelligent devices, connecting to the internet_ _challenges have grown manifold, such as ensuring communication security and keeping user credentials secret._ _Data integrity and user privacy have become crucial concerns in any ecosystem of advanced and interconnected_ _communications. Cryptographic hash functions have been extensively employed to ensure data integrity in insecure_ _environments. Hash functions are also combined with digital signatures to offer identity verification mechanisms_ _and non-repudiation services. The federal organization National Institute of Standards and Technology (NIST)_ _established the SHA to provide security and optimal performance over some time. The most well-known hashing_ _standards are SHA-1, SHA-2, and SHA-3. This paper discusses the background of hashing, followed by elaborating_ _on the evolution of the SHA family. The main goal is to present a comparative analysis of these hashing standards_ _and focus on their security strength, performance and limitations against common attacks. The complete assessment_ _was carried out using statistical analysis, performance analysis and extensive fault analysis over a defined test_ _environment. The study outcome showcases the issues of SHA-1 besides exploring the security benefits of all the_ _dominant variants of SHA-2 and SHA-3. The study also concludes that SHA-3 is the best option to mitigate novice_ _intruders while allowing better performance cost-effectively._ **_Keywords:_** **_Secured Hash Algorithms, Message Digest, Cryptographic Hashing, Statistical Analysis, Fault_** **_Analysis_** **1.0** **INTRODUCTION** The Internet is more critical in the current era than ever before, and almost everyone uses Internet services for different purposes. In most cases, the data traversing over the Internet comprises private or concealed information that everybody wants to protect [1], [2]. The rapid advancement in technologies has raised significant concerns about protecting our data from unauthorized access. With advanced software programs, malicious users can intercept data in transit or breach the confidentiality of data stored on distributed storage systems such as the cloud [3]. Therefore, protecting its information is pivotal in safeguarding any organization or individual's company and client assets. Security usually means using the best preventive and defensive measures to protect data from unauthorized access [4]. Cryptography is the science of security applications that offers a mechanism for keeping users' information secure with privacy and confidentiality [5]. As an imperative sub-area of fast communication, cryptography allows a protected communication process between different users so that malicious users cannot access the contents inside the file [6]. However, this secure communication area has a long history of achievements and failures. Several encoding messages have appeared over the eras and are continually broken after a while [7]. A cryptographic hash algorithm aims to provide secure communication using the digest of messages to generate a hash value of data to detect unauthorized attempts to data. Hash algorithms are widely used in many applications, such as data integrity and corruption verification, ownership protection, authentication, and many more [8]. The system of cryptocurrency also depends on the hashing algorithms and functions. 179 ----- The hashing algorithms are vital cryptographic primitives that accept input, process it and generate a digest [9]. The digest is a fixed-size alphanumeric string that humans cannot understand. Some older and newer versions of cryptographic hash functions are available such as the Message-Digest Algorithm 5 (MD5), MD4, SHA family, and many other hashing functions [10]. Hash is a computationally secure form of compression concerning security aspects [11]. There are different variants of SHA witnessed to date, e.g., SHA-1, SHA-2, SHA-256, SHA-348, and SHA-512. With the availability of discussion carried out by various research works and legal authorities, e.g., NIST, there are constantly evolving versions of secured encryption techniques [12], [13]. The research in this direction resulted in the formation of SHA-3, a new version. However, owing to the novel nature of the structure of this encryption technique, the degree of strength and weakness of this algorithm is still unknown. At present, there are various methods where SHA variants have been reportedly used for catering to different security demands, e.g., authentication of the electronic document [14], improving security system using chaotic map [15], secured message hiding [16], image security [17], improving anonymity [18], improving security by combining hashing with genetic science [19], strengthening security via blockchain [4], [20], authentication over the mobile network [21]. Apart from this, there is also literature [22], [23] on evaluating the strength of SHA. However, there is no explicit discussion to assess the strength of the three most essential SHA family members. The prime objective of this paper is to present a comparative assessment of SHA algorithms besides emphasizing the adoption of SHA-3. The specific goals of the paper are: - Assess the effectiveness of all the SHA versions over a similar test environment to chalk out a certain level of inference, and - Perform statistical analysis, performance analysis and extensive fault analysis to assess the strength of SHA approaches. The paper adopts a simple experimental design methodology considering three variants as well as sub-classes of SHA, i.e. SHA-1, SHA-2, and SHA-3. SHA-3 is exhibited to perform well in contrast to its counter-SHA1 and SHA2 versions with respect to minimal randomness of hashes, increased clock per byte, and maximal recovery of bits. Unlike any existing investigations, the proposed study contributes towards a generalized mathematical approach considering the standardized evaluation method to prove that SHA-3 is a robust and well-performing algorithm in a contemporary state. This outcome increases the higher adoption rate for any form of application that demands optimal security measures. The remaining sections of the paper are organized as follows: Section 2 provides a brief discussion on the hashing operation, while it is more elaborately discussed concerning its all-around aspect in section 3 with respect to the significant SHA variant. The proposed manuscript evaluates all the multiple variants of SHA via statistical analysis discussed in section 4, performance analysis addressed in section 5, and extensive fault analysis discussed in section 6. The outline of the study outcome is briefed in section 7, while section 8 provides conclusive remarks. References are listed at the end. **2.0** **STUDY BACKGROUND** The adoption of hashing mechanism is an integral part of the majority of security application designs. Technically, a hash function can be defined as a mathematical function that converts a numerical input value into another compressed numerical value. An arbitrary length is considered the input of a hash function, while a fixed length always characterizes hashing. Fig. 1 highlights the usage of messages of random length, which results in a fixedlength hash value after being subjected to a hash function. 180 ----- #### Message of arbitrary #### Hash #### Hash Value of fixed length #### 15 |arbitrary length|Col2|Function (H)|fixed lengt (h)|xed lengt (h)|Col6| |---|---|---|---|---|---| |John Smith|||00 01 02 03 04 05 - - 15|00|| ||||||| |||||01|| |Sandra Dee|||||| |||||15|| Fig. 1: Conventional Hashing Mechanism To develop it as an efficient cryptographic tool, specific essential properties must feature a standard hash function. The first important property is preimage resistance, which induces a computational challenge to reverse the hash function operation [24]. This property is essential as it safeguards from an intruder who has possession of hash value only and attempts to obtain the input data. The second important property is called second preimage resistance, which ensures a higher degree of complexity in getting hash from different inputs for a given input and hash [25]. This security property of hash ensures safety from an intruder who has both input value and its respective hash value. In contrast, the intruder is interested in replacing the alternative value as a legitimate value in the position of the source input value. The third property of hashing is Collision Resistance, which ensures complexity in obtaining two different input values of any length yielding the same hash [26]. This property provides that the collision is highly complex, making it difficult for the intruder to explore two input values to possess the same hash. Using these security properties makes it feasible to construct multiple security applications [27], [28]. One of the essential applications of hashing is to carry out password storage, while another frequently adopted application is data integrity. Most hashing can be utilized for authentication systems because the aforementioned two standards are frequently employed in applications [29]. The construction mechanism of hashing is also quite simplified as well as unique. Referring to Fig. 1, it can be seen that a mathematical function is quite essential to generating hashes. The sample message of arbitrary length, i.e., "John Smith," "Lisa Smith," "Sam Doe" and "Sandra Dee," will formulate blocks of data that further result in hash codes as an outcome. Usually, the size of such a data block resides in a range of 128-512 bits. At the same time, the hash function of different types constitutes to become a hashing algorithm. The processing of hashing operation consists of multiple rounds similar to block cipher. The process rounds continue until the complete messages are subjected to hash. 181 ----- |Seed Value|Col2| |---|---| ||| |Message Block-1|Col2| |---|---| ||| |Hash Function|Col2| |---|---| ||| |Message Block-2|Col2| |---|---| ||| ### - ### Block-n |Message Block-n|Col2| |---|---| ||| Fig. 2: Construction Process of Hash The process flow shown in Fig. 2 is also called as avalanche effect of hashing [30]. Due to the fact that the hash value associated with the first block of the message acts as an input file for the second hash operation and impacts the consecutive function, this effect generates unique hash values for messages without any form of mapping relation with each other. As a result, the two hash values differ by at least one bit of data. The construction process of hashing also has unique differences between hash algorithms and hash functions. The hash algorithm is defined as a complete process that formulates breaking down the message block. It also establishes the result obtained from the previous block of the message and their connectivity with the other generated blocks of messages. The hash function generates a hash code subjected to multiple blocks associated with a fixed-length of binary data as an outcome. Therefore, a hash function needs to map the anticipated input over its range of outcomes as evenly as feasible. The robustness of hashing is carried out considering theoretical and practical approaches. Theoretical methods will mainly compute the probability of mapping all the keys in a single slot, whereas the practical process will consist of assessing the longest probe sequence. Usually, the practical method consists of uniform hashing, which means that any key value can be used to map specific slots with maximum probability. However, such probability of a real-time attack is significantly small, and hence, hashing is one of the cost-effective mechanisms adopted in network security. **3.0** **SHA FAMILY** The adoption of hashing is witnessed in improving security over different application variants [31], [32], [33], [34], [35]. The hashing algorithm effectively identifies and validates that a legal person forwards the data received by a user, or the received information is authentic and not altered [36]. Many algorithms have been introduced to generate hash values, some of which were rejected, and some have become standards. Message Digest (MD) and SHA (SHA-1 and SHA-2) are the popularly accepted standards for generating hash values [37], [38]. It is always necessary to upgrade to a new technique or replace an existing one to meet the latest requirements with technology developments. A hash algorithm is an explicit cryptographic operation that alters an arbitrary-length input message to a fixed compressed numerical value called a hash value. The fundamental attributes of the hash algorithms are that it is difficult to find two different messages with the same hash value [39]. A typical process of the hash algorithm is exhibited in Fig. 3. 182 ----- # f(x) ### Plain Message # Hashing Algorithm f(x) ### Hash Function Fig. 3: A Typical Process of Hash Algorithm ### Hashed Message Value A hash function is a mathematical operation that takes in plain message data of variable length and provides a fixed size of the hash message value. Numerically, the process of a hash algorithm can be expressed as follows: 𝐻(𝑓𝑥): {0,1}[∗] →{0,1}𝑛 (1) In expression (1), {0,1}[∗] indicates the set of arbitrary-length elements, including the empty string, and {0,1}𝑛 refers to elements with length 𝑛. Therefore, a hash function maps a set of fixed-length elements to arbitrary-length elements. However, the length of the output hashed values mainly depends on the type of hash algorithm used. Generally, the size of the hash algorithms ranges between 160 bits and 1088 bits. To obtain a preset fixed-length hash value, the input message must be divided into fixed-size data blocks because the hash function receives data with a fixed length, as shown in Fig. 4. ### Hash Function ### Hashed Message Value ### Data Blocks Fig. 4: A Typical Process of Fixed-Size Data Blocks The data block sizes vary depending on the type of algorithm used. For instance, let's consider the case of SHA-1, which takes plain text messages in the block size of 512 bits. If the plain text message's length is 512-bit, the hash function 𝐻(𝑓𝑥) executes 80 rounds (80 rounds means one-time execution). If the plain text message's length is 1024-bit, then the message will be divided into two fixed-length block sizes of 512-bit and the hash function executing two times. However, the size of plain text messages is rarely in multiple of 512-bit. A technique called padding is considered to divide input plain text messages into fixed-length data blocks for other cases. The process of the fixed-length block is demonstrated in Fig. 4. The plain text message is first divided into multiple fixed-length block sizes, and the output of the first data block and the consecutive data block is fed to the hash function. Therefore, the final output is the shared value of all blocks. If the alteration of one bit anywhere in the text message, the overall hashed value will be changed, called the avalanche effect [40], [41]. **3.1** **Properties of Hash Function** A hash function should satisfy the following properties [42], [43]. 183 ----- - Resistance to Collision Attacks: This attribute indicates that two different messages should not have the same hash value. An adversary can't identify the same hash value 𝐻(𝑚) for two messages (𝑚1, 𝑚2). Collision attacks refer to a process when a pair of different messages with the same hash 𝐻(𝑚) conflict attacks occur. The hash function must have the exact property of not constructing the same hashed value for two different messages. - Resistance to Preimage Attacks: This property refers to the function of a hash algorithm strong enough that it will be challenging to generate the corresponding message for a given hash value. Therefore, the hash function should have preimage resistance. Preimage refers to a message that hashed to a provided value. This attack generally considers that at least one message is hashed to a given value. Therefore, an adversary cannot obtain the original data from the given hash value. - Resistance to Second Preimage Attacks: Given a message, it must not be possible to identify another message that will construct the same hash value as the previous message. Therefore, the hash function must have the ability to resist the second preimage attack. A second preimage refers to a message that hashes to the same value as a given message. **3.2** **Secure Hash Algorithms** SHA is a cluster of hash mathematical functions released by NIST as a US Federal Information Processing Standard. NIST first proposed the SHA-1 hash algorithm in 1995. The design principle of this algorithm is based on Merkle Damgard's structure as MD5 (Fig. 5). The algorithm uses a string of any length and provides a 160-bit compressed MD for variable-length input messages. Usually, in this scheme, the message is padded with 1; the required 0's are added to make the message length equal to 64 bits (less than an even multiple of 512). Sixty-four bits representing the length of the original message are added to the end of the padded message, which is processed in 512-bit data blocks. The following are the functional steps involved in SHA-1: The first step is padding bits to the end of the message length. The next step is subjected to the process of length appending. The third step is about dividing the input message into 512-bit data blocks. In this step of the hashing algorithm, chaining variables are initialized, i.e., internal state size. This means that number of bits is carried over to the next block. Fig. 5 shows the operation of SHA-1. In SHA-1, five chaining variables of 32 bits are taken, i.e., each equals 160 bits of the total. The final step is block processing. In this step, chaining variables are copied, and the data block of 512 bits is divided into 16 sub-blocks processed for 80 rounds for single-time execution. SHA-1 is used in a wide range of security applications and protocols, including Transport Layer Security and SSL, PGP, SSH, S/MIME and IPsec. However, this hash algorithm is vulnerable to collision attacks due to cryptographic flaws. In the past few years, confirmed cases of collision attacks have been reported for this algorithm. As a result, after 2010, most encryption users no longer recognize this standard. NIST introduced the SHA-2 hashing algorithm in 2002 with hash code lengths of 224, 256, 348 and 512. The design principle of this algorithm is based on the similar approach of the MD5 and SHA-1. However, this algorithm is more robust than SHA-1 due to the inclusion of a nonlinear function in the compression module [31]. Fig. 6 shows the operation of SHA-2. 184 ----- |A|B|C|D|E| |---|---|---|---|---| **Wt** **Kt** |A|B|C|D|E| |---|---|---|---|---| Fig. 5: One SHA-1 Round [37] ## E F G E F G Fig. 6: One SHA-2 Round [22] **Wt** |A|B|C|D|E|F|G|H| |---|---|---|---|---|---|---|---| **Kt** |A|B|C|D|E|F|G|H| |---|---|---|---|---|---|---|---| In SHA-2, the message is padded with 1; the required number of 0's are added to make its length equal to sixty-four bits (less than an even multiple of 512). Sixty-four bits representing the length of the original message are added to the end of the padded message, which is processed in 512-bit data blocks. Each data block contains 1024 bits as a sequence of 64-bit words and supports SHA-512 and SHA-384. However, SHA-2 is not preferred to ensure integrity, as it has less time-efficiency than SHA-1 due to the absence of multithreading. Besides, SHA-256 for hash cash is primarily used in cryptocurrency to achieve security on the transactions among peers in the Bitcoin network. Both SHA-512 and SHA-256 are new functions that use different constants and shift amounts and differ in the number of rounds. A research study carried out has reported that SHA-2 is vulnerable to attacks. Following is the overview of the computing steps involved in SHA-2: - The padded length - The total message length to be hashed should be a multiple of 1024 bits. 185 ----- - The padding is initiated by adding 1 as the first bit followed by the required number of 0's with 128 bits message {𝑚1, 𝑚2 … 𝑚𝑛} - Initial hash value, 𝐻[(0)] 𝑡𝑜 𝐻[(𝑖)] = 𝐻[(𝑖−1)] + 𝐶𝑀(𝑖)(𝐻(𝑖−1)) - Where C indicates compression function and 𝐻[(𝑁)] is the hash value of the message (𝑚). - The output obtained from SHA-256 is the 64 bits of a message, and the six logical attributes function as 𝑎, 𝑏, 𝑐 with a 64-bit message. After a successful collision attack in SHA-1, an open competition was announced by NIST to develop SHA-3 (a new hash algorithm). A few years later, on October 2, 2012, the Keccak (as SHA-3) was announced as the competition winner. In 2015, NIST recognized SHA-3 as a standard hash function [32], [33]. Its operation is shown in Fig. 7. p0 p1 p1 z0 z1 r 0 c _f_ _f_ _f_ _f_ _f_ Fig. 7: One SHA-3 Round [36] |Col1|p0|Col3|Col4| |---|---|---|---| |0||f|| ||||| |0|||| ||||| |f|Col2| |---|---| ||| ||| |f|Col2|Col3| |---|---|---| |||| |||| |f|Col2| |---|---| ||| ||| |f|Col2| |---|---| ||| ||| SHA-3 utilizes a sponge structure that consists of two phases, the first is the absorbing phase, and the second is the squeeze phase. The message block should be XORed into sub-space in the absorption phase and converted. In the squeezing step, output blocks are analyzed from the same sub-space and transformed with state transitions. The different variants of SHA-3 consisting of SHA-3:224, SHA-3:256, SHA-3:348, SHA-3:512 generate 224, 256, 348, and 512-bit message digests, respectively. It uses a block size of 1152, 1088, 832 and 576 bits, respectively. This algorithm shows relatively low software performance compared with other hash functions. The comparative analysis of different SHA variants is demonstrated in Table 1 and Table 2 with their essential parameters. Table 1 presents a comparative analysis of different secure hashing algorithms. The study is carried out concerning the functional parameters of each technique discussed. In Table 2, a comparative analysis is given concerning the security aspect and computational parameters. Table 1: Comparative Analysis of SHA concerning their parameters **Block** **Secure Hashing** **Output Size** **Internal State** **Size** **Rounds** **Operation** **Algorithm** **(Bits)** **Size (Bits)** **(Bits)** AND, XOR, OR, Add(mod 232), SHA-1 160 160(5x32) 512 80 Rotate with no carry SHA-224 224 AND, XOR, OR, Add(mod 232), 256(8x32) 512 64 Rotate with no SHA-256 256 carry, Right Logical Shift SHA-2 SHA-384 384 AND, XOR, OR, SHA-512 512 Add(mod 264), 512(6x64) 1024 80 SHA-512/224 224 Rotate with no carry SHA-512/256 256 186 |Secure Hashing Algorithm|Col2|Output Size (Bits)|Internal State Size (Bits)|Block Size (Bits)|Rounds|Operation| |---|---|---|---|---|---|---| |SHA-1||160|160(5x32)|512|80|AND, XOR, OR, Add(mod 232), Rotate with no carry| |SHA-2|SHA-224|224|256(8x32)|512|64|AND, XOR, OR, Add(mod 232), Rotate with no carry, Right Logical Shift| ||SHA-256|256||||| ||SHA-384|384|512(6x64)|1024|80|AND, XOR, OR, Add(mod 264), Rotate with no carry| ||SHA-512|512||||| ||SHA-512/224|224||||| ||SHA-512/256|256||||| ----- |SHA-3|SHA-3-224|224|1600(5x5x64)|1152|24|AND, XOR, NOT, Rotate with no carry| |---|---|---|---|---|---|---| ||SHA-3-256|256||1088||| ||SHA-3-512|384||832||| ||SHA-3-512|512||576||| ||SHAKE128|d(arbitrary)||1344||| ||SHAKE256|d(arbitrary)||1088||| Table 2: Comparative Analysis of SHA on Security and Computational Performance **Capacity** **Performance on** **Security (in Bits** **Secure Hashing** **Against** **Skylake** **Year of** **Against Collision** **Algorithm** **Extensive** **Release** **Attack)** **Long** **Length Attacks** **8 bytes** **Message** SHA-1 <63 (collusion found) 0 3.47 52 1995 SHA-224 112 32 7.62 84.50 2004 SHA-256 128 0 7.63 85.25 2001 SHA-384 192 128(≤384) 5.12 135.75 SHA-2 2001 SHA-512 256 0 5.06 135.50 SHA-512/224 112 288 5.12 135.75 2012 SHA-512/256 128 256 5.12 135.75 SHA-3-224 112 448 8.12 154.25 SHA-3-256 128 512 8.59 155.50 SHA-3-384 192 768 11.06 164.00 SHA-3 2015 SHA-3-512 256 1024 15.88 164.00 SHAKE128 min(d/2,128) 256 7.08 155.25 SHAKE256 min(d/2,256) 512 8.59 155.50 **3.3** **Issues in SHA-1 and SHA-2** SHA-1 and SHA-2 were both introduced by the NSA and issued as patents for public use. Although SHA-1 and SHA-2 are not identical, they are designed based on the same mathematical concept, which comprises the same flaws. However, the reason that makes SHA-2 more secure than SHA-1 is that SHA-2 uses more considerable input and output, and it generates hash values with increased length. Since SHA-1 and SHA-2 are different, they hold similar algorithmic logic and eventually, specific hash lengths may be vulnerable to a similar collision attack. Since 2008, there have been public attacks on SHA-2 like SHA-1, and attacks against SHA-2 are getting severe over time. Some of the recent threats to SHA-2 were publicly declared in 2016. Likewise, it is expected that the existing hash will be attacked and become weaker over time. The federal authority of NIST selected SHA-3 as an improved hash standard, which is different from the SHA family and can be used when required. In 2010, Keccak Hash was mainly chosen as the only finalist. NIST released the standard in 2015, and SHA-3 became the certified standard. However, SHA-1 has promoted the entire world's development, and due to cryptography flaws, the significant shift towards SHA-2 took place in late 2016 and 2017. The deadline for the cut-off for obtaining SHA-1 certificates was on December 31, 2017. However, SHA-3 adoption is still in its infancy, and the reasons are listed: - The prime reason is that implementation of SHA-3 is limited to the research domain only, and most of the existing software and hardware are yet not fully ready to support it. It also requires a customized code or program for each device to support it. - If the recommended guidelines were to shift from SHA-1 to SHA-3, the hash vendors would have done this rather than moving to SHA-2 due to similar effort and cost. - Most importantly, SHA-3 is a relatively new standardized hashing technique and was released when the SHA-2 migration schemes were still being explored. 187 |Secure Hashing Algorithm|Col2|Security (in Bits Against Collision Attack)|Capacity Against Extensive Length Attacks|Performance on Skylake|Col6|Year of Release| |---|---|---|---|---|---|---| |||||Long Message|8 bytes|| |SHA-1||<63 (collusion found)|0|3.47|52|1995| |SHA-2|SHA-224|112|32|7.62|84.50|2004| ||SHA-256|128|0|7.63|85.25|2001| ||SHA-384|192|128(≤384)|5.12|135.75|2001| ||SHA-512|256|0|5.06|135.50|| ||SHA-512/224|112|288|5.12|135.75|2012| ||SHA-512/256|128|256|5.12|135.75|| |SHA-3|SHA-3-224|112|448|8.12|154.25|2015| ||SHA-3-256|128|512|8.59|155.50|| ||SHA-3-384|192|768|11.06|164.00|| ||SHA-3-512|256|1024|15.88|164.00|| ||SHAKE128|min(d/2,128)|256|7.08|155.25|| ||SHAKE256|min(d/2,256)|512|8.59|155.50|| ----- - Another reason is that SHA-2 is an improved version of SHA-1, so it is not much vulnerable to collision attacks as SHA-1. Implementation of any version of the SHA-2 hashing algorithm is sufficient in the current scenario of today's digital world. Many research studies also explored that SHA-3 is much slower than SHA-2. So, why shift over SHA-3, which is slower? **3.4** **Reason for Shifting Towards SHA-3** The advancement of technology has laid out specific requirements that drive an upgrade to a new hashing technique to cope with the requirements needed in futuristic computing. SHA-3 is a robust technique compared to the existing hashing algorithms. In upcoming years, every system will undoubtedly shift to SHA-3. However, it depends on how the actual threats and security attacks on SHA-2 keep happening. In terms of software, the existing hashing schemes on windows machines are faster than SHA-3. Nonetheless, in terms of hardware, it easily defeated existing SHA-1 and SHA-2 hashing schemes. The cryptographic procedures are gradually controlled by hardware components in the future technology like an ecosystem of the Internet of Cyber-Physical Things (IoCPT). With the advancement of technologies, the CPU or processor is getting faster and faster. So, in most scenarios, the time required to hash will not cause much burden. Furthermore, few researchers [44], [45] have explored various ways to improve the speed of SHA-3 in software. In the context of the security strength of SHA-3 few significant points are highlighted as follows: - SHA-3 provides a secure one-way hash function. More particularly, it is not susceptible to length expansion attacks. In this case, the input cannot be reconstructed using only the hash output, nor the input data can be altered by changing the hash value. Although SHA-3 offers the updated secure hash algorithm, the existing hashing algorithm SHA-2is still feasible for some applications. - NIST still believes that only two identical hash functions (i.e., SHA-256 and SHA-512) of SHA-2 are secure. - However, the newly released SHA-3 algorithm complements the existing SHA-2 while offering large varieties. SHA-3 offers various functions with different digest bit lengths, including SHA-3-224, SHA-3-256, SHA-3-384, and SHA-3-512. It also offers two flexible output functions, SHAKE-128 and SHAKE-256, where 128 and 256 extensions are security robustness factors, which can be utilized to attain global and randomized hashing, hashbased message authentication code and even for performing stream encryption. **4.0** **STATISTICAL ANALYSIS** This section discusses the statistical analysis being carried out over different variants of the SHA family. An opensource platform of Java is considered for experimenting with a standard 64-bit Windows machine. The proposed analysis considers Java Cryptography API to generate the hashes [46], assuming a sample binary string of ten thousand random messages. The digest size for input and output is retained for the more straightforward analysis. The statistical analysis is carried out for SHA-1 with 160 bits and SHA-2 and SHA-3 with 512 bits. Although there are various SHA-2 and SHA-3, the proposed analysis chooses a 512-bit variant of both. The statistical analysis in the proposed study is carried out using the series test, bits probability test, and Hamming distance test. **4.1** **Series Test** The proposed system carries out a non-parametric test that considers random samples from multiple populations associated with the cumulative distribution of continuous data. The prime idea of this test is to measure the state of the randomness of the hashes with clear insight towards studying the dependencies of each hash function. The mathematical expression of assessing test statistics ∅ is as follows: ∅= ∆𝜓 (2) 𝜎 In expression (2), Δ𝜓 is evaluated by the difference between 𝜓 and 𝜓1, which represents series cardinality depicted by one hash and anticipated cardinality of series. The variable σ represents the standard deviation. Further, the computation of 𝜓1 depends upon two parameters of the cardinality of hashes, i.e., favourable cardinality 𝑐𝑓 and total cardinality 𝑐𝑡, while their empirical relationship is as follows: 188 ----- 𝜓1 = 2𝑐𝑓 (3) 𝑐𝑡 [+ 1] In expression (3), 𝑐𝑓 and 𝑐𝑡 are computed as (𝑐0 ∗𝑐) and (𝑐0 + 𝑐), where 𝑐0 and 𝑐 represent the cardinality of subsequences which has all the 0 and 1, respectively. Considering the significance level of 5%, the proposed system carries out the series test. It will mean that the hypothesis for random creation of hash will be proven effective if the absolute value of test statistics ∅ is more than test statistics corresponding to 0.975, which will be 1.96. Table 3: Accomplished Numerical Outcome of Series Test **Hash Function** **Item** **SHA-1** **SHA-2** **SHA-3** Maximum 5.34 5.15 5.41 Minimum 0 0 0 Average 0.81 0.91 0.81 Standard Deviation ±0.71 ±0.72 ±0.61 From the above numerical outcomes (Table 3), SHA-3 is the lowest value of standard deviation that is statistically significant to exhibit the best test outcome compared to the SHA-1 and SHA-2 family of hashes. **4.2** **Bits Probability Test** This is the second statistical test carried out in the proposed system to assess the predictability of the bits present in the MD. As a result, the study computes the probability of one second for every position of the message bit. Accordingly, the ideal condition will confirm 0 in 50% probability while 1 in another 50% probability. Mathematically, it can be represented as follows: |Item|Hash Function|Col3|Col4| |---|---|---|---| ||SHA-1|SHA-2|SHA-3| |Maximum|5.34|5.15|5.41| |Minimum|0|0|0| |Average|0.81|0.91|0.81| |Standard Deviation|±0.71|±0.72|±0.61| 𝑃𝑟𝑜𝑏𝑖(𝑖) = ∑𝑚𝑎𝑥𝑗=1 ℎ[𝑗][𝑖] (4) 𝑚𝑎𝑥 In the above expression (4), the computation of probability 𝑃𝑟𝑜𝑏 is carried out by considering the maximum position of bit 𝑚𝑎𝑥 and a maximum length of hash. The expression also represents ℎ as the hash values of the table associated with yielded MD. The assessment considers the max value to be 10000 in expression (3). The proposed analysis is carried out by considering test statistics |∅| discussed in expression (4) in the following sub-section. With an anticipated outcome of 50% probability, the assessment is regarded as a pass if the value |∅| is found to be less than 1.96. The numerical outcome is shown in Table 4. Table 4: Accomplished Numerical Outcome of Bit Prediction **Probability** **Item** **SHA-1** **SHA-2** **SHA-3** Maximum 52.46 52.57 52.85 Minimum 49.31 49.51 49.87 Average 51.15 49.01 50.01 Standard Deviation ±0.53 ±0.54 ±0.52 A closer look at the above numerically tabulated value shows the average value in the proximity of 50% with a reduced standard deviation. It can also be noted that the performance of both SHA-1 and SHA-2 are nearly equivalent, where the analysis found 510 bits to possess a probability value that is different in comparison to 50%. While, the computation of absolute test statistics |∅| for SHA-1 and SHA-2 is found to be 1.05 and 0.82, respectively, representing the higher predictability property of both hash functions. On the other hand, the numerical outcome of SHA-2 is also found slightly to be equivalent to SHA-3. It was found that 507 bits out of 512 bits do not meet the condition of 50% probability with a minor difference. The analyzed test statistics value of |∅| is 0.328, demonstrating a lower prediction probability score than SHA-1 and SHA-2. Based on this test, it can be said that the proposed outcome exhibits lower predictability for SHA-3 than SHA-1 and SHA-2. 189 |Item|Probability|Col3|Col4| |---|---|---|---| ||SHA-1|SHA-2|SHA-3| |Maximum|52.46|52.57|52.85| |Minimum|49.31|49.51|49.87| |Average|51.15|49.01|50.01| |Standard Deviation|±0.53|±0.54|±0.52| ----- **4.3** **Hamming Distance Test** This is the third statistical analysis in the proposed evaluation to identify the significant impact of minor alteration in data towards outcome hash files. This assessment form involves applying a t-student test to compute the test statistics |∅|. This test constructs a hypothesis that the hash function can successfully pass through if the numerical score of the test statistic resides between 0 and 1.96 of the confidence intervals. The anticipated outcome of the test statistics is equivalent to half the hash size with 5% of the significance level. The mathematical formulation for this assessment of test statistics |∅| can be represented as shown in expression (5): |∅| = | 𝑣𝑎𝑣−𝑣𝜎 𝑒𝑥 √𝑠𝑠𝑖𝑧𝑒| (5) The prime purpose of this part of the analysis is to compute the Hamming distance between the hash. The operation carried out in this analysis is: considering 𝑋1 and 𝑋2 are primary and secondary bits of information, then the length of 𝑋1 is equivalent to the size of 𝑋2. Therefore, Hamming distance computation is carried out as 𝑋3 = 𝑋1 ⊕𝑋2. This distance represents the cardinality of position with discretely different values for both 𝑋1 and 𝑋2 . The numerical outcomes of Hamming distance are shown in Table 5. Table 5: Accomplished Numerical Outcome of Series Test **Hash Function** **Item** **SHA-1** **SHA-2** **SHA-3** Maximum 5.34 5.15 5.41 Minimum 0 0 0 Average 0.81 0.91 0.81 Standard Deviation ±0.71 ±0.72 ±0.61 From the numerical outcomes of Hamming distance in Table 5, it can be seen that absolute test statistics |∅| for SHA-1 is about 1.28, which will mean that a minor alteration in the input data will affect 50% of the hashes of SHA1. Regarding SHA-2, the hash variation is sometimes closer to and lower than 50%, while the average test statistic |∅| is about 0.159. This outcome shows SHA-2 to perform better than SHA-1. From the viewpoint of SHA-3, in the majority of the cases, the critical values of test statistics are found to be 0.44, which exhibits SHA-2 is still the best outcome. At the same time, SHA-3 and SHA-1 occupy the third and second positions, respectively. **5.0** **PERFORMANCE ANALYSIS** The performance analysis of the SHA family is carried out considering the number of cycles utilized for data (byte) processing. Unlike conventional analysis mechanisms using data transmitted per unit time, the proposed analysis considers the processor's clock cycles to perform data processing. To compute the number of cycles utilized in processing one unit of data associated with the encrypted hash function, the amount of the data processed is divided by the cycles it consumes to process the data. It is to be noted that this performance parameter could significantly differ from one to another encryption option. The prime justification behind adopting this performance parameter is that it offers the possibility to compute total duration, which will mean that the processor with maximal frequency will exhibit better performance. Furthermore, this performance parameter will directly indicate the effective processing capability of all kinds of cryptographic operations. The computation of the cycles per data is carried out as follows: |Item|Hash Function|Col3|Col4| |---|---|---|---| ||SHA-1|SHA-2|SHA-3| |Maximum|5.34|5.15|5.41| |Minimum|0|0|0| |Average|0.81|0.91|0.81| |Standard Deviation|±0.71|±0.72|±0.61| 𝑐𝑦𝑐𝑙𝑒𝑠𝑝𝑒𝑟𝑑𝑎𝑡𝑎= (𝐷ℎ∗𝜆) (6) 𝑙𝑒𝑛 Expression (6) shows that cycles per data depend on the duration of hashing operation 𝐷ℎ, CPU frequency 𝜆, and length of input message 𝑙𝑒𝑛. To carry out this performance analysis, the proposed study considers cycles per data (or byte) as the prime observational values in the presence of 1 kB, 1 MB, and 64 MB input message size in a typical 64-bit windows environment. This evaluation method will give a complete idea of the scalability performance of the SHA approach over a similar test environment. 190 ----- (a) (b) 191 ----- (c) Fig. 8: Performance Analysis for a) 1 KB, b) 1 MB, and c) 64 MB input data The outcome exhibited in Fig. 8 highlights three different test cases to assess the clock usage per byte involved for three different variants of the hash function. Fig. 8(a) shows that SHA-1 has the lower occupation of around 200 cycles per byte, which is the best result compared to the other two variants of the hash function, i.e., SHA-2 and SHA-3. A closer look into SHA-2 shows no significant increase in cycles ranging between 700-800 cycles per byte. At the same time, SHA-3 and its three variants range between 890-1100 cycles per byte. A similar trend is observed in Fig. 8(b) and Fig. 8(c). Another observation among the three graphical outcomes shows an increase in cycles per byte with the input message size. Despite the reduced occupation of cycles per byte by SHA-1, it cannot be considered adequate concerning its security operation. It is known that SHA-1 is anticipated to be operational in no two segments that operate through the internal process, which is again expected to be equivalent to the same hash. It is also known that the hash of SHA-1 usually is 160 bits long (i.e., a string of 160 zeroes and ones), referring to the presence of 2160 or 1.4 quindecillion possibilities of hash. This is significantly less when compared with SHA-2 variants. Another significant observation is that theoretically, SHA-2 offers better performance than SHA-3 variants, especially regarding SHA-2-224 and SHA-2-512. SHA-2 variants utilize a structure of Davies-Meyer considering a block cipher that is constructed from MD4. On the other hand, SHA-3 variants make use of sponge structure considering the permutation of Keccak. This makes the complete internal structure different for both the variants of SHA-2 and SHA-3. There are no concrete criteria to confirm that anyone has more potential based on the outcome. SHA-1 possesses structural weakness, making it vulnerable to attacks like brute force. On the other hand, SHA-2 and SHA-3 are the same and cannot be considered to possess structural weaknesses. On the contrary, SHA-3 is slower than SHA-2, as exhibited by more cycles per byte in Fig. 8(b) and Fig. 8(c). Hence, from performance analysis based on cycles per byte, SHA-2 confirms to offer better performance. However, from a security viewpoint, there are few potential benefits of SHA-3 compared to SHA-2. The prime contribution of SHA-3 is its Keccak sponge that can be utilized both in the form of Media Access Control (MAC) and hash, unlike SHA-2, which also results in an increase of cycles per byte in the presented outcome of Fig. 8. It is used as a function that can derive the secret key costeffectively. Although SHA-3 has more cycle per byte inclusion and will demand slightly more resources, they are still cost-effective security solutions. This outcome shows the higher applicability of SHA-3 on Internet of Things (IoT) systems, which requires using the low-powered device with cost-effectiveness. As the construction of SHA-3 completely differs from SHA-2 variants, it is unlikely to apply SHA-3 in case of new intruder breaks in SHA-2 or vice-versa. Therefore, both SHA-2 and SHA-3 must be emphasized equally for making a more straightforward transition to SHA-3 until SHA-3 is proven to be 100% effective from a performance and security viewpoint. **6.0** **EXTENSIVE FAULT ANALYSIS** An extensive fault analysis using differential and Algebraic Fault Analysis (AFA) strategies to evaluate the strength of all three variants of the SHA algorithm was carried out. This part of the analysis considered the message size for 192 ----- SHA-2 and SHA-3 limited to 224 and 256 only as there were no significant differences in the outcome achieved in 256 and 512 message sizes for two variants of SHA, i.e., SHA-2 and SHA-3. The prime objective of the Differential Fault Analysis (DFA) was to retrieve the information of an inner state by using the difference in the correlation among the defective resultants and all intermediate parameters in contrast to the appropriate resultant. DFA was initially used to analyze block ciphers' strength, DES algorithm, hash functions and stream ciphers [47]. On the other hand, a different variant called Algebraic Fault Analysis (AFA) integrates algebraic crypto analysis with fault injection; and an algebraic expression is used over a finite field geometry to translate faults and cryptographic function. Technology of Satisfiability (SAT) Solver or Satisfiability Modulo Theories (SMT) is often used for recovering the message or secret key in such a mechanism [48]. It has been noted that analysis is more effective when carried out with AFA than DFA as the solver's complete propagation of fault reduces the dependency on human intervention to introduce the intrusion. Moreover, the application of AFA was found to automate the DFA mechanism toward hash function, stream ciphers and block ciphers [47]. The recovery problem associated with an internal state has been investigated in the proposed system concerning SHA-1, SHA-2 (224, 256), and SHA-3 (224, 256). This was carried out using the boolean expression for the operation, followed by using SAT solver to explore the solution for all parameters connected with confidential information. It was carried out by constructing an equation set for all hash functions associated with faulty and appropriate execution. The expression for a proper hash 𝐻 considering the input of 𝐿22𝑖 as: 𝐻= 𝐴. 𝐵 (7) In expression (7), the variable 𝐴= 𝑓(𝜓23, 𝐵, 𝑆, 𝑅, 𝐿) and 𝐵= 𝑓(𝜓22, 𝐵, 𝑆, 𝑅, 𝐿(𝐿22𝑖 )), where 𝜓23/ 𝜓22 represents binary XOR operation, 𝐵 represents a binary operation in rows over state bits, 𝑆 represents in-slice permutation over state bits, 𝑅 represents rotation over state bits 𝐿 represents linear operation with all input bits and single output bits. The above expression is slightly modified to include a fault in the form of 𝛥𝐿_𝑖^22 to generate a fault hash of: 𝐻1 = 𝐴. 𝐵1 (8) In expression (8), 𝐵1 represents 𝐵1 = 𝑓(𝜓22, 𝐵, 𝑆, 𝑅, 𝐿(𝐿_𝑖^22⨁𝛥𝐿_𝑖^22)). Considering the individual cases of all the variants of a hash function, the value of 𝐻 and 𝐻1 in expressions (1) and (2) can be 𝑥-bits. A closer look into this strategy will showcase that the value of 𝛥𝐵_𝑖^23 can be extracted using H and H1 differential while the same attacker can use 𝐵_𝑖^23 to launch an attack. 𝐵𝑜(𝑥, 𝑦, 𝑧) = 𝐵𝑖(𝑥, 𝑦, 𝑧) ⨁ 𝐵𝑖(𝑥+ 2, 𝑦, 𝑧) ⨁ 𝐵𝑖(𝑥+ 1, 𝑦, 𝑧). 𝐵𝑖(𝑥+ 2, 𝑦, 𝑧) (9) However, the expression (9) mentioned above differs for all the three hash functions, but it will perform faster for AFA, recovering 𝐵_𝑖^22 bits of data. It should be carefully noted that there is a significant level of interdependencies towards all the internal states with reversible functions. It will mean that a compromising attempt toward the hash algorithm can be carried out by extracting only one internal state. The solution to the proposed scheme of fault analysis will be just one uniquely recovered bit of information. Therefore, there is a need to explore 𝐵_𝑖^22 bits of information in the influence of fault injection. The SAT solver is executed to explore the initial rounds of solution and then reduce them to non-repeating bits. 193 ----- (a) DFA AFA **100** **90** **80** **70** **60** **50** **40** **30** **20** **10** **0** **SHA1** **SHA-224** **SHA-256** **SHA3-224** **SHA3-256** #### Hash Digest (b) Fig. 9: Comparative Analysis of Fault Ratio (a) For 8 Bits and (b) For 16 Bits The outcome shown in Fig. 9 highlights that both the variants of SHA-3 offer a higher value of fault ratio than the two variants of SHA-2 and SHA-1 when assessed using 8 and 16-bit values, respectively. An intruder can recognize the injected fault associated with a particular injection. The ratio of effective fault can be computed by considering a cumulative number of feasible faults for a specific number of positions with all possible fault values. This outcome showcases that adopting the SHA-3 variant for higher message size results in higher recovery of bits (Fig. 10) irrespective of AFA and DFA consideration. The outcome eventually infers that SHA-3 variants are always practical to adopt while dealing with a new generation of attacks; however, SHA-2 variants will also be capable enough to deal with previously reported attacks in literature. 194 ----- (a) SHA1 SHA-224 SHA-256 SHA3-224 SHA3-256 **0** **10** **20** **40** **60** **80** **100** #### Injected Faults (b) Fig. 10: Comparative Analysis of Recovered Bits: a) For AFA and b) For DFA 195 ----- (a) **8000** **7000** **6000** **5000** **4000** **3000** **2000** **1000** **0** **10** **20** **40** **60** **80** **100** **120** #### Injected Faults (b) Fig. 11: Comparative Analysis of Processing Time (a) For AFA and (b) For DFA The next part of the analysis investigates the recovered bits (Fig. 10). It is feasible for attackers to inject a different number of faults associated with the same input and execute all hash variant functions targeting the retrieval of the internal state of 𝐵_𝑖^22. The result is presented for the recovery process of SHA-1, SHA-224, SHA-256, SHA-3224, and SHA-3-256 to see the higher capability of AFA compared to DFA with more 𝐵_𝑖^22 bits. The outcome shows that SHA-3 offers lesser dependencies of less fault injection than other SHAs to recover the total value of the bit state. At the same time, analysis is carried out regarding the time required to perform cryptanalysis. (Fig. 11) showcases no significant difference between AFA and DFA performance for SHA-3 and SHA-2, but SHA-3 offers more time than SHA-2 variants. This is because the SAT solver required more time to recognize the faults, followed by recovering the internal states of the bits. Hence, the proposed investigation shows that SHA-3 offers better security features than other SHA variants. **7.0** **STUDY OUTCOME** After performing an extensive assessment of the SHA family, the following learning outcomes have been drawn: 196 ----- - The statistical analysis exhibits that SHA-3 has the lowest randomness of hashes than SHA-2 in the series test. Again, SHA-3 performance was found to have a lower predictability score, as seen from the bit probability test. This signifies that even with a less state of randomness, SHA-3 is hard to predict for a given number of bits present in the MD. This outcome makes SHA-3 more robust and resilient from unknown attackers. The Hamming distance test shows better results for SHA-2 compared to SHA-3. However, the sponge construction of SHA-3 as an internal structure will offer more security strength for novice forms of attackers. Hence, SHA-3 is the best complimentary solution on top of the SHA-2 approach. - The performance analysis exhibits that all the variants of SHA-3 offer increased clock per byte in contrast to SHA-1 and SHA-2. This outcome was justified by the internal structure of SHA-3, making it more secure due to the presence of MAC and the hash, which is not the case with SHA-2. Irrespective of more cycles per byte, SHA-3 is still the best option for offering authentication and data integrity. Due to its flexible structure, SHA-3 provides equivalent performance to non-anonymity, non-repudiation and privacy from maximum attacker forms. - The extensive fault analysis exhibits a higher recovery of bits for SHA-3, while the processing time for the SHA-3 is nearly the same for all of its respective variants. Although the processing time is slightly higher with increasing injected faults than SHA-2 and SHA-1, SHA-3 is still the preferred algorithm for resisting higher-end intruders with low faults in the cumulative outcome. Based on the above outcomes, it can be said that SHA-3 should be used in a specific environment where the attacker's strategy is less known. If the attacker strategy is well defined, SHA-2 will be quite enough to mitigate such attacks. SHA-1 is not recommended due to the limitation observed from the analysis discussed in prior sections. Hence, the outcome suggests wiser use of SHA-3 in a complex attack environment, whereas SHA-2 can still be used for normal to medium vulnerability. **8.0** **CONCLUSION** Hash functions play a vital role in network and communication security. The study and its contribution are manifold: - The analysis offers a justified outcome towards the usage of different variants of SHA, which has not been reported in existing research publications, - The implementation considers a uniform test environment subjected to SHA family using three different analysis mechanisms to offer validated outcome, - The outcome removes the myth that one variant of SHA is more effective than its counterparts instead, it concludes that `o` SHA-1 is less efficient in the majority of performance attribute while `o` SHA-2 and SHA-3 bears potential to resist the majority of the threats, - The study outcome continues to offer proof that SHA-3 is in a better position to mitigate the new definition of the attacker. At the same time, SHA-2 will be suitable enough to deal with the existing version of the known attacker, - The study also offers information that although SHA-3 has slightly more processing time, it can still be considered to look into its potential in other performance metrics. The main problem of any hashing algorithm is to check the integrity and authenticity of the data transmitted between two communicating nodes. This paper discussed the main highlights of cryptographic hash functions, the widespread use of various well-known hash functions and associated attacks. The study also carried out a comparative analysis of different hash algorithms and analyzed the trend of progress in this field. Studies on hash functions performed by other researchers were also discussed. However, most of them were limited to theoretical implementation and not tested against collision attacks. Based on the analytical findings, it has been shown that the current status and trend toward adopting SHA-3 were efficient, safe and capable of meeting the requirements of future network applications. Certain pitfalls were observed in the usage of the SHA-2 and SHA-3 families, which are associated with partial breakdown connected to bitwise operators. The beneficial point of SHA-3 is associated with its capability of faster response time and considering the maximum size of data for performing encryption. Hence, future works should address this problem by reducing clock cycles and quicker response time. **9.0** **ACKNOWLEDGEMENT** This research was partially funded by IIUM-UMP-UiTM Sustainable Research Collaboration Grant 2020 (SRCG) under Grant ID: SRCG20-003-0003 and Fundamental Research Grant Scheme (FRGS) under Grant ID FRGS19068-0676. The authors express their personal appreciation for the effort of Ms Gousia Nissar and Ms Manasha Saqib in proofreading, editing and formatting the paper. 197 ----- **REFERENCES** [1] T. Wang, M. Bhuiyan, G. Wang, L. Qi, J. Wu and T. Hayajneh, “Preserving Balance Between Privacy and Data Integrity in Edge-Assisted Internet of Things”, IEEE Internet of Things Journal, Vol. 7, No. 4, 2020, pp. 2679-2689, doi: 10.1109/jiot.2019.2951687. [2] D. Chen, P. Bovornkeeratiroj, D. Irwin and P. Shenoy, “Private Memoirs of IoT Devices: Safeguarding User Privacy in the IoT Era”, in 2018 _IEEE 38th International Conference on Distributed Computing Systems_ _(ICDCS), Vienna, Austria, 2018, pp. 1327-1336, doi: 10.1109/ICDCS.2018.00133._ [3] Z. A. Solangi, Y. A. Solangi, S. Chandio, M. bt. S. Abd. Aziz, M. S. bin Hamzah and A. Shah, “The Future of Data Privacy and Security Concerns in Internet of Things”, in _2018 IEEE International Conference on_ _Innovative_ _Research_ _and_ _Development_ _(ICIRD),_ Bangkok, Thailand, 2018, pp. 1-4, doi: 10.1109/ICIRD.2018.8376320. [4] I. Ochôa, L. Calbusch, K. Viecelli, J. de Paz, V. Leithardt and C. Zeferino, “Privacy in the Internet of Things: A Study to Protect User's Data in LPR Systems Using Blockchain”, in 2019 17th International Conference _on_ _Privacy,_ _Security_ _and_ _Trust_ _(PST),_ _Fredericton,_ NB, Canada, 2019, pp. 1-5, doi: 10.1109/PST47121.2019.8949076. [5] M. Khari, A. K. Garg, A. H. Gandomi, R. Gupta, R. Patan and B. Balusamy, “Securing Data in Internet of Things (IoT) Using Cryptography and Steganography Techniques”, _IEEE Transactions on Systems, Man,_ _and Cybernetics: Systems, Vol. 50, No. 1, 2020, pp. 73-80, doi: 10.1109/TSMC.2019.2903785._ [6] H. Al Hamid, S. Rahman, M. Hossain, A. Almogren and A. Alamri, “A Security Model for Preserving the Privacy of Medical Big Data in a Healthcare Cloud Using a Fog Computing Facility with Pairing-Based Cryptography”, IEEE Access, Vol. 5, 2017, pp. 22313-22328, doi: 10.1109/access.2017.2757844. [7] N. Sharma, H. Parveen Sultana, R. Singh and S. Patil, “Secure Hash Authentication in IoT based Applications”, Procedia Computer Science, Vol. 165, 2019, pp. 328-335, doi: 10.1016/j.procs.2020.01.042. [8] S. Suhail, R. Hussain, A. Khan and C. S. Hong, “On the Role of Hash-Based Signatures in Quantum-Safe Internet of Things: Current Solutions and Future Directions”, IEEE Internet of Things Journal, Vol. 8, No. 1, 2021, pp. 1-17, doi: 10.1109/JIOT.2020.3013019. [9] K. Saravanan and A. Senthilkumar, “Theoretical Survey on Secure Hash Functions and Issues”, International _Journal of Engineering Research & Technology (IJERT), Vol. 2, No. 10, 2013, pp. 1150-1153._ [10] A. Jurcut, T. Niculcea, P. Ranaweera and N. Le-Khac, “Security Considerations for Internet of Things: A Survey”, SN Computer Science, Vol. 1, No. 4, 2020, pp. 1-19, doi: 10.1007/s42979-020-00201-3. [11] H. Huang, Q. Huang, F. Xiao, W. Wang, Q. Li and T. Dai, “An Improved Broadcast Authentication Protocol for Wireless Sensor Networks Based on the Self-Reinitializable Hash Chains”, Security and Communication _Networks, Vol. 2020, 2020, pp. 1-17, doi: 10.1155/2020/8897282._ [12] M. Stevens, E. Bursztein, P. Karpman, A. Albertini and Y. Markov, “The First Collision for Full SHA-1”, in _Annual International Cryptology Conference, Santa Barbara, USA, 2017, pp. 570-596, doi: 10.1007/978-3-_ 319-63688-7_19. [13] R. Martino and A. Cilardo, “SHA2 Acceleration Meeting the Needs of Emerging Applications: A Comparative Survey”, IEEE Access, Vol. 8, 2020, pp. 28415-28436, doi: 10.1109/access.2019.2920089. [14] A. Mohammed Ali and A. Kadhim Farhan, “A Novel Improvement with an Effective Expansion to Enhance the MD5 Hash Function for Verification of a Secure E-Document”, IEEE Access, Vol. 8, 2020, pp. 8029080304, doi: 10.1109/ACCESS.2020.2989050. [15] H. Liu, A. Kadir and J. Liu, “Keyed Hash Function Using Hyper Chaotic System with Time-Varying Parameters Perturbation”, _IEEE_ _Access,_ Vol. 7, 2019, pp. 37211-37219, doi: 10.1109/ACCESS.2019.2896661. [16] M. Rathor and A. Sengupta, “IP Core Steganography Using Switch Based Key-Driven Hash-Chaining and Encoding for Securing DSP Kernels Used in CE Systems”, _IEEE Transactions on Consumer Electronics,_ Vol. 66, No. 3, 2020, pp. 251-260, doi: 10.1109/TCE.2020.3006050. 198 ----- [17] J. Ouyang, X. Zhang and X. Wen, “Robust Hashing Based on Quaternion Gyrator Transform for Image Authentication”, IEEE Access, Vol. 8, 2020, pp. 220585-220594, doi: 10.1109/ACCESS.2020.3043111. [18] Y. Zhou, B. Yang, Z. Xia, Y. Mu and T. Wang, “Anonymous and Updatable Identity-Based Hash Proof System”, IEEE Systems Journal, Vol. 13, No. 3, 2019, pp. 2818-2829, doi: 10.1109/JS2018.2878215. [19] D. I. Nassr, “Secure Hash Algorithm-2 formed on DNA”, _Journal of the Egyptian Mathematical Society,_ Article No. 24, 2019, pp. 1-20, doi: 10.1186/s42787-019-0037-6. [20] Y. Lee, S. Rathore, J. H. Park, and J. H. Park, “A Blockchain-Based Smart Home Gateway Architecture for Preventing Data Forgery”, Human-centric Computing and Information Science, Vol. 10, No. 1, 2020, pp. 114, doi: 10.1186/s13673-020-0214-5. [21] M. A. Rezazadeh Baee, L. Simpson, X. Boyen, E. Foo, and J. Pieprzyk, “Authentication Strategies in Vehicular Communications: A Taxonomy and Framework”, EURASIP Journal on Wireless Communication _and Networking, Vol. 2021, No. 1, 2021, pp. 1-50, doi: 10.1186/s13638-021-01968-6._ [22] R. Martino and A. Cilardo, “SHA2 Acceleration Meeting the Needs of Emerging Applications: A Comparative Survey”, IEEE Access, Vol. 8, 2020, pp. 28415-28436, doi: 10.1109/ACCESS.2020.2972265. [23] I. E. Salem, A. M. Salman and M. M. Mijwil, “A Survey: Cryptographic Hash Functions for Digital Stamping”, _Journal of Southwest Jiaotong University, Vol. 54, No. 6, 2019, pp. 1-11, doi:_ 10.35741/issn.0258-2724.54.6.2. [24] Jianhua Mo, Xiawen Xiao, Meixia Tao and Nanrun Zhou, “Hash Function Mapping Design Utilizing Probability Distribution for Preimage Resistance”, in _2012 IEEE Global Communications Conference_ _(GLOBECOM), Anaheim, CA, 2012, pp. 862-867, doi: 10.1109/GLOCOM.2012.6503221._ [25] B. Preneel, “Second Preimage Resistance”, _Encyclopedia of Cryptography and Security, Boston, MA,_ Springer, 2005, doi: 10.1007/978-3-540-25937-4_24. [26] A. Maetouq and S. M. Daud, “HMNT: Hash Function Based on New Mersenne Number Transform”, _IEEE_ _Access, Vol. 8, 2020, pp. 80395-80407, doi: 10.1109/ACCESS.2020.2989820._ [27] W. Jing, D. Zhang and H. Song, “An Application of Ternary Hash Retrieval Method for Remote Sensing Images in Panoramic Video”, _IEEE_ _Access,_ Vol. 8, 2020, pp. 140822-140830, doi: 10.1109/ACCESS.2020.3006103. [28] H. Cui, L. Zhu, J. Li, Y. Yang and L. Nie, “Scalable Deep Hashing for Large-Scale Social Image Retrieval”, _IEEE Transactions on Image Processing, Vol. 29, 2020, pp. 1271-1284, doi: 10.1109/TIP.2019.2940693._ [29] Y. Zheng, Y. Cao and C. H. Chang, “UDhashing: Physical Unclonable Function-Based User-Device Hash for Endpoint Authentication”, IEEE Transactions on Industrial Electronics, Vol. 66, No. 12, 2019, pp. 95599570, doi: 10.1109/TIE.2019.2893831. [30] A. Biswas, A. Majumdar, S. Nath, A. Dutta and K. L. Baishnab, “LRBC: A Lightweight Block Cipher Design for Resource Constrained IoT Devices”, Journal of Ambient Intelligence and Humanized Computing, 2020, pp. 1-15, doi: 10.1007/s12652-020-01694-9. [31] P. Chanal and M. Kakkasageri, “Security and Privacy in IoT: A Survey”, _Wireless Personal_ _Communications, Vol. 115, No. 2, 2020, pp. 1667-1693, doi: 10.1007/s11277-020-07649-9._ [32] E. Molina and E. Jacob, “Software-Defined Networking in Cyber-Physical System: A Survey”, Computers & _Electrical Engineering, Vol. 66, 2018, pp. 407-419, doi: 10.1016/j.compeleceng.2017.05.013._ [33] I. Graja, S. Kallel, N. Guermouche, S. Cheikhrouhou and A. Hadj Kacem, “A Comprehensive Survey on Modeling of Cyber‐Physical Systems”, Concurrency and Computation: Practice and Experience, Vol. 32, No. 5, 2018, pp. 1-18, doi: 10.1002/cpe.4850. [34] B. U. I. Khan, R. F. Olanrewaju, F. Anwar and M. Yaacob, “Offline OTP Based Solution for Secure Internet Banking Access”, in 2018 IEEE Conference on e-Learning, e-Management and e-Services (IC3e), Langkawi, Malaysia, 2018, pp. 167-172, doi: 10.1109/IC3e.2018.8632643. 199 ----- [35] B. U. I. Khan, R. F. Olanrewaju, F. Anwar, R. N. Mir and A. Najeeb, “A Critical Insight into the Effectiveness of Research Methods Evolved to Secure IoT Ecosystem”, International Journal of Information _and Computer Security, Vol. 11, No. 45, 2019, pp. 332-354, doi: 10.1504/ijics.2019.10023470._ [36] C. Jin, “Cryptographic Solutions for Cyber-Physical System Security”, Doctoral Dissertation, University of Connecticut, Storrs, 2019. [37] A. Tawalbeh and H. Tawalbeh, “Lightweight Crypto and Security”, Security and Privacy in Cyber-Physical _Systems, John Wiley & Sons, 2017, pp. 243-261, doi: 10.1002/9781119226079.ch12._ [38] G. Sabaliauskaite and A. Mathur, “Aligning Cyber-Physical System Safety and Security”, Complex System _Design & Management, Cham, Springer, 2021, pp. 41-53, doi: 10.1007/978-3-319-12544-2_4._ [39] J. Wang, T. Zhang, J. Song, N. Sebe and H. Shen, “A Survey on Learning to Hash”, IEEE Transactions on _Pattern_ _Analysis_ _and_ _Machine_ _Intelligence,_ Vol. 40, No. 4, 2018, pp. 769-790, doi: 10.1109/tpami.2017.2699960. [40] M. Ebrahim, S. Khan and U. B. Khalid, “Symmetric Algorithm Survey: A Comparative Analysis”, _International Journal of Computer Applications, Vol. 61, No. 20, 2013, pp. 12-19, doi: 10.5120/10195-4887._ [41] M. J. Moayed, A. A. A. Ghani and R. Mahmod, “A Survey on Cryptography Algorithms in Security of Voting System Approaches”, in 2008 International Conference on Computational Sciences and its _Applications, Perugia, Italy, 2008, pp. 190-200, doi: 10.1109/ICCSA.2008.42._ [42] D. Lee, “Hash Function Vulnerability Index and Hash Chain Attacks”, in 2007 3rd IEEE Workshop on _Secure Network Protocols, Beijing, China, 2007, pp. 1-6, doi: 10.1109/NPSEC.2007.4371616._ [43] F. Breitinger and H. Baier, “Properties of a Similarity Preserving Hash Function and Their Realization in SDhash”, in 2012 Information Security for South Africa, Johannesberg, South Africa, 2012, pp. 1-8, doi: 10.1109/ISSA.2012.6320445. [44] S. El Moumni, M. Fettach and A. Tragha, “High Throughput Implementation of SHA3 Hash Algorithm on Field Programmable Gate Array (FPGA)”, _Microelectronics Journal, Vol. 93, 2019, pp. 1-8, doi:_ 10.1016/j.mejo.2019.104615. [45] H. Choi and S. Seo, “Fast Implementation of SHA-3 in GPU Environment”, _IEEE Access, Vol. 9, pp._ 144574-144586, 2021, doi: 10.1109/access.2021.3122466. [46] S. Rahaman, N. Meng and D. Yao, “Tutorial: Principles and Practices of Secure Crypto Coding in Java”, in _2018 IEEE Cybersecurity Development (SecDev), Cambridge, MA, 2018, pp. 122-123, doi:_ 10.1109/SecDev.2018.00024s. [47] A. Baksi, S. Bhasin, J. Breier, D. Jap and D. Saha, “Fault Attacks in Symmetric Key Cryptosystems”, IACR _Cryptology ePrint Archive, 2020, pp. 1-24._ [48] P. K. Gundaram, A. Naidu Tentu and N. B. Muppalaneni, “Performance of Various SMT Solvers in Cryptanalysis”, in _2021 International Conference on Computing, Communication, and Intelligent Systems_ _(ICCCIS), Greater Noida, India, 2021, pp. 298-303, doi: 10.1109/ICCCIS51004.2021.9397110._ 200 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.22452/mjcs.vol35no3.1?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.22452/mjcs.vol35no3.1, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://ejournal.um.edu.my/index.php/MJCS/article/download/35355/14869" }
2,022
[]
true
2022-07-27T00:00:00
[]
18,628
en
[ { "category": "Engineering", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00eb6f6ebb68f199f12507a6bd585718f464d75d
[ "Engineering" ]
0.82397
Revision and Enhancement of Two Three Party Key Agreement Protocols Vulnerable to KCI Attacks
00eb6f6ebb68f199f12507a6bd585718f464d75d
[ { "authorId": "3111122", "name": "M. A. Strangio" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
In two recent papers, Zuowen Tan (Secu- rity and Communication Networks) and Chih-ho Chou et al. (Computers and Electronics) published three party key agreement protocols especially convenient for protecting communications in mobile-centric environ- ments such as e-payments, vehicular mobile networks (VMN), RFID applications, etc. For his protocol, Tan provides a formal security proof developed in a model of distributed computing based on the seminal work of Bellare and Rogaway. In this paper, we show that both protocols are vulnerable to KCI attacks. We suggest modifications to both protocols that fix the vulnerability at the expense of a small decrease in their computational efficiency.
# Revision and Enhancement of Two Three Party Key Agreement Protocols Vulnerable to KCI Attacks ### Maurizio Adriano Strangio Department of Mathematics, University of Rome “Roma Tre”, Rome, Italy _∗Corresponding Author: strangio@mat.uniroma3.it_ Copyright c⃝2014 Horizon Research Publishing All rights reserved. ### Abstract In two recent papers, Zuowen Tan (Security and Communication Networks) and Chih-ho Chou _et al._ (Computers and Electronics) published three party key agreement protocols especially convenient for protecting communications in mobile-centric environments such as e-payments, vehicular mobile networks (VMN), RFID applications, etc. For his protocol, Tan provides a formal security proof developed in a model of distributed computing based on the seminal work of Bellare and Rogaway. In this paper, we show that both protocols are vulnerable to KCI attacks. We suggest modifications to both protocols that fix the vulnerability at the expense of a small decrease in their computational efficiency. ### Keywords Three Party Key Agreement, Keycompromise Impersonation, Mobile Network Communications ## 1 Introduction In two recent papers, Tan [2] and Chou et al. [5] present three party key agreement protocols especially convenient for protecting communications in mobilecentric environments such as e-payments, vehicular mobile networks (VMN), RFID applications, etc. For his protocol, Tan provides a formal security proof in a model of distributed computing based on the work of Bellare and Rogaway [6, 20] and Abdalla et al. [1]. Unfortunately, both protocols are vulnerable to a particular man-in-the-middle attack known as Key Compromise Impersonation (KCI) [12, 15]. In such attacks, an adversary that has obtained the private key of party _A can impersonate a legitimate peer B; if the attack_ is successful A will share a session key with the adversary (instead of B). This is a subtle attack that can have drastic consequences since the adversary may ob material (e.g. passwords) or private data (e.g. credit card numbers). With three party protocols the adversary may impersonate a peer (A or B) or the trusted third party (S). In this paper, we describe successful KCI attacks against the aforementioned protocols and also suggest modifications to fix the vulnerabilities at the expense of a small decrease in their computational efficiency. The rest of the paper is organized as follows. We describe KCI attacks against the protocols of TAN and CHOU et al. respectively in Sections 2 and 3. In Section 4, we suggest modifications to the above protocols in order to prevent those attacks. Finally, Section 5 contains our concluding remarks. ## 2 A KCI attack against Tan’s protocol In this section we review Tan’s three-party key agreement protocol and describe how a successful KCI attack can be conducted by an adversary that has compromised the private keying material of an honest party. ### 2.1 Review of the protocol specification The protocol consists of an initialization phase (wherein A and B register with the trusted server S and obtain an dA = H(IDA∥x) and dB = H(IDB∥x) respectively, where x ∈ _Fq is the master key held by S)_ and the subsequent authenticated key agreement phase (Figure 1). System parameters are defined by the following tuple: ΦT AN = (p, n, E, G, P, Ek(·), Dk(·), h(·)) were (i) p, n are prime numbers; (ii) E(Fp) is an elliptic curve defined by the equation _y[2]_ = x[3] + ax + b over the finite field Fp, where 3 2 ----- (iii) master key x ∈R Zn[∗] [, the later is the set of residues] modulo n; (iv) group G and point P with order n over E(Fp); (v) symmetric encryption/decryption algorithms Ek(·), _Dk(·);_ (vi) hash function h( ). _·_ _A[dA], B[dB], S[dA, dB, x]_ _A →_ _B:_ _{IDA, request}_ _R_ _A :_ _a_ _←_ _Zn∗_ _R1 ←_ _aP_, e1 ← _h(R1∥IDA∥IDB)_ _e2 ←_ _EdA_ (R1∥IDA∥IDB∥e1) _A →_ _S:_ _{IDA, IDB, e2}_ _R_ _B :_ _b_ _←_ _Zn∗_ _R2 ←_ _bP_, e3 ← _h(R2∥IDB∥IDA)_ _e4 ←_ _EdB_ (R2∥IDB∥IDA∥e3) _B →_ _S:_ _{IDB, IDA, e4}_ _S :_ _dRA1′_ _←[∥][ID]h[′]([A]ID[∥][ID]A∥[′][B]x),[∥][e] d[′][1]B[=] ←[ D]h[d](AID[(][e][2]B[)]∥x)_ _R[′]2∥ID[′]B∥ID[′]A∥e[′]3 = DdB_ (e4) if e[′]1 ̸= h(R[′]1∥IDA∥IDB) abort if e[′]3 ̸= h(R[′]2∥IDB∥IDA) abort _e ←_ _h(R[′]1∥R[′]2∥IDA∥IDB)_ _Q1 ←_ _EdA_ (R[′]1∥R[′]2∥e) _Q2 ←_ _EdB_ (R[′]2∥R[′]1∥e) _S →_ _A:_ _{Q1}_ _S →_ _B:_ _{Q2}_ _BA : :_ _RififskRififsk R e R e12′′′′ ← ←[∥][∥]′′′′12[R][R] ̸ ̸ ̸ ̸====21′′′′hh R R[∥][∥](( R RaRbR[e][e]11′′′′′′′′12′′′′ = =[∥][∥]1′′2′′[R][R][abort][abort][∥][∥] D D[ID]22′′′′[ID][∥][∥]dd[ID][ID][A]AB[A][∥](([∥]Q[A][A]Q[ID][ID][∥][∥]12[ID][ID]))[B][B][)][)][B][B]_ [abort][abort] **Figure 1. TAN’s Protocol** ### 2.2 Description of the KCI attack sce- nario Below we provide a detailed description of a successful KCI attack against Tan’s protocol: 1. Adversary _A_ obtains _A’s_ private key _dA_ = _h(IDA∥x);_ 2. A generates a random nonce a _∈_ _Zn[∗][,]_ computes R1 = aP, e1 = h(R1∥IDA∥IDB), e2 = _EdA_ (R1∥IDA∥IDB∥e1) and sends {IDA, request} and {IDA, IDB, e2} to B and S respectively to initiate a protocol instance with B; 3. A intercepts the messages {IDA, request} and _{RID1′_ _[∥][ID]A, ID[′][A][∥]B[ID], e2[′][B]}, decrypts the ciphertext[∥][e][′][1]_ [chooses a random nonce] DdA (e[ c]2) =[ ∈] _Z_ _[∗]_ _R_ _P_ _h(R[′]_ _RIDID )_ 4. on receipt of _Q1,_ _A_ computes _DdA_ (Q1) = _R1′′_ 2′′ _′′_ . If the equations R1 = R1′′ [and][ e]′′ = _R1′′_ _[∥][∥][R][R]2′′_ _[∥][∥][e][ID][A][∥][ID][B]_ [are verified (indeed they are] since the transcripts are indistinguishable from those exchanged by honest parties) A terminates with the session key skA = h(aR2′′ _[∥][ID][A][∥][ID][B][);]_ 5. A computes sk′ = h(cR1′ _[∥][ID][A][∥][ID][B][) and will be]_ able to establish a communication session with A since sk′ = skA. ## 3 A KCI attack against Chou et al.’s protocol In this section we review Chou et al.’s three-party key agreement protocol and describe how a successful KCI attack can be conducted by an adversary. ### 3.1 Review of the protocol specification This protocol also comprises an initialization phase (wherein users A and B register with the trusted server _S and obtain YA = skAPKS and YB = skBPKS) and_ an authenticated key agreement phase (Figure 2). System parameters are defined by the following tuple: ΦCHOU = (p, n, E, G, P, Ek(·), Dk(·), h(·)) were each parameter is defined as in Section 2. _A[skA, PKA], B[skB, PKB], S[skS, PKS]_ _R_ _A :_ _rA_ _←_ _Zp∗_ _RA_ _rAPKA + YA_ _←_ _KA_ _rAYA = (KA.x, KA.y)_ _←_ _CA_ _EKA,x_ (RA, IDA, IDB, TA) _←_ _A →_ _B:_ _{IDA, request}_ _A →_ _S:_ _{IDA, RA, CA, TA}_ _R_ _B :_ _rB_ _←_ _Zp∗_ _RB_ _rBPKB + YB_ _←_ _KB_ _rBYB = (KB.x, KB.y)_ _←_ _CB_ _EKB.x_ (RB, IDB, IDA, TB) _←_ _B →_ _A:_ _{IDB, response}_ _B →_ _S:_ _{IDB, RB, CB, TB}_ _S :_ _KA ←_ _skS(RA-skSPKA) = (KA.x, KA.y)_ _KB ←_ _skS(RB-skSPKB) = (KB.x, KB.y)_ _RA, IDA, IDB, TA = DKA.x_ (CA) _RB, IDB, IDA, TB = DKB.x_ (CB) check TA, TB, RA, RB _CSA_ _EKA.x_ (RA, KB, IDA, IDS, TS) _←_ _CSB_ _EKB.x_ (RB, KA, IDB, IDS, TS) _←_ _S →_ _A:_ _{CSA, TS}_ _S →_ _B:_ _{CSB, TS}_ _A :_ _RA, KB, IDA, IDS, TS = DKA.x_ (CSA) check RA, TS _sk ←_ _rAskAKB_ _B :_ _RB, KA, IDB, IDS, TS = DKB,x_ (CSB) check RB, TS _sk ←_ _rBskBKA_ **Figure 2** Chou et al ’s Protocol ----- ### 3.2 Description of the KCI attack sce- nario Below we provide a detailed description of a successful KCI attack against Chou’s protocol: 1. Adversary A obtains A’s private key skA; 2. A executes all operations according to the protocol specification and sends {IDA, request} and _{IDA, RA, CA, TA} to B and S respectively to ini-_ tiate a protocol instance with B; 3. A intercepts the message _{IDA, RA, CA, TA},_ chooses a random nonce rE ∈ _Zp[∗]_ [and sends][ C][SA] [=] _EKA.x_ (RA, KB, IDB, IDS, TS) where KB = rEP to _A;_ 4. on receipt of CSA, A follows the protocol specification (RA, TS will pass the verification step) and terminates with the session key sk = rAskAKB = _rAskArEP = rErAPKA = rE(RA_ _YA);_ _−_ 5. can establish a communication session with A by _A_ computing sk = rE(RA _−YA) where YA = skAPKS_ and thus is easily computed by the adversary. ## 4 Revisiting the protocols to eliminate the vulnerability to KCI attacks In this section we illustrate modifications to the protocols of Tan and Chou et al. that do not allow a malicious party to perform successfull KCI attacks. ### 4.1 A KCI-resilient version of Tan’s pro- tocol It is interesting to notice that Tan’s protocol cannot withstand KCI attacks despite the fact that a formal security proof was provided by the author (see Theorems 1, 2 in [2]). The arguments used by the author to support the proof of KCI-resilience (Theorem 2) require that the adversary must be able to forge the transcript e4; however, this assumption misses the point since the adversary does not need to faithfully reproduce the protocol actions of B but must simply generate message transcripts (in this particular case, a Diffie-Hellman ephemeral key R2 = cP ) that are indistinguishable (for _A) from the real ones. In general, for the sake of protocol_ security analysis one assumes that a man-in-the-middle attacker has total control over the network (i.e. she can insert, delete, modify messages flowing across the network) and is allowed to achieve her goals by arbitrarily diverging from the protocol specification. To prevent the attack Tan’s protocol can be modified as follows: the trusted third party S shall generate keys for each peer DA = dAP, DB = dBP (eventually during the initialization stage) and compute e = _h(R[′]1∥R[′]2∥IDA∥IDB∥DA∥DB), Q1 = EdA(R[′]1∥R[′]2∥e)_ and Qi 2 =k _EdB_ (kR[′]2∥hR( Dd R[′]1∥e); finally′′ _IDID A shall compute its ) ( i_ il l With the preceding modifications, Tan’s key exchange can now withstand KCI attacks. Indeed, the session key is computed using the method introduced in the MTI/A0 [9] protocol, which is immune from KCI attacks; the adversary must now obtain both the private key dA and the random nonce a to compute the session key of A. A significant difference with the MTI protocol is due to the fact that the keys DA, DB do not necessarily have to be pre-distributed to A, B respectively and therefore do not require public key certificates. The revised protocol enjoys KCI attack resilience at the expense of an increased computational workload with respect to the original version. In particular, the trusted first party S must now compute two additional scalar multiplications to generate the keys _DA, DB._ However, for reasons of efficiency the computation can be performed prior to online executions of the protocol. Each peer A, B is required to compute an additional scalar multiplication (aDB, bDA) at runtime to compute the session key. ### 4.2 A KCI-resilient version of Chou et al.’s protocol It is worth mentioning that the procotol of Chou et al. is actually a revised version of a flawed protocol previously proposed by Zuowen Tan [3] which was yet another version of the protocol originally published by Yang and Chang [4] (the later also contained a vulnerability to impersonation attacks). Chou et al.’s protocol can withstand KCI attacks with the following modifications. The trusted third party computes CSA = _EKA.x_ (RA, KB, YB, IDA, IDS, TS), _CSB = EKB.x_ (RB, KA, YA, IDB, IDS, TS) and sends these transcripts to A, B respectively; with respect to the original specification the terms YB, YA are included in CSA, CSB respectively. On receipt of CSA, A computes the session key sk = h(rAYB∥skAKB) (similarly for B) where h is an appropriate hash function. The revised protocol (similarly to Tan’s protocol) enjoys KCI attack resilience at the expense of a slightly increased computational workload with respect to the original version. In this case, each peer A, B needs an additional scalar multiplication (rAYB, rBYA, IDA, IDB) at runtime to compute the session key. ## 5 Concluding remarks We have shown that the three party key agreement protocols recently published in the literature by Tan [2] and Chou et al. [5] are not resilient to KCI attacks although their authors claim the contrary. Designing secure key agreement protocols is far from being a simple task, such protocols involve so many details and complicated interactions between different cryptographic primitives that it is nearly impossible to establish beyond doubt that they are infallible. Indeed, in the history of this subject there is an abundance of protocols that were more or less trivially broken regardless of whether formal or heuristic arguments were proid d t t it l i ----- is that few people are capable of verifying them because the maths can be quite sophisticated and may be therefore difficult to understand for the average practitioner. In the last two decades many formal security models have been proposed (which often differ by a few details) but it still is not quite clear how to ”transfer” a security property proved in one model to a different model and to the real world (Choo et al. [14] have examined this issue). Furthermore, provable security of cryptographic primitives is largely based on the use of computational assumptions, these have proliferated in the last decade [16] making it difficult to differentiate between their relative strengths (an interesting classification was proposed by Naor [17]. It is also important to analyse the security of cryptographic protocols not only from an theoretical point of view (e.g. adversary interacts with parties in a black-box mode) but in more realistic models wherein an opponent may physically attack an honest peer (e.g. tampering with devices [18]). In any case, the publication of protocol specifications in specialised literature and conferences is of fundamental importance since the peer review process and the subsequent period of public scrutiny can increase the confidence in the security of the protocol either because vulnerabilities are not discovered or if there are proposals to fix them will be immediately published. A topic for future research is the development of a concrete implementation of the protocols discussed in this paper to evaluate real world security issues and efficiency. In particular, we plan to develop an implementation with the Java Cryptography Architecture framework [19] which uses a ”‘provider”’-based approach and contains a set of APIs for various purposes (e.g. encryption, key generation and management, secure random number generation, certificate validation, etc). ## REFERENCES [1] Abdalla M., Fouque P.A., Pointcheval D. 2005 _Password- based authenticated key exchange in the_ _three-party setting Proceedings of the PKC05, LNCS,_ 3386: 65-84 [2] Tan A. 2012. A Communication and computation_efficient three party authenticated key agreement pro-_ _tocol Security Comm. Networks DOI: 10.1002/sec.622_ [3] Tan Z., 2010. An enhanced three-party authentication _key exchange protocol for mobile commerce environ-_ _ments J.Commun., DOI:10.4304/jcm.5.5.436-443_ [4] Yang J.H., Chang C.C., 2009. An efficient three_party authenticated key exchange protocol using elliptic_ _curve cryptography for mobile-commerce environments_ J. Syst. Software, DOI:10.1016/j.jss.2009.03.075 [5] Chou Chih-ho, Tsai Kuo-yu, Wu Tzong-chen, Yeh Kuohu. 2013. Efficient and secure three-party authenti _cated key exchange protocol for mobile environments,_ 14(5):347-355 [6] Bellare M. and Rogaway P. 1993 Entity Authentica_tion and Key Distribution Advances in Cryptology-_ CRYPTO 93, pp. 110-125, Springer-Verlag. [7] Canetti R. and Krawczyk H. 2001 Analysis of key ex_change protocols and their use for building secure chan-_ _nels Proc. of Eurocrypt’01, LNCS 2045, pp. 453-474._ [8] Hankerson D., Menezes A.J. and Vanstone S.A. 2004 _Guide to Elliptic Curve Cryptography Springer Profes-_ sional Computing, New York [9] Matsumoto T., Takashima Y, and Imai H. 1986 On _seeking smart public-key distribution systems Transac-_ tions of IEICE, VolE69:99-106 [10] Lamacchia B., Lauter B. and Mityagin A. 2007 Stronger _Security of Authenticated Key Exchange LNCS 4784,_ pp. 1-16. [11] Mohammad Z. and Chi-Chun Lo 2009 Vulnerability of _an improved Elliptic Curve Diffie-Hellmann Key Agree-_ _ment and its Enhancement Proc. of EBISS’09, pp. 1-5_ [12] Strangio M.A. 2006 _On_ _the_ _Resilience_ _of_ _Key_ _Agreement_ _Protocols_ _to_ _Key_ _Compro-_ _mise_ _Impersonation_ Cryptology ePrint Archive, http://eprint.iacr.org/2006/252.pdf. [13] Wang S, Cao Z, Strangio M.A and Wang B 2009 Crypt_analysis and Improvement of an Elliptic Curve Diffie-_ _Hellmann Key Agreement Protocol IEEE Communica-_ tion Letters, Vol. 12, No. 2, pp. 149-151. [14] Choo K.R, Boyd C, Hitchcock Y 2005 _Examin-_ _ing Indistinguishability-Based Proof Models for Key_ _Establishment_ _Protocols,_ Advances in CryptologyASIACRYPT, Springer Berlin/Heidelberg, pp. 585-604 [15] Boyd C, Mathuria A 2003 Protocols for Authentication _and Key Establishment, Springer Berlin/Heidelberg_ [16] ECRYPT 2013 _Final_ _Report_ _on_ _Main_ _Computa-_ _tional Assumptions in Cryptography,_ ECRYPT II, http://www.ecrypt.eu.org/documents/D.MAYA.6.pdf [17] Naor M. 2003 On Cryptographic Assumptions and Chal_lenges, Advances in Cryptology, Lecture Notes in Com-_ puter Science Volume 2729, pp 96-109 [18] Anderson R, 2001 Security Engineering: _A guide to_ _Building Dependeable Distributed Systems, Wiley_ [19] Oracle, 2013 _Java_ _Cryptography_ _Architecture,_ http://docs.oracle.com/javase/7/docs/technotes/guides /security/crypto/CryptoSpec.html [20] Bellare M., Rogaway P. 1995 Provably secure session _key distribution: the three party case, Proceedings of_ the ACM Symposium on the Theory of Computing (STOC95), pp. 5766 -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.13189/WJCAT.2014.020204?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.13189/WJCAT.2014.020204, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "http://www.hrpub.org/download/20140105/WJCAT4-13701794.pdf" }
2,014
[]
true
null
[ { "paperId": "b99219fa436dcc8ec75ced6830b11c555439bf09", "title": "A communication and computation-efficient three-party authenticated key agreement protocol" }, { "paperId": "2784a03c76d2597eab0a3c75a526b8a68ec8b347", "title": "Efficient and secure three-party authenticated key exchange protocol for mobile environments" }, { "paperId": "f3c4fccc31a335d1d5b931331177015f4cd8246f", "title": "Efficient and secure three-party authenticated key exchange protocol for mobile environments" }, { "paperId": "c8daa77f4e613cbf07378ae6614a51d0e2554325", "title": "An Enhanced Three-Party Authentication Key Exchange Protocol for Mobile Commerce Environments" }, { "paperId": "31a99b99dbd97f5304a3a4329cb6f8f5aed00224", "title": "An efficient three-party authenticated key exchange protocol using elliptic curve cryptography for mobile-commerce environments" }, { "paperId": "eec113a2cd5c92f4596c0c55646e184b80fc59b8", "title": "Vulnerability of an Improved Elliptic Curve Diffie-Hellman Key Agreement and its Enhancement" }, { "paperId": "2d349ea5ae75426cac8a6928fe46248b2201ef7a", "title": "Cryptanalysis and improvement of an elliptic curve Diffie-Hellman key agreement protocol" }, { "paperId": "5ace5930cd7f2bf8be646ba3fc950b52727fe838", "title": "On the Resilience of Key Agreement Protocols to Key Compromise Impersonation" }, { "paperId": "a5b87921abef3f4c65b1c7f1f9d152bddb833aa2", "title": "Stronger Security of Authenticated Key Exchange" }, { "paperId": "1f4f1f0cf6649a26cf046b16a7ec1aac72d0d58f", "title": "Examining Indistinguishability-Based Proof Models for Key Establishment Protocols" }, { "paperId": "9859517d9c9a539227506e3f576ce40d1ed123ca", "title": "Password-Based Authenticated Key Exchange in the Three-Party Setting" }, { "paperId": "a4013e651c0927d70f3c95474bb00869231a9557", "title": "Protocols for Authentication and Key Establishment" }, { "paperId": "9c9ecf582e5b66e8e2cb171f95be9daf64fe784b", "title": "On Cryptographic Assumptions and Challenges" }, { "paperId": "c54359ab93e998efdf4b7b4ae931a58a8955fb72", "title": "Protocols for Key Establishment and Authentication" }, { "paperId": "e56f9dc2ef53c9479d3d94f593b66a40e6dba52b", "title": "Analysis of Key-Exchange Protocols and Their Use for Building Secure Channels" }, { "paperId": "ca7ecb6e0625bb175bda5eb0655ca609b6b5a6f9", "title": "Provably secure session key distribution: the three party case" }, { "paperId": "d4d2b30a290228436c570c5d274e8202f45a6988", "title": "Entity Authentication and Key Distribution" }, { "paperId": "aa0c3b62c6f1e0d990bb58f52e488af986f065ee", "title": "ON SEEKING SMART PUBLIC-KEY-DISTRIBUTION SYSTEMS." }, { "paperId": "7936ae14d2d0d4e68d5e68f522349f25e1fc79c9", "title": "A Key Compromise Impersonation attack against Wang's Provably Secure Identity-based Key Agreement Protocol" }, { "paperId": null, "title": "Efficient and secure three-party authenticated" }, { "paperId": null, "title": "Java Cryptography Architecture" }, { "paperId": "27eb6c1f53e041c80801d24f59485c40cb2304ad", "title": "Final Report on" }, { "paperId": "3bfdca5ffd79bdfb8b9110c84a23470d88d774a6", "title": "Guide to Elliptic Curve Cryptography" }, { "paperId": "f0773786c41ae59cfd316fa0203bbfb6442c2ffc", "title": "Security Engineering: A Guide to Building Dependable Distributed Systems" }, { "paperId": null, "title": "work-load with respect to the original version" }, { "paperId": null, "title": "Adversary A obtains A ’s private key d A = h ( ID A ∥ x )" }, { "paperId": null, "title": "master key x" }, { "paperId": null, "title": "A generates a random nonce a ∈ Z ∗ n" }, { "paperId": null, "title": "A intercepts the messages" }, { "paperId": null, "title": "(ii) E( F p ) is an elliptic curve defined by the equation y 2 = x 3 + ax + b over the finite field F p , where 4" }, { "paperId": null, "title": "on receipt of C SA , A follows the protocol specifi-cation ( R A ; T S will pass the verification step) and terminates with the session key" }, { "paperId": null, "title": "on receipt of Q 1 ," }, { "paperId": null, "title": "A executes all operations" } ]
5,858
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Mathematics", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00eb7acc1610d210d538de47fd9ca0cb75faf055
[ "Computer Science" ]
0.880526
Many-to-One Trapdoor Functions and Their Ralation to Public-Key Cryptosystems
00eb7acc1610d210d538de47fd9ca0cb75faf055
Annual International Cryptology Conference
[ { "authorId": "1703441", "name": "M. Bellare" }, { "authorId": "1808458", "name": "S. Halevi" }, { "authorId": "1695851", "name": "A. Sahai" }, { "authorId": "1723744", "name": "S. Vadhan" } ]
{ "alternate_issns": null, "alternate_names": [ "Int Cryptol Conf", "Annu Int Cryptol Conf", "CRYPTO", "International Cryptology Conference" ], "alternate_urls": null, "id": "212b6868-c374-4ba2-ad32-19fde8004623", "issn": null, "name": "Annual International Cryptology Conference", "type": "conference", "url": "http://www.iacr.org/" }
null
## Many-to-One Trapdoor Functions and Their Relation to Public-Key Cryptosystems Mihir Bellare 1 , Shai Halevi 2, Amit Sahai 3, and Salil Vadhan 3 1 Dept. of Computer Science & Engineering, University of California at San Diego, 9500 Gilman Drive, La Jolla, CA 92093, USA. E-Mail: mihir@cs.ucsd, edu. ``` UI%L: http://w~-cse, ucsd. edu/users/mihir. ``` 2 IBM T. J. Watson Research Center, P.O. Box 704, Yorktown Heights, NY 10598, USA. E-Maih `shaih@watson, ibm. com.` 3 MIT Laboratory for Computer Science, 545 Technology Square, Cambridge, MA 02139, USA. E-Malh amitsQtheory, lcs .mit. edu, salil~math.mit, edu. **URL:** http ://www-math. mit. edu/~ salil. **Abstract.** The heart of the task of building public key cryptosystems is viewed as that of "making trapdoors;" in fact, public key cryptosys- terns and trapdoor functions are often discussed as synonymous. How accurate is this view? In this paper we endeavor to get a better under- standing of the nature of "trapdoorness" and its relation to public key cryptosystems, by broadening the scope of the investigation: we look at general trapdoor functions; that is, functions that are not necessarily in- jective (ie., one-to-one). Our first result is somewhat surprising: we show that non-injective trapdoor functions (with super-polynomial pre-image size) can be constructed from any one-way function (and hence it is un- likely that they suffice for public key encryption). On the other hand, we show that trapdoor functions with polynomial pre-image size are suffi- cient for public key encryption. Together, these two results indicate that the pre-image size is a fundamental parameter of trapdoor functions. We then turn our attention to the converse, asking what kinds of trapdoor functions can be constructed from public key cryptosystems. We take a first step by showing that in the random-oracle model one can construct injective trapdoor functions from any public key cryptosystem. 1 Introduction A major dividing line in the realm of cryptographic primitives is that between "one-way" and "trapdoor" primitives. The former effectively means the primi- tives of private key cryptography, while the latter are typically viewed as tied to public key cryptosystems. Indeed, the understanding is that the problem of building public key cryptosystems is the problem of "making trapdoors." Is it really? It is well known that injective (ie. one-to-one) trapdoor functions suffice for public key cryptography [Ya,GoMi]. We ask: is the converse true as ----- **284** a closer look at the notion of a trapdoor, in particular from the point of view of how it relates to semantically secure encryption schemes, and discover some curious things. Amongst these are that "trapdoor one-way functions" are not necessarily hard to build, and their relation to public key encryption is more subtle than it might seem. 1.1 Background The main notions discussed and related in this paper are one-way functions [DiHe], trapdoor (one-way) functions [DiHe], semantically secure encryption schemes [GoMi], and unapproximable trapdoor predicates [GoMi]. Roughly, a "one-way function" means a family of functions where each partic- ular function is easy to compute, but most are hard to invert; trapdoor functions are the same with the additional feature that associated to each particular func- tion is some "trapdoor" information, possession of which permits easy inversion. (See Section 2 for formal definitions.) In the study of one-way functions, it is well appreciated that the functions need not be injective: careful distinctions are made between "(general) one- way functions", "injective one-way functions," or "one-way permutations." In principle, the distinction applies equally well to trapdoor one-way functions. (In the non-injective case, knowledge of the trapdoor permits recovery of some pre- image of any given range point [DiHe].) However, all attention in the literature has focused on injective trapdoor functions, perhaps out of the sense that this is what is necessary for constructing encryption schemes: the injectivity of the trapdoor function guarantees the unique decryptability of the encryption scheme. This paper investigates general (ie. not necessarily injective) trapdoor one- way functions and how they relate to other primitives. Our goal is to understand exactly what kinds of trapdoor one-way functions are necessary and sufficient for building semantically secure public key encryption schemes; in particular, is injectivity actually necessary? Among non-injective trapdoor functions, we make a further distinction based on "the amount of non-injectivity', measured by pre-image size. A (trapdoor, one-way) function is said to have pre-image size _Q(k)_ (where k is the security parameter) if the number of pre-images of any range point is at most _Q(k)._ We show that pre-image size is a crucial parameter with regard to building public- key cryptosystems out of a trapdoor function. Rather than directly working with public-key cryptosystems, it will be more convenient to work with a more basic primitive called an unapproximable trap- door predicate. Unapproximable trapdoor predicates are equivalent to semanti- cally secure public key schemes for encrypting a single bit, and these in turn are equivalent to general semantically secure cryptosystems [GoMi]. **1.2** **Results** We have three main results. They are displayed in Fig. 1 together with known ----- **285** semantically secure Injective public-key cryptosystems trapdoor functions _Theorem 3 ~. ~_ [GoMi] I ....~,,JtYal _l trivial_ Unapproximable .q Trapdoor functions with trapdoor ~redicates **_Theorem 2_** **poly-bound~ pre-image size** [lmLu] _l td'~al_ One-way _Theorem 1_ Trapdoor functions with functions IP super-poly pm-irnage size Fig. 1. Illustrating our results: _Solid lines are standard implications; the dotted line_ _is an implication in the random oracle model._ _One-way functions imply trapdoor functions._ Our first result, given in Theorem 1, may seem surprising at first glance: we show that one-way functions imply trapdoor functions. We present a general construction which, given an arbitrary one-way function, yields a trapdoor (non-injective) one-way function. Put in other words, we show that trapdoor functions are not necessarily hard to build; it is the combination of trapdoorness with "structural" properties like injectivity that may be hard to achieve. Thus the "curtain" between one-way and trapdoor primitives is not quite as opaque as it may seem. What does this mean for public key cryptography? Impagliazzo and Rudich [ImRu] show that it would be very hard, or unlikely, to get a proof that one-way functions (even if injective) imply public key cryptosystems. Hence, our result shows that it is unlikely that any known technique can be used to construct public key encryption schemes from generic, non-injective, trapdoor functions. As one might guess given [ImRu], our construction does not preserve injectivity, so even if the starting one-way function is injective, the resulting trapdoor one- way function is not. _Trapdoor functions with poly pre-image size yield eryptosystems. In_ light of the above, one might still imagine that injectivity of the trapdoor func- tions is required to obtain public key encrypti0n. Still, we ask whether the in- jectivity condition can be rela~ed somewhat. Specifically, the trapdoor one-way functions which we construct from one-way functions have super-polynomial pre-image size. This leads us to ask about trapdoor functions with polynomially bounded pre-image size. Our second result, Theorem 2, shows that trapdoor functions with poly- nomially bounded pre-image size suffice to construct unapproximable trapdoor predicates, and hence yield public key cryptosystems. This belies the impression ----- 286 a public key cryptosystem from it, and also suggests that the super-polynomial pre-image size in the construction of Theorem 1 is necessary. #### From trapdoor predicates to trapdoor functions, We then turn to the other side of the coin and ask what kinds of trapdoor functions must necessarily exist to have a public key cryptosystem. Since unapproximable trapdoor pred- icates and semantically secure public key cryptosystems are equivalent [GoMi] we consider the question of whether unapproximable trapdoor predicates imply injective trapdoor functions. In fact whether or not semantically secure public key cryptosystems imply injective trapdoor functions is not only an open question, but seems a hard one. (In particular, a positive answer would imply injective trapdoor functions based on the Diffie-Hellman assumption, a long standing open problem.) In order to get some insight and possible approaches to it, we consider it in a random oracle model (cf. [ImRu,BeRo]). Theorem 3 says that here the answer is aff~mative: given an arbitrary secure public key cryptosystem, we present a function that has access to an oracle H, and prove the function is injective, trapdoor, and one-way when H is random. The construction of Theorem 3 is quite simple, and the natural next question is whether the random oracle H can be replaced by some constructible crypto- graphic primitive. In the full version of the paper [BHSV], we show that this may be difficult, by showing that a cryptographically strong pseudorandom bit generator [B1Mi,Ya], which seems like a natural choice for this construction, does not suffice. The next step may be to follow the approach initiated by Canetti [Ca]: find an appropriate cryptographic notion which, if satisfied by H, would suffice for the correctness of the construction, and then try to implement H via a small family of functions. However, one should keep in mind that replacement of a random oracle by a suitable constructible function is not always possible [CGH]. Thus, our last result should be interpreted with care. 1.3 Discussion and implications Theorems 1 and 2 indicate that pre-image size is a crucial parameter when con- sidering the power of trapdoor functions, particularly with respect to construct- ing public-key cryptosystems. The significance and interpretation of Theorem 3, however, requires a bit more discussion. At first glance, it may seem that public key cryptosystems "obviously im- ply" injective trapdoor functions. After all, a public key cryptosystem permits unique decryptability; doesn't this mean the encryption algorithm is injective? No, because, as per [GoMi], it is a _probabilistic algorithm, and thus not a func-_ tion. To make it a function, you must consider it a function of two arguments, the message and the coins, and then it may no longer be injective, because two coin sequences could give rise to the same ciphertext for a given message. More- over, it may no longer have a (full) trapdoor, since it may not be possible to recover the randomness from the ciphertext. (Public key cryptosystems in the ----- 287 the authors remark, but that's because encryption there is deterministic. It is now understood that secure encryption must be probabilistic [GoMi].) Theorem 3 has several corollaries. (Caveat: All in the random oracle model). First, by applying a transformation of [BeRo], it follows that we can construct non-malleable and chosen-ciphertext secure encryption schemes based on the Ajtal-Dwork cryptosystem [AjDw]. Second, combining Theorems 3 and 2, the existence of trapdoor functions with polynomially bounded pre-image size im- plies the existence of injective trapdoor functions. (With high probability over the choice of oracle. See Remark 5.) Third, if the Decisional Diflie-Hellman prob- lem is hard (this means the E1 Gamal [E1G] cryptosystem is semantically secure) then there exists an injective trapdoor function. Note that in the random oracle model, it is trivial to construct (almost) injective one-way functions: a random oracle mapping, say, n bits to 3n bits, is itself an injective one-way function except with probability 2 -n over the choice of the oracle. However, random oracles do not directly or naturally give rise to trapdoors [ImRu]. Thus, it is interesting to note that our construction in Theorem 3 uses the oracle to "amplify" a trapdoor property: we convert the weak trapdoor property of a cryptosystem (in which one can only recover the message) to a strong one (in which one can recover both the message and the randomness used). Another interpretation of Theorem 3 is as a demonstration that there ex- ists a model in which semantically secure encryption implies injective trapdoor functions, and hence it may be hard to prove a separation result, in the style of [ImRu], between injective trapdoor functions and probabilistic encryption schemes. #### 2 Definitions We present definitions for one-way functions, trapdoor functions, and unapprox- imable trapdoor predicates. PRELIMINARIES. If S is any probability distribution then x +- S denotes the operation of selecting an element uniformly at random according to S, and [S] is the support of S, namely the set of all points having non-zero probability under S. If S is a set we view it as imbued with the uniform distribution and write x ~ S. If A is a probabilistic algorithm or function then A(x, y,... ; R) denotes the output of A on inputs x, y,... and coins R, while A(x, Y,...) is the probability distribution assigning to each string the probability, over R, that it is output. For deterministic algorithms or functions A, we write z:=A(x, y,...) to mean that the output of A(x, Y,...) is assigned to z. The notation Pr [ E : R1 ; R2 ; ... ; Rk ] refers to the probability of event E after the random processes R1,..., Rk are performed in order. If x and Y are strings we write their concatenation as _xll y_ or just _xy._ "Polynomial time" means time polynomial in the security parame- ter k, PPT stands for "probabilistic, polynomial time", and "efficient" means computable in polynomial time or PPT ----- **288** **2.1** **One-way and trapdoor function families** We first define families of functions, then say what it means for them to be one-way or trapdoor. FAMILIES OF FUNCTIONS. A _family of functions_ is a collection F = {Fk}keN where each Fk is probability distribution over a set of functions. Each f E [Fk] has an associated domain Dom(f) and range Range(f). We require three properties of the family: �9 Can generate: The operation f +-- Fk can be efficiently implemented, mean- ing there is a PPT _generation_ algorithm _F-Gen_ that on input i k outputs a "description" of a function f distributed according to _Fk._ This algorithm might also output some auxiliary information aux associated to this function (this is in order to later model trapdoors). _�9 Can sample:_ Dora(f) is efficiently samplable, meaning there is a PPT algo- rithm _F-Stop that given f E [Fk] returns a uniformly distributed element of_ ``` Dora(f). ``` _�9 Can evaluate: f_ is efficiently computable, meaning there is a polynomial time evaluation algorithm _F-Eval_ that given f E Fk and x E Dora(f) returns #### f(x). For an element y E Range(f) we denote the set of pre-images of y under f by _f-l(y) = { x e Dom(f)_ : _f(x) = y } ._ We say that F is _injeetive_ if f is injective (ie. one-to-one) for every f E [Fk]. If in addition Dom(f) = Range(f) then we say that F is a family of permutations. We measure the amount of "non-injectivity" by looking at the maximum pre- image size. Specifically we say that F has _pre-image size_ bounded by _Q(k)_ if ]f-l(y)l < _Q(k)_ for _all f e [Fk], all y e_ Range(f) and all k e N. We say that _F has polynomiaUy bounded pre-image size if there is a polynomial_ _Q(k)_ which bounds the pre-image size of F. ONE-WAYNESS. Let F be a family of functions as above. The _inverting probability_ of an algorithm I(.,-) with respect to F is a function of the security parameter k, defined as InvProbf(I, k) d=ef Pr Ix' e _f-l(y) : f +._ Fk ; x +- Dom(f) ; y +-- f(x);_ _x' ~ I(f,y) ] ._ F is _one-way_ if InvProbF (I, k) is negligible for any PPT algorithm I. TRAPDOORNESS. A family of functions is said to be trapdoor if it is possible, while generating an instance f, to simultaneously generate as auxiliary output "trapdoor information" _tp, knowledge of which permits inversion of f. Formally,_ a family of functions F is _trapdoor_ if _F-Gen_ outputs pairs _if, tp)_ where f is the "description" of a function as in any family of functions and _tp is auxiliary_ _trapdoor information._ We require that there exists a probabilistic polynomial time algorithm _F-Inv such that for all k, all (f, tp) E [F-Gen(lk)],_ and all points y e Range(f), the algorithm _F-Inv(f, tp, y)_ outputs an element of f-Z(y) with probability 1. A family of trapdoor functions is said to be one-way if it is also a family of one-way functions ----- **289** A good (candidate) example of a trapdoor, one-way function family which **is** non-injective is the _Rabin family_ [l~ab]: here each function in Fk is four to one. (Traditionally, this function is used as the basis of a public key cryptosystem by first modifying it to be injective.) _Remark 1._ It is well known that one can define one-way functions either in terms of function families (as above), or in terms of a single function, and the two are equivalent. However, for trapdoor functions, one must talk of families. To maintain consistency, we use the family view of one-way functions as well. 2.2 Trapdoor Predicate Families We define unapproximable trapdoor predicate families [GoMi]. Recall that such a family is equivalent to a semantically secure public-key encryption scheme for a single bit [GoMi]. _A predicate_ in our context means a probabilistic function with domain {0, 1}, meaning a predicate p takes a bit b and flips coins r to generate some output _y = p(b; r). In a_ _trapdoor predicate family P =_ {P~}keN, each Pk is a probability distribution over a set of predicates, meaning each p E [P~] is a predicate as above. We require: �9 Can generate: There is a generation algorithm _P-Gen_ which on input 1 k outputs (p, tp) where p is distributed randomly according to Pk and _tp_ is trapdoor information associated to p. In particular the operation p +-- Pk can be efficiently implemented. �9 Can evaluate: There is a PPT algorithm _P-Eval_ that given p and b E {0,1} flips coins to output y distributed according to p(b). We say P has _decryption_ error J(k) if there is a PPT algorithm _P-Inv_ who, with knowledge of the trapdoor, fails to decrypt only with this probability, namely DecErrp ( P-Inv, k) def__~ Pr [ b' r b : p ~ Pk ; b ~- {0, 1} ; y +- p(b) ; b' e- P-Inv(p, tp, y) ] (1) is at most J(k). If we say nothing it is to be assumed that the decryption error is zero, but sometimes we want to discuss families with non-zero (and even large) decryption error. UNAPPROXIMABILITY. Let P be a family of trapdoor predicates as above. The _predicting advantage_ of an algorithm I(., .) with respect to P is a function of the security parameter k, defined as PredAdvp(I, k) **def** 1 Pr[b' _=b : p+-P~; b+--_ {0,1}; y +-p(b); b'e-I(p,y)]- _2"_ We say that P is _unapproximable_ if PredAdvp(I, k) is negligible for any PPT algorithm I. 3 From one-way functions to trapdoor **functions** ----- **290** **Theorem** 1. _Suppose there exists a family of one-way functions. Then there_ _exists a family of trapdoor, one-way functions._ _J_ This is proved by taking an arbitrary family F of one-way functions and "em- bedding" a trapdoor to get a family G of trapdoor functions. The rest of this section is devoted to the proof. **3.1** **Proof sketch of Theorem 1** Given a family F = {Fk}keN of one-way functions we show how to construct a family G = {Gk}keN of trapdoor one-way functions. Let us first sketch the idea. Given f E Fk we want to construct g which "mimics" f but somehow embeds a trapdoor. The idea is that the trapdoor is a particular point c~ in the domain of f. Function g will usually just evaluate f, except if it detects that its input contains the trapdoor; in that case it will do something trivial, making g easy to invert given knowledge of the trapdoor. (This will not happen often in normal execution because it is unlikely that a randomly chosen input contains the trapdoor.) But how exactly can g "detect" the trapdoor? The first idea would be to include a in the description of g so that it can check whether its input contains the trapdoor, but then g would no longer be one-way. So instead the description of g will include/3 = f(a), an image of the trapdoor under the original function f, and g will run f on a candidate trapdoor to see whether the result matches/3. (Note that we do not in fact necessarily detect the real trapdoor a; the trivial action is taken whenever some pre-image of/5 under f is detected. But that turns out to be OK.) In the actual construction, g has three inputs, y, x, v, where v plays the role of the "normal" input to f; x plays the role of the candidate trapdoor; and y is the "trivial" answer returned in case the trapdoor is detected. We now formally specify the construction and sketch a prof that it is correct. A particular function g E [Gk] will be described by a pair (f,/3) where f E [Fk] and/3 E Range(f). It is defined on inputs y, x, v by _g(y, x, v) = { y_ if f(x) =/3 (2) _f(v)_ otherwise. Here _x,v E_ Dora(f), and we draw y from some samplable superset S I of Range(f). (To be specific, we set Sf to the set of all strings of length at most p(k) where p(k) is a polynomial that bounds the lengths of all strings in Range(f).) So the domain of g is Dom(g) = S I • Dom(f) • Dora(f). We now give an intuitive explanation of why G is one-way and trapdoor. First note that for any z it is the case that (z, c~, a) is a preimage of z under g, so knowing a enables one to invert in a trivial manner, hence G is trapdoor. For one-wayness, notice that if g(y, x, v) = z then either _f(v) = z_ or _f(x) =/3._ Thus, producing an element of g-1 (z) requires inverting f at either z or/3, both of which are hard by the one-wayness of F. A formal proof that G satisfies the definition of a family of one-way trapdoor functions can be found in the full version of this paper [BHSV]. _Remark 2._ One can verify that the trapdoor functions g produced in the above ----- 291 if the original one-way functions f are regular. Thus, adding regularity as a requirement is _not_ likely to suffice for making public-key cryptosystems. 4 From trapdoor functions to cryptosystems Theorem 1 coupled with [ImRu] says that it is unlikely that general trapdoor functions will yield semantically secure public-key cryptosystems. However, in our construction of Section 3.1 the resulting trapdoor function was "very non- injective" in the sense that the pre-image size was exponential in the security parameter. So, we next ask, what is the power of trapdoor function families with polynomially bounded pre-image size? We show a positive result: **Theorem** 2. _If there exist trapdoor one-way function families with polynomially_ _bounded pre-image size, then there exists a family of unapproximable trapdoor_ _predicates with exponentially small deeryption error._ Theorem 2 extends the well-known result of [Ya,GoMi] that injective trapdoor functions yield semantically secure public-key cryptosystems, by showing that the injectivity requirement can be relaxed. Coupled with [ImRu] this also implies that it is unlikely that the analogue of Theorem 1 can be shown for trapdoor functions with polynomially bounded pre-image sizes. 4.1 Proof of Theorem 2 Let F = {Fk}keN be a family of trapdoor one-way functions with pre-image size bounded by a polynomial Q. The construction is in two steps. We first build an unapproximable family of trapdoor predicates P with decryption error 1/2 - 1/poly(k), and then reduce the decryption error by repetition to get the family claimed in the theorem. The first step uses the Goldreich-Levin inner-product construction [GoLe]. This construction says that if f is a one-way function, one can securely encrypt a bit b via f(x), r, a where a = b $ (x | r) with r a random string, x E Dom(f), and | denoting the inner-product mod 2. Now, if f is an _injeetive trapdoor func-_ tion, then with the trapdoor information, one can recover b from f(x), r, and a by finding x and computing b = a $ (x | r). If instead f has polynomial-size pre-images, the "correct" x will only be recovered with an inverse polynomial probability. However, we will show that the rest of the time, the success proba- 1 bility is exactly 50%. This gives a noticeable (�89 + pol-'~'~) bias towards the right value of b. Now, this slight bias needs to be amplified, which is done by repeat- ing the construction many times in parallel and having the decryptor take the majority of its guesses to the bit in the different coordinates. A full description and proof follow. We may assume wlog that there is a polynomial _l(k)_ such that Range(f) C {0, 1} i(k) for all f e [F~] and all k E N. We now describe how to use the Goldreich-Levin inner-product construction [GoLe] to build P = {Pk}keN. We ----- 292 Predicate _p(b)_ _// Takes input a bit b_ x ~ Dom(f) _// Choose x at random from the domain of f_ r r {0,1} l(k) _//Choose a random l(k)-bit string_ a := _b$ (x@r)_ _//XOR b with the GL bit_ Output _(f(x), r, a)_ Here (9 denotes XOR (ie. addition rood 2) and @ denotes the inner-product rood 2. The generator algorithm for P will choose _(f, tp) ~ F-Gen(1 k)_ and then output (p, tp) with p defined as above. Notice that p is computable in PPT if f is. The inversion algorithm _P-Inv_ is given p, the trapdoor _tp,_ and a triple (y, r, a). It first runs the inversion algorithm _F-Inv_ of F on inputs _f, tp, y_ to obtain x I, and then outputs the bit b' = a (9 (x'@ r). It is clear that the inversion algorithm is not always successful, but in the next claim we prove that it is successful appreciably more often than random guessing. _Claim. P_ is an unapproximable trapdoor predicate family, with decryption error at most (1/2) - 1/[2Q(k)]. _Proof._ We know that F is one-way. Thus, the inner product is a hardcore bit for F [GoLe]. This implies that P is unapproximable. It is left to show that the decryption error of P is as claimed, namely that _DecErrp(P-Inv, k) (as defined_ in Equation (1)) is at most (1/2) - _1/[2Q(k)]._ Fix _f, tp, b,_ let x, r be chosen at random as by p(b), let y = _f(x),_ let a = b ~ (x @ r), let x ~ ~- _F-Inv(f, tp, y),_ and let U = a (9 (x' @ r). Notice that if x' = x then b ~ = b, but if x ' ~ x then the random choice of r guarantees that b ~ -- b with probability at most 1/2. (Because _F-Inv,_ who generates x ~, gets no information about r.) The chance that x = x' is at least _1/Q(k)_ (because F-Inv gets no information about x other than that _f(x) = y)_ so _DecErrp(P-Inv, k) < __ (1-Q-~k)) "~ 1 as desired. [3 Now, we can iterate the construction _q(k) de=f O(kQ(k)2 )_ times independently and decrypt via a majority vote to reduce the decryption error to e -k. In more detail, our final predicate family Pq = {P~}keN is like this. An instancep q E [P~] is still described by a function f E [Fk] and defined as _pq(b) = p(b)ll.., lip(b),_ meaning it consists of q(k) repetitions of the original algorithm p on independent coins. The inversion algorithm _Pq-Inv is given the trapdoor_ _tp_ and a sequence of triples (Yl, rl, 0"1)[]""" [l(Yq(k) , rq(k) , ff q(k) ) . For i = 1,... _,q(k)_ it lets _b~ = P-Inv(p, tp, (yi, ri,ai))._ It outputs b ~ which is 1 if the majority of the values b~, ... ,bq(k) are 1, and 0 otherwise. Cher- noff bounds show that DecErrpq (Pq-Inv, k) < e -k. Furthermore standard "hy- brid"arguments [GoMi,Ya] show that _Pq_ inherits the unapproximability of P. ----- 293 #### Remark 3. Notice that Theorem 2 holds even if the family F only satisfies a very weak trapdoor property -- namely, that _F-Inv_ produces an element of f-1 (y) with probability at least _lip(k)_ for some polynomial p. Essentially the same proof will show that _P-Inv_ can guess b correctly with probability at least 1/2 + _1/[2Q(k)p(k)]._ ``` 5 From cryptosystems to trapdoor functions In this section we investigate the relation between semantically secure public key cryptosystems and injective trapdoor functions. It is known that the exis- tence of unapproximable trapdoor predicates is equivalent to the existence of semantically secure public-key encryption [GoMi]. It is also known that injective trapdoor one-way functions can be used to construct unapproximable trapdoor predicates ~Ya] (see also [GoLe]). In this section, we ask whether the converse is true: Question I. Can unapproximable trapdoor predicates be used to construct in- jective trapdoor one-way functions? Note the importance of the injectiveness condition in Question 1. We already know that non-injective trapdoor functions can be constructed from trapdoor predicates (whether the latter are injective or not) because trapdoor predicates imply one-way functions [ImLu] which in turn imply trapdoor functions by Theorem 1. We suggest a construction which requires an additional "random looking" function G and prove that the scheme is secure when G is implemented as a random oracle (to which the adversary also has access). Hence, IF it is possible to implement using one-way functions a function G with "sufficiently strong randomness properties" to maintain the security of this scheme, then Question 1 would have a positive answer (as one-way functions can be constructed from unapproximable trapdoor predicates [ImLu]). The key difference between trapdoor functions and trapdoor predicates is that predicates are probabilistic, in that their evaluation is a probabilistic process. Hence, our construction is essentially a de-randomization process. Suppose we have a family P of unapproximable trapdoor predicates, and we want to construct a family F of injective one-way trapdoor functions from P. A first approach would be to take an instance p of P and construct an instance f ``` ofF as #### f(blb2.., bkHrlH... ]Irk) = p(bl; rl)[]... [Ip(bk; r~), where k is the security parameter. Standard direct product arguments [Ya] im- ply that F constructed in this manner is one-way. However, F may fail to be trapdoor; the trapdoor information a~sociated with p only allows one to recover bl,..., bk, but not rl ,..., r~. Our approach to fixing this construction is to instead have rl,..., rk deter- mined by applying some "random-looking" function G to bl,..., bk: ----- 294 Since G must be length-increasing, an obvious choice for G is a pseudo-random generator. A somewhat circular intuitive argument can be made for the secu- rity of this construction: If one does not know bl,... ,bk, then rl,... ,rk "look random," and if rl,..., rk "look random," then it should be hard to recover _bl,..., bk by the unapproximability of P. In the full version of the paper [BHSV],_ we show that this argument is in fact false, in that there is a choice of an un- approximable trapdoor predicate P and a pseudorandom generator G for which the resulting scheme is insecure. However, it is still possible that there are choices of functions G that make the above secure. Below we show that the scheme is secure when G is implemented as a truly random function, ie. a random oracle (to which the adversary also has access). Intuitively, having access to the oracle does not help the adversary recover bl ... bk for the following reason: the values of the oracle are irrelevant except at bl .." bk, as they are just random strings that have nothing to do with bl... bk or _f(bl.., bk)._ The adversary's behavior is independent of the value of the oracle at bl--.ba unless the adversary queries the oracle at bl...bk. On the other hand, if the adversary queries the oracle at bl ... bk, it must already "know" bl ... bk. Specifically, if the adversary queries the oracle at _bl .." bk with_ non-negligible probability then it can invert / with non-negligible probability without making the oracle call, by outputting the query. We now proceed with a more formal description of the random oracle model and our result. THE RANDOM ORACLE MODEL. In any cryptographic scheme which operates in the random oracle model, all parties are given (in addition to their usual re- sources) the ability to make oracle queries [BeRo]. It is postulated that all oracle queries, independent of the party which makes them, are answered by a single function, denoted O, which is uniformly selected among all possible functions (where the set of possible functions is determined by the security parameter). The definitions of families of functions and predicates are adapted to the ran- dom oracle model in a straightforward manner: We associate some fixed poly- nomial Q with each family of functions or predicates, such that on security parameter k all the algorithms in the above definitions are given oracle access to a function (9 : {0, 1}* -~ {0, 1} Q(k). The probabilities in these definitions are then taken over the randomness of these algorithms and also over the choice of O uniformly at random among all such functions. **Theorem** 3. _If there exists a family of unapproximable trapdoor predicates, then_ _there exists a family of injective trapdoor one-way functions in the random oracle_ _model._ _Remark 4._ Theorem 3 still holds even if the hypothesis is weakened to only re- quire the existence of a family of unapproximable trapdoor predicates _in the_ _random oracle model._ To see that this hypothesis is weaker, note that a family of unapproximable trapdoor predicates (in the standard, non-oracle model) re- mains unapproximable in the random oracle model, as the oracle only provides randomness which the adversary can generate on its own ----- 295 See Sections 1.2 and 1.3 for a discussion of the interpretation of such a result. We now proceed to the proof. **5.1** **Proof of Theorem 3** Let P = {P~}keN be a family of unapproximable trapdoor predicates. Let _q(k)_ be a polynomial upper bound on the number of random bits used by any p E Pk. When used with security parameter k, we view the oracle as a function O: {0, 1}* -+ {0, 1} kq(k). We define a family F = {Fk}keN of trapdoor functions in the random oracle model as follows: We associate to any p E [P~] the function f defined on input bx...bk E {0,1} k by _f(bl"" bk) =_ p(bl; rl)ll" �9 IIp(b~; r~), where #### rill'" Ilrk = O(bl... bk), ri e {0,1} q(k) . The generator _F-Gen_ takes input 1 ~, runs _(p, tp) +-- P-Gen(1 k)_ and outputs (f, tp) where f is as defined above. It is clear that f can be evaluated in poly- nomial time using the evaluator _P-Eval_ for p. Notice that f can be inverted given the trapdoor information. Given _f, tp,_ and YllI"" [lYk = f ( bl . . . bk ) , inverter _F- Inv_ computes _b~ = P- Inv(p, tp, yi )_ for i -- 1,...,k, and outputs bl... b~. Furthermore, f is injective because P has zero decryption error: in this inversion process, _P-Inv_ correctly returns _bi,_ so we correctly recover the full input. It remains to show that F is one-way. _Claim. F_ is one-way. We prove this claim by describing several probabilistic experiments, modifying the role of the oracle with each experiment. The first arises from the definition of a family of one-way functions in the random oracle model. Let A be any PPT, let k be any positive integer, and let _q = q(k)._ _Experiment 1._ (1) Choose a random oracle O : {0, 1}* -+ {0, 1} kq(k). (2) Choose p ~- Pk (3) Select _bl,..., bk_ uniformly and independently from {0,1}. (4) Let rill"" Ilrk = O(bl... bk), where Iril _= q(k)_ for each i. (5) Let _x = p(bl;rl)[[... Hp(bk;rk)._ (6) Compute _z ~ AO(lk,p,x)._ We need to prove the following: _Claim._ For every PPT A, the probability that z = bt .-. b~ in Experiment 1 is a negligible function of k. To prove Claim 5.1, we first analyze what happens when the ri's are chosen independently of the oracle, as in the following experiment: Let A be any PPT, ----- 296 _Experiment 2._ (1)-(3) As in Experiment 1. (4) Select rl,..., rk uniformly and independently from {0, 1} q. (5)-(6) As in Experiment 1. _Claim._ For every PPT A, the probability that z -- bl .-. bk in Experiment 2 is a negligible function of k. Claim 5.1 follows from standard direct product arguments [Ya,GNW]. Specifi- cally, Claim 5.1 is a special case of the uniform complexity version of the Con- catenation Lemma in [GNW, Lemma 10]. _Claim._ For every PPT A, the probability that (9 is queried at point bl... _bk_ during the execution of _A~_ _x)_ in Step 6 of Experiment 2 is a negligible function of k. _Proof._ Suppose that the probability that (9 is queried at point bl." bk was greater that _1Is(k)_ for infinitely many k, where s is a polynomial. Then we could obtain a PPT A ~ that violates Claim 5.1 as follows. Let _t(k)_ be a polynomial bound on the running time of A. A ~ does the following on input (lk,p, x)" (1) Select i uniformly from {1,...,t(k)}. (2) Simulate A on input (lk,p, x), with the following changes: (1) Replace the oracle responses with strings randomly selected on-line, with the condition that multiple queries at the same point give the same answer. (2) Halt the simulation at the i'th oracle query and let w be this query. (3) Output w. Then A ~, when used in Experiment 2, outputs bx ... bk with probability greater that _1/(s(k)t(k))_ for infinitely many k, which contradicts Claim 5.1. [] In order to deduce Claim 5.1 from Claims 5.1 and 5.1, we give an equivalent reformulation of Experiment 1: Let A be any PPT, let k be any positive integer, and let _q = q(k)._ _Experiment 3._ (1)-(3) As in Experiment 1. (4) Select rl,..., rk uniformly and independently from {0, 1} q. (5) Let _x = p(51; rl)ll.., liP (bk; rk)._ (6) Modify (9 at location bl... bk to have value rill"" Ilrk. (7) Compute _z +--A~_ We now argue that Experiment 3 is equivalent to Experiment 1. In Experiment 1, rl,..., rk are uniformly and independently distributed in {0, 1} q and after Step 5 of Experiment 1 the only information about the oracle that has been used is that _rl H"" [Irk = (9(bl.." bk)._ Thus, the final distribution on all random vari- ables are identical in the two experiments and it suffices to prove Claim 5.1 for ----- 297 _Proof._ Let E be the event that z = bl ... bk in Experiment 3. Let F be the event that O is queried at point bl". _bk_ during the execution of A ~ (p, x) in Step 7 of Experiment 3. To show that E occurs with negligible probability, it suffices to argue that both F and E A F occur with negligible probability. First we show that F occurs with negligible probability. Notice that whether or not _A ~_ queries O at bl.'. b~ in Experiment 3 will not change if Step 6 is removed. This is because its behavior cannot be affected by the change in _O(bl ... bk)_ until it has already queried that position of the oracle. If Step 6 is **removed from Experiment 3, we obtain Experiment 2. Hence, the probability of** F is negligible by Claim 5.1. Similarly, the probability that [z = bl ... bk and _A ~_ never queries the oracle at bl... bk] will not change if Step 6 is removed. Thus, the probability of E D F is bounded above by the probability that z -- bl... bh in Experiment 2, which is negligible by Claim 5.1. D _Remark 5._ If the family of unapproximable trapdoor predicates we start with has negligible decryption error, then the family of trapdoor functions we construct will in general also have negligible decryption error and may fail to be injective with some small probability. By first reducing the decryption error of the predicate family to exp(-~(k3)) as in the proof of Theorem 2 and then using the oracle to derandomize the inversion algorithm, one can produce an _injective family that has_ _zero decryption_ error with probability 1 - 2 -k (where the probability is just taken over the choice of the oracle). Acknowledgments The first author was supported by a 1996 Packard Foundation Fellowship in Science and Engineering, and by NSF CAREER Award CCR-9624439. The third and fourth authors were supported by DOD/NDSEG Graduate Fellowships and partially by DARPA grant DABT-96-C-0018. The starting point of this research was a question posed to us by Shaft Gold- wasser, namely whether trapdoor permutations could be built from the assump- tions underlying the Ajtai-Dwork cryptosystem. Thanks to Oded Goldreich and the members of the Crypto 98 program com- mittee for their comments on the paper. **References** `[AjDw]` **M.** AJTAI AND **C. DWORK. A** public-key cryptosystem with worst-case / average-case equivalence. _Proceedings_ of the 29th Annua/Symposium _on_ _the Theory of Computing,_ ACM, 1997. #### [AMM] ADLEMAN, ~VIANDERS AND MILLER. On taking roots in finite fields. Proceed- _ings of the_ 18th _Symposium on Foundations of Computer Science,_ IEEE, #### 1977. [BHSV] M. BELLARE, S. HALEVI, A. SAHAI, AND S. VADHAN. Many-to-one trapdoor functions and their relation to public-key cryptosystems. Full version of this ----- **298** [BeRo] `M. BELLARE AND P. ROGAWAY. Random oracles are practical: a paradigm` for designing efficient protocols. _Proceedings of the First Annual Conference_ ``` on Computer and Communications Security, ACM, 1993. ``` [Bel `E. BERLEKAMP. Factoring polynomials over large finite fields. Mathematics` ``` of Computation, Vol. 24, 1970, pp. 713-735. ``` [BIMi] `~]L BLUM AND S. MICALI. How to generate cryptographically strong se-` ``` quences of pseudo-random bits, SIAM Journal on Computing, Vol. 13, No. 4, 850-864, November 1984. ### [ca] R. CANETTI. Towards realizing random oracles: Hash functions that hide all ``` partial information. _Advances in Cryptology - Crypto 97_ _Proceedings,_ Lec- ture Notes in Computer Science Vol. 1294, B. Kaliski ed., Springer-Verlag, 1997. ### [ca R. CANETTI, O. GOLDREICH AND S. HALEVL The random oracle model, revisited. _Proceedings of the_ 30th _Annual Symposium on the Theory of_ _Computing,_ ACM, 1998. [DiHe] V~ r. DIFFIE AND M. HELLMAN. New directions in cryptography. IEEE Trans. _Info. Theory,_ Vol. IT-22, No. 6, November 1976, pp. 644-654. `[DDN]` D. DOLEV, C. DWORK, AND M. NAOR. Non-Malleable Cryptography. Pro- _ceedings of the 23rd Annual Symposium on the Theory of Computing,_ ACM, 1991. ``` [EIO] T. EL GAMAL. A public key cryptosystem and a signature scheme based on discrete logarithms. IEEE Trans. Inform. Theory, Vol. 31, 1985, pp. 469-472. ### [GoLe] O. GOLDREICH AND L. LEVIN. A hard predicate for all one-way functions. ``` _Proceedings o[ the_ 21st _Annual Symposium on the Theory o[ Computing,_ ### ACM, 1989. [GoMi] S. GOLDWASSER AND S. MICALL Probabilistic Encryption. Journa/of Com- _puter and System Sciences, Vol. 28, April 1984, pp. 270-299._ ### [ swl O. GOLDREmH, N. NISAN, AND A. WIODERSON. On Yao's XOR Lemma. _Electronic Colloquium on Computational Complexity,_ TR95-050. March 1995. _http://r~w.eccc.uni-trier.de/eccc/_ [HILL] J. H],STAD, P~. IMPAOLIAZZO, L. Lv.VIN AND M. LUBY. Construction of a pseudo-random generator from any one-way function. Manuscript. Earlier versions in STOC 89 and STOC 90. ### [ImLu] R,. IMPAGLIAZZO AND M. LUBY. One-way Functions are Essential for Complexity-Based Cryptography. _Proceedings of the_ 30th _Symposium on_ _Foundations of Computer Science, IEEE, 1989._ #### [ImR.l R. IMPAGLIAZZO AND S. RUDICH. Limits on the provable consequences of one-way permutations. _Proceedings of the 21st Annual_ _Symposium on the_ _Theory of Computing,_ ACM, 1989. [NaYul M. NAOR AND 1VI. YUNG. Public-Key Cryptosystems Provably Secure against Chosen Ciphertext Attacks. _Proceedings of the 22nd Annual Sym-_ posium on _the Theory of Computing,_ ACM, 1990. `[Rab]` M. RABIN. Digitalized Signatures and Public Key Functions as Intractable as Factoring. _MIT/LCS/TR-212,_ 1979. **[Ya]** A. YAO. Theory and applications of trapdoor functions. _Proceedings of the_ ### 23rd Symposium on Foundations of Computer Science, IEEE, 1982. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/BFb0055735?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/BFb0055735, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://link.springer.com/content/pdf/10.1007/BFb0055735.pdf" }
1,998
[ "JournalArticle", "Conference" ]
true
1998-08-23T00:00:00
[]
12,581
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00eca0ba653718274f5b0593a06dda278759d534
[ "Computer Science", "Medicine" ]
0.841708
Adaptive Indoor Area Localization for Perpetual Crowdsourced Data Collection
00eca0ba653718274f5b0593a06dda278759d534
Italian National Conference on Sensors
[ { "authorId": "51183994", "name": "Marius Laska" }, { "authorId": "30601143", "name": "J. Blankenbach" }, { "authorId": "1790902", "name": "R. Klamma" } ]
{ "alternate_issns": null, "alternate_names": [ "SENSORS", "IEEE Sens", "Ital National Conf Sens", "IEEE Sensors", "Sensors" ], "alternate_urls": [ "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-142001", "http://www.mdpi.com/journal/sensors", "https://www.mdpi.com/journal/sensors" ], "id": "3dbf084c-ef47-4b74-9919-047b40704538", "issn": "1424-8220", "name": "Italian National Conference on Sensors", "type": "conference", "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-142001" }
The accuracy of fingerprinting-based indoor localization correlates with the quality and up-to-dateness of collected training data. Perpetual crowdsourced data collection reduces manual labeling effort and provides a fresh data base. However, the decentralized collection comes with the cost of heterogeneous data that causes performance degradation. In settings with imperfect data, area localization can provide higher positioning guarantees than exact position estimation. Existing area localization solutions employ a static segmentation into areas that is independent of the available training data. This approach is not applicable for crowdsoucred data collection, which features an unbalanced spatial training data distribution that evolves over time. A segmentation is required that utilizes the existing training data distribution and adapts once new data is accumulated. We propose an algorithm for data-aware floor plan segmentation and a selection metric that balances expressiveness (information gain) and performance (correctly classified examples) of area classifiers. We utilize supervised machine learning, in particular, deep learning, to train the area classifiers. We demonstrate how to regularly provide an area localization model that adapts its prediction space to the accumulating training data. The resulting models are shown to provide higher reliability compared to models that pinpoint the exact position.
# sensors _Article_ ## Adaptive Indoor Area Localization for Perpetual Crowdsourced Data Collection **Marius Laska** **[1,]*** **, Jörg Blankenbach** **[1]** **and Ralf Klamma** **[2]** 1 Geodetic Institute and Chair for Computing in Civil Engineering & Geo Information Systems, RWTH Aachen University, Mies-van-der-Rohe-Str. 1, 52074 Aachen, Germany; blankenbach@gia.rwth-aachen.de 2 Advanced Community Information Systems Group (ACIS), RWTH Aachen University, Lehrstuhl Informatik 5, Ahornstr. 55, 52074 Aachen, Germany; klamma@dbis.rwth-aachen.de ***** Correspondence: marius.laska@gia.rwth-aachen.de Received: 15 January 2020; Accepted: 3 March 2020; Published: 6 March 2020 [����������](https://www.mdpi.com/1424-8220/20/5/1443?type=check_update&version=2) **�������** **Abstract:** The accuracy of fingerprinting-based indoor localization correlates with the quality and up-to-dateness of collected training data. Perpetual crowdsourced data collection reduces manual labeling effort and provides a fresh data base. However, the decentralized collection comes with the cost of heterogeneous data that causes performance degradation. In settings with imperfect data, area localization can provide higher positioning guarantees than exact position estimation. Existing area localization solutions employ a static segmentation into areas that is independent of the available training data. This approach is not applicable for crowdsoucred data collection, which features an unbalanced spatial training data distribution that evolves over time. A segmentation is required that utilizes the existing training data distribution and adapts once new data is accumulated. We propose an algorithm for data-aware floor plan segmentation and a selection metric that balances expressiveness (information gain) and performance (correctly classified examples) of area classifiers. We utilize supervised machine learning, in particular, deep learning, to train the area classifiers. We demonstrate how to regularly provide an area localization model that adapts its prediction space to the accumulating training data. The resulting models are shown to provide higher reliability compared to models that pinpoint the exact position. **Keywords: indoor localization; area localization; crowdsourcing; fingerprinting; deep learning** **1. Introduction** In recent years, the usage of location-based services (LBS) has experienced substantial growth. This is mostly caused by the wide adoption of smartphones with the ability to reliably track a user’s location. Global Navigation Satellite Systems (GNSS), such as the Global Positioning System (GPS), are the dominant technology to enable LBS, since they offer accurate and reliable localization performance. However, GNSS do not provide sufficient availability and reliability inside buildings, since the satellite signals are attenuated and scattered by building features. This drawback has led to the development of various alternative indoor localization systems [1], which utilize a spectrum of techniques and technologies. Until today, there is not any gold standard for indoor localization, which can be stated as the main issue that has prevented indoor LBS from developing their full potential [2]. Indoor localization systems can serve different purposes. In monitor-based systems, the location of a user or entity is passively obtained relative to some anchor node [1]. This can be utilized, for example, to enhance the energy efficiency of buildings by automatically switching-off lighting and heating/cooling in empty rooms [3,4]. In contrast, in device-based systems, the location ----- _Sensors 2020, 20, 1443_ 2 of 26 information is obtained from a user-centric perspective [1], which can be utilized, for example, to enable navigation [5,6]. A variety of technologies and approaches are present in the field of indoor localization. Comprehensive overviews are given in [1,7–10]. In general, indoor localization systems can be grouped into (1) autonomous, (2) infrastructure-based and (3) hybrid systems. Autonomous systems apply inertial navigation [7]. In infrastructure-based systems, it can be differentiated between (2.1) analysis of signal propagation to dedicated transmitting stations and (2.2) scene analysis (fingerprinting) [11]. The former utilizes proximity, lateration or angulation measurements to estimate the user’s location. This requires line-of-sight and knowledge about the location of the stations. In contrast, fingerprinting does not rely on either. Instead, in an offline phase, the scene is scanned at certain reference points with a sensing device (e.g., smartphone). The observed sensor values at each reference point form so-called fingerprints. Using supervised machine learning (ML), a mapping from fingerprints to locations is learned, which is utilized to estimate the location for unseen fingerprints during online localization. Fingerprinting leverages existing infrastructure, which reduces upfront deployment cost. However, the accuracy of the system strongly depends on the quality of the offline site survey and the up-to-dateness of the fingerprint database. A crowdsourced site survey has been proposed to partition the collection among several participants and thus reduces the manual labeling effort [12–14]. Users either explicitly tag a fingerprint with a location, or the label is implicitly inferred by the system. The decentralized collection comes with the cost of heterogeneous data, which include among others device heterogeneity, labeling noise and an unequal spatial training data distribution [15]. Area localization can be applied in settings with imperfect data to achieve reliable positioning guarantees [16]. The problem is simplified such that the goal becomes to predict the right area instead of pinpointing the exact location. Existing area localization solutions employ a static segmentation into areas that is independent of the available training data [17–20]. This approach is not applicable for crowdsoucred data collection, since it features an unbalanced spatial training data distribution that changes over time. A segmentation is required that utilizes the existing training data distribution and adapts when new data is accumulated. The amount and shape of the areas, in particular, the richness of training data per area, affect the accuracy of classification models, which we subsequently call model performance. In addition, the expressive power is determined by the segmentation. If a model predicts one of few but large classes, the information gain of the user is lower compared to models that predict one of many smaller areas. We call the expressive power of the model that is determined by the segmentation expressiveness. Since crowdsourced data is expected to be generated continuously, the segmentation into areas as well as the successive classification model can be continuously improved. The challenge is, therefore, to continuously find a model with the right balance between expressiveness and performance given the most recent crowdsourced map coverage. The main contributions of this paper are summarized as follows: We introduce the concept of adaptive area localization to enable area classification for _•_ crowdsourced data that are continuously generated. We propose the idea of data-aware floor plan segmentation to compute segmentations that _•_ benefit subsequent classification. We present a clustering-based algorithm that determines such a segmentation with adjustable granularity. We formulate a metric to compare various area classifiers, such that the model, providing the _•_ optimal balance between expressiveness and performance, can be selected. This allows for automatic model building and selection in the setting of continuous crowdsourced data collection. We provide a comprehensive experimental study to validate the concepts on a self-generated and _•_ a publicly available crowdsourced data set. The rest of the paper is organized as follows: we introduce related work in Section 2 focusing on crowdsourced data collection, area classification and deep learning. Subsequently, Section 3 introduces the proposed concepts of adaptive area classification in detail. In Section 4 we present the locally _dense cluster expansion (LDCE) algorithm for computing floor plan segmentations with adjustable_ ----- _Sensors 2020, 20, 1443_ 3 of 26 granularities that are based on the available training data. Section 5 covers details regarding machine learning model building for area classification. In Section 6 the proposed concepts are evaluated on a self-generated as well as a publicly available crowdsourced data set. Finally, we discuss our findings in Section 7 and draw conclusion in Section 8. **2. Related Work** Fingerprinting-based indoor localization commonly utilizes a two stage approach. In the offline phase, radio frequency (RF) signals are collected at certain reference points and tagged with the position of collection. An algorithm is used to find a mapping from unknown fingerprints to locations. This algorithm is then applied during the online phase to localize an RF device [21]. The RF technology of choice for fingerprinting is commonly WLAN, however, solutions have been proposed that utilize alternative RF technologies such as LTE [22]. The most common approach for constructing a WLAN fingerprint is the received signal strength (RSS), which can be used either directly [23,24] or after feature extraction [25–28]. Recent studies on fingerprinting also incorporate channel state information (CSI) as input data in order to obtain more accurate prediction results [29–31]. The underlying assumption states that RSS values do not exploit the subcarriers in an orthogonal frequency-division multiplexing (OFDM). Therefore, CSI contains richer multipath information [32], which is beneficial for training complex models. However, obtaining CSI data is only achievable with certain Wi-Fi network interface cards (NIC) and thus is currently not suitable for smartphone based data collection like crowdsourcing. In this work, we focus on classical WLAN fingerprinting and utilize the RSS of scanned access points to construct the radio frequency map. _2.1. Crowdsourcing_ Several approaches have been proposed to reduce the manual labeling effort during crowdsourced data collection for Wi-Fi fingerprinting [12–14]. Rai et al. [12] were among the first to present a probabilistic model to infer the position of implicitly collected fingerprints. They periodically collected the RSS together with the timestamps of collection. Simultaneously, the system tracks the user utilizing a particle filter. After convergence, the path information is used to annotate the RSS measurements with a location. Radu and Marina [13] additionally integrated activity recognition and Wi-Fi fingerprinting via a particle filter to detect certain anchor points, such as elevator or stairs. He and Chan [33] utilized proximity information to Internet-of-things (IoT) sensing devices and the initially sparse RSS radio map to label fingerprints during implicit crowdsourcing. The IoT devices can be fixed, such as installed beacon transmitters or moving (smartphones of other participants). Santos et al. [14] utilized pedestrian dead reckoning (PDR) techniques to reconstruct the movements of users and classified the resulting trajectories using Wi-Fi measurements. Similar segments have been identified using an adaptive approach based on geomagnetic field distance. Finally, floor plans were reconstructed through a data fusion process and the collected Wi-Fi fingerprints were aligned to physical locations. Zhou et al. [34] abstracted the indoor maps as semantics graph. Crowdsourcing trajectories were mapped to the floor plan by applying activity detection and PDR. The annotated trajectories have been utilized to construct the radio map. Based on unfixed data collection, Jiang et al. [35] proposed the construction of a probabilistic radio map, where each cell was assigned a probability density function (PDF) instead of a mean value as in classical site survey approaches. Wei et al. [18] utilized the knowledge of location during the payment process inside the shops of a mall. They utilized this to annotate collected fingerprints with the current shop to build a hierarchical classification model that provides shop-level localization. In contrast to probabilistic fingerprint annotation, unsupervised learning can be utilized to obtain labeled Wi-Fi fingerprints [36,37]. Jung and Han [37] utilized unsupervised learning to infer the location of access points together with a path loss model and optimization algorithm, which they presented in [36]. They investigated how to adaptively recalibrate the resulting map to avoid performance degradation of downstream localization models. ----- _Sensors 2020, 20, 1443_ 4 of 26 Besides the reduction of labeling effort when collecting data via crowdsourcing, there are several additional challenges that have to be considered. Ye and Wang [15] identifed four major problems, which are: Inaccurate position tags for crowdsourced fingerprints that might occur during manual labeling _•_ of non-experts or are caused by automatic labeling via probabilistic models. The fluctuating dimensionality of RSS signals caused by varying numbers of hearable access _•_ points for various locations. The device heterogeneity that causes RSS to differ across various devices for the same _•_ measurement position. The nonuniform spatial data distribution, meaning that some areas feature a larger amount of _•_ data, while for others no data was collected. They constructed device-specific grid fingerprints utilizing clustering-based algorithms. For sparse areas fingerprints are interpolated and finally, the samples from several devices are fused to obtain device independent grid fingerprints. Yang et al. [25] additionally identified the short measurement time of crowdsourced sample collection as a typical problem. They utilized the fact that the most-recorded RSS does not differ much, irrespective of the length of measuring, to extract a characteristic fingerprint. In a follow up work, Kim et al. [26] evaluated the system in a case study and demonstrated its effectiveness. Pipelidis et al. [38] proposed an architecture for cross-device radio map construction via crowdsourcing. They utilized data labeled via a simultaneous localization and mapping (SLAM)-like algorithm. The RSS values between devices were calibrated via reference measurements at several landmarks. The data was clustered and subsequently used for classification of areas. _2.2. Area Localization_ In contrast to localization systems that aim at pinpointing the exact position of a user, the concept of area classification only focuses on estimating the current area of the user, such as the office room or the shop inside a mall. This is particularly suitable for large scale deployments or in situations where the data quality does not allow for accurate localization. Lopez Pastor et al. [17] evaluated a Wi-Fi fingerprinting-based indoor localization system inside a medium sized shopping mall. The system is meant for providing shop-level accuracy, while minimizing the deployment cost and effort. Data is collected by randomly walking in predefined areas, such that all data can be labeled with the corresponding shop. The authors claim that the achieved system performance is sufficiently independent of the device and does not deteriorate over time. Wei et al. [18] adopted a similar approach. They utilized the fact that during payment inside a shop, the location of the user is known. This can be used to annotate Wi-Fi fingerprints collected while paying. The obtained fingerprints can be utilized for shop-level position estimation. Rezgui et al. [19] proposed a variation of a support vector machine (SVM) (normalized rank based SVM) to address the problem of hardware variance and signal fluctuation of Wi-Fi based localization systems. The system achieves room level prediction accuracies. He et al. [16] compared the performance of various classification models, such as SVM, artificial neural network (ANN) and deep belief network (DBN) for various test sites. They addressed the identification of floors, indoor/outdoor and buildings. In a recent follow up work [39], they also tackled the inside/outside region decision problem and propose solutions for missing AP detection and fingerprint preprocessing. Liu et al. [20] proposed an algorithm for probability estimation over possible areas. By adopting the user’s trajectory and existing map information, they eliminate unreasonable results. The partitioning of the map into areas is done manually based on the different rooms and offices. _2.3. Deep Learning for Fingerprinting_ Fingerprinting-based indoor localization can be formulated as standard supervised learning problem. It can be modeled as regression problem with the goal to predict the exact position, or as a ----- _Sensors 2020, 20, 1443_ 5 of 26 classification task on predetermined areas. Due to the recent success of deep learning in areas such as image processing or speech recognition, the application of deep models for fingerprinting-based indoor localization has gained attention recently. Nowicki and Wietrzykowski [40] applied stacked autoencoders combined with a feed forward neural network for building and floor prediction. Xiao et al. [23] compared SVM and a deep neural network (DNN) on various publicly available data sets and propose a data augmentation schema as well as an approach for transfer learning. Adege et al. [28] applied regression analysis to fill missing RSS values and utilize linear discriminant analysis for dimensionality reduction. Finally, feed forward neural networks are applied to tackle the regression and classification problem. Kim et al. [41] formulated the problem as multi-label classification problem to predict the building, floor and position with a single network with minimal performance degradation. Mai et al. [42] utilized a convolutional neural network (CNN) on raw RSS data by applying the convolution on time-series data. The data is artificially constructed by combining measurements within a certain cell size that have been captured in temporal intervals not exceeding a certain threshold. By constructing an image of the RSS vector, CNNs that are predominantly used for image classification can be applied. Mittal et al. [27] filtered access point signals that have a low Pearson Correlation Coefficient (PCC) between the access point values and the location vector. The remaining RSS vector is transformed into an image matrix by multiplying each access point vector with the obtained correlation values and arranging as matrix with zero padding. Sinha et al. [24] simply arranged the RSS vector as a matrix to train a standard CNN image classifier. They proposed a data augmentation scheme where single values of the RSS vector are replaced by random values sampled from the interval of the difference of the actual value and the access point mean value. **3. Adaptive Area Classification for Crowdsourced Data** In the following section, we introduce our approach to adaptive area classification. We describe the concept overview and introduce relevant notations. Subsequently, a floor plan segmentation is formally defined and classification models for indoor localization are described. Finally, we propose a novel metric called ACS, which is utilized to select area classifiers with respect to the optimal balance between expressiveness and performance. _3.1. Concept Overview_ The performance of Wi-Fi fingerprinting-based indoor localization systems heavily relies on thorough and up-to-date site survey data. Crowdsourced training data collection continuously provides fresh data, but suffers from poor data quality. Several approaches suggest to maintain an up-to-date radio map, which stores a representative fingerprint or a probabilistic distribution for predefined locations, regions, or grid cells [33,35,43]. Missing data for certain locations prohibits equal radio map quality at all areas. This is solved by either enlarging the areas of the radio map or by interpolating fingerprints for sparsely covered areas [15]. The update of such a radio map is a complicated process, since its granularity is static. However, the spatial distribution of available training data is expected to shift over time. Therefore, instead of maintaining a radio map with characteristic fingerprints for predefined areas, we store the entire training data with the noisy position tags. At regular intervals, we dynamically subdivide the floor plan into areas based on the richness of available training data. The training data, which are originally annotated with noisy position tags, are labeled with the corresponding areas based on the computed floor plan segmentation. This enables training of standard supervised machine learning classifiers that predict the correct area. In order to quantify the gain of such an area classifier, two metrics can be utilized. The expressiveness measures the information gain of the user, which is mainly influenced by the _•_ extent of each individual area and the total coverage of the model. The performance indicates how reliably the model predicts a certain area. _•_ ----- _Sensors 2020, 20, 1443_ 6 of 26 The two metrics are inversely proportional. That means a fine segmentation (high expressiveness) negatively affects the performance of the model and vice versa. We assume that fresh crowdsourced training data is accumulated over time. This enables updates of the floor plan segmentation and the successive area classifier. The workflow for continuously providing area localization models, where the prediction space adapts to the new training data, is illustrated in Figure 1. Over time, the map gets covered with an increasing amount of training data, which is illustrated in the top row of Figure 1. At regular intervals, the goal is to provide an optimal indoor area classification model based on the current map coverage. This process includes the automatic floor plan segmentation into areas and the training of an ML model. Several floor plan segmentations can be determined that influence the expressiveness of the ML model and for each of these segmentations several ML models can be learned. For each epoch, the best combination of segmentation and model is selected. This is done with respect to a metric, called area classification score (ACS). The ACS balances expressiveness and performance and is introduced in Section 3.5. t1 time t2 |Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16|Col17|Col18|Col19|Col20|Col21|Col22|Col23|Col24|Col25|Col26|Col27|Col28|Col29|Col30|Col31|Col32|Col33|Col34|s|Col36|Col37|Col38|Col39|Col40|Col41|Col42|Col43|Col44|s|Col46|Col47|Col48|Col49|Col50|Col51|Col52|Col53|Col54|Col55|Col56|Col57|Col58|Col59|Col60|Col61|Col62|Col63|Col64|Col65|Col66|Col67| |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| |||||||||||||||||||Tr|ai|nin|g|||||||||||||ple||||||||||ple||||||||||||||||||||||| ||||||||||||||||||||da|ta||||||||||||||ng sam||||||||||ng sam||||||||||||||||||||||| |||||||||||||||||||||||||||||||||||# traini||||||||||# traini||||||||||||||||||||||| |Floor plan segmentation Expressiveness Model training & Model selection ACS = 0.4 ACS = 0.5||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||| **Figure 1. Concept of adaptive area classification for crowdsourced map coverage.** _3.2. Data Notations_ In the following, we introduce the formal notations that are subsequently used. We assume that at a certain point in time, a set of N labeled training data tuples (fingerprints) FP = { f pn = (xn, pn, tn)} for n = 1, ..., N has been collected for a given indoor map. Each fingerprint f p consists of a M-dimensional feature vector x = (x1, ..., xM)[T] and is tagged with a position pn = (px, py)[T] in two dimensions and the corresponding timestamp tn of collection. In the following we focus on Wi-Fi fingerprinting, such that each entry of the vector is the RSS value of the corresponding access point and M is equal to the total amount of access points that are observable for the map. Since not all access points are hearable at all locations, x contains missing entries, which have to be considered during further processing of the data. _3.3. Floor Plan Segmentation for Area Classification_ In order to train a classification model, we have to find a floor plan segmentation that assigns each fingerprint tuple (x, p, t) to one of the K areas or classes, Ck for k = 1, .., K. A floor plan segmentation determines a mapping SEG : Ck _Ak, where Ak might be any two-dimensional shape, such as a_ _→_ rectangle. Given such a mapping SEG, we can label each fingerprint with the class label of the area it is located in. For a given segmentation SEG, we obtain the transformed set FPSEG = {(xn, cn)}, where cn ∈{1, ..., K} and cn = k ⇔ **pn lies within Ak. The goal is now to find a classifier C : x →** _ck_ **Training** **data** ----- _Sensors 2020, 20, 1443_ 7 of 26 that determines the correct area of the floor plan segmentation for an unknown RSS fingerprint. We have now arrived at the standard formulation of a supervised learning problem, in particular, a classification problem. _3.4. ML Models for Area Classification_ Given the transformed set of fingerprints FPSEG = {(xn, cn)} for a segmentation SEG, we can utilize any standard ML classification model that learns to predict the unknown class ck for a fingerprint **x. We can either construct a discriminant function that directly assigns a class to an unknown** fingerprint, or we model the conditional probability distribution p(ck|x) [44]. SVMs depict a typical discriminant model used in the domain of indoor localization, while with DNNs, it is possible to model p(ck|x). Both models are utilized in the experimental study (Section 6) as classifiers for the transformed fingerprint sets FPSEG. _3.5. Area Classification Score_ In order to properly quantify the quality of the learned area localization model (combination of segmentation and trained classifier), we have to simultaneously investigate the model’s expressiveness as well as its performance. The expressiveness is influenced by the total extent of covered area as well as the size of each individual area. We state that the expressiveness of a model is higher if it predicts classes associated with smaller areas. However, the benefit of a narrow prediction area vanishes if the performance for that specific class, for example the accuracy, is poor. To capture this interplay, we have to look at each predicted class of the classifier individually. We define areak as the surface area of the area Ak that belongs to class Ck. On an individual class level, we define the expressiveness of class Ck as: _expλ(Ck) =_ _[area][min]_, (1) _area[λ]_ _k_ where areamin is the minimal extent that an area might have by definition (set to 1m[2] in the following) and λ is a parameter to adjust the slope of the function. Additionally, a performance metric is required, which measures the accuracy of the model on a class level. We choose the F1 score, since we are equally interested in precision and recall. Let F1(Ck) be the class-based F1 score for class Ck, evaluated on a separate test set. The chosen metrics for expressiveness and performance reside in the interval [0, 1], such that we can multiply them to obtain a value in [0, 1], which would be optimal, if the predicted class has the minimal extent of 1m[2] and a F1-score of 1 on the test set. In order to account for the total covered area, we take the weighted mean of the product of expressiveness and performance using the area of each class. We finally arrive at: 1 _ACS =_ _areatot_ _k_ ### ∑ F1(Ck)[µ] · expλ(Ck) · areak, (2) _k=1_ which we call area classification score (ACS) in the following. The expressiveness term (1) regulates how much the class score adds to the weighted mean. For λ = 0, the regularization term vanishes, such that the area size of the specific class has no influence on the amount that is added to the mean. This means that two localization models with constant class-wise classification performance F1 achieve the same score if they cover the same area areacov, independent of the amount of classes and their individual size: _ACSλ=0 =_ _[area][cov]_ _· F1 ._ (3) _areatot_ It follows that if λ approaches 0, the metric becomes less sensitive to the individual area sizes. With respect to models covering a similar extent of the map, those that provide a higher performance will be rated higher, independent of the number and individual size of their areas. The closer λ gets to 1, the higher is the influence of individual area sizes. High performance on broad areas will ----- _Sensors 2020, 20, 1443_ 8 of 26 not add much to the weighted mean, since they are downscaled by the expressiveness factor. As a consequence, models with finer segmentations score higher, since the influence of area regularization outweighs the performance factor. For λ = 1, the score is only sensitive to the amount of total segments. The ACS becomes 1 _k_ _ACSλ=1 =_ ∑ _F1(Ck)[µ]_, (4) _areatot_ _k=1_ which will be higher for finer segmentations given that the same total extent of the map is covered. The parameter µ can be utilized for fine tuning. By setting it larger than 1, models with overall low performance are penalized. We found that λ has a greater impact on the model selection and suffices for our use-cases. Therefore, µ is set to 1 during subsequent application of the ACS. Figure 2 emphasizes how the parameter choice of λ affects the ACS for three artificial segmentations (a–c) . The rectangular boxes represent the prediction areas of the classifier and the numbers show the class-wise F1 scores on a separate test set. We stated that λ influences the expressiveness. In particular, the closer the value gets to 1, the more each individual class score is downscaled by the size of its area. As a consequence, a low λ value targets high performant models with lower expressiveness and a high λ value selects models with high expressiveness and lower performance. Given the three segmentations (a–c), we plot the ACS for all possible choices of λ in Figure 2d to investigate which model achieves the highest score (illustrated by the color below the curve). As expected, the broad segmentation (a) is selected for low lambda values (0–0.13), the medium segmentation (b) is chosen for values (0.13–0.37) and the fine segmentation (c) is chosen for higher values (0.37–1). In practice, a pool of models is trained such as (a–c). The λ parameter is fixed, such that the best scoring model is determined. If the model does not adhere to the required use case requirements, λ can be adjusted accordingly, such that a different model is obtained. 16 16 0.91 0.89 0.91 0.88 14 14 12 12 10 0.93 10 0.93 8 8 6 6 4 0.95 4 0.9 0.92 2 2 0 16 14 12 10 8 6 4 2 0 0 0 10 20 0 30 10 40 20 50 30 60 16 0.91 0.89 0.91 0.88 14 12 0.93 10 0.93 8 6 0.95 4 0.9 0.92 2 (a) broad segmentation (b) medium segmentation |Col1|Col2|Col3|Col4|Col5|Col6|SEG broad|Col8| |---|---|---|---|---|---|---|---| |||||||medi fine|um| ||||||||| ||||||||| ||||||||| ||||||||| 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.6 SEG 0.68 0.81 0.7 0.73 0.84 broad 0.5 medium fine 0.4 0.72 0.85 0.3 0.2 0.8 0.76 0.54 0.88 0.1 0.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0 10 20 30 40 50 60 (c) fine segmentation (d) ACS for different λ values **Figure 2. Illustration of the impact of λ on the ACS for pool of example models.** ----- _Sensors 2020, 20, 1443_ 9 of 26 **4. Floor Plan Segmentation Algorithms** In order to train an area classification model, we have to determine a mapping from areas to classes that we defined as floor plan segmentation. If we neglect the underlying training data distribution, we end up with segmentations where certain classes feature few to zero fingerprint samples. This results in unsatisfying classification performance. The goal should be to leverage the knowledge about available training data to compute a segmentation that benefits subsequent classification but still provides the best possible expressiveness. We call such a segmentation data-aware _floor plan segmentation and present an algorithm for this purpose in the following._ _4.1. Locally Dense Cluster Expansion (LDCE)_ In the following, we introduce the LDCE algorithm that computes a floor plan segmentation, in particular, a mapping SEG : Ck → _Ak that assign each class Ck a shape Ak. Given SEG, we can label_ fingerprints (x, p, t) with the class that belongs to the area Ak in which p is located. Let FP = { f pn = (xn, pn)} for n = 1, ..., N be a set of training data, we cluster the observations and determine the shapes _Ak based on the position labels of the resulting cluster members._ Initially, we detect a set of locally dense base clusters. This serves two purposes: (1) observations that are densely connected to a certain degree should not be separated and (2) fingerprints that are not part of any initially dense cluster should be considered as noise. Both conditions are fulfilled when applying a standard density based clustering algorithm such as the density-based spatial clustering of applications with noise (DBSCAN) algorithm. The resulting base clusters are subsequently expanded. Each round the closest clusters are determined and merged. Resulting clusters that contain the required amount of stop_size members are deleted from the expansion set and added to the set of final clusters. This process is continued until either no clusters are present in the expansion set, or the smallest minimal distance exceeds the maximal allowed merging distance max_eps. Remaining clusters with fewer than stop_size members are postprocessed. By setting the minMembers parameters lower than stop_size, those clusters having at least minMember members are added to the set of final clusters. All other remaining clusters are added to the closest final cluster. This routine yields clusters with definable bounds for the amount of members. Since clusters with more than stop_size members are excluded from the merging phase, any merged cluster might have at most 2 · stop_size members. However, besides the amount of available training data per segment, we require a reasonable segmentation that adheres to the physical floor plan structure. In particular, segmentations should minimize spreads across multiple walls if possible. Furthermore, since the feature vector of subsequent classification consists of the RSS vector, the similarity in RSS signal space should be considered during the segmentation phase. The approach we propose achieves this by constructing a particular distance function between fingerprints and clusters of fingerprints that is used in the previously described algorithm. Given two fingerprints f pu = (xu, pu) and f pv = (xv, pv), we define their distance as: _dist( f pu, f pv) = ||pu −_ **pv||2 + θ · |Wpu,pv** _| + ζ · ||xu −_ **xv||2,** (5) where Wpu,pv is the set of walls between pu and pv. Note that the main distance factor is the Euclidean distance between the position labels, while the difference between RSS vectors and the number of conflicting walls are used to penalize this base distance. The distance between clusters is based on centroid distance. We add another penalty term to account for final clusters that might lie between merging clusters. Let Ci and Cj be two clusters, pi, pj the average position labels and xi, xj the average RSS vectors, the distance is then given by: _dist(Ci, Cj) = ||pi −_ **pj||2 + θ · |Wpi,pj** _| + ζ · ||xi −_ **xj||2 + η · |Ci,j|,** (6) ----- _Sensors 2020, 20, 1443_ 10 of 26 where Ci,j is the subset of final clusters, such that C _f ∈Ci,j ⇔∃_ _f p f ∈_ _C_ _f |p f within bounds(pi, pj)._ In order to prevent merging of far distant clusters, with respect to the penalized distance function, we set a threshold max_eps on the maximal allowed merging distance of two clusters. Note that the choice of max_eps determines the maximal amount of allowed walls between two merging clusters. If we choose max_eps = θ · x + δ, it holds that for any δ < θ, there will be at most x − 1 separating walls between any merging cluster. After we have determined the clustering, we have to construct the two-dimensional shapes that represent the floor plan segmentation. Those are obtained by using the position labels of the respective cluster members. We can construct the shape by taking the bounding box around the labels, or computing the convex or concave hull. Figure 3 shows stages of an example run of LDCE. The clusters merge over time (a–c) until all clusters have at least stop_size members (d). In the example, the final segments are obtained from the bounding boxes around the labels of the class members. The pseude code of the algorithm can be found in Algorithm 1. 15.0 15.0 12.5 12.5 10.0 10.0 7.5 7.5 5.0 5.0 2.5 2.5 0.0 0.0 (a) (b) 0 10 0 20 10 30 20 40 |Col1|Col2|Col3|Col4|15.0 12.5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16| |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| ||||||||||||||||| ||||||||||||||||| ||10.0 7.5||||||||||||||| ||||||||||||||||| ||||||||||||||||| ||5.0 2.5 0.0||||||||||||||| ||||||||||||||||| ||||||||||||||||| ||||||||||||||||| ||||||||||||||||| 15.0 12.5 10.0 7.5 5.0 2.5 15.0 12.5 10.0 7.5 5.0 2.5 0.0 15.0 12.5 10.0 7.5 5.0 2.5 (c) (d) 0 10 0 20 10 30 20 40 **Figure 3. Illustration of LDCE segmentation. Clusters expand over time (a–c) until all clusters have** reached a size greater than the stop_size threshold (d). |Col1|Col2|Col3|Col4|15.0 12.5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16| |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| ||||||||||||||||| ||||||||||||||||| ||10.0 7.5||||||||||||||| ||||||||||||||||| ||||||||||||||||| ||5.0 2.5 0.0||||||||||||||| ||||||||||||||||| ||||||||||||||||| ||||||||||||||||| ||||||||||||||||| ----- _Sensors 2020, 20, 1443_ 11 of 26 **Algorithm 1 LDCE floor plan segmentation** 1: Inputs: Fingerprints: FP = { f pn = (xn, pn)} _▷_ _n = 1, ..., N_ Walls: W = {(xsw, ysw, xew, yew )} _▷_ _w = 1, ..., W_ Main parameters: stop_size, max_eps Distance penalties: θ, η, ζ DBSCAN parameters: eps, minPts Postprocessing: minMembers 2: Initialize: _dist[ f pu, f pv] ←||pu −_ **pv||2 + θ · |Wpu,pv** _| + ζ · ||xu −_ **xv||2** _▷_ 1 ≤ _u, v ≤_ _N_ _C_ _f inal ←{}_ _Cexp ←_ _DBSCAN(dist, eps, minPts)_ _▷_ Main routine 3: while |Cexp| > 1 and min_dist < map_eps do 4: _C_dist[Ci, Cj] ←||pi −_ **pj||2 + θ · |Wpi,pj** _| + ζ · ||xi −_ **xj||2 + η · |Ci,j|** _▷_ 1 ≤ _i, j ≤|Cexp|_ 5: _min_dist ←_ _min(C_dist)_ 6: _Cm, Cn ←_ _argmin(C_dist)_ 7: _Cmerged ←_ _Cm ∪_ _Cn_ 8: _Cexp ←_ _Cexp \ Cx, Cy_ 9: **if |Cmerged| > stop_size then** 10: _C_ _f inal ←_ _C_ _f inal ∪_ _Cmerged_ 11: **else** 12: _Cexp ←_ _Cexp ∪_ _Cmerged_ 13: **end if** 14: end while _▷_ Postprocessing 15: for all C in Cexp do 16: **if |C| > minMembers then** 17: _C_ _f inal ←_ _C_ _f inal ∪_ _Cmerged_ 18: _Cexp ←_ _Cexp \ C_ 19: **end if** 20: end for 21: for all C in Cexp do 22: Add C to closest C _f ∈_ _C_ _f inal if closer than 2 · max_eps_ 23: end for _▷_ Determine final shapes 24: Pk = {pi|(pi, xi) ∈ _C_ _f inalk_ _}_ _▷_ _k = 1, ..., |C_ _f inal|_ 25: Ak = convex_hull(Pk) 26: return A **5. Machine Learning Model Building** The complete pipeline of ML model building comprises (1) preprocessing of the data, (2) model training and (3) model selection and evaluation. Each step is explained in the following. ----- _Sensors 2020, 20, 1443_ 12 of 26 _5.1. Preprocessing_ 5.1.1. Feature Preprocessing The applied machine learning models require inputs of fixed dimensions. Each access point that is observed during data collection represents one dimension of the input vector. Having observed a total amount of M access points, we can construct a feature vector xn = (x1, ..., xM)[T], where xi for _i = 1, ..., M and n = 1, ..., N represents the RSS value of the i-th access point of the n-th measurement._ Given a collected training sample, there is not a RSS value for each access point. This can have two reasons: (1) the access point cannot be observed at the measuring position because it is out of range, or (2) the access point is in general observable for the given location, however, its RSS value could not be recorded in that specific sample. The second reason is caused by the response rate of an access point, which is correlated with the average observable RSS value for a location [45]. For both causes of unobservable access points, an artificial value has to be chosen as entry for the feature vector. A common practice, which neglects the response rate of access points, is to simply set all missing values to a low RSS value, such as -110dB. This approach is adopted in our experiments. For gradient-based learning algorithms such as DNNs or distance-based algorithms such as k-nearest neighbor (k-NN), it is crucial to normalize or standardize each feature column [46]. This speeds up the learning phase and prevents features with a longer range to outweigh other features. It can be distinguished between feature scaling/normalization and feature standardization (z-score normalization). Scaling linearly transforms the data into the interval [0, 1], while standardization transforms the data to have zero mean and standard deviation equal to one. Standardization is especially useful if the range of the features are unknown or the feature contains many outliers. For choosing the right normalization technique, we have to investigate the influence of the given map coverage. Let APa and APb be two access points that are far away, such that there is no location where both can be observed simultaneously. Let areaa and areab be the areas where either signals of APa or APb are received. A map coverage that contains much more samples of areaa does only have few samples with signal of APb. When standardizing the data of the map coverage, we encode a strong bias into the preprocessed data, since the feature column of APb is strongly influenced by the vast amount of zero entries. Such a bias might be tolerable if the distribution of training data matches the test data distribution. However, during online localization, users might request their position mostly within areab, which would result in worse performance. In order to prevent this bias towards the given map coverage, we simply apply column-wise feature scaling. For each AP it is likely that a sample exist which could not register any signal strength for the AP. As a conclusion, the minimum RSS value for all columns is equal to the supplementary value for missing data. 5.1.2. Floor Plan Segmentation (Parameter Choice) To obtain the class labeled set FPSEG, we partition the floor plan with the introduced LDCE algorithm. The choice of certain parameters of the LDCE algorithm depends on the given floor plan and the spatial distribution of available training data. The parameters eps and minPts determine the starting clusters that result from the initial DBSCAN execution. They should be chosen empirically, such that the sizes of starting clusters do not exceed the stop_size member threshold and not too many observations are considered as noise. The value of max_eps and the wall penalty should also be chosen empirically based on the given floor plan dimensions and the amount of walls that should be allowed within segments. The penalty term η is set to 2, since higher values might yield overlapping clusters during the initial DBSCAN execution. ζ is set to the highest penalty value of 20 to avoid intersecting final clusters. After those parameters are fixed, we can vary the stop_size and minMembers parameters to obtain multiple segmentations with various granularities. An overview of the parameters can be found in Table 1. Those parameters that depend on the given test site are revisited in the corresponding Sections 6.2 and 6.3. ----- _Sensors 2020, 20, 1443_ 13 of 26 5.1.3. Label Preprocessing For training of regression models, the labels consist of the set of positions {pn}, n = 1, ..., N, where each label is a two-dimensional vector representing the position tag. In case of area classification, the labels {yn} with yn = (y1, ..., yK)[T], n = 1, ..., N, for the set FPSEG are the one-hot encoded areas of the floor plan segmentation, where yi = 1 ⇔ _i = cn and 0 at all other positions. K represents the_ amount of segments of the given floor plan segmentation FPSEG. **Table 1. Parameter choice of LDCE for experiments.** **Main** **Postprocessing** **Penalties** **DBSCAN** **Data Set** **stop_size** **max_eps** **minMembers** **_θ_** **_ζ_** **_η_** **eps** **minPts** RWTH Aachen {80, 50} 30 {40,20} 10 2 20 2 3 Tampere, Finnland {100, 60} 50 {60, 40} 5 2 20 5 3 _5.2. Model Training_ In the upcoming case study in Section 6, we focus on three types of supervised machine learning models that are suitable to predict the area of unknown fingerprints. After hyperparameter tuning we end up with a DNN model that has 3 hidden layers (HL) and 512 hidden units (HU) per layer and utilizes rectified linear unit (ReLU) as activation function between layers. In order to learn the conditional probability distribution p(y **x), we apply softmax activation function for the output layer** _|_ together with multiclass cross-entropy loss. This choice can be derived by following a maximum likelihood approach [47]. The Adam optimizer, a variant of stochastic gradient descent (SGD), is utilized for iterative learning of the weights. To prevent overfitting, we apply early stopping, which stops the training phase if the performance on a separate validation data set does not increase for a specified amount of epochs. Furthermore, weight regularization within the loss function and dropout are applied. The complete parameterization of the tuned DNN is given in Table 2. In addition, we train a CNN with similar hyperparameters as suggested by [24], which consists of two convolutional layers of size (16 16), a Maxpool layer of size (8 8), a convolutional layer of size (8 8) and a _×_ _×_ _×_ Maxpool layer of size (8 × 8). In-between layers, we add dropout layers with dropping probability of 0.25 and utilize ReLu as activation function. Finally, a fully connected dense layer of size 128 is used with output softmax activation function. We found that rearranging the RSS vector as matrix with zero padding outperforms the proposed preprocessing method of [27] that utilize the PCC to reduce the dimensionality and scale the data per access point. Furthermore, we fit a SVM with RBF kernel, which we utilize as discriminative model to directly predict y. Additionally, we select two regression models (k-NN and DNN(reg)). The DNN regression model has the same configuration as the DNN classifier but uses a linear output activation function and mean squared error as loss function. The k-NN models apply the weighted version of the algorithm and are evaluated for three values of k, namely, 2,3 and 5. To validate whether explicitly training a classifier provides valuable results, we label the regression outputs with the closest area of the floor plan segmentation during postprocessing and compare them to the output of the area classifiers. **Table 2. DNN model hyperparameter configuration.** **HU** **HL** **Dropout** **Reg. Penalty** **lr** **Batch** **Epochs** **Loss** **Activation** **Optimizer** 512 3 0.2 0.06 0.0007 32 200 Cat. cross-entropy ReLU Adam _5.3. Model Evaluation_ For model evaluation, we require a splitting strategy into training and test data as well as a metric that indicates how well a model performs. Those are introduced for the different model types in the following. ----- _Sensors 2020, 20, 1443_ 14 of 26 _Splitting strategy:_ Area classifiers: The training data is labeled according to the computed floor plan segmentations. _•_ We apply k-fold cross validation with k=5, such that we arrive at 20% test data per fold. We utilize the stratified version to obtain a good representative of the whole data set in each split. Regression models: We choose a subset of testing positions by applying DBSCAN on the position _•_ labels only. Based on the resulting clusters we apply 5-fold cross validation, such that 20% of the clusters are used as testing data in each fold. _Metric:_ As metrics, we compute error vectors for the vectors of predictions and ground truth labels. Those error vectors can be visualized via an empirical cumulative distribution function, which we will refer to as CDF in the following. Area classifiers: The error vector consists of the pairwise distances between the centers of _•_ the predicted areas and the ground truth areas, which is zero in case of a correct prediction. The y-intercept of the CDF corresponds to the machine learning accuracy metric (ACC). The curve yields additional knowledge about the significance of misclassification. Furthermore, we report the F1 score (F1). Regression models: In case of exact position estimation, the error vector consists of the pairwise _•_ distances between predictions and ground truth positions. Selection via ACS: During model selection, we utilize the ACS as metric. This requires computing _•_ the class-wise F1 scores of the predicted and ground truth areas. **6. Experimental Evaluation** The subsequent experimental case study targets two separate questions: 1. Does adaptive area localization based on a data-aware floor plan segmentation provide more robust results than the standard regression approach for exact position estimation? In particular, is it suited for arbitrarily collected training data via crowdsourcing? 2. When crowdsourced training data is generated continuously, the area classifier has to adapt to the current data basis. This is accomplished by recomputing the underlying floor plan segmentation and retraining a classification model on the data labeled with the corresponding areas. In this setting, is the proposed ACS suited for automatic model selection among a pool of models that provide varying performances and expressivenesses? _6.1. Study Design_ In order to answer these questions, we conduct two experiments. _•_ _Static performance analysis (Sections 6.2.1 and 6.3.1): we compute two floor plan segmentations with_ varying granularities for a snapshot of collected training data. For each segmentation we train and evaluate various classification models. In addition, the performance of the proposed area classifiers is compared to standard regression models that aim at pinpointing the exact location. _Model selection via ACS for continuous data collection (Sections 6.2.2 and 6.3.2): we subdivide all_ _•_ available training data into 5 epochs that contain roughly the same amount of additional data to simulate the continuous data collection. For each epoch we compute a pool of floor plan segmentations, where we choose the parameters stop_size and minMembers empirically to obtain segmentations with various granularities. Subsequently, we optimize a classifier on the data labeled with the areas. The parameter λ has to be chosen according to the use case requirements. We exemplarily choose the outer bounds (0 and 1), where 0 provides high performance and low expressiveness and 1 targets models with higher expressiveness. Furthermore, λ = 0.5 is chosen to select a balanced model. We demonstrate how to utilize the ACS to automatically select the optimal model for the given use case requirements. ----- _Sensors 2020, 20, 1443_ 15 of 26 Both experiments are conducted on two different data sets. The first one has been collected in our university building. The second one utilizes the publicly available benchmark dataset for indoor localization using crowdsourced data [48], which was captured in Tampere, Finland. In the following we report the results grouped by the different test sites. _6.2. Case Study: RWTH Aachen University Building_ The test environment for the data that we collected by ourselves is the 4th floor of the civil engineering building of the RWTH Aachen university, Germany. The floor contains several offices and a long hall. The total area is roughly 1500 m[2]. Two smartphones (Oneplus and LG) are used to collect labeled fingerprints with continuous position tags. In a period of 9 months (from December 2018 to August 2019), a total amount of above 1000 fingerprints have been collected. The initial performance analysis utilizes the entire training data as static data set. 6.2.1. Static Performance Analysis By applying the LDCE algorithm with two different parameterizations, we obtain two floor plan segmentations, which differ in granularity. The segmentations are shown in Figure 4, where the segments are represented by the shapes with black boundaries. The grey points represent fingerprint locations. We sum the amount of data per 2 2 m square and plot a heatmap to visualize the training data distribution. The initial DBSCAN _×_ is performed with eps = 2 and minPts = 3, which yields reasonably sized start clusters. We choose a wall penalty of 10 such that given max_eps = 30, there will be at most 2 separating walls between merging clusters. The first segmentation (Figure 4a) sets stop_size equal to 80, such that clusters are excluded from the expansion set when they reach more than 80 members. The second segmentation (Figure 4b) is obtained by setting stop_size to 50. 150 15.0 12.5 10.0 7.5 5.0 2.5 0.0 15.0 12.5 10.0 7.5 5.0 2.5 0.0 |0.98 0.97 0.92 0.98 0.98 0.98|Col2| |---|---| 0.98 0.97 0.92 0.98 0.98 0.98 0 10 20 30 40 50 60 70 80 (a) broad 0.94 0.96 0.97 0.91 0.9 0.95 0.97 0.91 0.57 1.0 0 10 20 30 40 50 60 70 80 (b) fine 100 50 20 10 5 1 150 100 50 20 10 5 1 |0.94 0.96 0.97 0.91 0.9 0.95 0.97 0.91 0.57 1.0|Col2| |---|---| 0.94 0.96 0.97 0.91 0.9 0.95 0.97 0.91 0.57 1.0 **Figure 4. Floor plan segmentations of RWTH Aachen university building. The black lined shapes** represent areas of classifier. The green numbers represent the class-wise F1-score of the best model. The grey dots are the fingerprint locations. The amount of training data per 2 × 2 m grid cell is illustrated via the heatmap color. We label the data set according to both segmentations and train the models described in Section 5.2 to predict the right area. The resulting CDF is illustrated in Figure 5. ----- _Sensors 2020, 20, 1443_ 16 of 26 SEG = broad SEG = fine 1.00 0.95 0.90 0.85 |Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|model CNN DNN SVM KNN(2) KNN(3) KNN(5) DNN(reg)| |---|---|---|---|---|---|---|---|---|---|---|---|---|---| ||||||||||||||| ||||||||||||||| 0 5 10 15 20 25 30 Distance [m] 0 5 10 15 20 25 30 Distance [m] **Figure 5. CDF of classification error. Error vector build from distances between centroids of true and** predicted areas. The CNN and the DNN achieve the best classification performance with an accuracy of above 97% on the broad segmentation and almost 95% on the finer segmentation. While the SVM achieves acceptable results for the broad segmentation its performance significantly decreases when using a finer segmentation. All regression model results are mapped to the closest class. They achieve lower performance than the CNN and DNN classifiers. A comprehensive overview of the model comparison can be found in Table 3. The lowest mean error is achieved by the DNN classifier with values of 0.43m and 0.66m respectively. For illustration purposes we plotted the class-wise F1 score of the best performing model as green numbers for each segment in Figure 4. **Table 3. Performance of classification models on both segmentations. The upper three models are** explicitly trained to predict one of the underlying areas, while the other models (reg->class) are regression models where we assign the closest area of the regression prediction during postprocessing. **Segmentation** **Model** **Parameter** **Area Center Error** **Classification** **Mean** **Std** **Min** **Max** **ACC** **F1** broad CNN 0.43 3.28 0.0 47.42 0.97 0.97 DNN 0.32 2.17 0.0 47.42 0.97 0.97 SVM 0.54 3.46 0.0 45.25 0.96 0.95 k-NN (reg- > class) k = 2 0.85 4.36 0.0 49.95 0.94 0.93 k-NN (reg- > class) k = 3 0.82 4.30 0.0 49.95 0.94 0.93 k-NN (reg- > class) k = 5 0.87 3.94 0.0 49.95 0.93 0.92 DNN (reg- > class) 0.56 2.88 0.0 25.09 0.95 0.95 fine CNN 0.66 4.18 0.0 55.12 0.95 0.91 DNN 0.54 3.74 0.0 59.90 0.95 0.91 SVM 1.12 4.80 0.0 59.90 0.88 0.79 k-NN (reg- > class) k = 2 1.15 5.47 0.0 59.90 0.91 0.84 k-NN (reg- > class) k = 3 0.99 4.94 0.0 59.90 0.91 0.87 k-NN (reg- > class) k = 5 1.00 4.54 0.0 48.34 0.91 0.86 DNN (reg- > class) 0.71 3.07 0.0 42.50 0.92 0.87 In addition, we evaluate the performance of training a standard regression model for exact position estimation. The results are presented in Figure 6. The best regression model (DNN) guarantees that in 95% of the cases, the estimated position will not differ more than 10 m. In comparison the area classification models guarantee a correct area prediction in 95% of the cases and thus achieve more ----- _Sensors 2020, 20, 1443_ 17 of 26 robust results. This is achieved by lowering the expressiveness and utilizing the knowledge about available training data. 1.0 0.9 0.8 0.7 model 0.6 DNN 0.5 KNN(2) 0.4 KNN(3) 0.3 KNN(5) 0.2 0.1 0.0 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Distance [m] DNN 4.76 4.09 0.01 49.89 k-NN(k = 2) 6.91 5.53 0.01 61.06 k-NN(k = 3) 6.50 4.24 0.31 40.71 k-NN(k = 5) 6.56 4.25 0.45 41.76 (b) **Model** **Values** **Mean** **Std** **Min** **Max** |Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10| |---|---|---|---|---|---|---|---|---|---| ||||||||||| ||||||||||| ||||||||m|odel|| ||||||||D|NN|| ||||||||K|NN(2)|| ||||||||K|NN(3)|| ||||||||K|NN(5)|| ||||||||||| ||||||||||| ||||||||||| ||||||||||| (a) **Figure 6. Performance of regression models. (a) shows the CDF of the prediction errors and (b) holds** mean, standard deviation, minimal and maximal error. 6.2.2. Model Selection via ACS In the following we present the results when applying the ACS for model selection as described in Section 6.1. Figure 7 shows the ACS score of the trained models on the pool of segmentations for the three choices of λ. The figure is interpreted by fixing a choice for λ depending on the use case. At each epoch, we can now deliver the model with the highest ACS, since it provides the best balance between expressiveness and performance. Note that for the first two epochs, the segmentations obtained from _stop_size =_ 60, 80 result in a single cluster, since too few data is available and are thus discarded. _{_ _}_ When inspecting the score for λ = 0.5, we see that at the second and third epoch, we would use the segmentation obtained by LDCE (5:20), while in epoch four the highest score is achieved on LDCE _(10:40). Finally, for the last epoch, the classifier that was optimized on LDCE (40:80) is selected._ 0.18 0.16 0.14 0.12 0.10 0.08 0.06 0.04 0.02 |Col1|SEG|Col3|Col4|Col5|Col6|Col7| |---|---|---|---|---|---|---| ||LDCE ( LDCE (|5:20) 10:40)||||| ||LDCE ( LDCE (|20:60) 40:80)||||| |||||||| |||||||| |||||||| |||||||| |||||||| 1 2 3 4 5 epoch (a) λ = 0 (b) λ = 0.5 (c) λ = 1 **Figure 7. Area classification score (ACS) for three choices of λ. Per epoch the model with the highest** score is chosen. The legend shows the (minMembers: stop_size) parameters used during segmentation. The other parameters of LDCE are chosen as presented in Table 1. The changes in ACS are discussed epoch-wise in the following. While epoch 1 contains only training data of the lower left offices, in epoch 2 additional training data along the hall has been collected. This allows for additional areas. LDCE (5:20) yields much more new segments among the hall, which causes the high increase for λ = 1. For λ = 0, those small segments do not affect the score, however, the achieved class-wise F1 score does, which is slightly lower for LDCE (10:40). Between epoch 2 and 3, only few new areas are covered, however, the lower left offices feature additional data. LDCE (20:60) and LDCE (40:80) are equal, which can also be observed from their similar ACS values. In LDCE (10:40), the lower offices have already been split in epoch 2, which yielded a bad performance. The additional data allows for improved model performance, which explains the ----- _Sensors 2020, 20, 1443_ 18 of 26 increased ACS. Between epoch 3 and 4, only data in previously uncovered areas is added. This causes an increased ACS value for all segmentations and λ values. For the broadest segmentation LDCE _(40:80), the previous areas remain the same, while the other segmentations adopt a finer granularity._ Therefore, the highest relative increase for λ = 0 is observed for LDCE (40:80). Between epoch 4 and 5, no additional areas are covered with training data. However, segmentation LDCE (40:80) rearranges its area shapes, such that the total covered area increases. While the class-wise F1 scores remain roughly the same, this causes the jump in ACS value for λ = 0. The other segmentations remain mainly unchanged, since only the F1 scores of the models slightly change. _6.3. Case Study: Tampere, Finland_ In addition to the data collected by ourselves, we evaluate our approach on a publicly available fingerprinting dataset that was generated via crowdsourcing [48]. The original dataset consists of 4648 fingerprints collected by 21 devices in a university building in Tampere, Finland. The fingerprints are distributed over five floors, while the 1st floor contains the highest sample density. Therefore, we select the data of the 1st floor as subset to conduct our experiments. 6.3.1. Static Performance Analysis Using the entire data collected on the 1st floor, we construct two floor plan segmentations based on the LDCE algorithm, which can be found in Figure 8. The initial DBSCAN is performed with _eps = 5 and minPts = 3. Note that in contrast to the other site, we slightly increase the eps parameter_ to obtain reasonably sized start clusters. This is justified because the overall training data distribution is more sparse and the map has more than 5 times the extent of the other test site. Following the same logic, we increase the max_eps parameter to 50. We use the same penalties as before but lowered the wall penalty to 5, since we want to allow clusters to span several office rooms. The remaining parameters can be found in Table 1. The broad segmentation was obtained by choosing a stop_size of 100 and for the fine segmentation we set stop_size equal to 60. The dataset is published with a predetermined train test split, which consists of 20% training data and 80% testing data. When plotting the training data of the 1st floor, we noted that only a single region contains training samples, which makes the proposed split impractical. Therefore, we apply the splitting strategy described in Section 5.3. The CDF of the class-wise error vectors is presented in Figure 9. Similar to the other dataset, the DNN classification models achieve the best results independent of the segmentation. On the broad segmentation, an accuracy of 89% is reached and in 97% of the cases the predicted centroid of the area is less than 30 m off from the centroid of the true area. A comprehensive overview of the individual model performance can be found in Table 4. The DNN achieves the lowest mean centroid error and has the lowest standard deviation. The prediction performance with respect to individual areas is illustrated in Figure 8. The green numbers represent the class-wise F1 scores that the best model achieved. For comparison with exact position estimation, we evaluate the performance of training standard regression models. The results are presented in Figure 10. While the DNN regression model achieves an error below 10 m with 90% probability, we achieve a correct area prediction in ~90% of the cases on the broad floor plan segmentation. Thus, for the goal of coarse localization the area classifiers provide higher guarantees. 6.3.2. Model Selection via ACS In the following we present the results when applying the ACS for model selection as described in Section 6.1. Figure 11 illustrates the obtained ACS scores of the trained models on the pool of segmentations for the three choices of λ. Using the ACS as selective feature, we can state the following observations. For λ = 0 (high performance), the model trained on LDCE (15:40) is chosen for the first epoch and LDCE (40:80) is selected for the second and third epoch. For the entire training data the classifier trained on LDCE (60:100) is chosen. For λ = 0.5 (balance between expressiveness and ----- _Sensors 2020, 20, 1443_ 19 of 26 performance), LDCE (15:40) provides the selected segmentation for the first four epochs and is replaced by the slightly broader segmentation LDCE (25:60) in the last epoch. The highest expressiveness is given for λ = 1, which selects the model trained on the finest segmentation LDCE (5:20) for all epochs. 80 70 60 50 20 10 40 30 20 10 0 80 5 1 |0.91 0.97 0.67 0.87 0.92 0.89 0.87 0.93|Col2| |---|---| 0.91 0.97 0.67 0.87 0.92 0.89 0.87 0.93 0 25 50 75 100 125 150 175 200 (a) broad 70 60 20 10 50 40 30 20 10 0 5 1 |0.89 0.88 0.91 0.81 0.87 0.71 0.85 0.66 0.47 0.86 0.8 0.9 0.79 0.87 0.92|Col2| |---|---| 0.89 0.88 0.91 0.81 0.87 0.71 0.85 0.66 0.47 0.86 0.8 0.9 0.79 0.87 0.92 0 25 50 75 100 125 150 175 200 (b) fine **Figure 8. Floor plan segmentations of 1st floor of public dataset [48]. The black lined shapes represent** areas of the classifier. The green numbers represent the class-wise F1-score of the best model. The grey dots are the fingerprint locations. The amount of training data per 4x4m grid cell is illustrated via the heatmap color. SEG = broad SEG = fine 1.00 0.95 0.90 0.85 0.80 0.75 0.70 |Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|model CNN DNN SVM KNN(2) KNN(3) KNN(5) DNN(reg)| |---|---|---|---|---|---|---|---|---|---| ||||||||||| ||||||||||| ||||||||||| ||||||||||| ||||||||||| 0 10 20 30 40 50 60 Distance [m] 0 10 20 30 40 50 60 Distance [m] **Figure 9. CDF of classification error. Error vector build from distances between centroids of true and** predicted areas. ----- _Sensors 2020, 20, 1443_ 20 of 26 **Table 4. Performance of classification models on both segmentations. The upper three models are** explicitly trained to predict one of the underlying areas, while the other models (reg->class) are regression models where we assign the closest area of the regression prediction during postprocessing. **Segmentation** **Model** **Parameter** **Area Center Error** **Classification** **Mean** **Std** **Min** **Max** **ACC** **F1** broad CNN 3.70 10.15 0.0 69.26 0.87 0.86 DNN 3.21 9.60 0.0 69.26 0.89 0.88 SVM 4.30 10.92 0.0 65.84 0.85 0.83 k-NN (reg- > class) k = 2 3.55 10.02 0.0 69.26 0.87 0.86 k-NN (reg- > class) k = 3 3.97 10.48 0.0 69.26 0.86 0.85 k-NN (reg- > class) k = 5 4.34 10.84 0.0 65.84 0.85 0.83 DNN (reg- > class) 4.62 11.17 0.0 65.84 0.83 0.81 fine CNN 3.65 9.11 0.0 90.47 0.83 0.81 DNN 3.53 9.00 0.0 91.75 0.84 0.81 SVM 7.00 12.36 0.0 100.44 0.71 0.56 k-NN (reg- > class) k = 2 3.72 9.12 0.0 90.47 0.82 0.79 k-NN (reg- > class) k = 3 3.95 9.30 0.0 90.47 0.81 0.77 k-NN (reg- > class) k = 5 4.25 9.55 0.0 69.79 0.80 0.76 DNN(reg) 5.00 10.07 0.0 69.79 0.76 0.72 1.0 0.9 0.8 0.7 model 0.6 DNN 0.5 KNN(2) 0.4 KNN(3) 0.3 KNN(5) 0.2 0.1 0.0 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Distance [m] DNN 5.34 4.00 0.05 45.72 k-NN(k = 2) 5.89 4.66 0.05 38.79 k-NN(k = 3) 5.79 4.76 0.15 38.30 k-NN(k = 5) 5.90 4.72 0.04 36.87 (b) **Model** **Values** **Mean** **Std** **Min** **Max** |Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9| |---|---|---|---|---|---|---|---|---| |||||||||| |||||||||| |||||||m|odel|| |||||||D|NN|| |||||||K|NN(2)|| |||||||K|NN(3)|| |||||||K|NN(5)|| |||||||||| |||||||||| |||||||||| |||||||||| (a) **Figure 10. Performance of regression models. (a) shows the CDF of the prediction errors and (b) holds** mean, standard deviation, minimal and maximal error. In the following the ACS graphs are analyzed epoch-wise. In the first epoch LDCE (25:60) and _LDCE (15:40) consist of only two broad segments, on which the models achieve the same class-wise F1_ scores. This can be observed, since both have the same scores for a fixed λ value. Since they cover a larger total area than the finer LDCE (5:20), they score higher for low and medium λ values. However, the larger number of segments of LDCE (5:20) causes the higher ACS value for λ = 1. In epoch 2 _LDCE (5:20) adds the most additional segments, while the number of added segments is the same_ for LDCE (15:40) and LDCE (25:60). This explains the scores observed for λ = 1. While for the three segmentations the number of segments increases, high class-wise F1 scores can be maintained for _LDCE (15:40) and LDCE (25:60). However, the finest segmentation LDCE (5:20) sacrifices performance_ for expressiveness and thus scores lower for λ = 0.5. For λ = 0 the score does not change much, since the total covered area remains mostly constant. However, LDCE (40:80), which is present in epoch 2 for first time, covers a much wider total area, since it only consists of few large segments and therefore scores considerably higher for λ = 0. Between epoch 2 and 3, data is collected in previously uncovered areas, which allows for finer segmentations independent of the chosen parameters. This can be observed by the significant increase in ACS for λ = {0.5, 1}. On the contrary, between epoch 3 and 4, mostly data within previously covered areas is collected, which allows for slightly higher performance. Finally, in the last epoch, the segmentations change again, while especially LDCE (60:100) computes a segmentation that covers a much larger total extent than the other segmentations. This explains the high increase in ACS value for λ = 0. ----- _Sensors 2020, 20, 1443_ 21 of 26 0.18 0.16 0.14 0.12 0.10 0.08 0.06 0.04 0.02 |Col1|SEG|Col3|LDCE (|25:60)|Col6|Col7| |---|---|---|---|---|---|---| ||LDCE ( LDCE (|5:20) 15:40)|LDCE ( LDCE (|40:80) 60:100)||| |||||||| |||||||| |||||||| |||||||| |||||||| |||||||| 1 2 3 4 5 epoch (a) λ = 0 (b) λ = 0.5 (c) λ = 1 **Figure 11. Area classification score (ACS) for three choices of λ. Per epoch the model with the highest** score is chosen. The legend shows the (minMembers: stop_size) parameters used during segmentation. The other parameters of LDCE are chosen as presented in Table 1. **7. Discussion** In the following the findings of our work are discussed. The results of the case study are analyzed with emphasis on the proposed concepts. Subsequently, the benefits of adaptive area localization are highlighted in comparison to existing solutions. And finally, potential applications of the proposed concept are described. _7.1. Case Study Results_ _Model performance:_ Independent of the test site, the DNN area classifiers outperformed all other models with respect to standard classification metrics, such as accuracy and F1 score. The F1 metric indicates that the model provides high precision and recall scores, which means that each individual area is detected properly and in case it is selected the prediction is trustable. CNN models are especially useful to learn tasks where inputs are locally connected, such as adjacent pixels in images [49]. When randomly arranging the access point vector as a matrix, it cannot be claimed that a comparable relation between adjacent matrix entries exists. Therefore, the additional feature extraction should not provide any benefits, which is empirically demonstrated by the results. The SVM model can only be used as multi-class classifier by training several individual classifiers and following a certain voting scheme. We applied the one-vs-one strategy, which results in K(K 1)/2 classifiers if we want to detect K areas. _−_ Besides, the high computational effort, the results are worse than a simple k-NN classifier, which is also observed in [50]. _LDCE floor plan segmentation algorithm:_ During the second experiment, it was demonstrated that the proposed LDCE algorithm is capable of providing a pool of segmentations with various granularities. Those can be utilized in combination with the proposed ACS to select the best area classifier with respect to the right balance between expressiveness and performance. The algorithm requires certain parameters to be chosen empirically based on the given site. _ACS model selection metric:_ The effect of λ on the ACS was theoretically evaluated and demonstrated for three values in the experiments. However, explicit values cannot be associated with qualitative terms, yet. In particular, it cannot be stated which exact value is optimal for a certain use case. However, the ACS is lazily computed. Once an area classifier has been trained, its ACS can be computed for several choices of _λ by utilizing the stored prediction and ground truth vectors. This means that it is computationally_ inexpensive to compute the ACS for a pool of trained models and a large set of λ values. An initial ----- _Sensors 2020, 20, 1443_ 22 of 26 _λ value is guessed. When the retrieved model does not meet the requirements, λ can be adjusted to_ match the right balance between expressiveness and performance. _7.2. Adaptive Area Localization_ Area localization has been proposed for large-scale deployments of fingerprinting-based solutions or when the data quality does not allow for exact position estimation. The objective is to provide higher positioning guarantees by lowering the expressiveness of the model. In related work, the segmentation during area classification features two characteristics [17–20]: It is determined independent of the available training data. _•_ It is statically determined, mostly prior to data collection. _•_ Both features are unfavorable when working with crowdsourced data that is continuously collected and solutions to apply area localization in such settings are currently missing in the literature. Crowdsourced data collection results in a spatially non-uniform data distribution [15]. Training a classifier on data where certain areas (classes) feature only few or no samples results in poor performance. A segmentation that is determined independent of the training data might result in such sparsely covered areas. Therefore, we introduce the concept of data-aware floor plan segmentation and propose the LDCE algorithm that computes such a segmentation. A data-aware floor plan segmentation introduces a trade-off between expressiveness and performance, which has not been quantified in the literature, yet. However, such a quantification is required to measure how well an area classifier performs given that the underlying segmentation is not static. Therefore, we propose the ACS that captures this trade-off. Furthermore, during crowdsourcing, data is accumulated over time. The segmentation determined for a given snapshot of data might become unfavorable once additional data has been collected. It is crucial to regularly recompute the segmentation into areas. In summary, our proposed concepts enable area localization for crowdsourced data and we empirically demonstrate that this achieves higher reliability than exact position estimation. The model adapts to the accumulating training data and finds the right balance between expressiveness and performance. _7.3. Potential Applications_ Depending on the use case, localization systems might have distinct requirements. A system with the objective to provide proximity based services (e.g., inside a shopping mall [17]) requires a coarse-grained position estimation with high guarantees. In contrast, a localization system utilized for navigation of people with visual impairments might benefit from a more fine-grained position estimation. Given a base of crowdsourced training data, our approach allows to automatically construct area localization models for any required tradeoff between expressiveness and performance. Furthermore, it adapts to the accumulating training data that results from continuous crowdsourced data collection. To the best of our knowledge, generating such adaptive localization models based on fingerprinting has not been proposed in the literature, yet. In addition, absolute location information can be merged with systems that iteratively determine the position of a user such as PDR. WLAN fingerprinting is already employed in sensor fusion solutions [51–53]. The granularity and level of guarantee of the fingerprinting model might impact initialization and convergence time of the fused model. With our approach, the fingerprinting-based localization model with the optimal granularity in that regards can be trained and deployed in the fused model. **8. Conclusions** In this work, we propose the concept of adaptive area localization to achieve reliable position estimations using crowdsourced data that is accumulated over time. Existing area localization solutions employ a static segmentation into areas that is independent of the available training data. This approach is not applicable for crowdsoucred data collection, since it features an unbalanced ----- _Sensors 2020, 20, 1443_ 23 of 26 spatial training data distribution that changes over time. To solve this, we propose the LDCE algorithm that computes data-aware floor plan segmentations with various granularities. The underlying segmentation influences the model performance as well as its expressiveness. We introduce the ACS to select the area classifier that provides the best trade-off between them. With those concepts, we can now regularly compute a pool of segmentations and train classifiers on the data labeled with the corresponding areas. We select the best model with the ACS and deploy it for localization. The proposed concepts are validated on a self-collected as well as on a publicly available crowdsourced data set. We demonstrate that the proposed area classifiers provide higher positioning guarantees than models for exact position estimation. Furthermore, we show that they adapt to the accumulating data base. In future work, we want to utilize PDR techniques and sensor fusion to automate the data collection process and to enhance the positioning quality during localization. In addition, our approach is not limited to WLAN RSS fingerprinting, but can be extended to support magnetic and light sensors [54,55] or bluetooth [56], which we want to demonstrate in future work. **Author Contributions: M.L., J.B. and R.K. designed the methodology; M.L. conceived and conducted the** experiments; J.B. and R.K. administrated and supervised the research project; M.L. wrote the paper, J.B. and R.K. reviewed the text and offered valuable suggestions for improving the manuscript. **Funding: This research received no external funding.** **Conflicts of Interest: The authors declare no conflict of interest.** **References** 1. Zafari, F.; Gkelias, A.; Leung, K.K. A Survey of Indoor Localization Systems and Technologies. Commun. Surv. _[Tutorials IEEE 2019, 21, 2568–2599. [CrossRef]](http://dx.doi.org/10.1109/COMST.2019.2911558)_ 2. Basiri, A.; Lohan, E.S.; Moore, T.; Winstanley, A.; Peltola, P.; Hill, C.; Amirian, P.; Figueiredo e Silva, P. Indoor location based services challenges, requirements and usability of current solutions. Comput. Sci. Rev. 2017, _[24, 1–12. [CrossRef]](http://dx.doi.org/10.1016/j.cosrev.2017.03.002)_ 3. Wang, Y.; Shao, L. Understanding occupancy pattern and improving building energy efficiency through [Wi-Fi based indoor positioning. Build. Environ. 2017, 114, 106–117. [CrossRef]](http://dx.doi.org/10.1016/j.buildenv.2016.12.015) 4. D’Aloia, M.; Cortone, F.; Cice, G.; Russo, R.; Rizzi, M.; Longo, A. Improving energy efficiency in building system using a novel people localization system. In Proceedings of the 2016 IEEE Workshop on Environmental, Energy, and Structural Monitoring Systems (EESMS), Bari, Italy, 13–14 June 2016; pp. 1–6. [[CrossRef]](http://dx.doi.org/10.1109/EESMS.2016.7504811) 5. Ahmetovic, D.; Murata, M.; Gleason, C.; Brady, E.; Takagi, H.; Kitani, K.; Asakawa, C. Achieving Practical and Accurate Indoor Navigation for People with Visual Impairments. In Proceedings of the 14th Web for _All Conference on The Future of Accessible Work—W4A ’17; ACM Press: New York, NY, USA, 2017; pp. 1–10._ [[CrossRef]](http://dx.doi.org/10.1145/3058555.3058560) 6. Ho, T.W.; Tsai, C.J.; Hsu, C.C.; Chang, Y.T.; Lai, F. Indoor navigation and physician-patient communication in emergency department. In Proceedings of the 3rd International Conference on Communication and Information _Processing—ICCIP ’17; Ben-Othman, J., Gang, F., Liu, J.S., Arai, M., Eds.; ACM Press: New York, NY, USA,_ [2017; pp. 92–98. [CrossRef]](http://dx.doi.org/10.1145/3162957.3162971) 7. Kárník, J.; Streit, J. Summary of available indoor location techniques. IFAC-PapersOnLine 2016, 49, 311–317. [[CrossRef]](http://dx.doi.org/10.1016/j.ifacol.2016.12.055) 8. He, S.; Chan, S.H.G. Wi-Fi Fingerprint-Based Indoor Positioning: Recent Advances and Comparisons. _[Commun. Surv. Tutorials IEEE 2016, 18, 466–490. [CrossRef]](http://dx.doi.org/10.1109/COMST.2015.2464084)_ 9. Xia, S.; Liu, Y.; Yuan, G.; Zhu, M.; Wang, Z. Indoor Fingerprint Positioning Based on Wi-Fi: An Overview. _[ISPRS Int. J. Geo-Inf. 2017, 6, 135. [CrossRef]](http://dx.doi.org/10.3390/ijgi6050135)_ 10. Batistic, L.; Tomic, M. Overview of indoor positioning system technologies. In Proceedings of the 2018 41st International Convention on Information and Communication Technology, Electronics and Microelectronics [(MIPRO), Opatija, Croatia, 21–25 May 2018; pp. 473–478.[CrossRef]](http://dx.doi.org/10.23919/MIPRO.2018.8400090) ----- _Sensors 2020, 20, 1443_ 24 of 26 11. Yassin, A.; Nasser, Y.; Awad, M.; Al-Dubai, A.; Liu, R.; Yuen, C.; Raulefs, R.; Aboutanios, E. Recent Advances in Indoor Localization: A Survey on Theoretical Approaches and Applications. Commun. Surv. Tutorials IEEE **[2017, 19, 1327–1346. [CrossRef]](http://dx.doi.org/10.1109/COMST.2016.2632427)** 12. Rai, A.; Chintalapudi, K.K.; Padmanabhan, V.N.; Sen, R. Zee: Zero-Effort Crowdsourcing for Indoor Localization. In Proceedings of the 18th annual international conference on Mobile computing and networking (Mobicom ’12), Istanbul, Turkey, 22–26 August 2012; pp. 293—304. 13. Radu, V.; Marina, M.K. HiMLoc: Indoor smartphone localization via activity aware Pedestrian Dead Reckoning with selective crowdsourced WiFi fingerprinting. In Proceedings of the International Conference on Indoor Positioning and Indoor Navigation, Montbeliard-Belfort, France, 28–31 October 2013; pp. 1–10. [[CrossRef]](http://dx.doi.org/10.1109/IPIN.2013.6817916) 14. Santos, R.; Barandas, M.; Leonardo, R.; Gamboa, H. Fingerprints and Floor Plans Construction for Indoor [Localisation Based on Crowdsourcing. Sensors 2019, 19, 919. [CrossRef]](http://dx.doi.org/10.3390/s19040919) 15. Ye, Y.; Wang, B. RMapCS: Radio Map Construction From Crowdsourced Samples for Indoor Localization. _[IEEE Access 2018, 6, 24224–24238. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2018.2830415)_ 16. He, S.; Tan, J.; Chan, S.H.G. Towards area classification for large-scale fingerprint-based system. In Proceedings _of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing - UbiComp ’16; Lukowicz,_ P.; Krüger, A.; Bulling, A.; Lim, Y.K.; Patel, S.N., Eds.; ACM Press: New York, NY, USA, 2016; pp. 232–243. [[CrossRef]](http://dx.doi.org/10.1145/2971648.2971689) 17. Lopez-Pastor, J.A.; Ruiz-Ruiz, A.J.; Martinez-Sala, A.S.; Luis Gomez-Tornero, J. Evaluation of an indoor positioning system for added-value services in a mall. In Proceedings of the 2019 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Pisa, Italy, 30 September 2019–3 October 2019; pp. 1–8. [[CrossRef]](http://dx.doi.org/10.1109/IPIN.2019.8911822) 18. Wei, J.; Zhou, X.; Zhao, F.; Luo, H.; Ye, L. Zero-cost and map-free shop-level localization algorithm based on crowdsourcing fingerprints. In Proceedings of the 2018 Ubiquitous Positioning, Indoor Navigation and [Location-Based Services (UPINLBS), Wuhan, China, 22–23 March 2018; pp. 1–10.[CrossRef]](http://dx.doi.org/10.1109/UPINLBS.2018.8559708) 19. Rezgui, Y.; Pei, L.; Chen, X.; Wen, F.; Han, C. An Efficient Normalized Rank Based SVM for Room Level [Indoor WiFi Localization with Diverse Devices. Mob. Inf. Syst. 2017, 2017, 1–19. [CrossRef]](http://dx.doi.org/10.1155/2017/6268797) 20. Liu, H.X.; Chen, B.A.; Tseng, P.H.; Feng, K.T.; Wang, T.S. Map-Aware Indoor Area Estimation with Shortest Path Based on RSS Fingerprinting. In Proceedings of the 2015 IEEE 81st Vehicular Technology Conference [(VTC Spring), Glasgow, UK, 11–14 May 2015; pp. 1–5.[CrossRef]](http://dx.doi.org/10.1109/VTCSpring.2015.7145926) 21. Torres-Solis, J.; Falk., T.; Chau, T. A Review of Indoor Localization Technologies: Towards Navigational Assistance for Topographical Disorientation. In Ambient Intelligence; Villanueva Molina, F.J., Ed.; IntechOpen: [Rijeka, Croatia, 2010. [CrossRef]](http://dx.doi.org/10.5772/8678) 22. Pecoraro, G.; Di Domenico, S.; Cianca, E.; de Sanctis, M. LTE signal fingerprinting localization based on CSI. In Proceedings of the 2017 IEEE 13th International Conference on Wireless and Mobile Computing, [Networking and Communications (WiMob), Rome, Italy, 9–11 October 2017; pp. 1–8.[CrossRef]](http://dx.doi.org/10.1109/WiMOB.2017.8115803) 23. Xiao, L.; Behboodi, A.; Mathar, R. A deep learning approach to fingerprinting indoor localization solutions. In Proceedings of the 2017 27th International Telecommunication Networks and Applications Conference [(ITNAC), Melbourne, VIC, Australia, 22–24 November 2017; pp. 1–7.[CrossRef]](http://dx.doi.org/10.1109/ATNAC.2017.8215428) 24. Sinha, R.S.; Lee, S.M.; Rim, M.; Hwang, S.H. Data Augmentation Schemes for Deep Learning in an Indoor [Positioning Application. Electronics 2019, 8, 554. [CrossRef]](http://dx.doi.org/10.3390/electronics8050554) 25. Yang, S.; Dessai, P.; Verma, M.; Gerla, M. FreeLoc: Calibration-free crowdsourced indoor localization. In Proceedings of the 2013 Proceedings IEEE INFOCOM, Turin, Italy, 14–19 April 2013; pp. 2481–2489. [[CrossRef]](http://dx.doi.org/10.1109/INFCOM.2013.6567054) 26. Kim, W.; Yang, S.; Gerla, M.; Lee, E.K. Crowdsource Based Indoor Localization by Uncalibrated [Heterogeneous Wi-Fi Devices. Mob. Inf. Syst. 2016, 2016, 1–18. [CrossRef]](http://dx.doi.org/10.1155/2016/4916563) 27. Mittal, A.; Tiku, S.; Pasricha, S. Adapting Convolutional Neural Networks for Indoor Localization with Smart Mobile Devices. In Proceedings of the 2018 on Great Lakes Symposium on VLSI—GLSVLSI ’18; Chen, D.; [Homayoun, H.; Taskin, B., Eds.; ACM Press: New York, NY, USA, 2018; pp. 117–122. [CrossRef]](http://dx.doi.org/10.1145/3194554.3194594) 28. Adege, A.; Lin, H.P.; Tarekegn, G.; Jeng, S.S. Applying Deep Neural Network (DNN) for Robust Indoor [Localization in Multi-Building Environment. Appl. Sci. 2018, 8, 1062. [CrossRef]](http://dx.doi.org/10.3390/app8071062) ----- _Sensors 2020, 20, 1443_ 25 of 26 29. Wang, X.; Gao, L.; Mao, S.; Pandey, S. DeepFi: Deep learning for indoor fingerprinting using channel state information. In Proceedings of the 2015 IEEE Wireless Communications and Networking Conference [(WCNC), New Orleans, LA, USA, 9–12 March 2015; pp. 1666–1671. [CrossRef]](http://dx.doi.org/10.1109/WCNC.2015.7127718) 30. Wang, X.; Gao, L.; Mao, S.; Pandey, S. CSI-based Fingerprinting for Indoor Localization: A Deep Learning [Approach. IEEE Trans. Veh. Technol. 2016, 66, 763–776. [CrossRef]](http://dx.doi.org/10.1109/TVT.2016.2545523) 31. Chen, H.; Zhang, Y.; Li, W.; Tao, X.; Zhang, P. ConFi: Convolutional Neural Networks Based Indoor Wi-Fi [Localization Using Channel State Information. IEEE Access 2017, 5, 18066–18074. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2017.2749516) 32. [Yang, Z.; Zhou, Z.; Liu, Y. From RSSI to CSI. ACM Comput. Surv. (CSUR) 2013, 46, 1–32. [CrossRef]](http://dx.doi.org/10.1145/2543581.2543592) 33. He, S.; Chan, S.H.G. Towards Crowdsourced Signal Map Construction via Implicit Interaction of IoT Devices. In Proceedings of the 2017 14th Annual IEEE International Conference on Sensing, Communication, and [Networking (SECON), San Diego, CA, USA, 12–14 June 2017; pp. 1–9.[CrossRef]](http://dx.doi.org/10.1109/SAHCN.2017.7964901) 34. Zhou, B.; Li, Q.; Mao, Q.; Tu, W. A Robust Crowdsourcing-Based Indoor Localization System. Sensors **[2017, 17, 864. [CrossRef]](http://dx.doi.org/10.3390/s17040864)** 35. Jiang, Q.; Ma, Y.; Liu, K.; Dou, Z. A Probabilistic Radio Map Construction Scheme for Crowdsourcing-Based [Fingerprinting Localization. IEEE Sensors J. 2016, 16, 3764–3774. [CrossRef]](http://dx.doi.org/10.1109/JSEN.2016.2535250) 36. Jung, S.h.; Moon, B.c.; Han, D. Unsupervised Learning for Crowdsourced Indoor Localization in Wireless [Networks. IEEE Trans. Mob. Comput. 2016, 15, 2892–2906. [CrossRef]](http://dx.doi.org/10.1109/TMC.2015.2506585) 37. Jung, S.h.; Han, D. Automated Construction and Maintenance of Wi-Fi Radio Maps for Crowdsourcing-Based [Indoor Positioning Systems. IEEE Access 2018, 6, 1764–1777. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2017.2780243) 38. Pipelidis, G.; Tsiamitros, N.; Ustaoglu, E.; Kienzler, R.; Nurmi, P.; Flores, H.; Prehofer, C. Cross-Device Radio Map Generation via Crowdsourcing. In Proceedings of the 2019 International Conference on Indoor [Positioning and Indoor Navigation (IPIN), Pisa, Italy, 30 September 2019–3 October 2019; pp. 1–8.[CrossRef]](http://dx.doi.org/10.1109/IPIN.2019.8911766) 39. Chow, K.H.; He, S.; Tan, J.; Chan, S.H.G. Efficient Locality Classification for Indoor Fingerprint-Based [Systems. IEEE Trans. Mob. Comput. 2019, 18, 290–304. [CrossRef]](http://dx.doi.org/10.1109/TMC.2018.2839112) 40. Nowicki, M.; Wietrzykowski, J. Low-Effort Place Recognition with WiFi Fingerprints Using Deep Learning. In Automation 2017, Advances in Intelligent Systems and Computing, Warsaw, Poland, 15–17 March 2017; Szewczyk, R.; Zieli´nski, C.; Kaliczy´nska, M., Eds.; Springer International Publishing: Cham, Switzerland, [2017; Volume 550, pp. 575–584.[CrossRef]](http://dx.doi.org/10.1007/978-3-319-54042-9_57) 41. Kim, K.S.; Lee, S.; Huang, K. A scalable deep neural network architecture for multi-building and multi-floor [indoor localization based on Wi-Fi fingerprinting. Big Data Anal. 2018, 3, 466. [CrossRef]](http://dx.doi.org/10.1186/s41044-018-0031-2) 42. Ibrahim, M.; Torki, M.; ElNainay, M. CNN based Indoor Localization using RSS Time-Series. In Proceedings of the 2018 IEEE Symposium on Computers and Communications (ISCC), Natal, Brazil, 25–28 June 2018; [pp. 01044–01049. [CrossRef]](http://dx.doi.org/10.1109/ISCC.2018.8538530) 43. Song, C.; Wang, J. WLAN Fingerprint Indoor Positioning Strategy Based on Implicit Crowdsourcing and [Semi-Supervised Learning. ISPRS Int. J. Geo-Inf. 2017, 6, 356. [CrossRef]](http://dx.doi.org/10.3390/ijgi6110356) 44. Bishop, C.M. Pattern Recognition and Machine Learning (Information Science and Statistics); Springer: Secaucus, NJ, USA, 2006. 45. Dong, K.; Ling, Z.; Xia, X.; Ye, H.; Wu, W.; Yang, M. Dealing with Insufficient Location Fingerprints in Wi-Fi [Based Indoor Location Fingerprinting. Wirel. Commun. Mob. Comput. 2017, 2017, 1–11. [CrossRef]](http://dx.doi.org/10.1155/2017/1268515) 46. Han, J.; Kamber, M.; Pei, J. Data Mining: Concepts and Techniques, online-ausg ed.; Morgan Kaufmann Series in Data Management Systems; Elsevier Science: Burlington, Vermont, 2011. 47. Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, USA, 2016. 48. Lohan, E.; Torres-Sospedra, J.; Leppäkoski, H.; Richter, P.; Peng, Z.; Huerta, J. Wi-Fi Crowdsourced [Fingerprinting Dataset for Indoor Positioning. Data 2017, 2, 32. [CrossRef]](http://dx.doi.org/10.3390/data2040032) 49. Rawat, W.; Wang, Z. Deep Convolutional Neural Networks for Image Classification: A Comprehensive [Review. Neural Comput. 2017, 29, 2352–2449. [CrossRef]](http://dx.doi.org/10.1162/NECO_a_00990) 50. Rana, S.P.; Prieto, J.; Dey, M.; Dudley, S.; Corchado, J.M. A Self Regulating and Crowdsourced Indoor [Positioning System through Wi-Fi Fingerprinting for Multi Storey Building. Sensors 2018, 18. [CrossRef]](http://dx.doi.org/10.3390/s18113766) 51. Chang, Q.; van de Velde, S.; Wang, W.; Li, Q.; Hou, H.; Heidi, S. Wi-Fi Fingerprint Positioning Updated by Pedestrian Dead Reckoning for Mobile Phone Indoor Localization. In China Satellite Navigation Conference _(CSNC) 2015 Proceedings: Volume III; Lecture Notes in Electrical Engineering; Sun, J., Liu, J., Fan, S., Lu, X.,_ [Eds.; Springer: Berlin/Heidelberg, Germany, 2015; Volume 342, pp. 729–739. [CrossRef]](http://dx.doi.org/10.1007/978-3-662-46632-2_63) ----- _Sensors 2020, 20, 1443_ 26 of 26 52. Zou, H.; Chen, Z.; Jiang, H.; Xie, L.; Spanos, C. Accurate indoor localization and tracking using mobile phone inertial sensors, WiFi and iBeacon. In Proceedings of the 2017 IEEE International Symposium on [Inertial Sensors and Systems (INERTIAL), Kauai, HI, USA, 27–30 March 2017; pp. 1–4.[CrossRef]](http://dx.doi.org/10.1109/ISISS.2017.7935650) 53. Jin, F.; Liu, K.; Zhang, H.; Feng, L.; Chen, C.; Wu, W. Towards Scalable Indoor Localization with Particle Filter and Wi-Fi Fingerprint. In Proceedings of the 2018 15th Annual IEEE International Conference on Sensing, [Communication, and Networking (SECON), Hong Kong, China, 11–13 June 2018; pp. 1–2.[CrossRef]](http://dx.doi.org/10.1109/SAHCN.2018.8397155) 54. Wang, X.; Yu, Z.; Mao, S. DeepML: Deep LSTM for Indoor Localization with Smartphone Magnetic and Light Sensors. In Proceedings of the 2018 IEEE International Conference on Communications (ICC), Kansas [City, MO, USA, 20–24 May 2018; pp. 1–6.[CrossRef]](http://dx.doi.org/10.1109/ICC.2018.8422562) 55. Zhang, W.; Sengupta, R.; Fodero, J.; Li, X. DeepPositioning: Intelligent Fusion of Pervasive Magnetic Field and WiFi Fingerprinting for Smartphone Indoor Localization via Deep Learning. In Proceedings of the 2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA), Cancun, Mexico, [18–21 December 2017; pp. 7–13.[CrossRef]](http://dx.doi.org/10.1109/ICMLA.2017.0-185) 56. Kanaris, L.; Kokkinis, A.; Liotta, A.; Stavrou, S. Fusing Bluetooth Beacon Data with Wi-Fi Radiomaps for [Improved Indoor Localization. Sensors 2017, 17, 812. [CrossRef]](http://dx.doi.org/10.3390/s17040812) _⃝c_ 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution [(CC BY) license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/.) -----
{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC7085741, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/1424-8220/20/5/1443/pdf?version=1584016953" }
2,020
[ "JournalArticle" ]
true
2020-03-01T00:00:00
[ { "paperId": "52c35300b9d1ed10cca3ea4725b7a9b426c5f2a4", "title": "Evaluation of an indoor positioning system for added-value services in a mall" }, { "paperId": "ecbbe52b36d1153e3c3cf378b0f14f6d82cee7ae", "title": "Cross-Device Radio Map Generation via Crowdsourcing" }, { "paperId": "3adcf3fa58b9cba53e6b3f14253fb9e7c76e352a", "title": "Data Augmentation Schemes for Deep Learning in an Indoor Positioning Application" }, { "paperId": "61596edb4e350da4a122f97bea766c305a2da387", "title": "Efficient Locality Classification for Indoor Fingerprint-Based Systems" }, { "paperId": "df2abfb0a6aaff84b58fe7de5bb694c4a0b50a3e", "title": "Fingerprints and Floor Plans Construction for Indoor Localisation Based on Crowdsourcing" }, { "paperId": "245fcdeed60543b52bf899147411859fbdb4dea0", "title": "A Self Regulating and Crowdsourced Indoor Positioning System through Wi-Fi Fingerprinting for Multi Storey Building" }, { "paperId": "75201f3ea629e09de1161bdad94a936b104e4a69", "title": "Applying Deep Neural Network (DNN) for Robust Indoor Localization in Multi-Building Environment" }, { "paperId": "5735804566ad31c37bb2cedd0f37a5e4aad6afa7", "title": "Towards Scalable Indoor Localization with Particle Filter and Wi-Fi Fingerprint" }, { "paperId": "1ea85d68947ffa02df67aa4c49d5fab6e4acc7db", "title": "CNN based Indoor Localization using RSS Time-Series" }, { "paperId": "79a4f79b89f31ad330f6e84f7f71510070122dd3", "title": "Adapting Convolutional Neural Networks for Indoor Localization with Smart Mobile Devices" }, { "paperId": "787d64c967bce7cb3ae909560d6f0f2f06fee731", "title": "DeepML: Deep LSTM for Indoor Localization with Smartphone Magnetic and Light Sensors" }, { "paperId": "049866640af88866accb7ae765ff54cb16347376", "title": "Zero-cost and map-free shop-level localization algorithm based on crowdsourcing fingerprints" }, { "paperId": "db4217c5c683b910fc02b68c2eb45b77403cabd7", "title": "A scalable deep neural network architecture for multi-building and multi-floor indoor localization based on Wi-Fi fingerprinting" }, { "paperId": "e5980ed8e21e334aed085a4a13b105b409fbc730", "title": "DeepPositioning: Intelligent Fusion of Pervasive Magnetic Field and WiFi Fingerprinting for Smartphone Indoor Localization via Deep Learning" }, { "paperId": "13199a8aa1fcadb99c005a876c971df95d91d7f5", "title": "Indoor navigation and physician-patient communication in emergency department" }, { "paperId": "03604284bcf4d1a87d25f1046ac8e1b86d92f9e2", "title": "WLAN Fingerprint Indoor Positioning Strategy Based on Implicit Crowdsourcing and Semi-Supervised Learning" }, { "paperId": "389d4d0d3e927b349642cc84dc4d00ee117b004a", "title": "A deep learning approach to fingerprinting indoor localization solutions" }, { "paperId": "c87ef7e41abd68293a7bb3a8c1e40405e3b93ba5", "title": "Wi-Fi Crowdsourced Fingerprinting Dataset for Indoor Positioning" }, { "paperId": "cd3033c432ac52a6c86ff1518b7d75f5f82b1249", "title": "LTE signal fingerprinting localization based on CSI" }, { "paperId": "5c17fc24c16d0bb9e4ba5b48e1c9fe88e39650d6", "title": "ConFi: Convolutional Neural Networks Based Indoor Wi-Fi Localization Using Channel State Information" }, { "paperId": "ea509ed799e7e45cd320d2d626a75ea163f226ec", "title": "A Survey of Indoor Localization Systems and Technologies" }, { "paperId": "c1f05b723e53ac4eb1133249b445c0011d42ca79", "title": "Deep Convolutional Neural Networks for Image Classification: A Comprehensive Review" }, { "paperId": "99b7e4a0a042d963f539d7af3578dc7eff03e045", "title": "An Efficient Normalized Rank Based SVM for Room Level Indoor WiFi Localization with Diverse Devices" }, { "paperId": "1152d59ee28eb11b137d5a39a3f4ecd427a69ec9", "title": "Towards Crowdsourced Signal Map Construction via Implicit Interaction of IoT Devices" }, { "paperId": "5576bc8173144aa8f1ee9e5314ee395d0820bc17", "title": "Indoor location based services challenges, requirements and usability of current solutions" }, { "paperId": "0394644b52f404d52d7f85164466dfcff5ee19db", "title": "Indoor Fingerprint Positioning Based on Wi-Fi: An Overview" }, { "paperId": "2c565cc70ad63a6325bb11e771c87b3a706cc3f5", "title": "Achieving Practical and Accurate Indoor Navigation for People with Visual Impairments" }, { "paperId": "53ef0c9a0605c8ef819e3f52669ca1018692c034", "title": "A Robust Crowdsourcing-Based Indoor Localization System" }, { "paperId": "acdf54252dac8fa27c48452e8442a971f0bff8a9", "title": "Fusing Bluetooth Beacon Data with Wi-Fi Radiomaps for Improved Indoor Localization" }, { "paperId": "32e9e97bb53e38b3a225f6ca5bd1df34db3f696c", "title": "Accurate indoor localization and tracking using mobile phone inertial sensors, WiFi and iBeacon" }, { "paperId": "d45987b100a328d554b052c709af897ea86c5a5c", "title": "Understanding occupancy pattern and improving building energy efficiency through Wi-Fi based indoor positioning" }, { "paperId": "3b98a19c1d44e3e7e43dd7fc58691927fa14d428", "title": "Recent Advances in Indoor Localization: A Survey on Theoretical Approaches and Applications" }, { "paperId": "0b4cf28ca8229bb36aa2fa73eeb7ed9c346ab3ae", "title": "Low-effort place recognition with WiFi fingerprints using deep learning" }, { "paperId": "2b50736bd76872ef562dd83de59bf2614f926177", "title": "Unsupervised Learning for Crowdsourced Indoor Localization in Wireless Networks" }, { "paperId": "1e93c836f0dbe4ef71bfa8b4488ab5e675fc79ac", "title": "Towards area classification for large-scale fingerprint-based system" }, { "paperId": "1c7a2b4a06abf90d2580a5fde510574daae1fb7d", "title": "Crowdsource Based Indoor Localization by Uncalibrated Heterogeneous Wi-Fi Devices" }, { "paperId": "ad142de141e224b6e5c4d731d212c6d7e0bdb31f", "title": "Improving energy efficiency in building system using a novel people localization system" }, { "paperId": "409ab99c91795436622ac063706d2fdf8f628c8b", "title": "A Probabilistic Radio Map Construction Scheme for Crowdsourcing-Based Fingerprinting Localization" }, { "paperId": "4016403a62b587a675198415b93b903e2a66264a", "title": "CSI-Based Fingerprinting for Indoor Localization: A Deep Learning Approach" }, { "paperId": "ca6261d8ab678d9f973fce5aaa5d29f9c25c6bd2", "title": "Wi-Fi Fingerprint-Based Indoor Positioning: Recent Advances and Comparisons" }, { "paperId": "42eee752cf03ca8e0ff176f47fef14b28c486be2", "title": "Map-Aware Indoor Area Estimation with Shortest Path Based on RSS Fingerprinting" }, { "paperId": "9713463f509ee274b95db430397a5bcd1cfe7c5b", "title": "DeepFi: Deep learning for indoor fingerprinting using channel state information" }, { "paperId": "7e5a6abb22ea859375a00b1ac40e51c548d51148", "title": "From RSSI to CSI" }, { "paperId": "e124ae6566af94e6ebbda6b6c3ad26b453179f3e", "title": "HiMLoc: Indoor smartphone localization via activity aware Pedestrian Dead Reckoning with selective crowdsourced WiFi fingerprinting" }, { "paperId": "b4121cae8d476fef80222d5ffa7718197a79264a", "title": "FreeLoc: Calibration-free crowdsourced indoor localization" }, { "paperId": "5b2c81e10064fc1cb148532e42d7b8db6096f186", "title": "Zee: zero-effort crowdsourcing for indoor localization" }, { "paperId": "cc7a8d51584e7ceb68962daca77837a309692876", "title": "A Review of Indoor Localization Technologies: towards Navigational Assistance for Topographical Disorientation" }, { "paperId": "e8a706b253be9692dacb4fc39c4b30342cd81c06", "title": "RMapCS: Radio Map Construction From Crowdsourced Samples for Indoor Localization" }, { "paperId": "9f584dbebbef64eeebd9d6542abaf1c23ce84b02", "title": "Overview of indoor positioning system technologies" }, { "paperId": "4b57637b58a125acf964397b83e57ff622835163", "title": "Automated Construction and Maintenance of Wi-Fi Radio Maps for Crowdsourcing-Based Indoor Positioning Systems" }, { "paperId": "4a35dc81e88d320194482313af7c30431a21e600", "title": "Dealing with Insufficient Location Fingerprints in Wi-Fi Based Indoor Location Fingerprinting" }, { "paperId": "6f6358bb586a9a253d12c59aeaf1277da7cabc01", "title": "Summary of available indoor location techniques" }, { "paperId": "8e3c872076750bcce868808f9d4d7a038f950040", "title": "Pattern Recognition And Machine Learning" }, { "paperId": "2913c2bf3f92b5ae369400a42b2d27cc5bc05ecb", "title": "Deep Learning" }, { "paperId": "c48200c7c0a2239cb5839da44a2a8dc0230a4bb1", "title": "Wi-Fi Fingerprint Positioning Updated by Pedestrian Dead Reckoning for Mobile Phone Indoor Localization" }, { "paperId": null, "title": "Data Mining: Concepts and Techniques, online-ausg ed.; Morgan Kaufmann Series in Data Management Systems" } ]
26,308
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00ed8311df8cd0e843ca8912bc76e6d365443859
[ "Computer Science" ]
0.853701
A comparison of TCP automatic tuning techniques for distributed computing
00ed8311df8cd0e843ca8912bc76e6d365443859
Proceedings 11th IEEE International Symposium on High Performance Distributed Computing
[ { "authorId": "2242631", "name": "E. Weigle" }, { "authorId": "145476815", "name": "Wu-chun Feng" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
Approved for public release; distribution is unlimited. _Title:_ A Comparison of TCP Automatic Tuning Techniques for ###### Distributed Computing _Author(s):_ Eric Weigle and Wu-Chun Feng _Submitted to:_ 1 1 th IEEE International Symposium on High-Performance ###### Distributed Computing Los Alamos NATIONAL LABORATORY Los Alamos National Laboratory, an affirmative actiodequal opportunity employer, is operated by the University of California for the U.S. Department of Energy under contract W-7405-ENG-36. By acceptance of this article, the publisher recognizes that the U.S. Government retains a nonexclusive, royalty-free license to publish or reproduce the published form of this contribution, orto allow others to do so, for U.S. Government purposes. Los Alamos National Laboratory requests that the publisher identify this article as work performed under the auspices of the U.S. Department of Energy. Los Alamos National Laboratory strongly supports academic freedom and a researcher's right to publish;asan institution,however,the Laboratorydoesnotendorsetheviewpointof a publicationor guaranteeitstechnicalcorrectness. ----- ###### A Comparison of TCP Automatic-Tuning Techniques for Distributed Computing Eric Weigle and Wu-chun Feng Research and Development in Advanced Network Technology Computer and Computational Sciences Division Los Alamos National Laboratory Los Alamos, NM 87545 ``` {ehw, feng}@lanl.gov Abstract In this paper, we compare and contrast these tuning ``` methods. First we explain each method, followed by an in- _Rather than paiizful, manual, static, per-connection op-_ depth discussion of their features. Next we discuss the ex- _timization of TCP buffer sizes siniply to achieve acceptable_ periments to fully characterize two particularly interesting _perfonnance for distributed applications 18, IO], inany re-_ methods (Linux 2.4 autotuning and Dynamic Right-Sizing). _searclzers liave proposed techniques to perforin this tuning_ We conclude with results and possible improvements. _automatically [4, 7,9,11,12,14]. This paperjrst discusses_ _tlie relative merits of the various approaches in tlaeory, and_ ###### 1.1. TCP Tuning and Distributed Computing _then provides substantial experimental data concerning two_ _coinpetiizg iniplerneiitations - the buffer autotuning already_ _present in Linux_ `2.4.x and “Dynainic Right-Sizing.”` `Tlzis` Computational grids such as the Information Power _paper reveals lieretofore unknown aspects_ `of the problem` Grid _[5], Particle Physics Data Grid_ **[l], and Earth System** _and current solutions, provides insiglit into tlie proper ap-_ Grid [3] all depend on TCP. This implies several things. _proach for various circumstances, and points toward ways_ First, bandwidth is often the bottleneck. Performance for _to further irnprove perfornzance._ distributed codes is crippled by using TCP over a WAN. An appropriately selected buffer tuning technique is one solu- **Keywords:** dynamic right-sizing, autotuning, high- tion to this problem. performance networking, TCP, flow control, wide-area net- work. Second, bandwidth and time are money. An OC-3 at 155Mbps can cost upwards of $50,000 a month and higher speeds cost even more. If an application can only utilize a ###### 1. Introduction few megabits per second, that money is being wasted. Time spent by people waiting for data, time spent hand-tuning TCP, for good or ill, is the only protocol widely avail- network parameters, time with under-utilized compute re- able for reliable end-to-end congestion-controlled network sources - also wasted money. Automatically tuned TCP communication, and thus it is the one used for almost all buffers more effectively utilize network resources and save distributed computing. that money, but an application designer must still choose from the many approaches. Unfortunately, TCP was not designed with high- performance computing in mind - its original design deci- Third, tuning is a pain. Ideally, network and protocol sions focused on long-term fairness first, with performance designers produce work so complete that those doing dis- a distant second. Thus users must often perform tortuous tributed or grid computing are not unduly pestered with the manual optimizations simply to achieve acceptable behav- “grungy” details. In the real world, application develop- ior. The most important and often most difficult task is de- ers must still make decisions in order to attain peak per- termining and setting appropriate buffer sizes. Because of formance. The results in this paper show the importance this, at least six ways of automatically setting these sizes of paying attention to the network and show one way to have been proposed. achieve maximal performance with minimal effort. ----- ###### 2. Buffer I’uning Techniques connections can increase their window size - performance improvements are an intentional side-effect. TCP buffer-tuning techniques balance memory demand Enable uses a daemon to perform the same tasks as a with the reality of limited resources - maximal TCP buffer human performing manual tuning. It gathers information space is useless if applications have no memory. Each tech- about every pair of hosts between which connections are nique discussed below uses different information and makes to be tuned and saves it in a database. Hosts then look up different trade-offs. All techniques are most useful for large this information when opening a connection and use it to data transfers (at least several times the bandwidth x _delay_ set their buffer sizes. Enable [ 113 reports performance im- product of the network). Short, small transmissions are provements over untuned connections by a factor of 10-20 dominated by latency, and window size is practically irrele- and above 2.4 autotuning by a factor of 2-3. vant. Auto-ncFTP also mimics the same sequence of events as a human manually tuning a connection. Here, it is per- formed once just before starting a data connection in FTP ###### 2.1. Current Tuning Techniques so the client can set buffer sizes appropriately. DRS FTP uses a new command added to the FTP control 1. Manual tuning [8,10] language to gain network information, which is used to tune 2. PSC’s Automatic TCP Buffer Tuning [9] buffers during the life of a connection. Tests of this method show performance improvements over stock FTP by a factor 3. Dynamic Right-Sizing (DRS) [4,14] of 6 with lOOms delay, with optimally tuned buffers giving an improvement by a factor of 8. 4. Linux 2.4 Auto-tuning [12] ###### 2.2. Comparison of Tuning Techniques 5. Enable tuning [ 111 6. NLANR’s Auto-tuned FI’P (in ncFTP) [7] **I** **Thing 0** I , **Level** I I **Changes -** 1 I **Band** I I **Visibilitv** 1 I 7. L A W S DRS FTP’(in WUFTP) I **PSC** I Kernel I Dvnamic I In I Transparent 1 **Linux2.4** I Kernel I Dynamic I In I Transparent Manual tuning is the baseline by which we measure au- totuning methods. To perform manual tuning, a human uses **DRS** I I Kernel I Dvnamic I In I Transparent **Enable** 1 User Static Out Visible tools such as ping and pathchar or pipechar to deter- **NLANRFTP** I User Static Out . ODaaue mine network latency and bandwidth. The results are mul- tiplied to get the bandwidth x _delay product, and buffers_ are generally set to twice that value. PSC’s tuning is a mostly sender-based approach. Here **Table 1. Comparison of** **Tuning Techniques** the sender uses TCP packet header information and times- tamps tb estimate the bandwidth x delay product of the net- work, which it uses to resize its send window. The receiver **User-level versus Kernel-level refers to whether the** simply advertises the maximal possible window. PSC’s pa- buffer tuning is accomplished as an application-level solu- per [9] presents results for a NetBSD 1.2 implementation, tion or as a change to the kernel (Linux, “BSD, etc.). showing improvement over stock by factors of 10-20 for Manual tuning tediously requires both types of changes. small numbers of connections. **An** ‘ideal’ solution would require only one type of change DRS is a mostly receiver-based buffer tuning approach - kernel-level for situations where many TCP-based pro- where the receiver tries to estimate the bandwidth x delay grams require high performance, user-level where only a product of the network and the congestion-control state of single TCP-based program (such as FTP) requires high per- the sender, again using TCP packet header information and formance. timestamps. The receiver then advertises a window large Kernel-level implementations will always be more effi- enough that the sender is not flow-window limited. cient, as more network and high-resolution timing informa- Linux autotuning refers to a memory management tion is available, but they are complicated and non-portable. technique used in the stable Linux kernel, version 2.4. Whether this is worth the 20- 100% performance improve- This technique does not attempt any estimates of the ment is open to debate. _bandwidth_ _x_ _delay_ product of a connection. Instead, it **Static versus Dynamic refers to whether the buffer tun-** simply increases and decreases buffer sizes depending on ing is set to a constant at the start of a connection, or if it available system memory and available socket buffer space. can change with network “weather” during the lifetime of a By increasing buffer sizes when they are full of data, TCP connection. 2 ----- Generally a dynamic solution is preferable - it adapts **3.1. Varied Experimental Parameters** itself to changes in network state, which some work has shown to have multi-fractal congestion characteristics [6, 131. Static buffer sizes are always too large or small Our experiments consider the following parameters: given “live” networks. Yet, static connections often have **Tuning** (None, 2.4-auto, DRS): We compare a Linux smoother application-level performance than dynamic con- 2.2.20 kernel which has no autotuning, a 2.4.17 kernel nections, which is desirable. which has Linux autotuning, and a 2.4.17 kernel which also Unfortunately, both static and dynamic solutions have has Dynamic Right-Sizing. We will refer to these three as problems. Dynamic changes in buffer sizes imply changes 2.2.20-None, 2.4.17-Auto, and 2.4.17-DRS. in the advertised window, which if improperly implemented **Buffer Sizes (32KB to 32MB): Initial buffer size con-** can break TCP semantics (data legally sent for a given win- figuration is required even for autotuning implementations. dow is in-ff ight when the window is reduced, thus causing There are three cases: the data to be dropped at the end host). Current dynamic tuning methods monotonically increase window sizes to avoid this - possibly wasting memory. No user or kernel tuning; buffer sizes at defaults. Gives **In-Band versus Out-of-Band** refers to whether baseline for comparison with tuned results. _bmdwzdth_ _x_ _delay_ information is obtained from the connection itself or is gathered separately from the data transmission to be tuned. An ideal solution would be Kernel-only tuning; configure maximal buffer sizes in-band to minimize user inconvenience and ensure the only. Gives results for kernel autotuning implemen- correct time-dependent and path-dependent information is tations. being gathered. DRS FTP is ‘both’ because data is gathered over the con- trol channel; usually this channel uses the same path as the User and kernel tuning; use `setsockopt ( ) to con-` figure buffer sizes manually. Gives results for manu- data channel, but in some ‘third-party’ cases the two chan- ally tuned connections. nels are between different hosts entirely. `In the first case` data collection is ‘in-band’, while in the second not only is it out of band, it measures characteristics of the wrong con- **Network Delay** (zOSms, 25ms, 50ms, 100ms): We nection! Auto-ncFTP suffers from the same ‘third-party’ vary the delay from 0.5ms to lOOms to show the perfor- problem. mance differences between LAN and WAN environments. **Transparent versus Visible refers to user inconvenience** We use TICKET [15] to perform WAN emulation. This ###### - how easily can a user tell if they are using a tuning emulator can route at line rate (up to lGbps in our case) in- method, how many changes are required, etc. An ideal so- troducing a delay between 200 microseconds and 200 mil- lution would be transparent after the initial install and con- liseconds. figuration required by all techniques. The kernel approaches are transparent; other than im- **Parallel Streams** (1, 2, 4, 8): We use up to 8 paral- proved performance they are essentially invisible to aver- lel streams to test the effectiveness of this commonly-used age users. The FTP programs are ‘opaque’ because they technique with autotuning techniques. This also shows how can generate detectable out-of-band data, and require some well a given tuning technique scales with increasing num- start-up time to effectively tune buffer sizes. Enable is com- bers of flows. When measuring performance, we time from pletely visible. It requires a daemon and database sepa- the start of the first process to the finish of the last process. rate from any network program to be tuned, generates fre- quent detectable network benchmarking traffic, and requires ###### 3.2. Constant Experimental Parameters changes to each network program that wishes to utilize its functionality. **Topology: Figure 1 shows the generic topology we use** ###### 3. Experiments in our tests. We have some number of network source (S) processes sending data to another set of destination (D) pro- These experiments shift our focus to the methods of di- cesses through a pair of bottleneck routers (R) connected rect interest: manual tuning, Linux 2.4 autotuning, and via some WAN cloud. The “WAN cloud” may be a direct Dynamic Right-Sizing under Linux. The remaining ap- long-haul connection or through some arbitrarily complex proaches are not discussed further because such analysis is network. (In the simplest case, both routers and the “WAN available in the referenced papers. cloud” could be a single very high-bandwidth LAN switch.) ----- ###### n **Hardware:** Tests are run between two machines with ## . dual 933MHz Pentium I11 processors, an Alteon Tigon I1 Gigabit Ethernet card on a 64-bit 66-MHz PCI bus, and _5 12MB of memory._ ###### D5 ``` 4 4. Results and Analysis ``` **Figure 1. Generic Topology** Our experiments place all processes (parallel streams) on a single host. The results of more complicated one- We present data in order of increasing delay. With con- to-many or many-to-one experiments (common in scatter- stant bandwidth (Gigabit Ethernet), this will show how well gather computation, or for web servers) can be inferred by each approach scales as pipes get “fatter.” observing memory and CPU utilization on the hosts. This information shows the scalability of the sender and receiver tuning and whether one end’s behavior characterizes the performance of the connection. This distinction is critical **4.1.** **First Case, zO.5rns Delay** for one-to-many relationships, as the “one” machine must split its resources among many flows while each of the “many” machines can dedicate more resources to the one flow. With delays on the order of half a millisecond, we ex- **Unidirectional Transfers: Although TCP is inherently** pect that even very high bandwidth links can be saturated a full duplex protocol, the majority of traffic generally flows with small windows - the default 64KB buffers should be in one direction. TCP protocol dynamics do not signifi- sufficient. cantly differ between one flow with bidirectional traffic and two unidirectional flows sending in opposite directions [ 161. Figure 2 shows the performance using neither user nor ###### Loss: Our WAN emulator is configured to emulate no kernel tuning. With a completely default configuration, the loss (although loss may still occur due to senderheceiver Linux 2.4 stack with autotuning outperforms the Linux 2.2 buffer over-runs). All experiments are intended to be the stack without autotuning by lOOMbps or more (as well as best-case scenario. The artificial inclusion of loss adds showing more stable behavior). Similarly, 2.4.17-DRS out- nothing to the discussion, as congestion control for TCP performs 2.4.17-Auto by a smaller margin of 30-50Mbps. Reno/SACK under Linux is a constant for all experiments. This is due to more appropriate use of TCP’s advertised **Data Transfer: Rather than using some of the available** window field and faster growth to the best buffer size possi- benchmarking programs we chose to write a simple TCP ble. based program to mimic message-passing traffic. This pro- gram tries to send large (1MB) messages between hosts as Unexpectedly for such a low-delay case, all kernels ben- fast as possible. A total of 128 messages are sent, a number efit from the use of parallel streams, with improvements chosen because: in performance from 5570%. When a single data flow is striped among multiple TCP streams, it effectively obtains 128MB transfers are large enough to allow the conges- a super-exponential slow-start phase and additive increase tion window to fully open. by N (the number of TCP streams). In this case, that behav- ior improves performance. 128MB transfers are small enough to occur commonly in practice ’. Note that limitations in the firmware of our Gigabit Eth- ernet NICs limit performance to 800Mbps or below, so we Longer transfers do not help differentiate among tun- simply consider 800Mbps ideal. ing techniques (tested, but results omitted). It is evenly divisible among all numbers of parallel streams. *Custom firmware solutions can improve throughput, but such results ‘“In the long run we are all dead.” -John Maynard Keynes are neither portable nor relevant to this study. 4 ----- **I000** **2.2.20-none .... --*-----** **2.4.17-Auto** `----e----` **2.4.17-Auto** `----8----` **800** - **2.4.17-DRS** **2.4.17-DRS** #### - - **U** .- _[e ]D_ ###### s ..* .............................................................. **_4_** `1 -` =........ _... ............................................................. **B** _9_ ................... t -a 3 **400 1 ..............** B `m` m **2oo I** **2oo t** O L I 0 ‘ I **1** **2** **3** **4** **5** **6** **7** **8** **1** **2** **3** **4** **5** **6** **7** **8** **Number of Parallcl Processes** **Number of Parallel Processes** **Figure 2.** **No Tuning, 0.5ms** **Figure 3. Kernel-Only Tuning, 0.5ms** Figure 4 shows the results for hand-tuned connections. DRS obeys the user when buffers are set by `setsock-` ###### opt ( 1 , so 2.4.17-Auto and 2.4.17-DRS use the same buffer sizes and perform essentially the same. The perfor- mance difference between 2.2.20 and 2.4.17 is due to stack improvements in Linux 2.4. Figure 3 shows the performance with kernel tuning only; that is, increasing the maximum amount of memory that the kernel can allocate to a connection. 2.2.20-none -----*.---- **2.4.17-Aut0** `----e----` **2.4.17-DRS** #### - As expected, results for 2.2.20-None (which does no au- totuning) mirror the results from the prior test. ....... .............................. - .............................................................. ....... Goo t 2.4.17-Auto connections perform 30-50Mbps better than in the untuned case, showing that the default 64KB buffers **400 i** were insufficient. **2oo t** 0 ‘ J **1** **2** **3** **4** **5** **6** **7** **8** 2.4.17-DRS connections also perform better with one or **Number of Parallel Processes** two processes, but as the number of processes increases, DRS actually performs worse! DRS is more aggressive in allocating buffer space; with such low delay, it overallocates **Figure 4. User/Kernel Tuning with Ideal Sizes,** memory, and performance suffers (see Figure 5’s discussion **0.5ms** below). The “ideal” buffer sizes in the prior graph (Figure 4) are Furthermore, performance is measured at the termina- larger than one might expect; Figure _5_ shows the perfor- tion of the entire transfer (when the final parallel process mance of 2.4.17-Auto with buffer sizes per process between completes). Large numbers of parallel streams can lead to 8KB and 64MB. We achieve peak performance with sizes the starvation of one or more processes due to TCP conges- on the order of 1MB - much larger than the calculated ideal tion control, so the parallelized transfer suffers. Yet this can of 64KB, the bandwidthx delay of the network. The differ- be a good thing - parallel flows can induce chaotic network ence is due to the interaction and feedback between several behavior and be unfair in some cases; by pcnalizing users factors, the most important of which are TCP congestion of heavily parallel flows, DRS could induce more network control and process scheduling. fairness while still providing good performance. To keep the pipe full, one must buffer enough data to ----- avoid transmission “bubbles.” However, with multi-fractal burstiness caused by TCP congestion control [6,13], occa- **2.4.17-A~t0 .---*---** sionally the network is so overloaded that very large buffers **2.4.17-DRS** ###### t - are needed to accommodate it. Also, these buffers them- **_F_** **500** - .- selves can increase the effective delay (and therefore in- `3` creasing the buffering required) in a feedback loop only ter- **400** - minated by a lull later in the traffic stream. This buffering $ **300** - can occur either in the network routers or in the end hosts. `a` ###### 2 200 - Because of process scheduling, it is incorrect to divide the predicted “ideal” buffer size (the bandwidth x _delay)_ by the number of processes to determine the buffer size per process when using parallel streams. Because only one pro- **1** **2** **3** **4** **5** **6** **7** **8** cess can run on a given CPU at a given time, the kernel must **Number of Parallel Processes** buffer packets for the remaining processes until they can get a time slice. Thus, as the number of processes grows, the ef- fective delay experienced by those processes increases, and **Figure 6. No Tuning, 25ms** the amount of required buffering also grows. Beyond a cer- tain point, this feedback is great enough that the addition of Figure 7 shows results with Kernel-Only Tuning. The additional parallelism is actually detrimental. This is what performance of DRS improves dramatically while the per- we see with DRS in Figure 5. formance of simple autotuning and untuned connections is constant. As we increase the number of processes we again see the performance of DRS fall. 900 **1** . ’ I **1 process** This graph actually reveals a bug in the Linux 2.4 ker- **850** **2 process** **-----x--.-- -** nel series that our DRS patch fixes; the window scaling ad- vertised in SYN packets is based on initial (default) buffer size, not the maximal buffer size up to which Linux can tune. Thus with untuned default buffers, no window scaling is advertised - so even if the kernel is allowed to allocate multi-megabyte buffers, the size of those buffers cannot be represented in TCP packet headers. `IO000` 100000 I e 4 6 **1 e 4 7** **Buffer size** **_F_** **500** .- **a** ###### 3 400 **Figure** **5.** **Effect of Buffer Size on Perfor-** ``` 5 ``` **mance, 0.5ms** **_3_** **300** **a** **200** ##### : ’ ###### ...................................................... 4.2. Second Case, ~25rns Delay 100 ...................... ............ ......................... ....................... ~ ................. ............................. .................. I **1** **2** **3** **4** **5** **6** **7** **8** This case increases delay to values more in line with **Number of Parallel Processes** a network of moderate size, giving a bandwidth x _delay_ product of over 3MB. In this case, the default configura- tion is insufficient for high performance, giving less than **Figure 7. Kernel-Only Tuning, 25ms** 20Mbps for a single process with all kernels (Figure 6). As the number of processes increases, our effective flow win- With both user and kernel tuning, maximal performance dow increases, and we achieve a linear speed-up. In this increases for all kernels. However, as shown in Figure 8, case, simple autotuning outperforms DRS, as the memory- performance does fall for DRS in the two and four pro- management technique is more effective with small win- cess case - here we see that second-guessing the kernel can dows (it was designed for heavily loaded web servers). cause problems, and larger buffer sizes are not always de- ###### 6 ----- 2.2.20-none - - - . - . performance improvement. 2.4.17-Auto ----*---- 200 - 2.4.17-DRS --t 700 600 ###### g 500 ``` .3 ``` 400 .c 50 ~~ _c_ 4 .......... ###### 2 300 ....... ............ ........................ **U** ......................... ......................... ....................................... ..--- ....................................... ........... m 5 200 **1** 2 3 4 5 6 7 8 Number of Parallel Processes loo t **1** 2 3 4 5 6 7 8 **Figure 9. Kernel-Only Tuning, 1 OOms** Number of Parallel Processes Similar to Figure 8, the hand-tuned case in Figure 10 **Figure 8. User/Kernel Tuning with Ideal Sizes,** shows 2.4.17-DRS and 2.4.17-Auto performing identically **25ms** with 2.2.20-None performing slightly worse. Interestingly, at this high delay, the performance difference between DRS and autotuning is insignificant - the factors dominating per- ###### 4.3. Third and Fourth Cases, 50-1OOnis Delay formance are not buffer sizes but rather standard TCP slow- start, additive increase, and multiplicative decrease behav- The patterns observed in results for the 50ms and looms iors, and the 128MB transfer size is insufficient to differen- cases do not significantly differ (other than adjustments in tiate the flows. With latencies this high, very large (multi- scale) from those in the 25ms case - the factors dominat- gigabyte, minimum) transfer sizes would be required to ing behavior are the same. That is, at low delays (below more fully utilize the network. It would also help to use 20ms), one can find very interesting behavior as TCP inter- a modified version of TCP such as Vegas [2] or one of the acts with the operating system, NIC, and so forth. At higher plethora of other versions, because a multiplicative decrease delays (25ms and above), the time scales are large enough can take a ridiculous amount of time to recover on high- that TCP slow-start, additive increase, and multiplicative delay links. decrease behaviors are most important; interactions with the operating system and so forth become insignificant. **5. Guidelines on Selecting an Auto-Tuned TCP** In fact, the completely untuned cases differ so little that the following three equations (generated experimentally) , This section gives a few practical guidelines for a suffice to calculate the bandwidth in Mbps with error uni- prospective of an automatically tuned TCP. formly below 20%, given only the number of processes and the delay in milliseconds. 1. You have kernel-modification privileges to the ma- chine. So, you may use a kernel-level solution which `0 2.2.20-None: (processes x` 214)ldelay will generally provide the best performance. Currently, only NetJ3SD and Linux implementations exist, so for **_0_** 2.4.17-Auto: `(processes x` 467)ldelay other operating systems, you must either wait or use a `0 2.4.17-DRS: (processes x 355)ldelay` user-level solution. **_0_** If you want to use NetBSD, you must use PSC's As in Figure 7, the kernel-only tuning case in Figure 9 tuning. shows 2.4.17-DRS significantly outperforming 2.4.17-Auto (by a factor of 5 to 15). DRS at 50ms delay with 8 processes **_0_** Linux 2.4 autotuning is appropriate for achieves 310Mbps (graph omitted), and at lOOms with 8 large numbers of small connections, such ----- ###### 6. Conclusion **2.2.20-none** --=- - - **2.4.17-Auto** `----e----` **200** - **2.4.17-DRS -** We have presented a detailed discussion on the various techniques for automatic TCP buffer tuning, showing the benefits and problcms with each approach. We have pre- sented experimental evidence showing the superiority of Dynamic Right-Sizing over simple autotuning as found in Linux 2.4. We have also uncovered several unexpected as- pects of the problem (such as the calculated “ideal” buffers performing more poorly than somewhat larger buffers). Fi- **50 t** nally, the discussion has provided insight into which solu- ``` 0 ``` **1** **2** **3** **4** **5** **6** **7** **8** tions are appropriate for which circumstances, and why. **Number of Parallel Processes** ###### References **Figure** **10.** **User/Kernel Tuning with** **Ideal** ANL, CalTech, LBL, SLAC, JF, U. Wisconsin, BNL, **Sizes, 100ms** FNL, and SDSC. Thc Particle Physics Data Grid. h ttp://www.cacr.cal tech.cdu/ppdg/. L. Brakmo and L. Peterson. TCP Vegas: End to End Con- as web/media servers, or machines where users gestion Avoidance on a Global Internet. **_IEEE Journal_** **_on_** **_Selected Areas_** **_iri_** **_Cornnirrnication, 13(8): 1465-1480,_** Octo- are willing to tune parallel streams. ber **1995.** **_0_** DRS is appropriate for smaller numbers of large W. Feng, I. Fostcr, S. Haminond, B. Hibbard, C. Kesselman, connections, such as FTP or bulk data transfers, A. Shoshani, B. Tierney, and D. Williams. Prototyping an ###### or machines where users are not willing to tune Earth System Grid. http://\vww.scd.ucar.edulcss/esg/. parallel streams. M. Fisk and W. Feng. Dynamic Adjustment of TCP Window Sizes. Technical Report Los Alamos Unclas- 2. You do not have kernel-modification privileges to the sified Report (LA-UR) **00-3221,** Los Alamos National machine or are unwilling to make changes, forcing a Laboratory, July **2000.** See http://www.lanl.gov/radi- user-level solution. All user-level solutions perform `ant/website/pubs/hptcp/tcpwindow.pdf.` comparably, so the choice between them is based on W. E. Johnston, D. Gannon, and B. Nitzberg. Grids as Production Computing Environments: The Engineering As- features. pects of NASA‘s Information Power Grid. In **_Proceedirigs_** **_0_** If all you need is FTP, LANL‘s DRS FTP or of **_8th IEEE lnteniational S)wiposium on Higlt-Per/oniiaiice_** NLANR’s Auto-tuned FTP will be the easiest **_Distributed_** **_Coriipufirzg, August 1999._** W. Leland, M. Taqqu, W. Willinger, ~ and D. Wilson. On plug-in solutions. Obviously, we are biased in the Self-similar Nature of Ethemet Traffic (Extended Ver- favor of LANL’s implementation, which dynam- sion). **_IEEWACM Trailsactioils on_** **_Networking,_** 2( 1): 1-1 **5,** ically adjusts the window, over NLANR’s imple- February 1994. mentation, which does not. G. Navlakha and J. Ferguson. Automatic **_0_** If you require multiple applications, then the En- TCP Window Tuning and Applications. able [ll] service may fit your needs. This will, `http://dast.nlanr.net/Projects/Autobuf/autotcp.html. April` however, require source-code level changes to **2001.** Pittsburgh Supercomputing Center. Enabling each program you wish to use. High-Performance Data Transfers on Hosts. In all cases, initial tuning should be performed to `http://www.psc.edu/networking/perf-tune.htm1.` J. Semke, J. Mahdavi, , and M. Mathis. Automatic TCP Ensure TCP window scaling, timestamps, SACK op- Buffer Tuning. ACM SIGCOMM 1998,28(4), October 1998. tions are enabled. B. Tiemey. **TCP** Tuning Guide for Distributed Applica- tions on Wide-Area Networks. In **_USENIX & SAGE Login,_** Set the maximum memory available to allocate per http://www-didc.lbl .gov/tcp-wan.htm1, February 2001. connection or for user-level tuning. B. L. Tierney, D. Gunter, J. Lee, and M. Stoufer. Enabling Network-Aware Applications. In Proceedings of **_IEEE Iizter-_** Set ranges for Linux 2.4 autotuning. **_national Syniposirun on Nigh Perfoniiaiice Distrrtbted Com-_** **_puting, August 2001._** (Optional) Flush caches in between runs so inappropri- L. Torvalds and The Free Software Community. The Linux ately set slow-start thresholds are not re-used. Kernel, September 1991. http://www.kemel.org/. 8 ----- [13] **A** Veres and M. Boda. The Chaotic Nature of TCP Conges- tion Control. In Proceedings of IEEE Iizfocom 2000, March 2000. [14] E. Weigle and W. Feng. Dynamic Right-Sizing: **A** Simu- lation Study. In **_Proceedings of_** _IEEE_ **_Inlernational Con-_** **_ference_** **_on Computer Coninzunicutioris and Networks, 2001._** http://public.lanl.gov/ehw/papers/ICCCN-2001 -DRS.ps. [I51 E. Weigle and W. Feng. TICKETing High-speed Traffic with Commodity Hardware and Software. In **_Proceedings_** **_of the Third Annual Passive and Active Meusurenzent Work-_** _shop (PAM2002), March 2002._ [16] L. Zhang, S. Shenker, and D. D. Clark. Observations on the Dynamics of a Congestion Control Algorithm: The Ef- fects of Two-way Traffic. In Proceedings of ACM SigCornm **_1991, September 1991._** -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/HPDC.2002.1029926?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/HPDC.2002.1029926, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "other-oa", "status": "GREEN", "url": "https://digital.library.unt.edu/ark:/67531/metadc927872/m2/1/high_res_d/976162.pdf" }
2,002
[ "JournalArticle" ]
true
2002-07-24T00:00:00
[]
8,926
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00ee11fd044876642ee1440432a33d7faa6d3292
[ "Computer Science" ]
0.862692
SCI networking for shared-memory computing in UPC: blueprints of the GASNet SCI conduit
00ee11fd044876642ee1440432a33d7faa6d3292
29th Annual IEEE International Conference on Local Computer Networks
[ { "authorId": "3094819", "name": "H. Su" }, { "authorId": "36900480", "name": "B. Gordon" }, { "authorId": "1770398", "name": "S. Oral" }, { "authorId": "48081786", "name": "A. George" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
# SCI Networking for Shared-Memory Computing in UPC: Blueprints of the GASNet SCI Conduit ## H. Su, B. Gordon, S. Oral, A. George {su, gordon, oral, george}@hcs.ufl.edu _High-performance Computing and Simulation (HCS) Research Lab, Dept. of Electrical and Computer Engineering,_ _University of Florida, Gainesville, Florida 32611-6200_ ### Abstract learning curve for people with C experience to begin _Unified Parallel C (UPC) is a programming model for_ creating parallel programs and often results in tighter and _shared-memory parallel computing on shared- and_ more efficient code. _distributed-memory systems. The Berkeley UPC_ One recent development in UPC is the interest in _software, which operates on top of their Global_ providing a means for executing UPC over _Addressing Space Networking (GASNet) communication_ commercial-off-the-shelf (COTS) clusters. The _system, is a portable, high-performance implementation_ Berkeley UPC runtime system [3], developed by U.C. _of UPC for large-scale clusters. The Scalable Coherent_ Berkeley and LBNL, is a promising tool now available to _Interface (SCI), a torus-based system-area network (SAN),_ support this endeavor. An underlying key to this system _is known for its ability to provide very low latency_ is the Global Addressing Space Networking (GASNet) _transfers as well as its direct support for both_ communication system [4-5]. GASNet defines a _shared-memory and message-passing communications._ standard application interface that can be implemented _High-speed clusters constructed around SCI promise to be_ over a wide variety of standard and high-performance _a potent platform for large-scale UPC applications._ networks such as Ethernet, InfiniBand, Myrinet, and _This paper introduces the design of the Core API for the_ Quadrics. _new SCI conduit for GASNet and UPC, which is based on_ In this study, we present the design of a new GASNet _Active Messages (AM). Latency and bandwidth data_ conduit operating over the Scalable Coherent Interface _were collected and are compared with raw SCI results_ (SCI) network [6]. Benchmarks were executed on the _and with other existing GASNet conduits. The outcome_ newly developed conduit and compared against the raw _shows that the new GASNet SCI conduit is able to provide_ performance of SCI, the GASNet Myrinet conduit, and _promising performance in support of UPC applications._ GASNet MPI conduit on SCI using Scali’s ScaMPI [7] to evaluate various strengths and weaknesses. ### Keywords: Scalable Coherent Interface, Global Address The next section of this paper briefly describes the _Space Networking, Unified Parallel C, Active Messages._ architecture of SCI and GASNet. In Section 3, we discuss related research. Section 4 describes the design overview ### 1 Introduction of the GASNet/SCI conduit. Section 5 presents the performance results and analyses. Finally, Section 6 Many scientific as well as commercial endeavors rely presents conclusions and directions for future research. on the ability to solve complex problems in a quick and efficient manner. One of the dominant solutions to this **2 Background** problem has been the advent of parallel computing. To supplement the architectural improvements in this area, In the following subsections we present an overview parallel programming models have emerged to provide of the SCI high-performance network. Also included is a programmers alternate ways in solving complex and brief introduction to the GASNet communication system. computationally intensive problems. Such models ### 2.1 SCI include message passing, shared memory, and distributed shared memory. SCI is an ANSI/ISO/IEEE standard (1596-1992) that While message passing and shared memory are the describes a packet-based protocol [8] for system-area two most popular ways to implement parallel programs, networking. SCI was initially developed as an attempt to distributed shared memory is quickly gaining momentum. address the problems associated with buses for use with One of the reasons for this development is the growing many processors. It has evolved to become a acceptance of Unified Parallel C (UPC) [1-2] and other high-performance interconnect for SANs and embedded models like it. UPC is a parallel extension to the ANSI systems. SCI uses point-to-point links, maintaining low C standard that gives programmers the ability to create latency while achieving high data rates between nodes. It parallel programs that can target a variety of parallel features a shared-memory mentality so that memory on architecture platforms while maintaining a familiar each node can be addressable by every other node on the C-style structure. This approach allows a smaller network. SCI uses 64 bits in its addressing. The ----- most-significant 16 bits are used to specify the node in the network, and the remaining 48 bits are used for addresses within each node. With this scheme, the SCI environment can support up to 64K nodes with 256TB of addressable space. SCI offers many advantages for the unique nature of parallel computing demands. Perhaps the most significant of these advantages is its low-latency performance, typically (based on current commercial products from Dolphin) on the order of single-digit microseconds for remote-write operations and tens of microseconds for remote-read operations. Based on the latest technology, SCI offers a link data rate of 5.3 Gb/s with topologies including 1D (ring), 2D, or 3D torus. The Dolphin SISCI API [9] is a standard set of API calls allowing users to access and control SCI hardware behavior directly based on a shared-memory paradigm. To enable inter-node communication, the receiver must set aside a portion of its physical memory (global memory region) for use by the SCI network. The sender then imports the memory region into its virtual address space and is thus able to read and write the receiver’s memory region by way of either PIO (shared-memory operation) or DMA (zero-copy operation) transfer modes. The SCI hardware automatically converts accesses in SCI-mapped virtual address space to network transfers. ### 2.2 GASNet Global Addressing Space Networking (GASNet), developed at UCB/LBNL, is a language-independent, low-level communications layer that provides network-independent, high-performance communication primitives aimed at supporting parallel shared-memory programming languages such as UPC and Titanium, a parallel dialect of Java. The system is divided into two layers, the GASNet Core API and the GASNet Extended API (Figure 1). The Core API is a narrow interface based on Active Messages (AM) [10] and the network-specific Firehose memory registration algorithm [11]. The Extended API is a network-independent interface that provides medium- and high-level operations on remote memory and collective operations. The GASNet segment is the location where most of the GASNet operations target. There are three ways the segment can be configured, as _fast,_ _large, or_ _everything._ Under the fast configuration, the size of the segment may be limited to provide faster transfers of GASNet operations. The _large configuration makes a large_ portion of memory available to the segment. The size may include all of the physical memory or more. The _everything configuration makes the whole virtual address_ space on every node available for GASNet operations. Currently, GASNet supports execution on UDP, MPI, Myrinet, Quadrics, InfiniBand and IBM LAPI. GASNet was first released on 1/29/2003 with the latest release as of this writing being Version 1.3. ### 3 Related research Since our GASNet Core API must provide for AM over SCI, Ibel’s paper [12] is useful as it discusses several possible ways to execute AM over SCI. However, his simple remote-queue implementation poses several limitations. First, with 1 buffer space for all AM replies, each node is restricted to having only 1 outstanding AM request to the whole network at any given time. Furthermore, the need for the receiver to copy bulk data (long AM payload) from the 4KB buffer to its appropriate memory location, and the cost of message polling (O(N), where _N_ denotes the system size), introduce additional overhead that significantly impacts system performance. With applications that exhibit frequent inter-node communication, system performance is degraded to a degree that the benefit of parallelization is no longer observed. Ibel briefly described a split remote-queue scheme that uses circular queues of _N−k (where_ _k is a constant)_ messages (one queue for each node able to hold _k_ messages) to allow parallel sending and receiving of messages. Unfortunately, this approach is not deadlock-free and the overhead for copying bulk data and message polling still remains. Additional research that was instrumental to this work consists of other existing GASNet conduits. The design documents and source code available on the GASNet website [4] were used as a guide in the design of the Core API for the new SCI conduit. ### 4 Core API design SCI hardware is designed such that remote writes are ~10 times faster than remote reads. This disparity is due to the inability to streamline reads through the memory PCI bridges. As a result, to obtain the best performance, only remote writes are used in our conduit design, as in Ibel’s approach. Additionally, due to driver limitations, only the GASNet fast segment configuration is supported by the SCI conduit. Future improvements will allow support of the other configurations. ### 4.1 Basic communication regions Instead of only 1 buffer space for both AM requests **Network** **Independent** **Language** **Independent** |work|Distributed Shared Memory Parallel Programming Language|Col3|Langu| |---|---|---|---| ||GASNet Extended API||| |endent|GASNet Core API|Direct Access|Indepe| ||Network||| Figure 1 - GASNet layers overview ----- and replies as in Ibel’s split-queue scheme, we divide the buffer (command region) into a request and a reply queue of equal size making the system deadlock free. Each request/reply buffer space is set to be the size of the longest AM header plus the maximum size of a medium payload. The request and reply are paired so that a node with an outstanding request to another node is guaranteed to have space to hold the reply for that particular request. Each node has a message queue reserved for it on every other node. This scheme allows each node to locally manage outgoing messages and guarantee no conflicts with other nodes (Figure 2). **Control** **Control X** **Segments** **Local (In use)** **(N total)** **Control X** **Command X-1** **Command X-1** **Physical** **Command** **Address** **...** **...** **Segments** **(N*N total)** **Command X-N** **Command X-N** **Payload X** **Payload** **Local (free)** **Payload X** **Segments** **(N total)** **Node X** **SCI Space** Figure 2a – Conceptual diagram of the segment exportation mechanism **Control 1** **Control** **Segments** **...** **(N total)** **Control 1** **Control N** **...** **Control N** **Virtual** **Command 1-X** **Address** **Command 1-X** **...** **Command** **Command N-X** **Segments** **...** **(N*N total)** **Command N-X** **Node X** **SCI Space** Figure 2b –Conceptual diagram of the segment importation mechanism Similar to Ibel’s approach, a message-ready flag is used to indicate if a particular message exists in a queue or not. However, rather than attaching the flag to the end of the AM message, these flags are separately placed in an array (control region) that is accessible by all other nodes. This method provides better data locality when checking for new messages, as all the message-ready flags now reside in one contiguous memory region. In addition, a single global message-exist flag is used to indicate the existence of any new messages. Finally, the size of the long AM payload region is significantly bigger and it corresponds to the range of remotely accessible memory as specified by the GASNet _fast segment configuration which the user defines, thus_ minimizing unnecessary data copying. Since the importing of regions occupies local virtual address space equal to the size of the segment, the large payload segments (payload regions) are not imported at initialization time so as to improve scalability. Fortunately, DMA transfer mode allows communication to take place without having to import the region into virtual memory space, but with added overhead. ### 4.2 AM communication The message sending and handling process is illustrated in Figure 3. In order to send a message from a sender node to a receiver node, the sender first prepares the AM header, which contains information such as the handler to be called, message type, payload size, etc. Once prepared, the header is then written to the receiver’s command region using a PIO transfer. For a medium AM message, another remote PIO write operation is used to transfer the medium payload to the same command region. The same sequence of operations is used for long AM transfers to handle the unaligned portion of the long payload (see Section 5.2.2 for further explanation). Otherwise, the data payload is sent directly to the payload region via a DMA transfer. |Local (In use) Control X Command X-1 sical ress ... Command X-N Payload X Local (free)|Col2|Control Control X Segments (N total)| |---|---|---| |sical ress|Local (In use)|| |||Command X-1 Command ... Segments (N*N total) Command X-N| ||Control X|| ||Command X-1|| ||...|| ||Command X-N|| ||Payload X|| |||Payload Payload X Segments (N total)| ||Local (free)|| |Control 1|Virt Add| |---|---| |...|| |Control N|| |Command 1-X|| |...|| |Command N-X|| |Control 1 Control Segments ... (N total) Control N|Col2| |---|---| |Command 1-X Command Segments ... (N*N total) Command N-X|| |Col1|AM Header Medium AM Payload Long AM Payload|Col3|Check Message Exist Flag Control P So tl ali rn tg Command Y-1 ... Command Y-X Ne Aw v aM ile ias bs la eg ?es ... Extract Command Y-N Message Information Payload Y Yes No Memory Process all new messages AM Reply or ack Polling Done Polling End| |---|---|---|---| |||n|| |W|ait for Completio||| ||Flags||| ||Other processing||| ||Process reply message||| ||||| **Node X** **Node Y** Figure 3 – High-level flowchart for inter-node communication Upon completion of these transfers, the sender writes the two message flags to the receiver’s control segment. The message-exist flag is used to tell the receiver that there is at least one new message available and the message-ready flag indicates that a particular message buffer contains a message. When the receiver calls the polling process, it checks the message-exist flag to see if there are any new messages that need to be handled. If there are, the receiver scans message-ready flags and handles the appropriate newly arrived messages. Using ----- this approach, the cost of an unsuccessful poll is O(1) and O(N) for a successful poll, leading to amortized costs for polling of only O(1). ### 5 Results and analysis In this section we present the latency and bandwidth results of the first full design and implementation of our Core API. These results are compared against Dolphin SISCI raw performance and two other existing GASNet conduits, namely the GM conduit for Myrinet and the MPI conduit, a core only implementation, on SCI using Scali’s ScaMPI. ScaMPI is a commercial MPI implementation for SCI, and it is considered the most efficient communication layer implemented to date for SCI. This comparison of results is used to evaluate the performance of our design. The GASNet system provides a reference-extended API implementation that is based on Core API functions. Consequently, a complete and fully functional GASNet conduit is created with the successful completion of the Core API. To complete the analysis of our design, we compared the results of the basic Extended API operations put and get for our native SCI conduit against the MPI conduit executing on top of ScaMPI. ### 5.1 Experimental setup Here we describe the environment and testing procedures used in obtaining performance measurements from each of the software environments. ### 5.1.1 Testbed Two sets of machines were used in this study. The first set consists of 16 server nodes, each with dual 2.4GHz Intel P4 Xeon CPUs with 256KB L2 cache, 1GB of DDR PC2100 (DDR266) RAM, and a 533MHz system bus. Each node is equipped with a Dolphin D339 3D SCI card and uses Linux Red Hat 9.0 with kernel 2.4.20-8smp and gcc version 3.3.2. These SCI nodes are wired and configured as two 4×2 2D torus networks. One torus uses the free open-source driver with SISCI API V2.2 provided by Dolphin, and the other uses the commercial Scali V4.0 driver with ScaMPI. Michigan Tech graciously provided access to their Myrinet 2000E cluster for this work. Their cluster consists of 16 server nodes, each with dual 2.2GHz Intel P4 Xeon CPUs with 256KB L2 cache, 2GB of DDR PC2100 (DDR266) RAM, and a 533MHz system bus. A 16-port Myrinet 2000 switch is used to connect these nodes. The Myrinet NIC in each node features an onboard 133MHz LANai 9.0 CPU with 2MB of on-card memory using GM V1.6.3. ### 5.1.2 Experiments Performance results for SCI Raw are obtained using _scipp (PIO benchmark, ping-pong) and dma_bench (DMA_ _benchmark, one-way), latency and bandwidth benchmarks_ provided by Dolphin for the SISCI API. Conduit results are obtained by executing a slightly modified version of _testam benchmark from the GASNet test suite. The_ _testam code was changed only to output the bandwidth_ measurements for AM long transfers. To test the latency of small-message _put/get_ operations in GASNet, we use the _testsmall benchmark_ from the GASNet test suite. It uses the gasnet_put() and _gasnet_get() functions to send data back and forth_ between nodes, obtaining the round-trip latency for these requests. Bandwidth is measured using the _testlarge_ benchmark available in the GASNet test suite. It uses the various bulk-data transfer functions available in the Extended API to send one-way data between two nodes. ### 5.2 Core API AM results and analysis Short, medium, and long AM latency, as well as long AM bandwidth results, are shown in this section. As short and medium AM transfers are typically small in size and do not transfer large amounts of data, bandwidth numbers for them are not included. Comparison and analysis of our SCI conduit’s performance versus the SCI Raw, the MPI/ScaMPI Conduit, and the Myrinet Conduit are also discussed. Unfortunately, direct comparisons between our results and those from Ibel’s work cannot be made due to vastly different hardware/software testbeds. ### 5.2.1 Short/Medium AM SCI Raw SCI Conduit MPI/ScaMPI Conduit Myrinet Conduit 50 0 Bytes Payload = Short 40 30 20 10 0 0 1 2 4 8 16 32 64 128 256 512 1024 Payload Size (Bytes) Figure 4 - Short/Medium AM ping-pong latency results Compared to SCI raw performance, our SCI conduit adds ~12us of overhead (Figure 4). The main cause is the overhead added to package and unpackage the AM header, obtaining free buffer space and system sanity checks. Our results are comparable to the Myrinet conduit, but somewhat lags behind the MPI/ScaMPI conduit. Other possible causes for the overhead and reason why MPI/ScaMPI has better performance is still under investigation. The transmission of medium AM messages can be performed in two ways. The header and payload can be copied into one contiguous memory location and then transmitted in one transfer to the receiver, or instead the header and payload can be transferred separately to the receiver (Figure 5). One would expect the first approach to perform better than the second given that network ----- communication cost is generally much higher than local processing cost. However, our testing indicates that using the 2 network transactions mechanism is slightly more efficient (Figure 6). One reason may have to do with the need to perform a memcpy(), which can sometimes be an expensive operation. Another part of the reason may be that SCI allows up to 16 outstanding transactions to be posted at once. Because of this, the second SCI transaction overhead is partially hidden from the user by the first transaction (i.e. overlapping transactions). **1 network** **AM Header** **AM Header** **transfer** **AM Header** **1 network** **Medium AM** **Copy** **Medium AM** **Medium AM** **mechanismtransaction** the high DMA engine setup overhead (~30µs), any long payloads that are less than 2048 bytes are treated as unaligned data and written to the command segment using PIO mode instead. In doing so, our conduit is able to achieve better performance for small long AM payloads and suffer lower overhead for unaligned data transfers (~13us). Future implementations of the SCI conduit might switch back to use the DMA engine directly, since Dolphin is currently working on improving their driver to reduce the mapping overhead, DMA engine start-up overhead, and the alignment requirement. Our long AM latency (Figure 7) and bandwidth results (Figure 8) follow the same growth trend as that of SCI Raw and are comparable to the Myrinet conduit. Although MPI/ScaMPI has better performance for smaller payload size, its maximum bandwidth is about 190 MB/s, mainly due to the fact that it uses PIO exclusively, whereas our conduit rises to 213 MB/s with payload size of 128K. |AM Header AM Header Copy Medium AM Medium AM Payload Payload Source|1 network transfer|AM Header Medium AM Payload| |---|---|---| **Node X** **Node Y** **2 network** **transactions** **mechanism** |Col1|Col2|Col3|Col4| |---|---|---|---| |AM Header Medium AM Payload Source|||AM Header Medium AM Payload| ||AM Header||| ||||| ||Medium AM Payload Source||| ||||| **Node X** **Node Y** Figure 5 –Conceptual diagram of "1 network transaction" and "2 network transactions" message delivery mechanisms 1 network transaction 2 network transactions 40 35 30 25 20 15 10 5 0 0 1 2 4 8 16 32 64 128 256 512 1024 Payload Size (Bytes) Figure 6 - Performance comparison of "1 network transaction" and "2 network transactions" message delivery mechanisms 1000 100 |SCI Raw result obtained by double the result obtained from dma_bench|Col2| |---|---| ||| 10 Payload Size (Bytes) Figure 7 - Long AM ping-pong latency results ### 5.2.2 Long AM The SISCI API requires any DMA transfer to have 8-byte alignment between the source and the target segment (both starting address and transfer size). Sending of unaligned data thus became a problem as costly dynamic mapping (~200us overhead) and unmapping of the target segment is needed. To overcome this shortcoming, the request/reply buffer region reserved for medium payload is used as a bounce buffer for the unaligned portion of the long payload, which is later copied to the appropriate payload address when handled by the receiver. Furthermore, because of Figure 8 - Long AM bandwidth results ### 5.3 Put/Get There are two modes of testsmall, transfers to within and without the main GASNet segment. Since all small and medium AM transactions take place through buffers, the results for both modes are the same and only the graph for transfers within the segment is shown. Figure 9 ----- shows the results of testsmall for our SCI conduit and the MPI conduit on ScaMPI. Since the Extended API implementation of these two conduits is based on AM transactions in their Core APIs, the results correspond almost exactly to the latency gathered for the small and medium AM transfers in the Core API. Conduit Put (in) Conduit Get (in) MPI/ScaMPI Put (in) MPI/ScaMPI Get (in) 35 30 25 20 15 10 5 0 1 2 4 8 16 32 64 128 256 512 1024 Payload Size (Bytes) Figure 9 - Put/Get latency results The results for all blocking and non-blocking functions were the same, so only the results for _gasnet_put_bulk() and gasnet_get_bulk() are shown here._ Similar to _testsmall, there are two modes of transfer in_ _testlarge. Because our Core API currently supports only_ the _fast segment configuration, it is optimized for_ transfers to within the main GASNet segment. Therefore, only the results for one-way, in-segment transfers are shown in Figure 10. Conduit Put (in) Conduit Get (in) MPI/ScaMPI Put (in) MPI/ScaMPI Get (in) 250 200 150 100 50 0 Payload Size (Bytes) Figure 10 - Put/Get bandwidth results Similar to large AM transfers, the MPI conduit using ScaMPI achieves slightly better bandwidth for smaller transfer sizes. However, for transfers of 32KB and more, our SCI conduit shows better performance. ### 6 Conclusions GASNet is an important part of the push to expand UPC shared-memory computing capabilities to network-based systems like clusters. The GASNet conduits available on many networks allow UPC to be executed on a wide variety of platforms. SCI is a high-performance network that has many features that can be used to efficiently execute GASNet and UPC. By extending GASNet to SCI through the creation of an SCI conduit, the availability of UPC to parallel programmers increases. The creation of the GASNet Core API is an essential step in accomplishing this goal, as a complete Core API implementation is sufficient for a GASNet conduit. The tests conducted show that we have designed and created a complete and potent GASNet conduit design for SCI. The performance of our SCI conduit is shown to be comparable to the Myrinet conduit and slightly behind the MPI/ScaMPI conduit which uses proprietary SCI driver and MPI software. This outcome strengthens our belief that our SCI conduit is a promising extension to the GASNet system, as the driver used in the creation of the SCI conduit is free and open-source. Several ideas are under investigation which will further improve the performance of our conduit. Care is needed in balancing the many different aspects of network performance so that the SCI conduit can fully exploit the unique features available in the SCI network. Furthermore, currently the SCI conduit only supports GASNet global segment sizes up to 2MB, under Linux, without applying a large physical area patch. This requirement limits the usage of our conduit to those clusters whose system administrators are willing to patch the kernel on each SCI node. This patch requirement is primarily due to the limitation of the current SISCI driver where the size of each segment needs to be physically contiguous and relies on the underlying operating system to ensure continuity. We are currently working with Dolphin to resolve this issue and increase the ease of use of this conduit. Initial testing at the GASNet put/get level with our Core API again indicates that our conduit is comparable to other conduits. We are currently completing the implementation of an Extended API in order to improve the performance of our SCI conduit. Once complete, benchmarks at the UPC application level will be used to obtain a better assessment of the effectiveness of our SCI conduit from the communication to the application layer. ### Acknowledgements This work was supported in part by the U.S. Department of Defense and by equipment support of Dolphin Interconnect Solutions Inc. Also, we would like to express our thanks for the helpful suggestions and cooperation of Dan Bonachea and the UPC group members at UCB and LBNL, and to Hugo Kohmann and the support team at Dolphin for technical assistance. ### References 1. W. Carlson, J. Draper, D. Culler, K. Yelik, E. Brooks, K. Warren, “Introduction to UPC and Language Specification,” May 1999 http://www.gwu.edu/~upc/pubs.html 2. Official Unified Parallel C website ----- http://www.upc.gwu.edu/ 3. Official Berkeley UPC website http://upc.nersc.gov/ 4. Official GASNet website http://www.cs.berkeley.edu/~bonachea/gasnet 5. D. Bonachea, “GASNet Specification Version 1.3,” April 2003 http://www.cs.berkeley.edu/~bonachea/gasnet/dist/do cs/gasnet.pdf 6. D. Gustavson and Q. Li, “The Scalable Coherent Interface (SCI),” IEEE Communications, Vol. 34, No. 8, August 1996, pp. 52-63. 7. Scali, “ScaMPI – Design and Implementation,” http://www.scali.com/whitepaper/other/scampidesign. pdf 8. IEEE Service Center, “Scalable Coherent Interface, ANSI/IEEE Standard 1596-1992,” Piscataway, New Jersey, 1993. 9. Dolphin Inc., “SISCI API User Guide,” May 2001, http://www.dolphinics.com/support/documentation.ht ml 10. A. Mainwaring and E. Culler, “Active Messages: Organization and Applications Programming Interface,” Technical Document, 1995. 11. C. Bell and D. Bonachea, “A New DMA Registration Strategy for Pinning-Based High Performance Networks,” Workshop on Communication Architecture for Clusters (CAC'03), 2003. 12. M. Ibel, K.E. Schauser, C. J. Scheiman, and M. Weis, _“Implementing Active Messages and Split-C for SCI_ Clusters and Some Architectural Implications,” Sixth International Workshop on SCI-based Low-cost/High-performance Computing (SCIzzL-6), Santa Clara, CA, September 1996. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/LCN.2004.107?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/LCN.2004.107, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "http://gasnet.cs.berkeley.edu/SuGordon-HSLN.pdf" }
2,004
[ "Conference" ]
true
2004-11-16T00:00:00
[]
7,470
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Physics", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00ef77f1162f6eed2595e569d716f963c181de21
[ "Computer Science", "Physics" ]
0.860724
Classical zero-knowledge arguments for quantum computations
00ef77f1162f6eed2595e569d716f963c181de21
IACR Cryptology ePrint Archive
[ { "authorId": "2794473", "name": "Thomas Vidick" }, { "authorId": "2004707436", "name": "Tina Zhang" } ]
{ "alternate_issns": null, "alternate_names": [ "IACR Cryptol eprint Arch" ], "alternate_urls": null, "id": "166fd2b5-a928-4a98-a449-3b90935cc101", "issn": null, "name": "IACR Cryptology ePrint Archive", "type": "journal", "url": "http://eprint.iacr.org/" }
We show that every language in QMA admits a classical-verifier, quantum-prover zero-knowledge argument system which is sound against quantum polynomial-time provers and zero-knowledge for classical (and quantum) polynomial-time verifiers. The protocol builds upon two recent results: a computational zero-knowledge proof system for languages in QMA, with a quantum verifier, introduced by Broadbent et al. (FOCS 2016), and an argument system for languages in QMA, with a classical verifier, introduced by Mahadev (FOCS 2018).
# Classical zero-knowledge arguments for quantum computations #### Thomas Vidick[1] and Tina Zhang[2] 1Department of Computing and Mathematical Sciences, California Institute of Technology, USA 2Division of Physics, Mathematics and Astronomy, California Institute of Technology, USA We show that every language in QMA admits a classical-verifier, quantum-prover zero-knowledge argument system which is sound against quantum polynomial-time provers and zero-knowledge for classical (and quantum) polynomial-time verifiers. The protocol builds upon two recent results: a computational zero-knowledge proof system for languages in QMA, with a quantum verifier, introduced by Broadbent et al. (FOCS 2016), and an argument system for languages in QMA, with a classical verifier, introduced by Mahadev (FOCS 2018). ### 1 Introduction The paradigm of the interactive proof system is a versatile tool in complexity theory. Although traditional complexity classes are usually defined in terms of a single Turing machine—NP, for example, can be defined as the class of languages which a non-deterministic Turing machine is able to decide—many have reformulations in the language of interactive proofs, and such reformulations often inspire natural and fruitful variants on the traditional classes upon which they are based. (The class MA, for example, can be considered a natural extension of NP under the interactive-proof paradigm.) Intuitively speaking, an interactive proof system is a model of computation involving two entities, a verifier and a prover, the former of whom is computationally efficient, and the latter of whom is unbounded and untrusted. The verifier and the prover exchange messages, and the prover attempts to ‘convince’ the verifier that a certain problem instance is a yes-instance. We can define some particular complexity class as the set of languages for which there exists an interactive proof system that 1) is complete, 2) is sound, and 3) has certain other properties which vary depending on the class in question. Completeness means, in this case, that for any problem instance in the language, there is an interactive proof involving r messages in total that the prover can offer the verifier which will cause it to accept with at least some probability p; and soundness means that, for [Thomas Vidick: vidick@caltech.edu](mailto:vidick@caltech.edu) [Tina Zhang: tinazhang@caltech.edu](mailto:tinazhang@caltech.edu) 1 ----- any problem instance not in the language, no prover can cause the verifier to accept, except with some small probability q. For instance, if we require that the verifier is a deterministic polynomialtime Turing machine, and set r = 1, p = 1, and q = 0, the class that we obtain is of course the class NP. If we allow the verifier to be a probabilistic polynomial-time machine, and set r = 1, p = 3[2] [,] _q =_ [1]3 [, we have MA. Furthermore, if we allow the verifier to be an efficient][ quantum][ machine, and] we allow the prover to communicate with it quantumly, but we retain the parameter settings from MA, we obtain the class QMA. Finally, if we allow r to be any polynomial in n, where n is the size of the problem instance, but otherwise preserve the parameter settings from MA, we obtain the class IP. For every complexity class thus defined, there are two natural subclasses which consist of the languages that admit, respectively, a statistical and a computational zero-knowledge interactive proof system with otherwise the same properties. The notion of a zero-knowledge proof system was first considered by Goldwasser, Micali and Rackoff in [GMR89], and formalises the surprising but powerful idea that the prover may be able to prove statements to the verifier in such a way that the verifier learns nothing except that the statements are true. Informally, an interactive proof system is statistical zero-knowledge if an arbitrary malicious verifier is able to learn from an honest prover that a problem instance is a yes-instance, but can extract only negligible amounts of information from it otherwise; and the computational variant provides the same guarantee only for malicious polynomial-time verifiers. For IP in particular, the subclass of languages which admit a statistical zero-knowledge proof system that otherwise shares the same properties had by proof systems for languages in IP is known as SZK. Its computational sibling, meanwhile, is known as CZK. It is wellknown that, contingent upon the existence of one-way functions, NP CZK: computational zero_⊆_ knowledge proof systems have been known to exist for every language in NP since the early 1990s ([GMW91]). However, because these proof systems often relied upon intractability assumptions or techniques (e.g. ‘rewinding’) that failed in quantum settings, it was not obvious until recently how to obtain an analogous result for QMA. One design for a zero-knowledge proof system for promise problems in QMA was introduced by Broadbent, Ji, Song and Watrous in [BJSW16]. Their work establishes that, provided that a quantum computationally concealing, unconditionally binding commitment scheme exists, QMA QCZK. _⊆_ There are, of course, a myriad more variations on the theme of interactive proofs in the quantum setting, each of which defines another complexity class. For example, motivated partly by practical applications, one might also consider the class of languages which can be decided by an interactive proof system involving a classical verifier and a quantum prover communicating classically, in which the soundness condition still holds against arbitrary provers, but the honest prover can be implemented in quantum polynomial time. (For simplicity, we denote this class by IPBQP.) The motivation for this specific set of criteria is as follows: large-scale quantum devices are no longer so distant a dream as they seemed only a decade ago. If and when we have such devices, how will we verify, using our current generation of classical devices, that our new quantum computers can indeed decide problems in BQP? This problem—namely, the problem of showing that BQP _⊆_ IPBQP—is known informally as the problem of quantum verification. The problem of quantum verification has not yet seen a solution, but in recent years a number of strides have been made toward producing one. As of the time of writing, protocols are known for the following three variants on the problem: 2 ----- 1. It was shown in [ABE10, ABOEM17] that a classical verifier holding a quantum register consisting only of a constant number of qubits can decide languages in BQP by communicating quantumly with a single BQP prover. In [BFK09, FK17], this result was extended to classical verifiers with single-qubit quantum registers. All of these protocols are sound against arbitrary provers. 2. It was shown in [RUV13] that an entirely classical verifier can decide languages in BQP by interacting classically with two entangled, non-communicating QPT provers. This protocol is likewise sound against arbitrary provers. 3. It was shown in [Mah18] that an entirely classical verifier can decide languages in BQP by executing an argument system ([BCC88]) with a single BQP prover. An argument system differs from a proof system in that 1) its honest prover must be efficient, and 2) an argument system need not be sound against arbitrary provers, but only efficient ones. In this case, the argument system in [Mah18] is sound against quantum polynomial-time provers. (The class of languages for which there exists an argument system involving a classical probabilistic polynomial-time verifier and a quantum polynomial-time prover is referred to throughout [Mah18] as QPIP0.) The argument system introduced in [Mah18] is reliant upon cryptographic assumptions about the quantum intractability of Learning With Errors (LWE; see [Reg09]) for its soundness. For practical purposes, if this assumption holds true, the problem of verification can be considered solved. The last of these three results establishes that BQP ⊆ QPIP0, contingent upon the intractability of LWE. (As a matter of fact, the same result also establishes that QMA ⊆ QPIP0, provided the efficient quantum prover is given access to polynomially many copies of a quantum witness for the language to be verified, in the form of ground states of an associated local Hamiltonian.) In this work, we show that the protocol which [Mah18] introduces for this purpose can be combined with the zero-knowledge proof system for QMA presented in [BJSW16] in order to obtain a zero_knowledge argument system for QMA. It follows naturally that, if the LWE assumption holds, and_ quantum computationally hiding, unconditionally binding commitment schemes exist, [1] QMA _⊆_ CZK-QPIP0, where the latter refers to the class of languages for which there exists a computational _zero-knowledge interactive argument system involving a classical verifier and a quantum polynomial-_ time prover. Zero-knowledge protocols for languages in NP are an essential component of many cryptographic constructions, such as identification schemes, and are often used in general protocol design (for example, one can force a party to follow a prescribed protocol by requiring it to produce a zero-knowledge proof that it did so). Our result opens the door for the use of zero-knowledge proofs in protocols involving classical and quantum parties which interact classically in order to decide languages defined in terms of quantum information (for instance, to verify that one of the parties possesses a quantum state having certain properties). We now briefly describe our approach to the problem. The proof system for promise problems in QMA presented in [BJSW16] is almost classical, in the sense that the only quantum action which the honest verifier performs is to measure a quantum state after applying Clifford gates to it. The key contribution which [Mah18] makes to the problem of verification is to introduce a measurement _protocol which, intuitively, allows a classical verifier to obtain honest measurements of its prover’s_ 1It is known that quantum computationally hiding, unconditionally binding commitment schemes fitting our requirements can be constructed from LWE. See, for example, Section 2.4.2 in [CVZ19]. 3 ----- quantum state. The combining of the proof system from [BJSW16] and the measurement protocol from [Mah18] is therefore a fairly natural action. That the proof system of [BJSW16] is complete for problems in QMA follows from the QMAcompleteness of a problem which the authors term the 5-local Clifford Hamiltonian problem. However, the argument system which [Mah18] presents relies upon the QMA-completeness of the wellknown 2-local XZ Hamiltonian problem (see Definition 2.3). For this reason, the two results cannot be composed directly. Our first step is to make some modifications to the protocol introduced in [BJSW16] so it can be used to verify that an XZ Hamiltonian is satisfied, instead of verifying that a Clifford Hamiltonian is satisfied. We then introduce a composite protocol which replaces the quantum measurement in the protocol from [BJSW16] with an execution of the measurement protocol from [Mah18]. With the eventual object in mind of proving that the result is sound and zero-knowledge, we introduce a trapdoor check step into our composite protocol, and split the _coin-flipping protocol used in the proof system from [BJSW16] into two stages. We explain these_ decisions briefly here, after we present a summary of our protocol and its properties, and refer the reader to Sections 3, 5 and 6 for fuller expositions. **Protocol 1.1. Zero-knowledge, classical-verifier argument system for QMA (informal summary).** _Parties._ The protocol involves 1. A verifier, which runs in classical probabilistic polynomial time; 2. A prover, which runs in quantum polynomial time. _Inputs. The protocol requires the following primitives:_ A perfectly binding, quantum computationally concealing commitment protocol. _•_ A zero-knowledge proof system for NP. _•_ An extended trapdoor claw-free function family (ETCFF family), as defined in [Mah18]. _•_ Apart from the above cryptographic primitives, we assume that the verifier and the prover also receive the following inputs. 1. Input to the verifier: a 2-local XZ Hamiltonian H (see Definition 2.3), along with two numbers, _a and b, which define a promise about the ground energy of H._ Because the 2-local XZ Hamiltonian promise problem is complete for QMA, any input to any decision problem in QMA can be reduced to an instance of the 2-local XZ Hamiltonian problem. 2. Input to the prover: the Hamiltonian H, the numbers a and b, and the quantum state _ρ = σ[⊗][m], where σ is a ground state of the Hamiltonian H._ _Protocol._ 1. The prover applies an encoding process to ρ. Informally, the encoding can be thought of as a combination of an encryption scheme and an authentication scheme: it both hides the witness state ρ and ensures that the verifier cannot meaningfully tamper with the measurement results that it reports in step 5. Like most encryption and authentication schemes, this 4 ----- encoding scheme is keyed. For convenience, we refer to the encoding procedure determined by a particular encoding key K as EK.[2] 2. The prover commits to the encoding key K from the previous step using a classical commitment protocol, and sends the resulting commitment string z to the verifier. 3. The verifier and the prover jointly decide which random terms from the Hamiltonian H the verifier will check by executing a coin-flipping protocol. (‘Checking terms of H’ means that the verifier obtains measurements of the state EK(ρ) and checks that the outcomes are distributed a particular way—or, alternatively, asks the prover to prove to it that they are.) However, because it is important that the prover does not know which terms will be checked before the verifier can check them, the two parties only execute the first half of the coinflipping protocol at this stage. The verifier commits to its part of the random string, rv, and sends the resulting commitment string to the prover; the prover sends the verifier rp, its own part of the random string; and the verifier keeps the result of the protocol r = rv _rp secret_ _⊕_ for the time being. The random terms in the Hamiltonian which the verifier will check are determined by r. 4. The verifier and the prover execute the measurement protocol from [Mah18]. Informally, this allows the verifier to obtain honest measurements of the qubits of the prover’s encoded witness state, so that it can check the Hamiltonian term determined by r. The soundness guarantee of the measurement protocol prevents the prover from cheating, even though the prover, rather than the verifier, is physically performing the measurements. This soundness guarantee relies on the security properties of a family of trapdoor one-way functions termed an ETCFF family in [Mah18]. Throughout the measurement protocol, the verifier holds trapdoors for these one-way functions, but the prover does not, and this asymmetry is what allows the (intrinsically weaker) verifier to ensure that the prover does not cheat. 5. The verifier opens its commitment to rv, and also sends the prover its measurement outcomes _u and function trapdoors from the previous step._ 6. The prover checks, firstly, that the verifier’s trapdoors are valid, and that it did not tamper with the measurement outcomes u. (It can determine the latter by making use of the authentication-scheme-like properties of EK from step 1.) If both tests pass, it then proves the following statement to the verifier, using a zero-knowledge proof system for NP: There exists a string sp and an encoding key K such that z = commit(K, sp) and _Q(K, r, u) = 1._ The function Q is a predicate which, intuitively, takes the value 1 if and only if both the verifier and the prover were honest. In more specific (but still informal) terms, Q(K, r, u) takes the value 1 if u contains the outcomes of honest measurements of the state EK(ρ), where ρ is a state that passes the set of Hamiltonian energy tests determined by r. **Lemma 1.2 (soundness; informal). Assume that LWE is intractable for quantum computers. Then,** _in a no-instance execution of Protocol 1.1, the probability that the verifier accepts is at most a_ _function that is negligibly close to_ [3] 4 _[.]_ 2The notation used here for the encoding key is not consistent with that which is used later on; it is simplified for the purposes of exposition. 5 ----- **Lemma 1.3 (zero-knowledge; informal). Assume that LWE is intractable for quantum computers.** _In a yes-instance execution of Protocol 1.1, and for any classical probabilistic (resp. quantum)_ _polynomial-time verifier interacting with the honest prover, there exists a classical probabilistic_ _polynomial-time (resp._ _quantum polynomial-time) simulator such that the simulator’s output is_ _classical (resp. quantum) computationally indistinguishable from that of the verifier._ The reason we delay the verifier’s reveal of rv (rather than completing the coin-flipping in one step, as is done in the protocol in [BJSW16]) is fairly easily explained. In our classical-verifier protocol, the prover cannot physically send the quantum state EK(ρ) to its verifier before the random string _r is decided, as the prover of the protocol in [BJSW16] does. If we allow our prover to know r at_ the time when it performs measurements on the witness ρ, it will trivially be able to cheat. The trapdoor check, meanwhile, is an addition which we make because we wish to construct a _classical simulator for our protocol when we prove that it is zero-knowledge. Since our verifier is_ classical, we need to achieve a classical simulation of the protocol in order to prove that its execution (in yes-instances) does not impart to the verifier any knowledge it could not have generated itself. During the measurement protocol, however, the prover is required to perform quantum actions which no classical polynomial-time algorithm could simulate unless it had access to the verifier’s function trapdoors. Naturally, we cannot ask the verifier to reveal its trapdoors before the measurement protocol takes place. As such, we ask the verifier to reveal them immediately afterwards instead, and show in Section 6 that this (combined with the encryption-scheme properties of the prover’s encoding EK) allows us to construct a classical simulator for Protocol 1.1 in yes-instances. The organisation of the paper is as follows. 1. Section 2 (‘Ingredients’) outlines the other protocols which we use as building blocks. 2. Section 3 (‘The protocol’) introduces our argument system for QMA. 3. Section 4 (‘Completeness of protocol’) gives a completeness lemma for the argument system introduced in section 3. 4. Section 5 (‘Soundness of protocol’) proves that the argument system introduced in section 3 is sound against quantum polynomial-time provers. 5. Section 6 (‘Zero-knowledge property of protocol’) proves that the argument system is zeroknowledge (that yes-instance executions can be simulated classically). _Remark 1.4. As Broadbent et al. note in [BJSW16, Section 1.3], argument systems can often be_ made zero-knowledge by employing techniques from secure two-party computation (2PC). The essential idea of such an approach, applied to our particular problem, is as follows: the prover and the verifier would jointly simulate the classical verifier of the [Mah18] measurement protocol using a (classical) secure two-party computation protocol, and zero-knowledge would follow naturally from simulation security. (This technique is similar in spirit to those which are used in [BOGG[+]88] to show that any classical-verifier interactive proof system can be made zero-knowledge.) We think that the 2PC approach applied to our problem would have many advantages, including that it is more generally applicable than our approach; however, we also believe that our approach is a more direct and transparent solution to the particular problem at hand, and that it provides an early example of how two important results might be fruitfully combined. As such, we expect that our approach may more easily lead to extensions and improvements. 6 ----- **Related work.** Subsequent to the completion of this work, there have been several papers which explore other extensions and applications of the argument system from [Mah18], and also papers which propose zero-knowledge protocols (with different properties from ours) for QMA. Many of these works focus on decreasing the amount of interaction required to implement a proof or argument system for QMA. Although none of these works directly builds on or supersedes ours, we review them briefly for the reader’s convenience. In the category of extensions on the work of [Mah18], we mention [ACGH19], which proposes a non-interactive zero-knowledge variant of the Mahadev protocol and proves its security in the quantum random oracle model. (Of course, our protocol is interactive, and our analysis holds in the standard model.) In the category of ‘short’ proof and argument systems for QMA, we mention three independent works. In [BS19], the authors present a constant-round computationally zero-knowledge argument system for QMA. In [BG19] and [CVZ19] the authors present non-interactive zero-knowledge proof and argument systems, respectively, for QMA, with different types of setup phases. The main difference between all three of these new protocols and our protocol is that the three protocols mentioned all involve the exchange of quantum messages (although, in [CVZ19], only the setup phase requires quantum communication). _Acknowledgments. We thank Zvika Brakerski, Andru Gheorghiu, and Zhengfeng Ji for useful dis-_ cussions. We thank an anonymous referee for suggesting the approach based on secure 2PC sketched in Remark 1.4. Thomas Vidick is supported by NSF CAREER Grant CCF-1553477, AFOSR YIP award number FA9550-16-1-0495, MURI Grant FA9550-18-1-0161, a CIFAR Azrieli Global Scholar award, and the IQIM, an NSF Physics Frontiers Center (NSF Grant PHY-1125565) with support of the Gordon and Betty Moore Foundation (GBMF-12500028). Tina Zhang acknowledges support from the Richard G. Brewer Prize and Caltech’s Ph11 program. ### 2 Ingredients The protocol we present in section 3 combines techniques which were introduced in prior works for the design of protocols to solve related problems. In this section, we outline these protocols in order to introduce notation and groundwork which will prove useful in the remainder of the paper. We also provide formal definitions of QMA and of zero-knowledge. #### 2.1 Definitions **Definition 2.1 (QMA). The following definition is taken from [BJSW16].** A promise problem A = (Ayes, Ano) is contained in the complexity class QMAα,β if there exists a polynomial-time generated collection � � _Vx : x ∈_ _Ayes ∪_ _Ano_ (1) of quantum circuits and a polynomially bounded function p possessing the following properties: 1. For every string x ∈ _Ayes ∪_ _Ano, one has that Vx is a measurement circuit taking p(|x|) input_ qubits and outputting a single bit. 7 ----- 2. Completeness. For all x ∈ _Ayes, there exists a p(|x|)-qubit state ρ such that Pr(Vx(ρ) = 1) ≥_ _α._ 3. Soundness. For all x ∈ _Ano, and every p(|x|)-qubit state ρ, it holds that Pr(Vx(ρ) = 1) ≤_ _β._ In this definition, α, β [0, 1] may be constant values or functions of the length of the input _∈_ string x. When they are omitted, it is to be assumed that they are α = 2/3 and β = 1/3. Known error reduction methods [KSVV02, MW05] imply that a wide range of selections of α and β give rise to the same complexity class. In particular, QMA coincides with QMAα,β for α = 1 − 2[−][q][(][|][x][|][)] and β = 2[−][q][(][|][x][|][)], for any polynomially bounded function q. **Definition 2.2 (Zero-knowledge). Let (P, V ) be an interactive proof system (with a classical verifier** _V ) for a promise problem A = (Ayes, Ano). Assume that (possibly among other arguments) P and_ _V both take a problem instance x_ 0, 1 as input. (P, V ) is computational zero-knowledge if, for _∈{_ _}[∗]_ every probabilistic polynomial-time (PPT) V _[∗], there exists a polynomial-time generated simulator_ _S such that, when x ∈_ _Ayes, the distribution of V_ _[∗]’s final output after its interaction with the_ honest prover P is computationally indistinguishable from S’s output distribution. More precisely, let λ be a security parameter, let n be the length of x in bits, and and let {Dn,λ}n,λ and {Sn,λ}n,λ be the two distribution ensembles representing, respectively, the verifier V _[∗]’s output distribution_ after an interaction with the honest prover P on input x, and the simulator’s output distribution on input x. If (P, V ) is computationally zero-knowledge, we require that, for all PPT algorithms _A, the following holds:_ Pr Pr = µ(n)ν(λ), ����y←Dn,λ[[][A][(][y][) = 1]][ −] _y←Sn,λ[[][A][(][y][) = 1]]����_ where µ( ) and ν( ) are negligible functions. _·_ _·_ #### 2.2 Single-qubit-verifier proof system for QMA ([MF16]) Morimae and Fitzsimons ([MF16]) present a proof system for languages (or promise problems) in QMA whose verifier is classical except for a single-qubit quantum register, and which is sound against arbitrary quantum provers. The proof system relies on the QMA-completeness of the 2-local XZ Hamiltonian problem, which is defined as follows. **Definition 2.3 (2-local XZ Hamiltonian (promise) problem).** _Input. An input to the problem consists of a tuple x = (H, a, b), where_ 1. H = [�]s[S]=1 _[d][s][H][s][ is a Hamiltonian acting on][ n][ qubits, each term][ H][s][ of which]_ (a) has a weight ds which is a polynomially bounded rational number, (b) satisfies 0 ≤ _Hs ≤_ _I,_ (c) acts as the identity on all but a maximum of two qubits, (d) acts as the tensor product of Pauli observables in {σX _, σZ} on the qubits on which it_ acts nontrivially. 2. a and b are two real numbers such that 8 ----- (a) a < b, and (b) b − _a = Ω(_ poly1(|x|) [).] � � _Yes: There exists an n-qubit state σ such that_ _σ, H_ _a.[3]_ _≤_ � � _No: For every n-qubit state σ, it holds that_ _σ, H_ _b._ _≥_ � � _Remark 2.4. Given a Hamiltonian H, we call any state σ[∗]_ which causes _σ[∗], H_ to take its mini � � mum possible value a ground state of H, and we refer to the value _σ[∗], H_ as the ground energy of H. The following theorem is proven by Biamonte and Love in [BL08, Theorem 2]. **Theorem 2.5. The 2-local XZ Hamiltonian problem is complete for QMA.** We now describe an amplified version of the protocol presented in [MF16], and give a statement about its completeness and soundness which we will use. (See [MF16] for a more detailed presentation of the unamplified version of this protocol.) **Protocol 2.6 (Amplified variant of the single-qubit-verifier proof system for QMA from [MF16]).** _Notation. Let L = (Lyes, Lno) be any promise problem in QMA; let x ∈{0, 1}[∗]_ be an input; and let (H, a, b) be the instance of the 2-local XZ Hamiltonian problem to which x reduces. 1. If x ∈ _Lyes, the ground energy of H is at most a._ 2. if x ∈ _Lno, the ground energy of H is at least b._ 3. b − _a ≥_ poly1(|x|) [.] Let H = [�]s[S]=1 _[d][s][H][s][, as in Definition][ 2.3][. Define]_ _|ds|_ _πs =_ � _s_ _[|][d][s][|][ .]_ _Parties. The proof system involves_ 1. A verifier, who implements a classical probabilistic polynomial-time procedure with access to a one-qubit quantum register; and 2. A prover, who is potentially unbounded, but whose honest behaviour in yes-instances can be implemented in quantum polynomial time. The verifier and the prover communicate quantumly. _Inputs._ 3The angle brackets [�]·, ·[�] denote an inner product between two operators which is defined as follows: [�]A, B[�] = Tr(A[∗]B) for any A, B ∈ L(X _, Y), where the latter denotes the space of linear maps from a Hilbert space X to a_ Hilbert space . _Y_ 9 ----- 1. Input to the verifier: the Hamiltonian H and the numbers a and b. 2. Input to the prover: the Hamiltonian H, the numbers a and b, and the quantum state _ρ = σ[⊗][m], where σ is a ground state of the Hamiltonian H._ _Protocol._ 1. The verifier selects uniformly random coins r = (r1, . . ., rm). 2. For each j ∈{1, . . ., m}, the verifier uses rj to select a random sj ∈{1, . . ., S} according to the distribution D specified as follows: _D(s) = πs,_ for s ∈{1, . . ., S} . 3. The prover sends a state ρ to the verifier one qubit at a time. (The honest prover sends the state σ[⊗][m] that consists of m copies of the ground state of H.) 4. The verifier measures Hsj for j = 1, . . ., m, taking advantage of the fact that—if the prover is honest—it is given m copies of σ. (‘Measuring Hsj ’, in this case, entails performing at most two single-qubit measurements, in either the standard or the Hadamard basis, on qubits in _ρ, and then computing the product of the two measurement outcomes.)_ 5. The verifier initialises a variable Count to 0. For each j 1, . . ., m, if the jth product _∈{_ _}_ that it obtained in the previous step was equal to −sign(dj), the verifier adds one to Count. 6. If [Count] is closer to [1] _a_ _b_ _m_ 2 _[−]_ �s [2][|][d][s][|][ than to][ 1]2 _[−]_ �s [2][|][d][s][|] [, the verifier accepts. Otherwise, it rejects.] **Claim 2.7. Given an instance x = (H, a, b) of the 2-local XZ Hamiltonian problem, there is a** _polynomial P (depending only on a and b) such that, for any m = Ω(P_ ( _x_ )), the following holds. In _|_ _|_ _a yes-instance, the procedure of Protocol 2.6 accepts the state ρ = σ[⊗][m]_ _with probability exponentially_ _close (in_ _x_ _) to 1. In a no-instance, the probability that it accepts any state is exponentially small_ _|_ _|_ _in_ _x_ _._ _|_ _|_ _Proof. Consider the probability (over the choice of rj and the randomness arising from measure-_ ment) that the jth measurement from step 4 of Protocol 2.6, conditioned on previous measurement outcomes, yields −sign(dj). Denote this probability by qj. As shown in [MNS16, Section IV], it is not hard to verify that 1. when x ∈ _L, if the prover sends the honest witness σ[⊗][m], then qj ≥_ [1]2 _[−]_ �sa[2][|][d][s][|] [, and] 2. when x /∈ _L, for any witness that the prover sends, qj ≤_ [1]2 _[−]_ �sb[2][|][d][s][|] [.] The difference between the two cases is inverse polynomial in the size of the input to the 2-local XZ Hamiltonian problem. It is straightforward to show that, for an appropriate choice of m, this inverse polynomial gap can be amplified to an exponential one: see Appendix B. _Remark 2.8. It will be useful later to establish at this point that, if the string r from step 1 of_ Protocol 2.6 is fixed, it is simple to construct a state ρr which will pass the challenge determined by r with probability 1. One possible procedure is as follows. 10 ----- 1. For each j 1, . . ., m: _∈_ Suppose that Hsj = djP1P2, and that P1, P2 ∈{σX _, σZ} act on qubits ℓ1 and ℓ2, respectively._ (a) If −sign(dj) = 1, initialise the ((j − 1)n + ℓ1)th qubit to the +1 eigenstate of P1, and likewise, initialise the ((j − 1)n + ℓ2)th qubit to the +1 eigenstate of P2. (b) If −sign(dj) = −1, initialise the ((j − 1)n + ℓ1)th qubit to the +1 eigenstate of P1, and initialise the ((j − 1)n + ℓ2)th qubit to the −1 eigenstate of P2. 2. Initialise all remaining qubits to 0 . _|_ _⟩_ It is clear that the ρr produced by this procedure is a tensor product of |0⟩, |1⟩, |+⟩ and |−⟩ qubits. #### 2.3 Measurement protocol ([Mah18]) In [Mah18], Mahadev presents a measurement protocol between a quantum prover and a classical verifier which, intuitively, allows the verifier to obtain trustworthy standard and Hadamard basis measurements of the prover’s quantum state from purely classical interactions with it. The soundness of the measurement protocol relies upon the security properties of functions that [Mah18] terms noisy trapdoor claw-free functions and trapdoor injective functions, of which Mahadev provides explicit constructions presuming upon the hardness of LWE. (A high-level summary of these constructions can be found in Appendix A.) Here, we summarise the steps of the protocol, and state the soundness property that it has which we will use. **Protocol 2.9 (Classical-verifier, quantum-prover measurement protocol from [Mah18]).** _Parties. The proof system involves_ 1. A verifier, which implements a classical probabilistic polynomial-time procedure; and 2. A prover, which implements a quantum polynomial-time procedure. The verifier and the prover communicate classically. _Inputs._ 1. Input to the prover: an n-qubit quantum state ρ, whose qubits the verifier will attempt to derive honest measurements of in the standard and Hadamard bases. 2. Input to the verifier: (a) A string h 0, 1, which represents the bases (standard or Hadamard) in which it _∈{_ _}[n]_ will endeavour to measure the qubits of ρ. hi = 0 signifies that the verifier will attempt to obtain measurement outcomes of the ith qubit of ρ in the standard basis, and hi = 1 means that the verifier will attempt to obtain measurement outcomes of the ith qubit of ρ in the Hadamard basis. (b) An extended trapdoor claw-free function family (ETCFF family), as defined in Section 4 of [Mah18]. The description of an ETCFF family specifies a large number of algorithms, and we do not attempt to enumerate them. Instead, we proceed to describe the verifier’s prescribed actions at a level of detail which we believe to be sufficient for our purposes, and refer the reader to [Mah18] for a finer exposition. 11 ----- _Protocol._ 1. For each i 1, . . ., n (see ‘Inputs’ above for the definition of n), the verifier generates _∈{_ _}_ an ETCFF function key κi using algorithms provided by the ETCFF family, along with a trapdoor τκi for each function, and sends all of the keys κ to the prover. It keeps the trapdoors _τ to itself. If hi = 0, the ith key κi is a key for an injective function g, and if hi = 1, it is a key_ for a two-to-one function f known as a ‘noisy trapdoor claw-free function’. Intuitively, the _g functions are one-to-one trapdoor one-way functions, and the f functions are two-to-one_ trapdoor collision-resistant hash functions. The keys for f functions and those for g functions are computationally indistinguishable. (For convenience, we will from now on refer to the function specified by κi either as fκi or as gκi. Alternatively, we may refer to it as ηκi if we do not wish to designate its type.[4]) A brief outline of how these properties are achieved using LWE is given in Appendix A. We make two remarks about the functions ηκi which will become relevant later. (a) The functions ηκi always have domains of the form {0, 1} × X, where X ⊆{0, 1}[w] for some length parameter w. (b) The outputs of both the f and the g functions should be thought of not as strings but as probability distributions. The trapdoor τκi inverts the function specified by κi in the sense that, given a sample y from the distribution Y = ηκi(b∥x), along with the trapdoor _τκi, it is possible to recover b∥x, as well as any other b[′]∥x[′]_ which also maps to Y under _ηκi (should it exist)._ **Definition 2.10. Suppose that ηκi is the function specified by κi, whose output on each** input b _x in its domain_ 0, 1 is a probability distribution Y . Define a (deterministic) _∥_ _{_ _} × X_ function ηκ[∗]i[(][b][∥][x, e][) which takes as input an][ b][∥][x][ ∈{][0][,][ 1][} × X][ and a randomness][ e][ ∈E][, for] some well-defined finite set E, and returns a sample ye from the distribution Y = ηκi(b∥x). **Definition 2.11. Let ηκi be the function specified by κi, with domain {0, 1} × X** . Let y be a sample from one of the distributions Y ∈Y, where Y is the range of ηκi. It is guaranteed that the distributions in the range of ηκi have compact support, and we call b∥x ∈{0, 1} × X a preimage of y if y is in the support of the probability distribution ηκi(b∥x). 2. The prover uses the function keys κ1, . . ., κn to ‘commit’ to the quantum state of each of the _n qubits of ρ, and sends n commitment strings y1, . . ., yn to the verifier. We direct the reader_ to Sections 2.2 and 5.1 of [Mah18] for a high-level, and then a more detailed, description of the commitment procedure, and explanations of how such a procedure will allow the verifier to extract trustworthy standard and Hadamard basis measurements of the qubits of ρ. For convenience, we summarise the procedure without justification here. For each i, the prover concatenates to the qubit indexed by i (which we call, following [Mah18], the ith ‘committed qubit’, and whose state we denote by[5] _|ψi⟩_ = γi |0⟩ + _δi |1⟩) a register (the_ ‘preimage register’) containing a uniform superposition over all x (where 0, 1 is the _∈X_ _{_ _}×X_ domain of the function ηκi). It also concatenates to it a register containing a superposition 4The letter η has been chosen because it bears some resemblance to the Latin letter h. 5Strictly speaking, the state of the ith commited qubit may not be a pure state, but we ignore this fact for simplicity of presentation. 12 ----- over all e ∈E, with E defined as it is in Definition 2.10. It then applies the function ηκ[∗]i [(see] Definition 2.10) in superposition to _x_ [�] _[|][ψ][i][⟩|][x][⟩|][e][⟩][, and collects the outputs in a new register]_ (the ‘output register’), obtaining the quantum state � _γi |0⟩|x⟩|e⟩|ηκ[∗]i[(0][∥][x, e][)][⟩]_ [+][ δ][i] _[|][1][⟩|][x][⟩|][e][⟩|][η]κ[∗]i[(1][∥][x, e][)][⟩]_ _[.]_ _x∈{0,1}[w]_ Finally, the prover measures the output register, obtains a sample yi, and sends yi to the verifier as its ith commitment string. 3. The verifier then chooses at random to run either a test round or a Hadamard round. Each option is chosen with probability [1] 2 [.] (a) If a test round is chosen, the verifier requests standard basis measurements of each committed qubit |ψi⟩ and its associated preimage register (recall that, in the previous step, only the output registers were measured), and receives measurement results βi, xi for i 1, . . ., n from the prover. It then checks, for each i 1, . . ., n, whether _∈{_ _}_ _∈{_ _}_ the ith measurement result is a preimage of yi. (See Definition 2.11 for a definition of the phrase ‘preimage of y’.) More precisely: if hi = 0, the verifier accepts iff yi is in the support of gκi(βi∥xi) (recalling that gκi(βi∥xi) is a probability distribution); and if _hi = 1, the verifier accepts iff yi is in the support of fκi(βi∥xi)._ (b) If a Hadamard round is chosen, the verifier requests Hadamard basis measurements of each committed qubit and its associated preimage register, and receives βi, di for _i_ 1, . . ., n from the prover. (We note, for future reference, that the distribution of _∈{_ _}_ the strings di sent by the honest prover is the uniform distribution over {0, 1}[w].) It then obtains the measurement results it desires using the procedure outlined in Section 2.2 of [Mah18]: i. If hi = 0, the verifier ignores the measurement results βi, di, and uses the trapdoor _τκi to invert yi. If the inverse of yi exists, the inversion procedure outputs a preimage_ (mi, xmi,yi), and the verifier stores mi as the standard-basis measurement outcome. If the inverse does not exist, the verifier stores a random bit as mi and rejects. The specification that the verifier stores a random bit is made so that the n-bit string of measurement results, m, is well-defined even when the verifier rejects. ii. If hi = 1, the verifier computes the two preimages x0,i and x1,i to yi under fκi using the trapdoor τκi. It then stores βi _di_ (x0,i _x1,i) as the Hadamard basis_ _⊕_ _·_ _⊕_ measurement outcome. Once again, if either inverse does not exist, the verifier stores a random bit as mi and rejects. The soundness property of this measurement protocol (Claim 7.1 in [Mah18]) is stated below. Before we state it, we briefly introduce the notation used in [Mah18]. (For a more detailed exposition, we refer the reader to Section 5.3 of [Mah18].) DP,h refers to the distribution over measurement results _m_ 0, 1 that the verifier obtains when it executes a Hadamard round with the prover labelled _∈{_ _}[n]_ P on the basis choice h. DP[C],h [is the same distribution, but conditioned on the verifier accepting] (in a Hadamard round). Dξ,h is the distribution over measurement outcomes in {0, 1}[n] that would result from directly measuring the quantum state ξ in the bases determined by h. ph,T and ph,H are defined so that the verifier’s probability of accepting (on basis choice h) in a test and a Hadamard round, respectively, are 1 _−_ _ph,T and 1_ _−_ _ph,H_ . ∥·∥TV denotes the total variation norm, and A ≈c B indicates that two distributions A and B are (quantum) computationally indistinguishable. 13 ----- **Claim 2.12. Assume that the Learning With Errors problem (with the same choices of parameters** _as those made in [Mah18, Section 9]) is quantum computationally intractable._ _Then, for any_ _arbitrary quantum polynomial-time prover P who executes the measurement protocol (Protocol 2.9)_ _with the honest verifier V, there exists a quantum state ξ, a prover P[′]_ _and a negligible function µ_ _such that_ _∥DP[C],h_ _[−]_ _[D][P][′][,h][∥][TV]_ _[≤√][p][h,T]_ [+][ p][h,H] [+][ µ] _and_ _DP′,h_ _c Dξ,h ._ _≈_ #### 2.4 Zero-knowledge proof system for QMA ([BJSW16]) In [BJSW16], Broadbent, Ji, Song and Watrous describe a protocol involving a quantum polynomialtime verifier and an unbounded prover, interacting quantumly, which constitutes a zero-knowledge proof system for promise problems in QMA. (Although it is sound against arbitrary provers, the system in fact only requires an honest prover to perform quantum polynomial-time computations.) We summarise the steps of their protocol below. For details and fuller explanations, we refer the reader to [BJSW16, Section 3]. **Protocol 2.13 (Zero-knowledge proof system for QMA from [BJSW16]).** _Notation. Let L be any promise problem in QMA. For a definition of the k-local Clifford Hamilto-_ _nian problem, see [BJSW16, Section 2]. The k-local Clifford Hamiltonian problem is QMA-complete_ for k = 5; therefore, for all possible inputs x, there exists a 5-local Clifford Hamiltonian H (which can be computed efficiently from x) whose terms are all operators of the form C[∗] 0[k] 0[k] _C for_ _|_ _⟩⟨_ _|_ some Clifford operator C, and such that 1. if x ∈ _Lyes, the ground energy of H is ≤_ 2[−][p], 2. if x ∈ _Lno, the ground energy of H is ≥_ [1]q [,] for some positive integers p and q. _Parties. The proof system involves_ 1. A verifier, who implements a quantum polynomial-time procedure; 2. A prover, who is unbounded, but who is only required by the protocol to implement a quantum polynomial-time procedure. The verifier and the prover communicate quantumly. _Inputs._ 1. Input to the verifier: (a) The Hamiltonian H. (b) A quantum computationally concealing, perfectly binding (classical) commitment protocol. (c) A proof system for NP sound against arbitrary quantum provers. 14 ----- 2. Input to the prover: (a) The Hamiltonian H. (b) The n-qubit quantum state ρ, where ρ is a ground state of the Hamiltonian H. (c) A quantum computationally concealing, perfectly binding (classical) commitment protocol. (d) A zero-knowledge proof system for NP sound against arbitrary quantum provers. _Protocol._ 1. The prover’s encoding step. The prover firstly applies a concatenated Steane code (which maps every one qubit to N qubits) to each qubit in ρ. (For details on the concatenated Steane code, see [BJSW16, Appendix A.6]. It will be important to Broadbent et al.’s purposes—and ours, in Section 3—that this code admits transversal applications of Clifford operations.) It then executes the following steps: (a) Concatenate N trap qubits to the end of each logical qubit (alternatively, to the end of each N -qubit block) in the result of applying the concatenated Steane code to ρ. Each trap qubit is initialised uniformly at random to one of |0⟩ _, |+⟩_ _, |+y⟩.[6]_ Denote the string that records the choices of trap qubits for all n logical qubits by t = t1, . . ., tn. (b) Permute each 2N -tuple of qubits in the result of (a) according to a uniformly random permutation π ∈ _S2N_ . (Note that the same permutation π is applied to every 2N -tuple.) (c) Apply a Pauli one-time pad X _[a]Z[b], for uniformly random a, b_ 0, 1, to the entire _∈{_ _}[2][nN]_ 2nN -qubit state. _Remark 2.14. The number N functions as a security parameter, both for this protocol and_ for the protocol that we construct in section 3. Accordingly, we assume that N is polynomial in the size of the verifier’s input. The prover’s encoding applied to ρ is denoted by E(ρ), and the procedure E is fully determined by the encoding key (t, π, a, b) which the prover chose to use. At this point, the prover sends the state E(ρ) to the verifier, along with a commitment (using some perfectly binding, computationally concealing classical commitment protocol) to the tuple (π, a, b). (A commitment to the sequence of trap qubits t is unnecessary because, in a sense, the trap qubits exist only to check the verifier.) Let the prover’s commitment string be denoted z. 2. Coin-flipping protocol. The prover and the verifier execute a coin-flipping protocol, choosing a string r of fixed length uniformly at random. This random string r determines a local Hamiltonian term Hr = Cr[∗] _[|][0][k][⟩⟨][0][k][|][ C][r]_ [that is to be tested. (This step can be implemented,] of course, using the same classical commitment protocol that the prover employed in the previous step.) 3. Verifier’s challenge. The verifier applies the Clifford Cr transversally to the qubits on which the k-local Hamiltonian term Hr acts nontrivially, and measures them in the standard basis. It then sends the measurement results ui1, . . ., uik which it obtained to the prover. (Each ui is a 2N -bit string, and i1, . . ., ik are the indices of the logical qubits on which the term Hr acts nontrivially.) 6|+y⟩ here refers to the state _√12_ [(][|][0][⟩] [+][ i][ |][1][⟩][).] 15 ----- 4. Prover’s response. The prover receives the verifier’s measurement results u, and firstly checks whether they cause a predicate Q(t, π, a, b, r, u) to be satisfied. (We will explain the predicate _Q in more detail shortly. Intuitively, Q is satisfied if and only if both verifier and prover_ behaved honestly.) If Q is not satisfied, the prover aborts, causing the verifier to reject. If Q is satisfied, then the prover proves to the verifier, using an NP zero-knowledge protocol, that there exists randomness sp and an encoding key (t, π, a, b) such that z = commit((π, a, b), sp) and Q(t, π, a, b, r, u) = 1. We now describe the predicate Q in precise terms. For convenience, Broadbent et al. define a predicate Rr, which represents the prover’s check after it reverses the effects on u of the one-time pad X _[a]Z[b], and then proceed to define Q in terms of Rr. Since we will shortly_ have cause to alter the definition of Rr, we quote the definition of Rr used in [BJSW16] for comparison. (Note that we have altered their notation slightly: the strings that they call yi we have called qi, to avoid a conflict.) **Definition 2.15 (Definition of Rr). Let ui1, . . ., uik be the measurement results that the** verifier sent to the prover in step 3. For each i ∈{i1, . . ., ik}, define strings qi, zi in {0, 1}[N] such that π(qi∥zi) = ui (alternatively: π[−][1](ui) = qi∥zi). The predicate Rr(t, π, u) takes the value 1 if and only if the following two conditions are met: (a) qi ∈DN for every i ∈{i1, . . ., ik}, and qi ∈DN[1] [for at least one index][ i][ ∈{][i][1][, . . ., i][k][}][.] (DN = DN[0] _N_ [is the set of all valid classical][ N] [-bit codewords of the concatenated] _[∪D][1]_ Steane code. We note, as an aside, that DN[0] [and][ D]N[1] [are both sets of size 8][t][, where][ t][ is] an even positive integer such that 7[t] = N ; as such, DN is polynomially sized.) (b) ⟨zi1 · · · zik _| Cr[⊗][n]_ _|ti1 · · · tik_ _⟩̸= 0._ Now we define the predicate Q(t, π, a, b, r, u) in terms of Rr: **Definition 2.16 (Definition of Q). Let c1, . . ., cn, d1, . . ., dn ∈{0, 1}[2][N]** be the unique strings such that _Cr[⊗][2][N]_ (X _[a][1]Z[b][1]_ _⊗· · · ⊗_ _X_ _[a][n]Z[b][n]) = α(X_ _[c][1]Z[d][1]_ _⊗· · · ⊗_ _X_ _[c][n]Z[d][n])Cr[⊗][2][N]_ for some α ∈{1, i, −1, −i}. (It is possible to efficiently compute c = c1, . . ., cn and d = _d1, . . ., dn given a, b and Cr.) The predicate Q is then defined by_ _Q(t, π, a, b, r, u) = Rr(t, π, u ⊕_ _ci1 · · · cik_ ) . #### 2.5 Replacing Clifford verification with XZ verification in Protocol 2.13 The authors of [BJSW16] introduce a zero-knowledge proof system which allows the verifier to determine whether the prover holds a state that has sufficiently low energy with respect to a klocal Clifford Hamiltonian (see Section 2 of [BJSW16]). In this section, we modify their proof system so that it applies to an input encoded as an instance of the XZ local Hamiltonian problem (Definition 2.3) rather than as an instance of the Clifford Hamiltonian problem. Before we introduce our modifications, we explain why it is necessary in the first place to alter the proof system presented in [BJSW16]. Modulo the encoding E which the prover applies to its state 16 ----- in Protocol 2.13, the quantum verifier from the same protocol is required to perform a projective measurement of the form Π = C[∗] 0[k] 0[k] _C, Id_ Π of the state that the prover sends it (where _{_ _|_ _⟩⟨_ _|_ _−_ _}_ _C is a Clifford unitary acting on k qubits) and reject if it obtains the first of the two possible_ outcomes. Due to the properties of Clifford unitaries, this action is equivalent to measuring k commuting k-qubit Pauli observables C[∗]ZiC for i ∈{1, . . ., k} (where Zi is a Pauli σZ observable acting on the ith qubit), and rejecting if all of said measurements result in the outcome +1. Our goal is to replace the quantum component of the verifier’s actions in Protocol 2.13—a component which, fortunately, consists entirely of performing the projective measurement just described—with the measurement protocol introduced in [Mah18] (summarized as Protocol 2.9). Unfortunately, the latter protocol 1. only allows for standard and Hadamard basis measurements, and 2. does not accommodate a verifier who wishes to perform multiple successive measurements on the same qubit: for each qubit that the verifier wants to measure, it must decide on a measurement basis (standard or Hadamard) prior to the execution of the protocol, and once made its choices are fixed for the duration of its interaction with the prover. This allows the verifier to, for example, obtain the outcome of a measurement of the observable C[∗]ZiC for some particular i, by requesting measurement outcomes of all k qubits in the appropriate basis and taking the product of the outcomes obtained. However, it is not obvious how the same verifier could request the outcome of measuring a k-tuple of commuting Pauli observables which all act on the same k qubits. To circumvent this technical issue, we replace the Clifford Hamiltonian problem used in [BJSW16] with the QMA-complete XZ Hamiltonian problem. The advantage of this modification is that it becomes straightforward to implement the required energy measurements using the measurement protocol from [Mah18]. In order to make the change, we require that the verifier’s measurements act on a linear, rather than a constant, number of qubits with respect to the size of the problem input. A different potentially viable modification to the proof system of [BJSW16] is as follows. Instead of replacing Clifford Hamiltonian verification with XZ Hamiltonian verification, we could also repeat the original Clifford-Hamiltonian-based protocol a polynomial number of times. In such a scheme, the honest prover would hold m copies of the witness state (as it does in Protocol 2.6). The verifier, meanwhile, would firstly choose a random term Cr[∗] _[|][0][k][⟩⟨][0][k][|][ C][r]_ [from the Clifford Hamiltonian, and] then select m random Pauli observables of the form Cr[∗][Z][i][C][r][—where][ C][r] [is the particular][ C][r] [which] it picked—to measure. (For each repetition, i would be chosen independently and uniformly at random from the set 1, . . ., k .) The verifier would accept if and only if the number of times _{_ _}_ it obtains −1 from said Pauli measurements is at least 2[m]k [. This approach is very similar to the] approach we take for XZ Hamiltonians (which we explain below), and in particular also fails to preserve the perfect completeness of the original protocol in [BJSW16]. For simplicity, we choose the XZ approach. We now introduce the alterations which are necessary in order to make it viable. Firstly, we require that the honest prover possesses polynomially many copies of the witness state _σ, instead of one. We do this because we want the honest verifier to accept the honest prover_ with probability exponentially close to 1, which is not naturally true in the verification procedure for 2-local XZ Hamiltonians presented by Morimae and Fitzsimons in [MF16], but which is true in our amplified variant, Protocol 2.6. Secondly, we need to modify the verifier’s conditions for acceptance. In [BJSW16], as we have mentioned, these conditions are represented by a predicate _Q (that in turn evaluates a predicate Rr; see Definitions 2.15 and 2.16)._ 17 ----- We now describe our alternative proof system for QMA, and claim that it is zero-knowledge. Because the protocol is very similar to the protocol from [BJSW16], this can be seen by following the proof of zero-knowledge in [BJSW16], and noting where our deviations require modifications to the reasoning. On the other hand, we do not argue that the proof system is complete and sound, as we do not need to make explicit use of these properties. (Intuitively, however, the completeness and the soundness of the proof system follow from those of Protocol 2.6, and the soundness of the latter is a property which we will use.) **Protocol 2.17 (Alternative proof system for QMA).** _Notation. Refer to notation section of Protocol 2.6._ _Parties. The proof system involves_ 1. A verifier, who implements a quantum polynomial-time procedure; 2. A prover, who is unbounded, but who is only required by the protocol to implement a quantum polynomial-time procedure. The verifier and the prover communicate quantumly. _Inputs._ 1. Input to the verifier: (a) The Hamiltonian H, and the numbers a and b. (b) A quantum computationally concealing, perfectly binding (classical) commitment protocol. (c) A proof system for NP sound against arbitrary quantum provers. 2. Input to the prover: (a) The Hamiltonian H, and the numbers a and b. (b) The n-qubit quantum state ρ = σ[⊗][n], where σ is the ground state of the Hamiltonian H. (c) A quantum computationally concealing, perfectly binding (classical) commitment protocol. (d) A zero-knowledge proof system for NP sound against arbitrary quantum provers. _Protocol._ 1. Prover’s encoding step: The same as the prover’s encoding step in Protocol 2.13, except that t ∈{0, +}[N] rather than {0, +, +y}[N] . (This change will be justified in the proof of Lemma 2.20.) 2. Coin flipping protocol: Unmodified from Protocol 2.13, except that r = (r1, . . ., rm) represents the choice of m terms from the 2-local XZ Hamiltonian H (with the choices being made as described in step 2 of Protocol 2.6) instead of a random term from a Clifford Hamiltonian. Note that r determines the indices of the 2m logical qubits which the verifier will measure in step 3. 3. Verifier’s challenge: The same as the verifier’s challenge in Protocol 2.13, except that the verifier now applies Ur transversally instead of Cr. (See item 2(c) in Definition 2.18 below for the definition of Ur.) 18 ----- 4. Prover’s response: The same as Protocol 2.13 (but note that the predicate Q, which the prover checks and then proves is satisfied, is the Q described in Definition 2.19 below). **Definition 2.18 (Redefinition of Rr). Let i1, . . ., i2m be the indices of the logical qubits which** were chosen for measurement in step 2 of Protocol 2.17, ordered by their corresponding js (so that _i1 and i2 are the qubits that were measured in order to determine whether Hs1 was satisfied, and_ so on). Let ui1, . . ., ui2m be the 2N -bit strings which the verifier claims are the classical states that remained after said measurements were performed, and for each i ∈{i1, . . ., i2m}, define N bit strings qi, zi such that π(qi||zi) = ui (alternatively: π[−][1](ui) = qi||zi). In Protocol 2.17, the predicate Rr(t, π, u) takes the value 1 if and only if the following conditions are met: 1. qi ∈DN for every i ∈{i1, . . ., i2m}. 2. The number [Count] (where Count is obtained by executing the following procedure) is closer _m_ to [1] _a_ _b_ 2 _[−]_ �s [2][|][d][s][|][ than to][ 1]2 _[−]_ �s [2][|][d][s][|] [.] (a) Initialise Count to 0. (b) For each j ∈{1, . . ., m}: Suppose that Hsj = djP1P2, for some P1, P2 ∈{σX _, σZ}. The_ tuple (P1, u2j−1, P2, u2j) determines a ‘logical’ measurement result that could equally have been obtained by measuring Hrj _σ, where σ is the unencoded witness state. We_ denote this measurement result by λ. If λ = −sign(dj), add one to Count. (c) Let Ur be the circuit obtained from the following procedure: i. For each j ∈{1, . . ., m}, replace any σX s in the term Hsj with H (Hadamard) gates, and replace any σZs in Hsj with I. (For example, if Hrj = πjσX,ℓ1σZ,ℓ2, where the second subscript denotes the index of the qubit on which the observable in question acts, then Uj = Hℓ1Iℓ2, where the subscripts ℓ1 and ℓ2 once again the denote the indices of the qubits on which the gates H and I act.) ii. Apply Uj to the qubits indexed (j − 1)n + 1 through jn. It must then be the case that ⟨zi1 · · · zi2m| Ur[⊗][N] _|ti1 · · · ti2m⟩̸= 0 (where each ti is an_ _N_ -bit string that represents the pattern of trap qubits which was concatenated to the _ith logical qubit during step 1 of Protocol 2.17)._ **Definition 2.19 (Redefinition of Q). Let c1, . . ., cn, d1, . . ., dn ∈{0, 1}[2][N]** be the unique strings such that _Ur[⊗][2][N]_ (X _[a][1]Z[b][1]_ _⊗· · · ⊗_ _X_ _[a][n]Z[b][n]) = α(X_ _[c][1]Z[d][1]_ _⊗· · · ⊗_ _X_ _[c][n]Z[d][n])Ur[⊗][2][N]_ for some α ∈{1, i, −1, −i}. (It is possible to efficiently compute c = c1, . . ., cn and d = d1, . . ., dn given a, b and Ur. In particular, recalling that Ur is a tensor product of H and I gates, we have that ci = ai and di = bi for all i such that the ith gate in Ur[⊗][2][N] is I, and ci = bi, di = ai for all i such that the ith gate in Ur[⊗][2][N] is H.) The predicate Q is then defined by _Q(t, π, a, b, r, u) = Rr(t, π, u ⊕_ _ci1 · · · cik_ ), where Rr is as in Definition 2.18. **Lemma 2.20. The modified proof system for QMA in Protocol 2.17 is computationally zero-** _knowledge for quantum polynomial-time verifiers._ 19 ----- _Proof. We follow the argument from [BJSW16, Section 5]. Steps 1 to 3 only make use of the security_ of the coin-flipping protocol, the security of the commitment scheme, and the zero-knowledge properties of the NP proof system, none of which we have modified. Step 4 replaces the real witness state ρ with a simulated witness ρr that is guaranteed to pass the challenge indexed by _r; this we can do also (see Remark 2.8). Step 5 uses the Pauli one-time-pad to twirl the cheating_ verifier, presuming that the honest verifier would have applied a Clifford term indexed by r before measuring. We note that, since Ur is a Clifford, the same reasoning applies to our modified proof system. Finally, using the fact that the Pauli twirl of step 5 restricts the cheating verifier to XOR attacks, step 6 from [BJSW16, Section 5] proves the following statement: if the difference |p0 − _p1| is_ negligible (where p0 and p1 are the probabilities that ρ and ρr respectively pass the verifier’s test in an honest prover-verifier interaction indexed by r), then the channels Ψ0 and Ψ1 implemented by the cheating verifier in each case are also quantum computationally indistinguishable. It follows from this statement that the protocol is zero-knowledge, since, in an honest verifier-prover interaction indexed by r, ρr would pass with probability 1, and ρ would pass with probability 1 − negl(N ). (This latter statement is true both in their original and in our modified protocol.) The argument presented in [BJSW16] considers two exclusive cases: the case when |v|1 < K, where v is the string that the cheating verifier XORs to the measurement results, |v|1 is the Hamming weight of that string, and K is the minimum Hamming weight of a nonzero codeword in DN ; and the case when _|v|1 ≥_ _K. The analysis in the former case translates to Protocol 2.17 without modification, but in_ the latter case it needs slight adjustment. In order to address the case when |v|1 ≥ _K, Broadbent et al. use a lemma which—informally—_ states that the action of a Clifford on k qubits, each of which is initialised uniformly at random to one of |0⟩ _, |+⟩, or |+⟩y, has at least a 3[−][k]_ chance of leaving at least one out of k qubits in a standard basis state. We may hesitate to replicate their reasoning directly, because our k (the number of qubits on which our Hamiltonian acts) is not a constant. While it is possible that a mild modification suffices to overcome this problem, we note that in our case there is a simpler argument for an analogous conclusion: since Ur is a tensor product of only H gates and I gates, it is straightforward to see that, if each of the 2m qubits on which it acts is initialised either to 0 _|_ _⟩_ or to +, then 1) each of the 2m qubits has exactly a 50% chance of being left in a standard basis _|_ _⟩_ state, and 2) the states of these 2m qubits are independent. Now we consider the situation where a string v = v1 v2 _v2m, of length 4mN and of Hamming_ _· · ·_ weight at least K, is permuted (‘permuted’, here, means that π ∈ _S2N is applied to each vi_ individually) and then XORed to the result of measuring 4mN qubits (2m blocks of 2N qubits each) in the standard basis after Ur has been transversally applied to those qubits. It is straightforward to see, by an application of the pigeonhole principle, that there must be at least one vi whose Hamming weight is _K_ _≥_ 2m [. Consider the result of XORing this][ v][i][ to its corresponding block of] measured qubits. Half of the 2N qubits in that block would originally have been encoding qubits, and half would have been trap qubits; half again of the latter, then, would have been trap qubits left in a standard basis state by the transversal action of Ur. As such, the probability that none of the 1-bits of vi are permuted into positions which are occupied by the latter kind of qubit is ( [3]4 [)][−] 2[K]m, which is negligibly small as long as K is made to be a higher-order polynomial in N than 2m is. The remainder of the argument in [BJSW16, Section 5] follows directly. 20 ----- ### 3 The protocol In this section, we present our construction of a zero-knowledge argument system for QMA. Our argument system allows a classical probabilistic polynomial-time verifier and a quantum polynomialtime prover to verify that any problem instance x belongs to any particular language L QMA, _∈_ provided that the prover has access to polynomially many copies of a valid quantum witness for an instance of the 2-local XZ local Hamiltonian problem to which x is mapped by the reduction implicit in Theorem 2.5. The argument system is sound (against quantum polynomial-time provers) under the following assumptions: **Assumptions 3.1.** 1. The Learning With Errors problem (LWE) [Reg09] is quantum computationally intractable. (Specifically, we make the same asssumption about the hardness of LWE that is made in [Mah18, Section 9] in order to prove the soundness of the measurement protocol.) 2. There exists a commitment scheme (gen, initiate, commit, reveal, verify) of the form described in Appendix C that is unconditionally binding and quantum computationally concealing. (This assumption is necessary to the soundness of the proof system presented in [BJSW16].) It is known that a commitment scheme with the properties required can be constructed assuming the quantum computational hardness of LWE [CVZ19], although the parameters may be somewhat different from those required for soundness. The following exposition of our protocol relies on definitions from Section 2, and we encourage the reader to read that section prior to approaching this one. We also direct the reader to Figures 1 and 2 for diagrams that chart the protocol’s structure. **Protocol 3.2. Zero-knowledge, classical-verifier argument system for QMA.** _Notation. Let L be any promise problem in QMA, and let (H =_ [�]s[S]=1 _[d][s][H][s][, a, b][) be an instance of]_ the 2-local XZ Hamiltonian problem to which L can be reduced (see Definition 2.3 and Theorem 2.5). Define _|ds|_ _πs =_ � _s_ _[|][d][s][|][ .]_ Following [BJSW16], we take the security parameter for this protocol to be N, the number of qubits in which the concatenated Steane code used during the encoding step of the protocol (step 1) encodes each logical qubit. We assume, accordingly, that N is polynomial in the size of the problem instance x. _Parties._ The protocol involves 1. A verifier, which runs in classical probabilistic polynomial time; 2. A prover, which runs in quantum polynomial time. _Inputs. The protocol requires the following primitives:_ 21 ----- A perfectly binding, quantum computationally concealing commitment protocol (gen, initiate, _•_ commit, reveal, verify) (which will be used twice: once for the prover’s commitment in step 2, and then again for the coin-flipping protocol in step 3). We assume that this commitment protocol is of the form described in Appendix C. A zero-knowledge proof system for NP. _•_ An extended trapdoor claw-free function family (ETCFF family), as defined in [Mah18]. _•_ (Note that we fall short of using the ETCFF family as a black box: for the trapdoor check of step 8, we rely on the specific properties of the LWE-based construction of an ETCFF family that [Mah18] provides. See Appendix A for details.) Apart from the above cryptographic primitives, we assume that the verifier and the prover also receive the following inputs. 1. Input to the verifier: the Hamiltonian H and the numbers a and b. 2. Input to the prover: the Hamiltonian H, the numbers a and b, and the quantum state _ρ = σ[⊗][m], where σ is a ground state of the Hamiltonian H._ _Protocol._ 1. The prover encodes the witness. The prover encodes the quantum witness ρ by applying the following steps: (a) Apply concatenated Steane code (b) Concatenate trap qubits _t_ _|_ _⟩_ (c) Apply a random permutation π (d) Apply a Pauli one-time-pad X _[a]Z[b]_ The encoding process here is the same as that from step 1 of Protocol 2.17; we direct the reader to Protocol 2.17, and the Protocol 2.13 to which it refers, for a more detailed explanation of the steps. Denote the application of the prover’s encoding to the state ρ by E(ρ). 2. The prover commits to its encoding keys. The prover commits to the strings (π, a, b) from the previous step, using randomness sp. Call the prover’s commitment string z, so that _z = commit((π, a, b), sp)._ 3. The verifier and the prover execute the first half of a two-stage coin-flipping protocol.[7] The verifier commits to rv, its part of the random string that will be used to determine which random terms in the Hamiltonian H it will check in subsequent stages of the protocol. Let _c = commit(rv, sv). The prover sends the verifier rp, which is its own part of the random_ string. The random terms will be determined by r = rv ⊕ _rp. (r is used to determine these_ terms in the same way that r is used in Protocol 2.6.) 4. The verifier initiates the measurement protocol. _(Refer to Protocol 2.9 for an outline of_ _the steps in said measurement protocol.) The verifier chooses the measurement bases h =_ 7We need to execute the coin-flipping protocol in two stages because, in our (classical-verifier) protocol, the prover cannot physically send the quantum state E(ρ) to its verifier before the random string r is decided, as the prover of Protocol 2.13 does. If we allow our prover to know r at the time when it performs measurements on the witness ρ, it will trivially be able to cheat. 22 ----- _h1_ _h2nN in which it wishes to measure the state E(ρ). 2kN out of the 2nN bits of h—_ _· · ·_ corresponding to k logical qubits—are chosen so that the verifier can determine whether σ satisfies the Hamiltonian terms specified by r = rv ⊕ _rp. In our particular case, k = 2m,_ where m is the number of Hamiltonian terms that the verifier will check are satisfied. For the remaining qubits i, the verifier sets hi to 0. The verifier sends the function keys κ = _κ1, . . ., κ2nN to the prover._ 5. The prover commits to its encoded witness state, as per the measurement protocol. The prover commits to the quantum state E(ρ) by concatenating a preimage register to each qubit in _E(ρ), applying the functions specified by κ1, . . ., κ2nN in superposition as Protocol 2.9 de-_ scribes, measuring the resulting output superpositions, and sending the outcomes y1, . . ., y2nN to the verifier. 6. The verifier chooses at random to run either a test round or a Hadamard round. Each option is chosen with probability [1] 2 [.] (a) If a test round is chosen, the verifier requests standard basis measurements of each committed qubit |ψi⟩ in E(ρ) and its associated preimage register, and receives measurement results βi, xi for i ∈{1, . . ., 2nN _} from the prover. It then checks, for each_ _i ∈{1, . . ., 2nN_ _}, whether the ith measurement result is a preimage of yi. (See Def-_ inition 2.11 for a definition of the phrase ‘preimage of y’.) More precisely: if hi = 0, the verifier accepts iff yi is in the support of gκi(βi∥xi) (recalling that gκi(βi∥xi) is a probability distribution); and if hi = 1, the verifier accepts iff yi is in the support of _fκi(βi∥xi)._ (b) If a Hadamard round is chosen, the verifier requests Hadamard basis measurements of each committed qubit and its associated preimage register, and receives βi, di for _i_ 1, . . ., 2nN from the prover. It then obtains the measurement results it desires _∈{_ _}_ using the procedure outlined in Section 2.2 of [Mah18]: i. If hi = 0, the verifier ignores the measurement results βi, di, and uses the trapdoor _τκi to invert yi. The inversion procedure outputs a preimage (mi, xmi,yi); the verifier_ stores mi as the standard-basis measurement outcome. (If the inverse does not exist, the verifier stores a random bit as mi and rejects.) ii. If hi = 1, the verifier computes the two preimages x0,i and x1,i to yi under fκi using the trapdoor τκi. It then stores βi _di_ (x0,i _x1,i) as the Hadamard basis_ _⊕_ _·_ _⊕_ measurement outcome. (Once again, if either inverse does not exist, the verifier stores a random bit as mi and rejects.) 7. The verifier opens its commitment, and sends the prover its measurement outcomes and func_tion trapdoors. If a test round was chosen in the previous step, the verifier now accepts the_ prover directly, provided that the preimage test was satisfied for all i. (If not all the preimage tests passed, the verifier directly rejects the prover.) If a Hadamard round was chosen, and the verifier instead opens its commitment to rv and sends its measurement outcomes _m = m1, . . ., m2nN to the prover, along with its trapdoors τκ1, . . ., τκ2nN ._ 8. The prover checks that the verifier’s measurement outcomes and trapdoors were consistent; it _then proves that the verifier’s measurement outcomes were consistent. The prover checks that_ _τκ1, . . ., τκ2nN are indeed trapdoors for the functions specified by κ1, . . ., κ2nN_, and also that _κ1, . . ., κ2nN are valid ETCFF keys, using the procedure described in Protocol A.2. It also_ 23 ----- defines u = ui1 _ui2m (each ui is 2N bits long) = mℓ1_ _mℓ4mN, where ℓ1, . . ., ℓ4mN are the_ _· · ·_ _· · ·_ indices of the qubits on which Ur[⊗][2][N] acts nontrivially, and checks that u causes the predicate _Q(t, π, a, b, r, u) to be satisfied. (The Q we refer to here is the Q of Definition 2.19. We define_ _Ur in the same way that Ur was defined in Definition 2.18.) If either of these tests fails, the_ prover aborts. If both tests pass, then the prover proves, using an NP zero-knowledge proof system,[8] that the verifier’s outcomes are consistent in the following sense: The verifier’s outcomes u are consistent if there exists a string sp and an encoding key (t, π, a, b) such that z = commit((π, a, b), sp) and Q(t, π, a, b, r, u) = 1. Figure 1: Diagrammatic representation of an honest execution of Protocol 3.2. We omit communication between the different parts of the prover for neatness, and we also omit the initial messages i (see Appendix C) from executions of the perfectly binding, quantum computationally concealing commitment protocol which we refer to in Assumptions 3.1. The blue parts of the diagram indicate what occurs in the case of a test round, and the red parts indicate what occurs in the case of a Hadamard round. 8It was shown in [Wat09] that the second item in Assumptions 3.1 suffices to guarantee the existence of a proof system for languages in NP that is zero-knowledge against quantum polynomial-time verifiers. Our proof that our protocol is zero-knowledge for classical verifiers only requires that the NP proof system used here is (likewise) zeroknowledge against classical verifiers; however, it becomes necessary to require post-quantum security of this proof system if we want our protocol also to be zero-knowledge for potentially quantum malicious verifiers. 24 ----- Figure 2: Diagrammatic representation of Protocol 3.2 with a cheating verifier. The cheating verifier V _[∗]_ may take some (classical) auxiliary input Z0, store auxiliary information (represented by Z1 and Z2), and produce a final output Z3 that deviates from that specified by the protocol. ### 4 Completeness of protocol **Lemma 4.1. Suppose that the instance x = (H, a, b) of the 2-local XZ Hamiltonian problem that is** _provided as input to the verifier and prover in Protocol 3.2 is a yes-instance, i.e. the ground energy_ _of H is smaller than a. Then, the probability that the honest verifier accepts after an interaction_ _with the honest prover in Protocol 3.2 is 1_ _µ(_ _x_ ), for some negligible function µ. _−_ _|_ _|_ _Proof. The measurement protocol outlined in section 2.3 has the properties that_ 1. for any n-qubit quantum state ρ and for any choice of measurement bases h, the honest prover is accepted by the honest verifier with probability 1 negl(n), and _−_ 2. the distribution of measurement outcomes obtained by the verifier from an execution of the measurement protocol (the measurement outcomes mi in step 6(b) of Protocol 3.2) is negligibly close in total variation distance to the distribution that would have been obtained by performing the appropriate measurements directly on ρ. These properties are stated in Claim 5.3 of [Mah18]. It is evident (assuming the NP zero-knowledge proof system has perfect completeness) that if the verifier of Protocol 3.2 had obtained the outcomes m through direct measurement of ρ, it would accept with exactly the same probability with which the verifier of Protocol 2.6 would accept ρ = σ[⊗][n]. By Claim 2.7, this latter probability is exponentially close to 1. Lemma 4.1 follows. 25 ----- ### 5 Soundness of protocol Let the honest verifier of the argument system in Protocol 3.2 be denoted V, and let an arbitrary quantum polynomial-time prover with which V interacts be denoted P. For this section, we will require notation from Section 5.3 of [Mah18], the proof of Theorem 8.6 of the same paper, and Section 4 of [BJSW16]. We will by and large introduce this notation as we proceed (and some of it has been introduced already in Sections 2.3 and 2.4, the sections containing outlines of the measurement protocol from [Mah18] and the zero-knowledge proof system from [BJSW16]), but the reader should refer to the above works if clarification is necessary. We begin by making some preliminary definitions and proving a claim, from which the soundness of Protocol 3.2 (Lemma 5.6) will naturally follow. Firstly, we introduce some notation from Section 4 of [BJSW16]: **Definition 5.1 (Projection operators Π0 and Π1). Define N as it is defined in Protocol 3.2. Let** _DN[0]_ [be the set of bitstrings][ x][ such that the encoding of][ |][0][⟩] [under the concatenated Steane code of] Protocol 2.13 (or of Protocol 3.2) is [�]x∈DN[0] _N_ [likewise be the set of bitstrings][ x][ such] _[|][x][⟩][, and let][ D][1]_ that the encoding of |1⟩ under the concatenated Steane code is [�]x∈DN[1] _[|][x][⟩][. (See Definition][ 2.15][,]_ and Section A.6 of [BJSW16], for details about the concatenated Steane code. The first condition in Definition 2.15 will provide some motivation for the following definitions of Π0 and Π1.) Define � Π0 = _x∈DN[0]_ � _|x⟩⟨x|,_ Π1 = _x∈DN[1]_ _x_ _x_ _._ _|_ _⟩⟨_ _|_ **Definition 5.2 (Projection operators ∆0 and ∆1). Define N as it is defined in Protocol 3.2. Let** ∆0 and ∆1 be the following projection operators: ∆0 = _[I]_ _[⊗][N][ +][ Z][⊗][N]_ _,_ ∆1 = _[I]_ _[⊗][N][ −]_ _[Z][⊗][N]_ _._ 2 2 ∆0 is the projection onto the space spanned by all even-parity computational basis states, and ∆1 is its equivalent for odd-parity basis states. Note that, since all the codewords in D0 have even parity, and all the codewords in D1 have odd parity, it holds that Π0 ≤ ∆0 and that Π1 ≤ ∆1. **Definition 5.3 (The quantum channel Ξ). Define a quantum channel mapping N qubits to one** qubit as follows: ΞN (σ) = _[⟨][I]_ _[⊗][N]_ _[, σ][⟩][I][ +][ ⟨][X]_ _[⊗][N]_ _[, σ][⟩][X][ +][ ⟨][Y][ ⊗][N]_ _[, σ][⟩][Y][ +][ ⟨][Z][⊗][N]_ _[, σ][⟩][Z]_ _._ 2 Loosely, ΞN can be thought of as a simplification of the decoding operator to the concatenated Steane code that the honest prover applies to its quantum witness in Protocol 2.13 (or in Protocol 3.2). Its adjoint is specified by Ξ[∗]N [(][σ][) =][ ⟨][I, σ][⟩][I] _[⊗][N][ +][ ⟨][X, σ][⟩][X]_ _[⊗][N][ +][ ⟨][Y, σ][⟩][Y][ ⊗][N][ +][ ⟨][Z, σ][⟩][Z][⊗][N]_ _,_ 2 and has the property that Ξ[∗]N [(][|][0][⟩⟨][0][|][) = ∆][0] _[,]_ Ξ[∗]N [(][|][1][⟩⟨][1][|][) = ∆][1] _[,]_ a property which we will shortly use. 26 ----- Let z be prover P’s commitment string from step 2 of Protocol 3.2. Because the commitment protocol is perfectly binding, there exists a unique, well-defined tuple (π, a, b) and a string sp such that z = commit((π, a, b), sp). **Definition 5.4. For notational convenience, we define a quantum procedure M on a 2nN** -qubit state ρ as follows: 1. Apply X _[a]Z[b]_ to ρ, to obtain a state ρ[′]. 2. Apply π[−][1] to each 2N -qubit block in the state ρ[′], to obtain a state ρ[′′]. 3. Discard the last N qubits of each 2N -qubit block in ρ[′′], to obtain a state ρ[′′′]. 4. To each N -qubit block in ρ[′′′], apply the map ΞN . We also define the procedure M[˜] as the application of the first three steps in M, again for notational convenience. Intuitively, we think of M as an inverse to the prover’s encoding procedure E. M may not actually invert the prover’s encoding procedure, if the prover lied about the encoding key that it used when it sent the verifier z = commit((π, a, b), sp); however, this is immaterial. We now prove a claim from which the soundness of Protocol 3.2 will follow. Before we do so, however, we make a remark about notation for clarity. When we write ‘V accepts the distribution _Dξ,h with probability p’ (or similar phrases), we mean that, in [Mah18]’s notation from section 8.2,_ � _vh(1 −_ _p˜h(Dξ,h)) = p._ _h∈{0,1}[2][nN]_ Here, h represents the verifier’s choice of measurement bases, as before; vh is the probability that the honest verifier will select the basis choice h, and 1 − _p˜h(D) is defined, for any distribution D_ over measurement outcomes m 0, 1, as the probability that the honest verifier will accept _∈{_ _}[2][nN]_ a string drawn from D on basis choice h. (When we refer to the latter probability, we assume, following [BJSW16, Section 4], that the prover behaves optimally—in terms of maximising the verifier’s eventual probability of acceptance—after the verifier sends it measurement outcomes at the end of step 6 in Protocol 3.2. For the purposes of the present soundness analysis, therefore, we can imagine that the verifier checks the predicate Q itself after step 6, instead of relying on the prover to prove to it during step 8 that Q is satisfied.) **Claim 5.5. Suppose there exists a quantum state ξ such that the honest verifier V accepts the** _distribution Dξ,h with probability p. Then the state M_ (ξ) is accepted by the verifier of Protocol 2.6 _with probability at least p._ _Proof. Fix a choice of r (see step 3 of Protocol 3.2 for a definition of r). Let Zr be the subset of_ 0, 1 such that the verifier of Protocol 2.6 accepts if and only if the n-bit string that results from _{_ _}[n]_ concatenating the measurement results it obtains in step 4 of said protocol is a member of Zr. It is unimportant to the analysis what Zr actually is; it matters only that it is well-defined. 27 ----- For this choice of r, we can express the probability that the verifier of Protocol 2.6 accepts a state _τ as_ � � � _Ur[∗]_ _[|][z][1][, . . ., z][n][⟩⟨][z][1][, . . ., z][n][|][ U][r][, τ]_ _._ _z∈Zr_ (Though only 2m of the n qubits in τ are relevant to Ur, we assume here for notational simplicity that Ur is a gate on n qubits, and that the verifier measures all n qubits of Urτ and ignores those measurement results which are irrelevant.) For the same choice of r, we can express the probability that the verifier V from Protocol 3.2 will eventually accept the distribution Dξ,h as � _pr =_ _z∈Zr_ � � (Ur[∗][)][⊗][N] [(Π][z]1 _[⊗· · · ⊗]_ [Π][z]n[)(][U][r][)][⊗][N] _[,][ ˜]M_ (ξ) _._ Following [BJSW16], we note that � _z∈Zr_ � � (Ur[∗][)][⊗][N] [(Π][z]1 _[⊗· · · ⊗]_ [Π][z]n[)(][U][r][)][⊗][N] _[,][ ˜]M_ (ξ) � _≤_ _z∈Zr_ � = _z∈Zr_ � = _z∈Zr_ � = _z∈Zr_ � � (Ur[∗][)][⊗][N] [(∆][z]1 _[⊗· · · ⊗]_ [∆][z]n[)(][U][r][)][⊗][N] _[,][ ˜]M_ (ξ) � � � � (Ur[∗][)][⊗][N] Ξ[∗]N [(][|][z][1][⟩⟨][z][1][|][)][ ⊗· · · ⊗] [Ξ][∗]N [(][|][z][n][⟩⟨][z][n][|][)] (Ur)[⊗][N] _,_ _M[˜]_ (ξ) � � (Ξ[⊗]N[n][)][∗][U]r[∗] _M_ (ξ) _[|][z][1][, . . ., z][n][⟩⟨][z][1][, . . ., z][n][|][ U][r][,][ ˜]_ � � _Ur[∗]_ _[|][z][1][, . . ., z][n][⟩⟨][z][1][, . . ., z][n][|][ U][r][, M]_ [(][ξ][)] _._ For the second-to-last equality above, we have used the observation that, for any n-qubit Clifford operation C, and every nN -qubit state σ, Ξ[⊗][n] _N_ [(][C][⊗][N] _[σ][(][C][⊗][N]_ [)][∗][) =][ C][Ξ]N[⊗][n][(][σ][)][C][∗][.] This is equation (35) in [BJSW16], and can be verified directly by considering the definition of ΞN . We conclude that, if the distribution Dξ,h is accepted by V with probability p = [�]r _[v][r][p][r][ =]_ � _h_ _[v][h][(1][ −]_ _[p][˜][h][(][D][ξ,h][)) (where][ v][r][ is the probability that a given][ r][ will be chosen, and the second]_ expression is simply a formulation in alternative notation of the first), the state M (ξ) is accepted by the verifier of Protocol 2.6 with probability at least p. Now we turn to arguing that Protocol 3.2 has a soundness parameter s which is negligibly close to 3 4 [.] 28 ----- **Lemma 5.6. Suppose that the instance x = (H, a, b) of the 2-local XZ Hamiltonian problem that is** _provided as input to the verifier and prover in Protocol 3.2 is a no-instance, i.e. the ground energy_ _of H is larger than b. Then, provided that Assumptions 3.1 hold, the probability that the honest_ _verifier V accepts in Protocol 3.2 after an interaction with any quantum polynomial-time prover P_ _is at most_ [3] 4 [+][ negl][(][|][x][|][)][.] _Proof. Claim 2.12 guarantees that, for any arbitrary quantum polynomial-time prover P who exe-_ cutes the measurement protocol with V, there exists a state ξ, a prover P[′] and a negligible function _µ such that_ _∥DP[C],h_ _[−]_ _[D][P][′][,h][∥][TV]_ _[≤√][p][h,T]_ [+][ p][h,H] [+][ µ,] and _DP′,h_ _c Dξ,h ._ (2) _≈_ (See the paragraph immediately above Claim 2.12 for relevant notation.) It follows from (2) that, if V accepts the distribution DP′,h with probability p, it must accept the distribution Dξ,h with probability p − negl(N ), because the two are computationally indistinguishable and the verifier V is efficient. Therefore (using Claim 5.5), if V accepts DP′,h with probability p, the verifier of Protocol 8.3 from [Mah18] accepts the state M (ξ) with probability at least p negl(N ). _−_ By the soundness of Protocol 2.6 (Claim 2.7), we conclude that p = negl(N ) when the problem Hamiltonian is a no-instance. We now apply a similar argument to that which is used in Section 8.2 of [Mah18] in order to establish an upper bound on the probability φ that V accepts P in a no-instance. Let EP[H],h [denote] the event that the verifier V does not reject the prover labelled P in a Hadamard round indexed by _h during the measurement protocol phase of Protocol 3.2. Let EP[T],h_ [denote the analogous event in a] test round. Furthermore, let EP,h denote the event that the verifier accepts the prover P in the last step of Protocol 3.2. The total probability that V accepts P is the average, over all possible basis choices h, of the probability that V accepts P after a test round indexed by h, plus the probability that V accepts P after a Hadamard round indexed by h. As such, � _φ =_ _vh( 2[1]_ _[Pr][[][E]P[T],h[] + 1]2_ _[Pr][[][E]P[H],h_ _[∩]_ _[E][P][,h][] )]_ _h∈{0,1}[2][nN]_ � = _vh( [1]2_ _[Pr][[][E]P[T],h[] + 1]2_ _[Pr][[][E]P[H],h[]][Pr][[][E][P][,h][|][E]P[H],h[] )]_ _h∈{0,1}[2][nN]_ � = _vh( 2[1]_ [(1][ −] _[p][h,T][ ) + 1]2_ [(1][ −] _[p][h,H]_ [)(1][ −] _[p][˜][h][(][D]P[C],h[)) )][ .]_ _h∈{0,1}[2][nN]_ Since Lemma 3.1 of [Mah18] and Claim 2.12 taken together yield the inequality _p˜h(DP′,h) −_ _p˜h(DP[C],h[)][ ≤∥][D]P[C],h_ _[−]_ _[D][P][,h][∥][TV]_ _[≤√][p][h,T]_ [+][ p][h,H] [+][ µ,] 29 ----- it follows that � _φ ≤_ _vh( [1]_ 2 [(1][ −] _[p][h,T][ ) + 1]2_ [(1][ −] _[p][h,H]_ [)(1][ −] _[p][˜][h][(][D][P][′][,h][) +][ √][p][h,T][ +][ p][h,H][ +][ µ][) )]_ _h∈{0,1}[2][nN]_ � � _≤_ [1] _vh(1 −_ _ph,T + (1 −_ _ph,H_ )(ph,H + _[√]ph,T )) + [1]_ _vh(1 −_ _p˜h(DP′,h))_ 2 _[µ][ + 1]2_ 2 _h∈{0,1}[2][nN]_ _h∈{0,1}[2][nN]_ _≤_ [1] 2 _[µ][ + 3]4 [+ 1]2_ _[p .]_ The upper bound of [3] 4 [in the last line can be obtained by straightforward calculation.][9][ We conclude] that Protocol 3.2 has a soundness parameter s which is negligibly close to [3]4 [.] ### 6 Zero-knowledge property of protocol In this section, we establish that Protocol 3.2 is zero-knowledge against arbitrary classical probabilistic polynomial time (PPT) verifiers. Specifically, we show the following: **Lemma 6.1. Suppose that the instance x = (H, a, b) of the 2-local XZ Hamiltonian problem that is** _provided as input to the verifier and prover in Protocol 3.2 is a yes-instance, i.e. the ground energy_ _of H is smaller than a._ _Then (provided that Assumptions 3.1 hold) there exists a polynomial-_ _time generated PPT simulator S such that, for any arbitrary PPT verifier V_ _[∗], the distribution_ _of V_ _[∗]’s final output after its interaction with the honest prover P in Protocol 3.2 is (classical)_ _computationally indistinguishable from S’s output distribution._ _Remark 6.2. Lemma 6.1 formulates the zero-knowledge property in terms of classical verifiers and_ computational indistinguishability against classical distinguishers, because this is the most natural setting for a protocol in which verifier and interaction are classical. However, the same proof can be adapted to show that, for any quantum polynomial-time verifier executing Protocol 3.2, there exists a quantum polynomial-time generated simulator whose output is QPT indistinguishable in yes-instances from that of the verifier. (In particular, the latter follows from the fact that the second item in Assumptions 3.1 implies an NP proof system which is zero-knowledge against quantum polynomial-time verifiers, an implication shown to be true in [Wat09].) We show that Protocol 3.2 is zero-knowledge by replacing the components of the honest prover with components of a simulator one at a time, and demonstrating that, when the input is a yes-instance, the dishonest verifier’s output after each replacement is made is at the least computationally indistinguishable from its output before. The argument proceeds in two stages. In the first, we show that the honest prover can be replaced by a quantum polynomial-time simulator that does not have access to the witness ρ. In the second, we de-quantise the simulator to show that the entire 9For example, one can obtain this bound by maximising the quantity f (ph,T, ph,H ) = 12 �1−ph,T +(1−ph,H )(ph,H + _√ph,T )[�]_ under the assumption that ph,T and ph,H lie in [0, 1]. The function f has one stationary point (ph,T = 1 9 _[, p][h,H][ =][ 1]3_ [) in [0][,][ 1]][2][; checking][ f][ at this point, in addition to its maxima on each of the boundaries of [0][,][ 1]][2][, reveals] that the choice of (ph,T, ph,H ) ∈ [0, 1][2] which yields the maximum value of f is ( [1]9 _[,][ 1]3_ [), giving][ f][ =] 32 [. Of course,] 2 3 _[<][ 3]4_ [; we use the bound of][ 3]4 [for consistency with [][Mah18][].] 30 ----- execution can be simulated by a classical simulator who likewise does not have access to ρ. (The latter is desirable because the verifier is a classical entity.) We begin with the protocol execution between the honest prover P and an arbitrary cheating verifier V _[∗], the latter of whom may take some (classical) auxiliary input Z0, store information_ (represented by Z1 and Z2), and produce an arbitrary final output Z3. A diagram representing the interaction between V _[∗]_ and P can be found in Figure 2. #### 6.1 Eliminating the coin-flipping protocol Our first step in constructing a simulator is to eliminate the coin-flipping protocol, which is designed to produce a trusted random string r, and replace it with the generation of a truly random string. (This step is entirely analogous to step 1 of Section 5 in [BJSW16], and we omit the analysis.) The new diagram is shown below. In this diagram, coins represents a trusted procedure that samples a uniformly random string r of the appropriate length. #### 6.2 Introducing an intermediary Our next step is to introduce an intermediary, denoted by I, which pretends—to the cheating verifier of Protocol 3.2—to be its prover P, while simultaneously playing the role of verifier to the prover from the zero-knowledge proof system of Protocol 2.17 [10]. (We denote the honest prover and honest verifier for the proof system of Protocol 2.17 by and, respectively, to distinguish them from _P_ _V_ the prover(s) P and verifier(s) V of the classical-verifier protocol currently under consideration.) We remark, for clarity, that I is a quantum polynomial-time procedure. The essential idea of this section is that I will behave so it is impossible for the classical verifier V to tell whether it is interacting with the intermediary or with its honest prover. (We achieve this simply by making I output exactly the same things that P would.) Given that this is so, the map that V implements 10Protocol 2.17 is identical in structure to the protocol presented in [BJSW16]. We refer the reader to Figure 4 in that paper for a diagram representing the appropriate interactions. 31 ----- from its input to its output, including its auxiliary registers, cannot possibly be different in the previous section as compared to this section. Figure 3: The intermediary interacting with the honest prover from the proof system of Protocol 2.17, denoted by P, and also with the cheating classical verifier V _[∗]. I1 receives the encoded quantum witness, which we have_ denoted by Y, from P, in addition to P’s commitment z. It then sends z to V1[∗][, along with][ Z][1][, the auxiliary] input that V1[∗] [is supposed to receive, and][ r][, the random string generated by][ coins][.][ I][2] [passes on any output][ V][ ∗]1 produces to V2[∗][, performs itself the procedure for committing to a quantum state from [][Mah18][], and executes] the measurement protocol with V2[∗][.][ I][3] [receives the measurement outcomes][ u][ and the trapdoors][ τ][ from][ V][ ∗]2 [, and] checks whether the trapdoors are valid. If they are invalid, it aborts directly; if they are valid, it sends u on to _P3 and passes Z2 to V3[∗][, so that][ P][3]_ [and][ V][ ∗]3 [can execute the NP zero-knowledge proof protocol. (Each part of][ I] should also send everything it knows to its successor, but we have omitted these communications for the sake of cleanliness, as we omitted the communication between parts of the prover in previous diagrams.) #### 6.3 Simulating the protocol with a quantum simulator We now note that Figure 3 looks exactly like Figure 4 from [BJSW16], if we consider the intermediary I and the cheating classical verifier V _[∗]_ taken together to be a cheating verifier V _[′]_ for the proof system of Protocol 2.17. 32 ----- Figure 4: Compare to Figure 4 of [BJSW16]. Note that S1 includes the behaviour of an arbitrary V1[′] [; the reason] it is called S1 and not V1[′] [is because][ V]1[′] [obtains][ r][ from a coin-flipping protocol, while][ S][1] [generates][ r][ using][ coins][.] In all other respects, S1 is the same as V1[′] [.] Using similar reasoning as in [BJSW16] (and recalling that, by Lemma 2.20, it still works when the Hamiltonian being verified is an XZ Hamiltonian), therefore, we conclude that we can replace ρ in Figure 4 with ρr—where ρr is a quantum state specifically designed to pass the challenge indexed by r—without affecting the verifier’s output distribution (to within computational indistinguishability). See Remark 2.8 for a procedure that explicitly constructs ρr. Note that, if our objective was to achieve a quantum simulation without knowing the witness state ρ, our task would already be finished at this step. However, our verifier is classical; therefore, in order to prove that our classical verifier’s interaction with its prover does not impart to it any knowledge (apart from the fact that the problem instance is a yes-instance) that it could not have generated itself, we need to achieve a classical simulation of the argument system. #### 6.4 Simulating the protocol with a classical simulator 6.4.1 Replacing P0 and I1 If we want to simulate the situation in Figure 4 classically, then we need to de-quantise P0, I1 and _I2. (I3 and P3 are already classical.) Our first step is to replace P0 and I1 with a single classical_ entity, I1[′] [.] _I1[′]_ [simply chooses encoding keys][ (][t, π, a, b][)][ and generates][ z][, a commitment to the encoding keys] (π, a, b). It then sends z, r and Z1 to V1[∗][, as][ I][1][ would have. Because][ I]1[′] [has exactly the same output] as I1, the verifier’s output in Figure 5 is the same as its output in Figure 4. (We assume that the still-quantum I2 now generates ρr for itself.) 33 ----- Figure 5: P0 and I1 have been replaced by I1[′] [.] 6.4.2 Some simplifications (which make it possible to de-quantise I2) Following [BJSW16], we make some alterations to Figure 5 that will allow us to eventually dequantise I2. The alterations are as follows: 1. Replace V3[∗] [and][ P][3][ with an efficient simulation][ S][3][. (An efficient simulation of the NP proof] protocol execution between V3[∗] [and][ P][3][ is guaranteed to exist because the NP proof protocol] is zero-knowledge.) Recall that the statement P3 is meant to prove to V3[∗] [in a zero-knowledge] way is as follows: ‘There exists a string sp and an encoding key (t, π, a, b) such that z = commit((π, a, b), sp) and Q(t, π, a, b, r, u) = 1.’ The zero-knowledge property of the NP proof system guarantees that, for yes-instances, the output of S3 is indistinguishable from the output of the protocol execution between V3[∗] [and][ P][3][. In our case,][ I]1[′] [always holds][ s][p][ and] (π, a, b) such that z = commit((π, a, b), sp), and the honest prover will abort the protocol if _Q(t, π, a, b, r, u) = 0. Therefore, whenever the prover does not abort, the output of S3 is_ computationally indistinguishable from that of V3[∗] [and][ P][3][. We assume, following [][BJSW16][],] that S3 also behaves as V3[∗] [would when the prover aborts.] If it does, then Figure 6 is computationally indistinguishable from Figure 5. 34 ----- Figure 6: V3[∗] [and][ P][3] [have been replaced by][ S][3][. Note that][ S][3] [does not require access to the witness][ (][s][p][, t, π, a, b][)][,] and so sp can be discarded immediately after I1[′] [is run.] 2. Replace the generation of the genuine commitment z with the generation of a commitment _z[′]_ = commit((π0, a0, b0), sp), where π0, a0 and b0 are fixed strings independent of the encoding key (t, π, a, b) that I1[′] [chooses. Because the commitment protocol is (computationally) con-] cealing, and the commitment is never opened (recall that sp is discarded after I1[′] [is run),][ V][ ∗]1 should not be able to tell (computationally speaking) that z has been replaced by z[′]. The genuine encoding key is still used to evaluate the predicate Q. Note that, because z has been replaced with z[′], the statement for which S3 must simulate the execution of a zero-knowledge proof between V3[∗] [and][ P][3][ is now as follows: ‘There exists a string][ s][p][ and] an encoding key (t, π, a, b) such that z[′] = commit((π, a, b), sp) and Q(t, π, a, b, r, u) = 1.’ This statement is, in general, no longer true, because the commitment protocol is perfectly binding. However, if the predicate Q is still satisfied for the encoding key (t, π, a, b) that I3 sent, then S3 will proceed to generate a transcript for the no-instance that is computationally indistinguishable from a transcript for a yes-instance. If Q is no longer satisfied, then S3 will abort, as before. In effect, therefore, the cheating verifier V _[∗]_ will not be able to tell (up to computational indistinguishability) that z has been replaced by z[′], and that the NP statement being ‘proven’ to it is no longer true. 6.4.3 De-quantising I2 We now replace I2 with a classical entity I2[′] [.] In the process, we require modifications to the behaviour of I3. Knowing r, I2[′] [can calculate for itself what][ ρ][r][ should be, though it cannot physically produce this] state. As we noted in Remark 2.8, ρr is a simple state: it is merely the tensor product of |0⟩ _, |1⟩_ _, |+⟩_ and |−⟩ qubits. Applying the concatenated Steane code to ρr will then result in a tensor product of N -qubit states that look like � � � � � � _x_ _,_ _x_ _,_ _x_ + _x_ _,_ and _x_ _x_ _,_ (3) _|_ _⟩_ _|_ _⟩_ _|_ _⟩_ _|_ _⟩_ _|_ _⟩−_ _|_ _⟩_ _x∈DN[0]_ _x∈DN[1]_ _x∈DN[0]_ _x∈DN[1]_ _x∈DN[0]_ _x∈DN[1]_ 35 ----- after appropriate normalisation. A brief argument will suffice to establish that it is possible to classically simulate standard or Hadamard basis measurements on the qubits in E(ρr). Each qubit of E(ρr) is either an encoding qubit or a trap qubit, up to the application of a random single-qubit Pauli operator. Simulating standard-basis measurements of encoding qubits is classically feasible, because DN[0] [and][ D]N[1] [are] polynomially sized, and the expressions in (3) only involve superpositions over those sets with equal-magnitude coefficients. Simulating standard-basis measurements of trap qubits, which are always initialised either to 0 or +, is trivially feasible. _|_ _⟩_ _|_ _⟩_ To simulate a Hadamard basis measurement, we can take advantage of the transversal properties of the encoding scheme, and apply H before we apply the concatenated Steane code. Denote the application of the concatenated Steane code to ρr by S(ρr). We have that _S(H_ _[⊗][n]ρrH_ _[⊗][n]) = H_ _[⊗][nN]_ _S(ρr)H_ _[⊗][nN]_ by transversality. To simulate a Hadamard basis measurement of E(ρr), we then 1. Apply H _[⊗][n]_ to ρr. This is easy to classically simulate, because ρr is a tensor product of 0 _,_ 1 _,_ + and qubits. _|_ _⟩_ _|_ _⟩_ _|_ _⟩_ _|−⟩_ 2. Apply the concatenated Steane code to H _[⊗][n]ρrH_ _[⊗][n]. Simulating this is classically feasible,_ by the same argument that we used for standard basis measurements, because H _[⊗][n]ρrH_ _[⊗][n]_ is still a tensor product of 0 _,_ 1 _,_ + and qubits. _|_ _⟩_ _|_ _⟩_ _|_ _⟩_ _|−⟩_ 3. Concatenate trap qubits to each N -qubit block in S(H _[⊗][n]ρrH_ _[⊗][n]) = H_ _[⊗][nN]_ _S(ρr)H_ _[⊗][nN]_ . Simulate the application of H to each trap qubit (which is, once again, classically easy to do because each trap qubit is initialised either to 0 or to + ). _|_ _⟩_ _|_ _⟩_ 4. Apply the permutation π to each 2N -tuple. 5. Simulate a standard basis measurement of the result. 6. XOR the string b to the measurement outcome (b was previously the Z-key for the Pauli one-time pad). Having established that it is possible to classically simulate standard and Hadamard basis measurements of the qubits in E(ρr), we now describe the procedure that the classical I2[′] [should follow] for each qubit i in the state E(ρr). 1. During the commitment phase, I2[′] [simulates a standard basis measurement on the][ i][th qubit,] obtains a simulated measurement result βi, and then chooses a uniformly random preimage _xi from the domain of the function specified by κi. It applies the function specified by κi to_ _βi∥xi and sets yi = ηκi(βi∥xi)._ 2. If the verifier requests a test round, I2[′] [sends][ β][i][∥][x][i][ to the verifier. This is exactly what the] quantum prover I2 would send in the case of a test round, so the verifier cannot tell that it is interacting with I2[′] [instead of][ I][2][.] 3. If the verifier requests a Hadamard round, I2[′] [sends a uniformly random string][ s][i][ ∈{][0][,][ 1][}][w][+1] to the verifier, where w is the length of the preimages. In the same situation, the quantum _I2 would have sent Hadamard basis measurements of the w + 1 qubits in the ith committed_ qubit in E(ρ) and its associated preimage register. 36 ----- (a) If hi = 0, the outcomes of these measurements are uniformly distributed and thus indistinguishable from the distribution of strings si reported by I2[′] [.] (b) Let |ψi⟩ be the state of the ith qubit of E(ρ), let x0,i and x1,i be the two preimages to _yi under the function fκi, and let bi be the ith bit of the one-time-pad Z-key b from I1[′]_ [’s] encoding key (t, π, a, b). If hi = 1, the outcomes of I2’s Hadamard basis measurements can be represented as a tuple (βi, di), where di is uniformly random, and _βi = di · (x0,i ⊕_ _x1,i) ⊕_ _bi ⊕_ Meas(H |ψi⟩) . (Meas here denotes a standard basis measurement.) Note that the distribution over (bi, βi, di) which one would obtain by measuring |ψi⟩ in the Hadamard basis, choosing di and bi uniformly at random, and letting _βi = di · (x0,i ⊕_ _x1,i) ⊕_ _bi ⊕_ Meas(H |ψi⟩) is equivalent to the one that one would obtain choosing a uniformly random si, measuring _|ψi⟩_ in the Hadamard basis, calculating _bi = si,1 ⊕_ _di · (x0,i ⊕_ _x1,i) ⊕_ Meas(H |ψi⟩), and finally setting βi = si,1, di = si,2 _si,w+1._ _· · ·_ The former set of actions is equivalent to the set of actions that I2 performs. The latter set of actions is (as we will shortly show) classically feasible provided that we have the verifier’s trapdoors. Note that I2[′] [only needs to send the verifier][ s][i][, and can rely on] its successor I3, who will have access to the verifier’s trapdoors, to calculate the bits bi retroactively. It follows that, given that I3 can produce correct bits bi (we will shortly show that it can), the distribution of strings reported by I _[′]_ is identical to the distribution of outcomes reported by I. Having established that I2[′] [and][ I][2][ are the same from][ V][ ∗]2 [’s perspective (meaning that it must have] the same behaviour that it did in Figure 5 after I2 is replaced with I2[′] [), it remains to ensure that] the choice of the one-time pad Z-key b is consistent with the si that I2[′] [picked. We relegate the task] of making this choice to I3[′] [, our new version of][ I][3][, because it has access to the verifier’s trapdoors] _τ_ . If any of the trapdoors that it receives from the verifier are invalid, or if any of the ETCFF keys κ which the verifier chose are invalid, I3[′] [aborts, as specified in Protocol][ 3.2][. (‘Validity’, here,] means the following: 1) all the κs which the verifier sent earlier well-formed, and 2) for each yi, the trapdoor τκi correctly inverts the function specified by κi. We expand on this notion of ‘validity’ in Appendix A.) Presuming upon valid keys and valid trapdoors, I3[′] [then deduces the verifier’s] choices of measurement basis, h, from τ . Given that the trapdoors are valid and that the keys are well-formed, I3[′] [can be confident that its deductions in this regard will lead it to behave in the same] way that the honest prover would, because (given valid keys and trapdoors) I3[′] [will know exactly] which superpositions the honest prover would have obtained during the measurement protocol after following the verifier’s instructions. For notational convenience, let _ψ[∗]_ denote the state obtained by applying the first three steps of _|_ _⟩_ _E, but not the last step, to ρr. I3[′]_ [subsequently executes the following procedure for all][ i][ such that] _hi = 1:_ 37 ----- 1. Set di to be the last w bits of si, and compute di (x0,i _x1,i) using the trapdoor τκi._ _·_ _⊕_ 2. Simulate a standard basis measurement of HX _[a][i]_ _|ψi[∗][⟩][. Denote the result by][ β][i][. (Here,][ a][i]_ refers to the ith bit of a, where a is taken from I1[′] [’s initial choice of one-time pad keys, and] _|ψi[∗][⟩]_ [denotes the][ i][th qubit of][ |][ψ][∗][⟩][.)] 3. Set b[′]i [(the][ i][th bit of][ b][′][, the new][ Z][-key for the one-time pad) to be equal to][ β][i] _[⊕][s][i,][1]_ _[⊕][d][i]_ _[·][(][x][0][,i]_ _[⊕]_ _x1,i) (where si,1 refers to the first bit of si). This will cause the equation Meas(H |ψi⟩) ⊕_ _di ·_ (x0,i _x1,i) = si,1 to be satisfied:_ _⊕_ Meas(H |ψi⟩) ⊕ _di · (x0,i ⊕_ _x1,i) = si,1_ _⇐⇒_ Meas(HZ[b][i]X _[a][i]_ _|ψi[∗][⟩][)][ ⊕]_ _[d][i]_ _[·][ (][x][0][,i]_ _[⊕]_ _[x][1][,i][) =][ s][i,][1]_ _⇐⇒_ _bi ⊕_ Meas(HX _[a][i]_ _|ψi[∗][⟩][)][ ⊕]_ _[d][i]_ _[·][ (][x][0][,i]_ _[⊕]_ _[x][1][,i][) =][ s][i,][1]_ _⇐⇒_ _bi ⊕_ _βi ⊕_ _di · (x0,i ⊕_ _x1,i) = si,1_ _⇐⇒_ _bi = βi ⊕_ _si,1 ⊕_ _di · (x0,i ⊕_ _x1,i)._ Having done this, I3[′] [then feeds][ (][t, π, a, b][′][)][ into][ Q][. (Note that replacing][ b][ with][ b][′][ cannot create] any conflict with the commitment string z[′] that the verifier will notice, because z[′] was already independent of the one-time-pad keys (a, b).) In all other respects I3[′] [behaves the same way that][ I][3] did. The final simulation will be as follows: Since all the entities in this simulation are classical and efficient, and none have access to information about the witness state ρ, it follows that the protocol is zero-knowledge. ### A LWE-based ETCFF family and efficient trapdoor check In order to explain the trapdoor check that the honest prover of Protocol 3.2 implements during step 8 of Protocol 3.2, we briefly outline, at a level of detail appropriate for us, how the LWE-based ETCFF family that is used in [Mah18] is constructed. 38 ----- We begin by introducing the instantiations of the keys κ and the trapdoors τ for noisy trapdoor claw-free (f ) and trapdoor injective (g) functions, whose properties we have relied upon in a blackbox way for the rest of this work. For details, we refer the reader to Section 9 of [Mah18]. The key (κ1, κ2) for an noisy two-to-one function f is (A, As + e), where A is a matrix in Z[n]q _[×][m]_ and e ∈ Z[n]q [is an error vector such that][ |][e][|][ < B][f] [for some small upper bound][ B][f] [. (The specific] properties that Bf should satisfy will be described later.) Here, n, m are integers, and q is a prime power modulus that should be chosen as explained in [Mah18]. In addition, in order to implement the trapdoor, we assume that the matrix A is generated using the efficient algorithm GenTrap which is described in Algorithm 1 in [MP11]. (For convenience, we use the ‘statistical instantiation’ of the procedure described in Section 5.2 of the paper.) GenTrap produces a matrix A that has the form A = [A|HG − _AR], for some publicly known G ∈_ Z[m]q _[×][w], n ≤_ _w ≤_ _m, some A ∈_ Zq[n][×][(][m][−][w][)], some invertible matrix H ∈ Z[n]q _[×][n], and some R ∈_ Z[(][m][−][w][)][×][w], where R = τA is the trapdoor to A. As shown in [MP11, Theorem 5.4], it is straightforward, given the matrix R, to verify that R is a ‘valid’ trapdoor, in the sense that it allows a secret vector s to be recovered from a tuple of the form (A, b = As + e) with certainty when e has magnitude smaller than some bound BInvert. Checking that R is a valid trapdoor involves computing the largest singular value of R and checking that A is indeed of the form A = [A _HG_ _AR] for some invertible H and for the publicly known G. Using_ _|_ _−_ any valid trapdoor, recovery can be performed via an algorithm Invert described in [MP11]. The key for an injective function g, meanwhile, is (A, u), where u is a random vector not of the form _As + e for any e of small enough magnitude. (Again, ‘small enough’ here refers to a specific upper_ bound, and what the bound is precisely will be described later. The distribution of u is uniform over all vectors that satisfy this latter requirement.) The trapdoor τA is still the R corresponding to the matrix A which is described in the preceding paragraph. The functions fκ and gκ both take as input a bit b and a vector x and output a probability distribution (to be more precise, a truncated Gaussian distribution of the kind defined in Section 2.3, equation 4 of [Mah18]). We clarify that, when we say that the functions output a probability distribution, we mean that they should be thought of as maps from the space of strings to the set of probability distributions, not that their outputs are randomised. Given a sample y from one such probability distribution Y, the trapdoor τA can be used to recover the tuple(s) (b, x) which are preimages of y under the function specified by κ. (See Definition 2.11 for a definition of the phrase ‘preimage of _y’.) The functions fκ and gκ can be defined (using notation explained in the paragraph below the_ definition) as follows: **Definition A.1 (Definition of trapdoor claw-free and trapdoor injective functions).** **(a)** _fκ(b, x) = Ax + e0 + b · (As + e),_ where e0 is distributed as a truncated Gaussian with bounded magnitude |e0|max **(b)** _gκ(b, x) = Ax + e0 + b · u ._ What the above notation means is that one samples from the distribution determined by the input (b, x) and the function key κ = (κ1, κ2) by sampling e0 from a truncated Gaussian centred at the origin and then computing κ1x + e0 + b · (κ2). A key feature of the f functions is that the output distributions given by fκ(0, x) and fκ(1, x − _s) are truncated Gaussians which overlap to a_ high degree (so that the statistical distance between the distributions fκ(0, x) and fκ(1, x − _s) is_ 39 ----- negligible). The g functions, meanwhile, are truly injective in the sense that g(b, x) and g(b[′], x[′]) never overlap for (b, x) = (b[′], x[′]). In order that these two things are true, we require that the e in _̸_ Definition A.1(a) is very small (Bf ≪|e0|max), and that the u in Definition A.1(b) is such that _u ̸= As + e for any |e| < Bg, where Bg > |e0|max. It follows from hardness of the (decisional) LWE_ assumption that the keys for the f functions and the keys for the g functions are computationally indistinguishable. The trapdoor check that the prover of protocol 3.2 executes in step 8 is as follows: **Protocol A.2 (Trapdoor and key check).** Let κi = (Ai, Aisi + ei). (Note that |ei| need not be smaller than any particular bound in this definition of κi.) For all i ∈{1, . . ., 2nN _}:_ 1. Check that τκi is a ‘valid’ trapdoor for Ai, in the sense that was explained in the third paragraph of this appendix. If it is not, abort. 2. For a choice of Bf, Bg and |e0|max such that Bf ≪|e0|max < Bg ≤ _BInvert and Bg −|e0|max >_ _|e0|max, check that one of the following three conditions hold:_ (a) Invert applied to κ2,i succeeds, and recovers an e such that |e| < Bf, or (b) Invert applied to κ2,i succeeds, and recovers an e such that Bg < |e|, or (c) Invert fails. Figure 7: Diagram illustrating one possible choice of parameters that satisfies the conditions in 2. above. When a circle is labelled with a number (such as Bf or Bg), the radius of the circle represents the size of that number. The conditions in step 2 above are intended to ensure that, for all i, κi is a key either for an f or a g function, and therefore well-formed. An ill-formed key would be of the form κbad = As + e 40 ----- for Bf < e < Bg; for some choices of Bf and Bg, a subset of κbads as just defined would behave neither like keys for f functions nor like keys for g functions, because the distributions ηκbad(0, x) and ηκbad(1, x−s) would overlap but not to a sufficient degree. The specifications on the parameters that are made in step 2 above, and the tests prescribed for the prover, are designed to ensure that _Bf and Bg are properly chosen and that the prover can check efficiently that the verifier’s κs are_ well-formed according to these appropriate choices of Bf and Bg. The following claim shows that, given a valid trapdoor (i.e. a matrix R that satisfies the efficiently verifiable conditions described in the third paragraph of this appendix), and a well-formed key _κ, the trapdoor can be used to successfully recover all the preimages under a function ηκ to any_ sample y from a distribution Y in the range of the function ηκ. This claim is needed to justify the correctness of the “de-quantized” simulator I2[′] [considered in Section][ 6.4.3][: if][ I]2[′] [can be sure that] it has recovered all the preimages to y, and no others, then it can successfully simulate the honest prover. **Claim A.3. Let A be a matrix in Z[n]q** _[×][m], let κ = (A, κ2), and let the function ηκ be defined by_ _ηκ(b, x) = Ax + e0 + b · κ2. (The output of ηκ is, as in Definition A.1, a probability distribution.)_ _Let τA be a purported trapdoor for A. Suppose that κ passes the test in step 2 of Protocol A.2, and_ _suppose that the trapdoor τA inverts the matrix A, in the sense that, given r = As + e for some e_ _of sufficiently small magnitude, τA can be used to recover the unique (s, e) such that As + e = r._ _Then one can use τA to efficiently recover all the preimages to any y sampled from any distribution_ _Y in the range of ηκ._ _Proof. By hypothesis, κ2 is either of the form As + e for some (s, e) (with |e| < Bf_ ), or it is not of the form As + e for any e such that |e| < Bg. We do not know a priori which of these is the case, but the procedure that we perform in order to recover the preimage(s) to y is the same in both cases: 1. Use the trapdoor τA to attempt to find (x1, e1) such that Ax1 + e1 = y. If such an (x1, e1) exists, and |e1| < |e0|max, record 0∥x1 as the first preimage. 2. Use the trapdoor τA to attempt to find (x2, e2) such that Ax2 + e2 = y − _κ2. If such an_ (x2, e2) exists, and |e2 < |e0|max, record 1∥x2 as the second preimage. If κ2 = As + e for some s and e such that |e| ≪ _Bf < |e0|max, then this procedure will return two_ preimages (except with negligible probability, which happens when y comes from the negligiblysized part of the support of a distribution fκ(b, x) which is not in the support of the distribution _fκ(¬b, x_ +(−1)[b]s); this can occur if y is a sample such that |e0| + _|e| > |e0|max, using notation from_ Definition A.1). Assuming that the latter is not the case, in step 1, the algorithm above will recover _x such that y = Ax + e0 for some e0 < |e0|max, because (under our assumption, and by linearity)_ _y is always of the form Ax + e0. In step 2, it will recover x[′]_ = x − _s, because x[′]_ = x − _s will satisfy_ the equation y − (As + e) = Ax[′] + e[′] for e[′] = e0 − _e, and |e0 −_ _e| < |e0| + |e| < |e0|max. We know_ that y has two preimages under our assumption, so we conclude that, when our assumption holds, the algorithm returns all of the preimages to y under ηκ and no others. In the negligible fraction of cases when y has only one preimage even though κ2 = As + e, the algorithm returns one preimage, which is also the correct number. 41 ----- It can be seen by similar reasoning that, when κ2 = u for u ̸= As + e for any e such that |e| < Bg, this procedure will return exactly one preimage, which is what we expect when κ2 = u. In the context of Protocol 3.2, the honest prover knows that ηκi has been evaluated correctly for all i, because the prover evaluated these functions for itself. Therefore, given Claim A.3, if our goal is to show that the honest prover can efficiently determine whether or not a purported trapdoor _τA[′]_ _i_ [can be used to recover all the preimages to][ y][i][ under][ η][κ]i[, with][ κ][i][ = (][A][i][, κ][2][,i][)][, it is sufficient] to show that a procedure exists to efficiently determine whether or not τA[′] _i_ [truly ‘inverts][ A][i][’, i.e.] recovers (s, e) correctly from all possible r = Ais + e with e having sufficiently small magnitude. This procedure exists in the form of Invert from [MP11]. ### B Completeness and soundness of Protocol 2.6 For notational convenience, define α = _a_ _b_ � � _s_ [2][|][d][s][|][ and][ β][ =] _s_ [2][|][d][s][|] [. Fix an arbitrary state][ ρ][ sent] by the prover. For j = 1, . . ., m let Xj be a Bernoulli random variable that is 1 if the j-th measurement from step 4 of Protocol 2.6 yields −sign(dj) and 0 otherwise. Let X = [�]j[m]=1 _[X][j][ and]_ _Bj = E[X|Xj, . . ., X1]. Then (B1, . . ., Bm) is a martingale. Applying Azuma’s inequality, for any_ _t_ 0 _≥_ � � Pr _X_ E[X] _t_ _e[−]_ 2[t]m[2] _._ _|_ _−_ _| ≥_ _≤_ In the case of an instance x /∈ _L, as mentioned in the main text E[Xj] ≤_ 21 _[−]_ _[β][.]_ Choosing _t =_ [1]2 _[m][(][β][ −]_ _[α][)][, it follows that in this case]_ � � Pr _X_ 2e[−][m][(][β][−][α][)][2][/][8] _._ _≤_ [1] _≤_ 2 _[m][(1][ −]_ _[β][ −]_ _[α][)]_ Since β _α is inverse polynomial, by [MNS16], the right-hand side can be made exponentially_ _−_ small by choosing m to be a sufficiently large constant times (β−|xα| )[2][ . The soundness of Protocol] 2.6 follows. Completeness follows immediately from a similar computation using the Chernoff bound, since in this case we can assume that the witness provided by the prover is in tensor product form. ### C Commitment scheme We provide an informal description of a generic form for a particular (and commonly seen) kind of commitment scheme. The protocol for making a commitment under this scheme requires three messages in total between the party making the commitment, whom we refer to as the committer, and the party receiving the commitment, whom we call the recipient. The first message is an initial message i from the recipient to the committer; the second is the commitment which the committer sends to the recipient; and the third message is a reveal message from the committer 42 ----- to the recipient. The scheme consists of a tuple of algorithms (gen, initiate, commit, reveal, verify) defined as follows: gen(1[ℓ]) takes as input a security parameter, and generates a public key pk. _•_ initiate(pk) takes as input a public key and generates an initial message i (which the recipient _•_ should send to the committer). commit(pk, i, m, s) takes as input a public key pk, an initial message i, a message m to which _•_ to commit, and a random string s, and produces a commitment string z. reveal(pk, i, z, m, s) outputs the inputs it is given. _•_ verify(pk, i, z, m, s) takes as argument an initial message i, along with a purported public key, _•_ commitment string, committed message and random string, evaluates commit(pk, i, m, s), and outputs 1 if and only if z = commit(pk, i, m, s). For brevity, we sometimes omit the public key pk and the initial message i as arguments in the body of the paper. The commitment schemes which we assume to exist in the paper have the following security properties: _Perfectly binding: If commit(pk, i, m, s) = commit(pk, i, m[′], s[′]), then (m, s) = (m[′], s[′])._ _•_ _(Quantum) computationally concealing: For any public key pk_ gen(1[ℓ]), fixed initial mes _•_ _←_ sage i, and any two messages m, m[′], the distributions over s of commit(pk, i, m, s) and commit(pk, i, m[′], s) are quantum computationally indistinguishable. It is known that a commitment scheme with the above form and security properties exists assuming the quantum hardness of LWE: see Section 2.4.2 of [CVZ19]. The commitment scheme outlined in that work is analysed in the common reference string (CRS) model, but the analysis can easily be adapted to the standard model when an initial message i is allowed to pass from the recipient to the committer. ### References [ABE10] Dorit Aharonov, Michael Ben-Or, and Elad Eban. Interactive proofs for quantum computations. In Andrew Chi-Chih Yao, editor, Innovations in Computer Science _- ICS 2010, Tsinghua University, Beijing, China, January 5-7, 2010. Proceedings,_ [pages 453–469. Tsinghua University Press, 2010. URL http://conference.itcs.](http://conference.itcs.tsinghua.edu.cn/ICS2010/content/papers/35.html) ``` tsinghua.edu.cn/ICS2010/content/papers/35.html. ``` [ABOEM17] Dorit Aharonov, Michael Ben-Or, Elad Eban, and Urmila Mahadev. Interactive proofs for quantum computations. arXiv preprint arXiv:1704.04487, 2017. [ACGH19] Gorjan Alagic, Andrew M. Childs, Alex B. Grilo, and Shih-Han Hung. Non-interactive classical verification of quantum computation. arXiv e-prints, page arXiv:1911.08101, [November 2019, 1911.08101.](http://arxiv.org/abs/1911.08101) [BCC88] Gilles Brassard, David Chaum, and Claude Cr´epeau. Minimum disclosure proofs of [knowledge. J. Comput. Syst. Sci., 37(2):156–189, October 1988. doi:10.1016/0022-](http://dx.doi.org/10.1016/0022-0000(88)90005-0) [0000(88)90005-0.](http://dx.doi.org/10.1016/0022-0000(88)90005-0) 43 ----- [BFK09] Anne Broadbent, Joseph Fitzsimons, and Elham Kashefi. Universal blind quantum computation. In Foundations of Computer Science, 2009. FOCS’09. 50th Annual _[IEEE Symposium on, pages 517–526. IEEE, 2009. doi:10.1109/focs.2009.36.](http://dx.doi.org/10.1109/focs.2009.36)_ [BG19] Anne Broadbent and Alex B Grilo. Zero-knowledge for QMA from locally simulatable proofs. arXiv preprint arXiv:1911.07782, 2019. [BJSW16] Anne Broadbent, Zhengfeng Ji, Fang Song, and John Watrous. Zero-knowledge proof systems for QMA. In Foundations of Computer Science (FOCS), 2016 IEEE 57th _[Annual Symposium on, pages 31–40. IEEE, 2016. doi:10.1109/focs.2016.13.](http://dx.doi.org/10.1109/focs.2016.13)_ [BL08] Jacob D. Biamonte and Peter J. Love. Realizable Hamiltonians for universal adiabatic quantum computers. _[Physical Review A, 78:012352, July 2008, 0704.1287.](http://arxiv.org/abs/0704.1287)_ [doi:10.1103/physreva.78.012352.](http://dx.doi.org/10.1103/physreva.78.012352) [BOGG[+]88] Michael Ben-Or, Oded Goldreich, Shafi Goldwasser, Johan H˚astad, Joe Kilian, Silvio Micali, and Phillip Rogaway. Everything provable is provable in zero-knowledge. [volume 403, pages 37–56, 08 1988. doi:10.1007/0-387-34799-2 4.](http://dx.doi.org/10.1007/0-387-34799-2_4) [BS19] Nir Bitansky and Omri Shmueli. Post-quantum zero knowledge in constant rounds. _arXiv preprint arXiv:1912.04769, 2019._ [CVZ19] Andrea Coladangelo, Thomas Vidick, and Tina Zhang. Non-interactive zeroknowledge arguments for QMA, with preprocessing. arXiv preprint arXiv:1911.07546, 2019. [FK17] Joseph F Fitzsimons and Elham Kashefi. Unconditionally verifiable blind quantum [computation. Physical Review A, 96(1):012303, 2017. doi:10.1103/physreva.96.012303.](http://dx.doi.org/10.1103/physreva.96.012303) [GMR89] Shafi Goldwasser, Silvio Micali, and Charles Rackoff. The knowledge complexity of interactive proof systems. _SIAM Journal on computing, 18(1):186–208, 1989._ [doi:10.1137/0218012.](http://dx.doi.org/10.1137/0218012) [GMW91] Oded Goldreich, Silvio Micali, and Avi Wigderson. Proofs that yield nothing but their validity or all languages in np have zero-knowledge proof systems. J. ACM, [38(3):690–728, July 1991. doi:10.1145/116825.116852.](http://dx.doi.org/10.1145/116825.116852) [KSVV02] Alexei Yu Kitaev, Alexander Shen, Mikhail N Vyalyi, and Mikhail N Vyalyi. Clas_sical and quantum computation._ Number 47. American Mathematical Soc., 2002. [doi:10.1090/gsm/047/10.](http://dx.doi.org/10.1090/gsm/047/10) [Mah18] Urmila Mahadev. Classical verification of quantum computations. In Foundations of _Computer Science (FOCS), 2018 IEEE 59th Annual Symposium on, pages 259–267,_ [Oct 2018. doi:10.1109/focs.2018.00033.](http://dx.doi.org/10.1109/focs.2018.00033) [MF16] Tomoyuki Morimae and Joseph F. Fitzsimons. Post hoc verification with a single prover. _arXiv_ _e-prints,_ March 2016, [1603.06046.](http://arxiv.org/abs/1603.06046) https://arxiv.org/pdf/1603.06046.pdf. [MNS16] Tomoyuki Morimae, Daniel Nagaj, and Norbert Schuch. Quantum proofs can be verified using only single-qubit measurements. Physical Review A, 93:022326, February [2016, 1510.06789. doi:10.1103/physreva.93.022326.](http://arxiv.org/abs/1510.06789) 44 ----- [MP11] Daniele Micciancio and Chris Peikert. Trapdoors for lattices: Simpler, tighter, faster, [smaller. Cryptology ePrint Archive, Report 2011/501, 2011. doi:10.1007/978-3-642-](http://dx.doi.org/10.1007/978-3-642-29011-4_41) [29011-4 41. https://eprint.iacr.org/2011/501.](http://dx.doi.org/10.1007/978-3-642-29011-4_41) [MW05] Chris Marriott and John Watrous. Quantum arthur–merlin games. Computational _[Complexity, 14(2):122–152, 2005. doi:10.1109/ccc.2004.1313850.](http://dx.doi.org/10.1109/ccc.2004.1313850)_ [Reg09] Oded Regev. On lattices, learning with errors, random linear codes, and cryptography. _[Journal of the ACM (JACM), 56(6):34, 2009. doi:10.1145/1568318.1568324.](http://dx.doi.org/10.1145/1568318.1568324)_ [RUV13] Ben W Reichardt, Falk Unger, and Umesh Vazirani. Classical command of quantum [systems. Nature, 496(7446):456, 2013. doi:10.1038/nature12035.](http://dx.doi.org/10.1038/nature12035) [Wat09] John Watrous. Zero-knowledge against quantum attacks. SIAM Journal on Comput_[ing, 39(1):25–58, 2009. doi:10.1137/060670997.](http://dx.doi.org/10.1137/060670997)_ 45 -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1902.05217, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://quantum-journal.org/papers/q-2020-05-14-266/pdf/" }
2,019
[ "JournalArticle" ]
true
2019-02-14T00:00:00
[ { "paperId": "10ec524f65c7ced123c8792febcec674feb3fcb2", "title": "Post-quantum zero knowledge in constant rounds" }, { "paperId": "fb78e69a942f8577b5a66e724c2a1d5a3a5b647f", "title": "Non-interactive classical verification of quantum computation" }, { "paperId": "bcca6d6326b1ac47e369d892fc6ba4798d757bf3", "title": "Non-interactive zero-knowledge arguments for QMA, with preprocessing" }, { "paperId": "84b62ec59c5574c516f4c63d2fff1dbaf110815a", "title": "Zero-Knowledge for QMA from Locally Simulatable Proofs" }, { "paperId": "8725345b364702c423fe28035087119b74bcffed", "title": "Classical Verification of Quantum Computations" }, { "paperId": "7371f5feb30bdf921c59262b8bb33217035d47c9", "title": "Zero-Knowledge Proof Systems for QMA" }, { "paperId": "2ca2430bef72150f81cd346f297807c6231d3461", "title": "Post hoc verification of quantum computation" }, { "paperId": "4db7200bdb3120fb462dff20293e1a83078fbbd3", "title": "Quantum proofs can be verified using only single qubit measurements" }, { "paperId": "dbc542602606daaa3de5d597e52da9b00af442cf", "title": "A Note on Quantum Security for Post-Quantum Cryptography" }, { "paperId": "b94e81164f5a5b9113f95fb6e684c5c1c5ed9cc1", "title": "Classical command of quantum systems" }, { "paperId": "07fed20167b9c14492e2ed1c42add42fd8593074", "title": "How to Construct Quantum Random Functions" }, { "paperId": "988ba838c2a4d601a5af5c06f73ee16bb32395ea", "title": "Trapdoors for Lattices: Simpler, Tighter, Faster, Smaller" }, { "paperId": "34b614f6f4156837a0179a5bd56d2a8159f4c057", "title": "Unconditionally verifiable blind quantum computation" }, { "paperId": "d47cc092877ba3c891bb013c68f469e80b770f3b", "title": "On lattices, learning with errors, random linear codes, and cryptography" }, { "paperId": "1fcad18c84c9a414b66137aab3e74839431e1518", "title": "Interactive Proofs For Quantum Computations" }, { "paperId": "ace672c3dddd5072e44cbef025e5739539620041", "title": "Universal Blind Quantum Computation" }, { "paperId": "0a6c31aa44fc829e59743c58cc4ca249ea2fd495", "title": "Realizable Hamiltonians for Universal Adiabatic Quantum Computers" }, { "paperId": "c3d1262160d759f424a83aae1d053b59d3352e7d", "title": "Zero-knowledge against quantum attacks" }, { "paperId": "db5e4a46def6dc05b47e594767c86e9cefe8d2df", "title": "Quantum Arthur–Merlin games" }, { "paperId": "68141fa85942b7adc3b5c58de6a93ba5e4aa994c", "title": "Correspondence between classical and quantum computation" }, { "paperId": "6bcd2606c8f0099a82bd03e520659058a5b12e51", "title": "Proofs that yield nothing but their validity or all languages in NP have zero-knowledge proof systems" }, { "paperId": "c6121161510bd96e4103319feb5f7c355f85beb9", "title": "Everything Provable is Provable in Zero-Knowledge" }, { "paperId": "9ccf7b6cb32cf89752a35bd910555adac54773e0", "title": "The Knowledge Complexity of Interactive Proof Systems" }, { "paperId": "2b1056b47daa0089e877da015910d279a12a6037", "title": "Minimum Disclosure Proofs of Knowledge" }, { "paperId": "ef1d731f28a649a5503c58630b3e26499e861f25", "title": "M ar 2 01 6 Post hoc verification with a single prover" }, { "paperId": null, "title": "new Z-key for the one-time pad) to be equal to β i ⊕s i,1 ⊕d i ·(x 0,i ⊕ x 1,i ) (where s i,1 refers to the first bit of s i )" }, { "paperId": null, "title": "Set d i to be the last w bits of s i , and compute d i · (x 0,i ⊕ x 1,i ) using the trapdoor τ κ i" }, { "paperId": null, "title": "Simulate a standard basis measurement of HX a i |ψ * i . Denote the result by β i . (Here, a i refers to the ith bit of a, where a is taken from I 1 's initial choice of one-time pad keys" }, { "paperId": null, "title": "a Hadamard round is chosen" }, { "paperId": null, "title": "that an entirely classical verifier can decide languages in BQP by interacting classically with two entangled, non-communicating QPT provers. This protocol is likewise sound against arbitrary provers" }, { "paperId": null, "title": "The verifier measures H s j for j = 1 , . . . , m , taking advantage of the fact that—if the prover is honest—it is given m copies of σ ." }, { "paperId": null, "title": "The verifier opens its commitment to r v , and also sends the prover its measurement outcomes u and function trapdoors from the previous step" }, { "paperId": null, "title": "an entirely classical verifier can decide languages in BQP by executing an argument system ([BCC88]) with a single BQP prover." }, { "paperId": null, "title": "The prover sends a state ρ to the verifier one qubit at a time" }, { "paperId": null, "title": "Prover’s encoding step: The same as the prover’s encoding step in Protocol 2.11, except that t ∈ { 0 , + } N rather than { 0 , + , + y } N" }, { "paperId": null, "title": "Verifier’s challenge" }, { "paperId": null, "title": "HZ b i X a i |ψ * i ) ⊕ d i · (x 0,i ⊕ x 1,i ) = s i" }, { "paperId": null, "title": "If the verifier requests a Hadamard round, I (cid:48)" }, { "paperId": null, "title": "decide which random terms from the Hamiltonian H the verifier will check by executing a coin-flipping protocol" } ]
38,339
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Mathematics", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00f0072a48291ce27d5dfe10e00847d98a915d76
[ "Computer Science" ]
0.846005
Conservative Linear Unbiased Estimation Under Partially Known Covariances
00f0072a48291ce27d5dfe10e00847d98a915d76
IEEE Transactions on Signal Processing
[ { "authorId": "15981931", "name": "Robin Forsling" }, { "authorId": "145298509", "name": "A. Hansson" }, { "authorId": "143710447", "name": "F. Gustafsson" }, { "authorId": "1850451", "name": "Zoran Sjanic" }, { "authorId": "2500955", "name": "Johan Löfberg" }, { "authorId": "2486923", "name": "Gustaf Hendeby" } ]
{ "alternate_issns": null, "alternate_names": [ "IEEE Trans Signal Process" ], "alternate_urls": [ "http://www.signalprocessingsociety.org/publications/periodicals/tsp/", "https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=78" ], "id": "1f6f3f05-6a23-42f0-8d31-98ab8089c1f2", "issn": "1053-587X", "name": "IEEE Transactions on Signal Processing", "type": "journal", "url": "http://ieeexplore.ieee.org/servlet/opac?punumber=78" }
Mean square error optimal estimation requires the full correlation structure to be available. Unfortunately, it is not always possible to maintain full knowledge about the correlations. One example is decentralized data fusion where the cross-correlations between estimates are unknown, partly due to information sharing. To avoid underestimating the covariance of an estimate in such situations, conservative estimation is one option. In this paper the conservative linear unbiased estimator is formalized including optimality criteria. Fundamental bounds of the optimal conservative linear unbiased estimator are derived. A main contribution is a general approach for computing the proposed estimator based on robust optimization. Furthermore, it is shown that several existing estimation algorithms are special cases of the optimal conservative linear unbiased estimator. An evaluation verifies the theoretical considerations and shows that the optimization based approach performs better than existing conservative estimation methods in certain cases.
### Conservative Linear Unbiased Estimation Under Partially Known Covariances #### Forsling Robin, Anders Hansson, Fredrik Gustafsson, Zoran Sjanic, Johan Löfberg and Gustaf Hendeby The self-archived postprint version of this journal article is available at Linköping University Institutional Repository (DiVA): http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-187807 N.B.: When citing this work, cite the original publication. Robin, F., Hansson, A., Gustafsson, F., Sjanic, Z., Löfberg, J., Hendeby, G., (2022), Conservative Linear Unbiased Estimation Under Partially Known Covariances, IEEE Transactions on Signal _Processing, 70, 3123-3135. https://doi.org/10.1109/tsp.2022.3179841_ #### Original publication available at: https://doi.org/10.1109/tsp.2022.3179841 Copyright: Institute of Electrical and Electronics Engineers http://www.ieee.org/index.html ©2022 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE. ----- ## Conservative Linear Unbiased Estimation Under Partially Known Covariances ##### Robin Forsling[∗], Anders Hansson[†], Fredrik Gustafsson[‡], Zoran Sjanic, Johan Löfberg[†], and Gustaf Hendeby[†] _∗_ Student Member, IEEE, † Senior Member, IEEE, ‡ Fellow, IEEE ##### Dept. of Electrical Engineering, Linköping University, Linköping, Sweden **_Abstract—Mean square error optimal estimation requires the_** **full correlation structure to be available. Unfortunately, it is** **not always possible to maintain full knowledge about the cor-** **relations. One example is decentralized data fusion where the** **cross-correlations between estimates are unknown, partly due to** **information sharing. To avoid underestimating the covariance** **of an estimate in such situations, conservative estimation is one** **option. In this paper the conservative linear unbiased estimator is** **formalized including optimality criteria. Fundamental bounds of** **the optimal conservative linear unbiased estimator are derived.** **A main contribution is a general approach for computing the** **proposed estimator based on robust optimization. Furthermore,** **it is shown that several existing estimation algorithms are special** **cases of the optimal conservative linear unbiased estimator. An** **evaluation verifies the theoretical considerations and shows that** **the optimization based approach performs better than existing** **conservative estimation methods in certain cases.** **_Index Terms—Conservative estimation, robust optimization,_** **unknown cross-correlations, covariance intersection, decentral-** **ized estimation.** I. INTRODUCTION PTIMAL ESTIMATION of parameters in a linear regression is a well studied subject. Minimum variance unbiased # O estimators such as the best linear unbiased estimator (BLUE) require full knowledge about the measurement covariance [1]. If the covariance structure is only partially known one solution is to use a conservative estimator that does not provide a too optimistic uncertainty. That is, a conservative estimator guarantees that the covariance of an estimate is not underestimated. In the literature this property is often denoted _consistency or covariance consistency [2]. Estimation in a_ regression with unknown or partially known covariances goes at least as far back as [3]. Real-world examples of when the covariance structure is only partially known are found in [4, 5]. A comprehensive survey of estimation under unknown crosscorrelations is provided in [6]. The conservative estimation problem has earlier been studied in a fusion context only, see, e.g., [7–12]. Fusion denotes the estimation problem where multiple estimates of the same parameters are merged into an improved estimate [13]. This can be seen as a special case of a general linear regression This work has been supported by the Industry Excellence Center LINKSIC funded by The Swedish Governmental Agency for Innovation Systems (VINNOVA) and Saab AB, and by the project Scalable Kalman filters funded by the Swedish Research Council (VR). G. Hendeby has received funding from the Center for Industrial Information Technology at Linköping University (CENIIT) t 17 12 |Col1|R2|Col3| |---|---|---| |||| |||| |||| Fig. 1. Merging of two estimates with covariances R1 and R2. Multiple BLUEs for different cross-correlations are provided. A conservative bound is also shown. where two direct observations of the same unknown parameter vector are available. Optimal fusion of two estimates with covariance R1 and R2 requires the cross-correlations R12 to be known. If so, the BLUE yields optimal fusion. Under unknown R12 a conservative estimator has to consider all possible instances of R12. A geometrical interpretation of this is given in Fig. 1, where R1 and R2 are represented as ellipses, and the task is to compute a new ellipse that summarizes the total information without overestimating it, i.e., we seek a conservative bound. The BLUE covariance ellipses for zero, maximum and several nonzero correlations are illustrated in Fig. 1. Since a conservative bound must take into account all possible values of R12 it must be simultaneously larger than all the individual BLUEs and hence, for instance, cannot be the BLUE assuming either zero correlations or maximum correlations. The literature describes basically three different methods corresponding to slightly different assumptions on _R12. These are covariance intersection (CI, [7]), the largest_ _ellipsoid (LE, [9]) method and inverse covariance intersection_ (ICI, [12]). In this paper we generalize the existing theory to the general linear regression framework based on the work initiated in [14]. The BLUE approach is formulated as an optimization problem. This enables us to formulate the conservative linear _unbiased estimator (CLUE) as a similar optimization problem._ A major point is that standard optimization software can be applied to compute a CLUE in general cases, which is something that is currently not possible [6]. To evaluate both the theory and optimization algorithms, we study a number of fusion problems, which means that we can compare with the existing methods from literature. For that comparison reason, we also provide a review of these fusion methods in the same notational framework and show that these are special cases of _R1_ BLUE under zero correlations BLUE under maximum correlations BLUE under nonzero correlations conservative bound ----- the CLUE. The following are the main contributions: _• A framework which unifies existing conservative es-_ timation methods, facilitates the development of new methods and is able to serve as a tool in the analysis of conservative estimation problems. _• A number of derived properties of the general linear_ regression CLUE. _• A methodology for computing a CLUE in general cases_ using standard optimization software. _• A theorem that states under which conditions the LE_ method is an optimal CLUE. II. BACKGROUND Necessary mathematical notation and theory of linear estimation are introduced below. This is followed by a literature review of decentralized and distributed estimation. _A. Notation_ Let R, R[n] and R[m][×][n] denote the set of real numbers, the set of real-valued n-dimensional vectors and the set of real-valued _m × n matrices, respectively. Let S[n]+_ [and][ S]++[n] [denote the set] of n _n symmetric positive semidefinite (PSD) matrices and_ _×_ the set of n _n symmetric positive definite (PD) matrices,_ _×_ respectively. For A, B ∈ **S[n]+[, the inequalities][ A][ ⪰]** _[B][ and]_ _A ≻_ _B are equivalent to (A −_ _B) ∈_ **S[n]+** [and][ (][A][ −] _[B][)][ ∈]_ **[S]++[n]** [,] respectively. The ellipsoid of a matrix A ∈ **S[n]++** [is given by] the set of points (A) = _x_ **R[n]** _x[T]A[−][1]x_ 1 . If A is a _E_ _{_ _∈_ _|_ _≤_ _}_ covariance matrix then (A) describes an uncertainty. A larger _E_ ellipsoid means a larger uncertainty and for A, B ∈ **S[n]++** [it is] also true that _A_ _B_ (A) (B), (1a) _≻_ _⇐⇒E_ _⊃E_ _A_ _B_ (A) (B). (1b) _⪰_ _⇐⇒E_ _⊇E_ Appendix A provides a summary of matrix properties used in this paper. _B. Preliminaries_ The Fisherian estimation philosophy is adopted which means that the state x[0] to be estimated is deterministic. The overall problem is to derive an estimate ˆx of x[0] **R[n]** from _∈_ measurements y **R[m]** given as a linear regression _∈_ _y = Hx[0]_ + v, (2) where H **R[m][×][n]** and v is zero-mean random noise. It is _∈_ assumed m _n and rank(H) = n. If rank(H) < n then_ _≥_ _y is insufficient for the considered problem in the sense that_ it is not possible to estimate all components of x[0]. In terms of a least squares estimator ˆx there would be infinitely many solutions ˆx in the rank deficient case [15]. The true covariance of y is given by R[0] = cov(y) = E(y E y)(y E y)[T], where _−_ _−_ E is the expectation operator. The covariance of ˆx is denoted by P . In linear estimation[1] _xˆ = Ky where K_ **R[n][×][m]** is the _∈_ estimation gain. An estimator is unbiased if E ˆx = x[0]. For a 1Since E v = 0 only linear estimators ˆx = Ky are considered and hence th l ffi ti t ˆ _K_ + b i t i l d d i thi linear estimator with y as in (2) this implies KH = I, where _I is the identity matrix of appropriate dimension. The true_ covariance of ˆx = Ky is given by cov(ˆx) = E(ˆx _x[0])(ˆx_ _−_ _−_ _x[0])[T]_ = KR[0]K [T]. A linear estimator is completely defined by (K, P ). Fusion is a subclass of estimation problems where the components of y are estimates y1, y2, . . . to be merged into a more accurate estimate. Here it is assumed that yi = Hix[0] +vi and vi is zero-mean noise. In linear fusion an estimate is computed from N estimates y1, . . ., yN according to _xˆ =_ �K1 _. . ._ _KN_ ��y1 _. . ._ _yN_ �T = Ky, (3a)  _R1[0]_ _. . ._ _R1[0]N_   _K1_  _P =_ �K1 _. . ._ _KN_ � ... ... ... ...     _RN[0]_ 1 _. . ._ _RN[0]_ _KN_ = KR[0]K [T], (3b) where Ri[0] [= cov(][y][i][)][ and][ R]ij[0] [= cov(][y][i][, y][j][)][ is the cross-] covariance between yi and yj. When speaking of crosscorrelations we mean Rij[0] [. The linear fusion problem of (3) is] structurally equivalent to the linear estimation problem. _C. Related Research_ Two estimates y1 and y2 are optimally fused using the Bar_Shalom-Campo [16] formulas if the cross-correlations R12[0]_ [is] available. Cross-correlations are in general unknown [7, 17], but nevertheless need to be handled carefully. Otherwise there is an immediate risk of double counting information. In [18] the cross-correlations are compensated for by subtracting previously accounted information. However, this requires some sort of bookkeeping mechanism which is not always possible in practice. A related concept is the channel filter [19] which allows for compensation of cross-correlations in certain sensor network topologies. In [20–23] several consensus-based methods are proposed. These approaches also make specific assumptions about the sensor network topology, and then use averaging to drive the estimates towards consensus. Another class of methods are distributed Kalman filtering algorithms— see, e.g., [24–26]—which restrict the estimates to be merged to follow specific filtering schemes. The methods described above are useful in a vast amount of applications given that their conditions are met. A problem is that these conditions are not always satisfied. For instance, there are situations where: (i) there is no knowledge about the sensor network topology; (ii) the history of exchanged data is unavailable; (iii) the filtering schemes, if any, deployed by the other nodes in the sensor network are unknown. If some or all of (i)–(iii) hold then the problem is more or less structureless and we are basically forced to rely upon _conservative estimation [27]. Detailed descriptions of the three_ main methods of conservative linear estimation are provided in Sec. VI. Theoretical aspects of conservative linear estimation have been studied to some extent. See, e.g., [8, 17, 28, 29] for theoretical work on CI. ICI is further examined in [30, 31]. A main aspect of conservative estimation is partial knowledge about cross-correlation. Exploiting partial knowledge is studied in [10 12 32 34]         ----- This paper considers estimation in a linear regression with an uncertainty in the model. Here we use P cov(ˆx) as the _⪰_ necessary condition for an estimator given by ˆx and P to be called conservative. In [35] a closely related problem is studied where the authors instead uses tr(P ) tr (cov(ˆx)) as the _≥_ necessary condition for a conservative estimator. The resulting algorithm is a minimax optimization method that computes a worst-case estimate given the uncertainty in the model. Other minimax formulation for estimation under model uncertainties are derived in [36–39]. To be able to apply a worst-case approach a worst-case element must exist. Unfortunately, for the general problem considered in this paper a worst-case element is not unambiguously defined. To ensure P cov(ˆx), _⪰_ a conservative estimator must instead simultaneously consider multiple elements which may all be worst-case in different senses. We further elaborate on this topic in Sec. III-D to illustrate why a minimax formulation is not possible for the general problem studied in this paper. III. LINEAR UNBIASED ESTIMATION AS OPTIMIZATION In this section the problem is formulated. First, the CLUE is defined. The BLUE is then introduced using an optimization formulation. We finally generalize the BLUE concept for the optimal CLUE problem. _A. Conservative Linear Unbiased Estimation_ In conservative estimation R[0] is generally not fully known. Instead R[0] belongs to a set A where A ⊂ **S[m]++** [is known.] As a result, cov(ˆx) cannot be computed. The approach is then to bound cov(ˆx) from above, i.e., to find K and P such that _P_ cov(ˆx) with ˆx = Ky. An estimator which is able to _⪰_ guarantee P cov(ˆx), or equivalently P _KRK_ [T], _R_, _⪰_ _⪰_ _∀_ _∈A_ is conservative. It is assumed that the elements of have _A_ only finite eigenvalues, which means that R[0] has only finite eigenvalues. A CLUE which computes an estimate ˆx and covariance P is characterized by the following properties _xˆ = Ky_ _KH = I_ _P_ _KRK_ [T], _R_ _,_ (4) _∧_ _∧_ _⪰_ _∀_ _∈A_ where y is defined according to (2). The problem studied in this paper can now be formulated as: For y given according to (2) and a given set, derive a CLUE (K, P ) where P is _A_ as small as possible. As an example of a conservative estimator and a nonconservative estimator, consider a case where x[0] **R[2]** _∈_ and A = {R[a], R[b], R[c], R[d], R[e]} ⊂ **S[4]++[. Let][ (][K, P]** [)][ and] (K _[′], P_ _[′]) be two estimators, where P and P_ _[′]_ are given by their ellipses in Fig. 2. Since (P ) (KRK [T]), _R_ _E_ _⊇E_ _∀_ _∈A_ and hence P _KRK_ [T], _R_ we conclude that (K, P ) _⪰_ _∀_ _∈A_ is conservative given . By a similar geometrical reasoning _A_ we see that P _[′]_ _⪰_ _K_ _[′]R(K_ _[′])[T]_ cannot hold for all R ∈A and we therefore conclude that (K _[′], P_ _[′]) is not a conservative_ estimator. _Remark 1. The authors of [8, 12] use the notion admissibility_ where admissible matrices are those R that are permitted given the problem formulation. Here is used to represent the set _A_ of all admissible matrices. The optimization problem in (5) is easily solved and a closed-form solution exists, and therefore this optimization formulation is seldom used. The reason we write the BLUE in this way is made clear shortly. If R[0] is invertible, then the BLUE is given by K _[⋆]_ = �H [T](R[0])[−][1]H�−1 H T(R0)−1 [1] and hence _xˆ[⋆]_ = �H [T](R[0])[−][1]H�−1 H T(R0)−1y, (6a) _P_ _[⋆]_ = �H [T](R[0])[−][1]H�−1 . (6b) According to the Gauss-Markov theorem [15] KR[0]K [T] _P_ _[⋆]_ _⪰_ for any K such that KH = I. Hence, the same solution _P_ _[⋆]_ is obtained irrespective of the choice of matrix increasing function J. _C. Best Conservative Linear Unbiased Estimation_ In the general case where is not a singleton the BLUE is _A_ not well-defined. The typical reason for this is that the crosscorrelations, e.g., R12[0] [, are unknown. Still it is desirable to] design an estimator similar to the BLUE as in Definition 1. A best CLUE is defined in Definition 2. A best CLUE reduces to the BLUE if = _R[0]_ . Similar formulations have _A_ _{_ _}_ been proposed in [17, 28, 29]. **Definition 2 (Best Conservative Linear Unbiased Estimator).** Let y = Hx[0] + v. Assume cov(y) = R[0] _∈A ⊂_ **S[m]++[.]** An estimator reporting ˆx[⋆] _K_ _[⋆]y and P_ _[⋆]_ is called a best _KRK_ [T], R ∈A _P_ _K_ _[′]R(K_ _[′])[T], R ∈A_ _P_ _[′]_ Fig. 2. An example of a conservative estimator (K, P ) and nonconservative estimator (K _[′], P_ _[′])._ _B. Best Linear Unbiased Estimation_ In the classical setting, which is a special case of a CLUE with = _R[0]_, it is well-known that a linear unbiased _A_ _{_ _}_ estimator with the smallest mean square error is given by the BLUE [1]. The BLUE is defined in Definition 1, where also a loss function J : R[n][×][n] **R is introduced. Throughout** _→_ this work it is assumed that J is matrix increasing [40], see Appendix A. **Definition 1 (Best Linear Unbiased Estimator). Let y =** _Hx[0]_ + v where cov(y) = R[0]. An estimator ˆx[⋆] = K _[⋆]y where_ _P_ _[⋆]_ = K _[⋆]R[0](K_ _[⋆])[T]_ is called the best linear unbiased estimator if K _[⋆]_ is the solution to minimize _J(P_ ) _K,P_ subject to _KH = I_ _P = KR[0]K_ [T], for a given matrix increasing function J. (5) ----- **Conservative Linear Unbiased Estimation** Assumptions: y = Hx[0] + v, cov(y) = R[0] _∈A_ Estimator: ˆx = Ky, KH = I, P ⪰ cov(Ky) _A = {R[0]}_ _A = {R[a], R[b], . . ., R[0], . . . }_ **optimal estimator: BLUE** **optimal estimator: best CLUE** _exact and unique closed-form solution_ _fundamental bounds_ _general CLUE using_ _special cases_ _[⋆]_ = P _[⋆]H_ [T](R[0])[−][1], P _[⋆]_ = �H [T](R[0])[−][1]H�−1 _robust optimization_ Fig. 3. Overview of the conservative linear unbiased estimation problem. The special case with A = {R[0]} is illustrated to the left and the general case is illustrated to the right. The green box to the right visualizes the scope of this paper. |optima|A = {R0 }|Col3| |---|---|---| ||l estim|ator:| |||| _conservative linear unbiased estimator if (K_ _[⋆], P_ _[⋆]) is the_ solution to minimize _J(P_ ) _K,P_ _J(P_ ) = J(KRK [T]) is minimized for a worst-case element _R_ . However, as seen in the example below, the solution _∈A_ to this minimax problem is not necessarily feasible w.r.t. the original problem in (2). Let J( ) = tr( ). Let H = �I _I[�][T]_ and = _R[a], R[b]_, _·_ _·_ _A_ _{_ _}_ where subject to _KH = I_ _P_ _KRK_ [T], _R_ _,_ _⪰_ _∀_ _∈A_ for a given matrix increasing function J. (7) 2 0 0 1 0 4 1 0 0 1 4 0  1 0 0 2  _._   _,_ _R[b]_ =  2 0 2 0 0 4 0 2 2 0 4 0  0 2 0 2 A solution to (7) is given by a pair (K _[⋆], P_ _[⋆]) which is_ one example of a feasible point [40] as this pair satisfies all constraints of the problem. The set of all feasible points is called the feasible set. In particular, the feasible set of the problem in (7) equals the set of all CLUEs. The BLUE problem has a unique solution, the same is not true for the best CLUE problem for which the choice of loss function J is crucial. While the BLUE finds a minimum element P _[⋆]_ of the feasible set to the problem in Definition 1, the best CLUE P _[⋆]_ is a minimal element of the feasible set to the problem in Definition 2. See Appendix A for the definitions of minimum and minimal elements including procedures for how they can be found. Since the best CLUE problem boils down to finding a minimal element the natural loss function is tr(W ), where _·_ _W ∈_ **S[n]++[, see Appendix A. The reason for using a more]** general matrix increasing J is that it includes, e.g., the determinant, which is a common loss function in the literature and is not obviously related to tr(W ). However, it is shown in _·_ Appendix A that any matrix increasing function can be used to find minimal elements. The literature suggests that trace and the determinant are the most commonly used loss functions. Minimizing the trace is related to minimizing the variance, and minimizing the determinant is related to minimizing the entropy [41]. _D. Relation to Minimax Optimization_ In Sec. II-C it was discussed that related problems are solved in [35–39] using minimax formulations. These papers consider scalar loss functions, cf. J(P ), but they do not impose the PSD constraint P _KRK_ [T], _R_ which is a necessary _⪰_ _∀_ _∈A_ condition for a CLUE. Using a relaxed constraint J(P ) _≥_ _J(KRK_ [T]), _R_, e.g., as is used in [35] with J( ) = tr( ), _∀_ _∈A_ _·_ _·_ it is possible to derive a minimax optimization problem where _R[a]_ = The two BLUEs for each of the elements in are given by _A_ _Ka = PaH_ [T](R[a])[−][1], _Pa =_ �H [T](R[a])[−][1]H�−1, _Kb = PbH_ [T](R[b])[−][1], _Pb =_ �H [T](R[b])[−][1]H�−1 . For a solution (K, P ) to satisfy tr(P ) tr(KR[a]K [T]) and _≥_ tr(P ) tr(KR[b]K [T]) it must necessarily satisfy _≥_ tr(P ) ≥ max (tr(Pa), tr(Pb)), as a consequence of the Gauss-Markov theorem [15]. In this case we have tr(Pa) = tr(KaR[a]Ka[T][) = 4][,] tr(KaR[b]Ka[T][) = 4][,] tr(Pb) = tr(KbR[b]Kb[T][) = 2][.][63][,] tr(KaR[b]Ka[T][) = 4][.][41][,] i.e., tr(Pa) = tr(KaR[b]Ka[T][) = 4][ which is as small as it pos-] sible can be. We hence have that (Pa, Ka) is a solution to the minimax problem suggested by relaxing P _KRK_ [T], _R_ _⪰_ _∀_ _∈_ into tr(P ) tr(KRK [T]), _R_ . Meanwhile _A_ _≥_ _∀_ _∈A_ � 0 1� _−_ _Pa −_ _KaR[b]Ka[T]_ [=] 1 0 _̸⪰_ 0, _−_ which violates P _⪰_ _KRK_ [T], ∀R ∈A, and (Pa, Ka) is therefore not a feasible solution w.r.t. (2). This counterexample shows that a minimax formulation does not in general apply for the best CLUE problem in (2). The minimax formulation is therefore not pursued further. _E. Proposed Framework_ Definition 2 is the backbone of the proposed framework for conservative linear unbiased estimation. A similar optimization formulation has been proposed in e g [29] for the CI case In ----- this paper we further develop this concept to reach a general framework for conservative linear unbiased estimation. The rest of this paper is devoted to the CLUE concept: Problem properties are analyzed in Sec. IV. A major motivation for the optimization formulation of a CLUE is that standard optimization software can be applied to compute a CLUE and in some cases guarantee a best CLUE. This is the topic of Sec. V. In Sec. VI some of the existing conservative estimation methods are shown to be CLUE and sometimes even best CLUE under certain assumptions on . To evaluate both _A_ theory and optimization algorithms, we study a number of fusion problems in Sec. VII. Fig. 3 illustrates conservative linear unbiased estimation and the special case of linear unbiased estimation. IV. PROBLEM PROPERTIES The optimization problems to find the BLUE and best CLUE are very similar. However, while a closed form solution is available for the BLUE, the additional uncertainty in the best CLUE formulation makes the problem much harder to solve and no general solution procedure is available. This section highlights differences between the two optimization problems, and derives a simplified optimization problem providing a lower bound Pl of the obtainable covariance of the CLUE. Also an upper bound Pu is provided implying that P _[⋆]_ lies in an interval 0 ≺ _Pl ⪯_ _P_ _[⋆]_ _⪯_ _Pu, where Pl and Pu depend on_ . _A_ _A. Lower Bound on Best CLUE_ The CLUE cannot be better than the BLUE for any R . _∈A_ As a consequence, a lower bound of the CLUE covariance can therefore be computed as the smallest covariance larger than all BLUE covariances, which is an easier optimization problem than the best CLUE problem. In analogy to the Cramér-Rao _lower bound [42] this lower bound can be used as a guideline_ for system design, e.g., to a tradeoff between communication bandwidth and performance. It should be noted that this formulation relaxes the constraints, and hence there is no guarantee that a gain K achieving this covariance exists in the general case. Below we derive a lower bound Pl for the best CLUE covariance P _[⋆], where subscript l refers to quantities_ related to the lower bound. It is shown that J(P _[⋆]) ≥_ _J(Pl)._ If a CLUE (K, P ) satisfies J(P ) = J(Pl) then this CLUE is also a best CLUE. Instead of the best CLUE, consider the problem minimize _J(P_ ) _P_ (8) subject to _P_ �H [T]R[−][1]H�−1, _R_ _._ _⪰_ _∀_ _∈A_ For a given J a solution Pl to (8) is a lower bound[2] on P _[⋆]._ **Theorem 1 (Best CLUE Lower Bound). Let (K** _[⋆], P_ _[⋆]) be_ _given by (7) and let Pl be given by (8). Then J(P_ _[⋆]) ≥_ _J(Pl)._ 2S VII D id l h th l b d i t i t _Proof. By assumption the same matrix increasing J and the_ same are used in both (7) and (8). Since (K _[⋆], P_ _[⋆]) solves_ _A_ (7) we have for each R that _∈A_ _P_ _[⋆]_ _K_ _[⋆]R(K_ _[⋆])[T]_ (H [T]R[−][1]H)[−][1], _⪰_ _⪰_ where the last inequality follows from the Gauss-Markov theorem [15] as we have K _[⋆]H = I. Hence, P_ _[⋆]_ satisfies the constraints in (8) and therefore J(P _[⋆]) ≥_ _J(Pl)._ An implication of the theorem above is that if Pl given by (8) is obtained by a CLUE, then this estimator is a best CLUE. We now derive an estimator with the gain Kl computed from _Pl, and give sufficient conditions for when (Kl, Pl) is a best_ CLUE. Start by solving for Rl in _Pl =_ �H [T]Rl[−][1]H�−1, (9) which has a solution since Pl ∈ **S[n]++** [and since][ H] [T][ has full] column rank. Then define _Kl =_ �H [T]Rl[−][1]H�−1 H TRl−1, (10) which yields an unbiased estimator since KlH = I. For (Kl, Pl) to be a CLUE, cf. (4), it must hold that Pl ⪰ _KlRKl[T][,][ ∀][R][ ∈A][. As it follows from (10) that][ K][l][R][l][K]l[T]_ [=][ P][l] with Rl according to (9), a sufficient condition for (Kl, Pl) to be a CLUE is that Rl ⪰ _R, ∀R ∈A since then_ _Pl = KlRlKl[T]_ _l_ _[,][ ∀][R][ ∈A][.]_ _[⪰]_ _[K][l][RK]_ [T] The results above are summarized in the following theorem. **Theorem 2 (Best CLUE From Lower Bound). Assume Pl** _solves (8) and that Kl is according to (10) with Rl implicitly_ _given by (9) . If Rl ⪰_ _R, ∀R ∈A, then (Kl, Pl) is a best_ _CLUE._ As will be seen in Sec. VII-B3 it is possible to not satisfy _Rl ⪰_ _R, ∀R ∈A while still satisfying Pl ⪰_ _KlRKl[T][,][ ∀][R][ ∈]_ . _A_ _B. Upper Bound on Best CLUE_ Assume C is given, where C _R,_ _R_ . Then it is _⪰_ _∀_ _∈A_ possible to construct a CLUE as (K, P ) with P = KCK [T] for any K subject to KH = I. In particular a CLUE can be derived by first finding a C _R,_ _R_, and then _⪰_ _∀_ _∈A_ compute the BLUE w.r.t. this C. Finding a smallest covariance larger than all possible R is a simpler problem than _∈A_ the best CLUE problem. However, approaching the problem in this way restricts the feasible set, and therefore a CLUE, but not necessarily a best CLUE, is obtained. Nevertheless, it is sometimes useful to bound tightly using C, see, e.g., _A_ [35, 43–45]. In such cases the closed-form expression of (12) below can be used to compute a CLUE. Next we derive an upper bound Pu on P _[⋆]_ of (7). Subscript u refers to quantities related to the upper bound. Introduce the set _C_ **S[m]** _C_ _R_ _R_ (11) _B_ _{_ _∈_ _|_ _⪰_ _∀_ _∈A}_ ----- which contains all matrices C ∈ **S[m]++** [that are larger than all] elements R ∈A. A CLUE (Ku, Pu) is then given by _Ku =_ �H [T]C _[−][1]H�−1 H_ TC _−1,_ (12a) _Pu =_ �H [T]C _[−][1]H�−1,_ (12b) where C ∈B and KuH = I. By a similar reasoning that leads up to Theorem 2, (Ku, Pu) according to (12) is a CLUE for any C . _∈B_ If C is a minimal element of in (11) it is also called _B_ a minimal bound on A since there exists no element R[′] _⪯_ _C, R[′]_ = C for which R[′] _R,_ _R_ . In general (12) is too _̸_ _⪰_ _∀_ _∈A_ conservative to be a best CLUE, even if C is a minimal bound[3] on . This fact is also discussed in [29]. We summarize the _A_ results in the following theorem. **Theorem 3 (Best CLUE Upper Bound). Let (K** _[⋆], P_ _[⋆]) be_ _given by (7) and let (Ku, Pu) be given by (12), where C ∈B_ _with B as in (11). Then J(Pu) ≥_ _J(P_ _[⋆])._ _C. Summary_ If the approach of Sec. IV-A yields a CLUE then (Kl, Pl) is also a best CLUE, otherwise (Kl, Pl) is a too optimistic estimator. On the other hand, the estimator (Ku, Pu) given by (12) is generally too pessimistic to be a best CLUE. If the lower and upper bounds coincide, then, as a consequence of Theorem 2, a best CLUE is trivially found. V. GENERAL CONSERVATIVE LINEAR UNBIASED ESTIMATION USING ROBUST OPTIMIZATION In this section it is shown how robust optimization (RO, [46]) can be used to solve general CLUE problems. Other cases where RO is used to solve estimation problems are studied in, for instance, [46, 47]. We begin by showing that our problem fits into the ro_bust semidefinite programming optimization framework [48]._ Tractability and optimality are then discussed. With tractability we mean that a solution can be found within a reasonable amount of time. Finally, an implementation of conservative estimation using RO in YALMIP [49] is provided. _A. Robust Semidefinite Optimization_ Since we deal with optimization problems having semidefinite constraints our focus is on a class of problems called _semidefinite programs (SDP). Let ∆_ be an uncertain _∈D_ optimization parameter only known to reside in an uncertainty _set_ **R[d]. In RO none of the constraints are allowed to be** _D ⊂_ violated for any value ∆ [50]. A generic SDP with an _∈D_ inequality constraint uncertainty can be stated as [48] minimize _f_ (z) _z_ (13) subject to **F(z, ∆)** 0, ∆ _,_ _⪰_ _∈D_ for a loss function f ( ). In (13) z is an optimization variable _·_ and F(z, ∆) is a matrix-valued function that depends on z and 3A l f thi i id d i S VII B3 ∆. The constraint in (13) is a linear matrix inequality (LMI) for any fixed ∆. The best CLUE problem of Definition 2 is aligned with the formulation in (13). To see this, replace z by (K, P ), by, _D_ _A_ ∆ by R and f (z) by J(P ). Then rewrite the matrix inequality of (7) using Schur complement [51] as � _P_ _KR_ **F(K, P, R) =** _RK_ [T] _R_ � 0. (14) _⪰_ Since (14) is equivalent to P 0 _P_ _KRK_ [T] 0 we have _⪰_ _∧_ _−_ _⪰_ retrieved the problem in Definition 2. _B. Tractability And Optimality_ There are a few cases where the computed solution to the problem in (13), with =, is tractable and optimal in the _D_ _A_ sense that a minimal element is found when the RO problem is solved [52]. If is a finite set then tractability follows trivially _A_ since the uncertainty is replaced by a finite set of LMIs. The problem is also tractable if is the convex hull [40] of a finite _A_ set, see below for an example on convex hulls. The convex hull of a set is another set which contains all _V_ convex combinations of the elements of [40]. For example, _V_ consider a set of covariances V = {A, B} ⊂ **S[2]+** [where][ A][ =] � _−42_ _−12_ � and B = [ [4 2]2 1 []][. The convex hull of][ V][ is] _θA + (1_ _θ)B_ _θ_ [0, 1] _{_ _−_ _|_ _∈_ _}_ = �� _−4(2θθ+2(1+1−−θ)θ)_ _−2θθ+2(1+1−θ−θ)_ ���� _θ ∈_ [0, 1]� = �� 2−44θ 2−14θ ��� _θ ∈_ [0, 1]� _,_ which is equivalent to {[ [4]c 1[ c] []][ |][ c][ ∈] [[][−][2][,][ 2]][}][. In case the] unknown cross-covariance is not a scalar it is generally not possible to express as the convex hull of a finite set. _A_ For general uncertainty sets, i.e., general, there are only a _A_ few constructive results on robust counterparts for the problem in (13). The case treated here where both the RO problem and the uncertainty set is defined by semidefinite constraints is largely untreated in the literature. Not only are exact solutions absent in contrast to the simple example above, but also general tractable conservative approximations are missing. _C. Robust Estimation Using YALMIP_ YALMIP is a MATLAB[®] toolbox developed to model and solve optimization problems [49], and it has the ability to derive RO problems [53]. The strategies in [53] focus on cases where exact robust counterparts can be derived which rules out problems according to the model in (13). However, theory has recently been developed and added to the YALMIP toolbox to support problems according to (13), i.e., uncertainty structures involving arbitrary intersections of conic-representable sets. These additions[4] are described in the forthcoming [54]. The key feature for us is realized by a function called uncertain(), which enables the uncertainty imposed by _R[0]_ to be handled. Next, we illustrate conservative _∈A_ estimation using RO in YALMIP with an example. Consider 4It should be noted that these additions are already implemented in YALMIP b t th d t ti d ibi th i tl bli h d ----- the task of computig a conservative estimate ˆx = Ky and P where tr(P ) is minimized. Let y = [ _[y]y[1]2_ [] =] � _HH12_ � _x[0]_ +v where _x[0]_ _∈_ **R[2], H1 = H2 = [** [1 0]0 1 []][,][ R]1[0] [= [][ 1 0]0 4 []][ and][ R]2[0] [= [][ 4 0]0 1 []][.] Assume R12[0] [is completely unknown and that][ R][0][ ⪰] [0][.] The problem is translated into MATLAB[®] syntax using YALMIP in Listing 1. YALMIP functions are highlighted in **orange. The result is K = [** [0]0[.][8] 00.2 00.2 00.8 []][ and][ P][ = 1][.][6 [][ 1 0]0 1 []][.] Listing 1: A simple YALMIP example. H = [eye(2) ; eye(2)]; R1 = diag([1 4]); R2 = diag([4 1]); K = sdpvar(2,4); % Declare SDP variable P = sdpvar(2); R12 = sdpvar(2,2,’full’); R = [R1 R12 ; R12’ R2]; F = [uncertain(R12), K*H == eye(2), [P K*R ; R*K’ R] >= 0, R >= 0]; % Constraints J = trace(P); **optimize(F, J) % Solve problem** In this example YALMIP finds a best CLUE. However, in general the solution is approximative and the only guarantee is that the solution is a CLUE. VI. SPECIAL CASES OF CONSERVATIVE LINEAR UNBIASED ESTIMATION In this section it is shown that several existing conservative estimation methods are best CLUE under different assumptions on . Common for all methods is that the diagonal _A_ blocks of R[0] are known while the off-diagonal blocks, e.g., _R12[0]_ _[,][ are unknown. What differs between the methods is the]_ assumptions on the off-diagonal blocks. For instance, it could be that we have some extra knowledge that R12[0] [is diagonal] or that the eigenvalues of R12[0] [(][R]12[0] [)][T][ are smaller than][ a >][ 0][.] Benefits of exploiting any extra structure on the otherwise unknown cross-correlations are illustrated using an example. Assume R[0] = [ [4]c 1[ c] []][ where][ c][ is unknown. If it is only known] that R[0] 0 then c [ 2, 2] and hence R[0] could be repre_⪰_ _∈_ _−_ sented by any ellipse encloses in the rectangle of Fig. 4(a). A minimal bound on is in this case given by C. If on the _A_ other hand it is known that c ∈ [− 2[1] _[,][ 1]2_ []][, then it is possible] to find an even smaller minimal bound C _[′], see Fig. 4(b). In_ Fig. 4 we have also illustrated A = {[ [4]c 1[ c] []][ |][ c][ ∈] [[][−][2][,][ 2]][}][ and] _A[′]_ = {[ [4]c 1[ c] []][ |][ c][ ∈] [[][−] 2[1] _[,][ 1]2_ []][} ⊂A][.] Below we consider CI, ICI and LE. Recalling the assumptions made in Sec. II-B, with N denoting the number of estimates to be merged, it is now assumed that (yi, Ri[0][)][ are] available, where i = 1, . . ., N, and that Rij[0] [, with][ i][ ̸][=][ j][,] are unknown. It should be emphasized that CI, ICI and LE give different solutions to a problem since they are related to different assumptions on . Among the described methods _A_ LE makes the most restrictive assumptions on while CI _A_ makes the least restrictive assumptions on . Hence LE in _A_ general is less conservative than ICI while ICI in general is less conservative than CI. _A. Covariance Intersection_ CI was originally proposed in [7] for the fusion of two correlated estimates CI is based on completely unknown _C_ _C_ _[′]_ |Col1|A|A′| |---|---|---| (a) (b) Fig. 4. Illustration of the benefits of utilizing any extra structure on the unknown parts of R[0]. (a) C is a minimal bound on A. (b) A smaller minimal bound C _[′]_ _≺_ _C can be found for A[′]_ _⊂A._ _cross-correlations for which the only condition on the cross-_ correlation is that R[0] 0, and therefore _≻_ _A =_ �� _RR12[T]1 R[R]12[2]_ � _∈_ **S[m]++���R1 = R10[, R][2]** [=][ R]2[0] � _._ (15) Let y = �y1[T] _. . ._ _yN[T]_ �T, H = �H1[T] _. . ._ _HN[T]_ �Tand C = diag � _Rω11[0]_ _[, . . .,][ R]ωNN[0]_ �. CI is given by[5] _P_ _[−][1]_ = H [T]C _[−][1]H =_ _N_ � _ωiHi[T][(][R]i[0][)][−][1][H][i][,]_ (16a) _i=1_ _P_ _[−][1]xˆ = H_ [T]C _[−][1]y =_ _N_ � _ωiHi[T][(][R]i[0][)][−][1][y][i][,]_ (16b) _i=1_ where ωi ∈ [0, 1] and [�]i[N]=1 _[ω][i][ = 1][. The free parameters]_ _ω1, . . ., ωN are found by minimizing J(P_ ). The CI gain K is given by _K = P_ �ω1H1([T]R1[0][)][−][1] _. . ._ _ωN_ _HN[T]_ [(][R]N[0] [)][−][1][�] _._ (17) Concerning optimality of CI, let N = 2 and H1 = H2 = I. In this case we only have one free parameter ω since it is possible to define ω1 = ω and ω2 = 1 − _ω. Let the optimal_ value of ω given J(P ) be denoted by ω[⋆]. Further, let an arbitrary CLUE be given by (K _[′], P_ _[′]). In [29] it is shown that_ if ω[⋆] is obtained by minimizing J(P ) w.r.t. ω with P given by (16a), then for (K _[⋆], P_ _[⋆]) given by_ _K_ _[⋆]_ = �ω⋆P ⋆(R10[)][−][1] (1 − _ω[⋆])P_ _[⋆](R2[0][)][−][1][�]_ _,_ (18a) _P_ _[⋆]_ = �ω[⋆](R1[0][)][−][1][ + (1][ −] _[ω][⋆][)(][R]2[0][)][−][1][�][−][1][,]_ (18b) it holds that _P_ _[′]_ _⪯_ _P_ _[⋆]_ =⇒ _P_ _[′]_ = P _[⋆]._ This means P _[⋆]_ is a minimal element of the feasible set. Hence, (K _[⋆], P_ _[⋆]) according to (18) constitute a best CLUE, provided_ that N = 2 and that the cross-correlations are completely unknown. _B. Inverse Covariance Intersection_ ICI is derived in [12] for the case where N = 2 and H1 = _H2 = I. ICI is less conservative than CI since it utilizes a_ certain structure on R12[0] [called][ common information][ [12].] We introduce Γ[−][1] _∈_ **S[n]++** [to denote common information] included in both (R1[0][)][−][1][ and][ (][R]2[0][)][−][1][, and][ ˆ][γ][ to denote the] 5Thi t ti i l k th i f _ti_ _f_ ----- corresponding estimate for which Γ = cov(ˆγ). The common information structure is then defined as (Ri[0][)][−][1][ = (][R]i[e][)][−][1][ + Γ][−][1][,] (19a) (Ri[0][)][−][1][y][i] [= (][R]i[e][)][−][1][y]i[e] [+ Γ][−][1][γ,][ˆ] (19b) for i = 1, 2, where (Ri[e][)][−][1][ and][ y]i[e] [are the exclusive informa-] tion and the exclusive estimate of the ith estimate, respectively. The resulting cross-covariance becomes [12] _R12[0]_ [=][ R]1[0][Γ][−][1][R]2[0][.] (20) An implication of (19a) is that (R1[0][)][−][1][,][ (][R]2[0][)][−][1][ ⪰] [Γ][−][1][. The] set is now given by _A_ **Algorithm 1 Largest Ellipsoid Method** **Input: (y1, R1[0][)][ and][ (][y]2[, R]2[0][)]** 1: Factorize R1[0] [=][ U]1[D]1[U][ T]1 [and let][ T]1 [=][ D]1− 2[1] _U1[T][. Factorize]_ _T1R2[0][T][ T]1_ [=][ U]2[D]2[U][ T]2 [and let][ T]2 [=][ U][ T]2 [.] 2: Transform using T = T2T1 according to _y1[′]_ [=][ Ty]1[,] _D1[′]_ [=][ TR]1[0][T][ T][ =][ I,] _y2[′]_ [=][ Ty]2[,] _D2[′]_ [=][ TR]2[0][T][ T][.] 3: For each i = 1, . . ., n, of the ˆx[′] and diagonal P _[′], compute_ �[ˆx[′]]i, [P _[′]]ii�_ = � ([y1[′] []]i[,][ 1)][,] if 1 ≤ [D2[′] []]ii[,] ([y2[′] []]i[,][ [][D]2[′] []]ii[)][,] if 1 > [D2[′] []]ii[.] **Output: T** _[−][1]xˆ[′], T_ _[−][1]P_ _[′]T_ _[−][T]_ where D2 and D12 are diagonal. The LE method is outlined in Algorithm 1. The condition in (24) is equivalent to   (21)  _[.]_ ������ _R1 = R1[0][, R][2][ =][ R]2[0]_ _R12 = R1Γ[−][1]R2_ _R1[−][1][, R]2[−][1]_ _⪰_ Γ[−][1] = _A_    � _R1 R12_ _R12[T]_ _[R][2]_ � _∈_ **S[m]++** An estimate is computed using ICI according to _P_ _[−][1]_ = (R1[0][)][−][1][ + (][R]2[0][)][−][1][ −] �ωR1[0] [+ (1][ −] _[ω][)][R]2[0]�−1,_ (22a) _P_ _[−][1]xˆ =_ �(R1[0][)][−][1][ −] _[ω]_ �ωR1[0] [+ (1][ −] _[ω][)][R]2[0]�−1[�]_ _y1_ + �(R2[0][)][−][1][ −] [(1][ −] _[ω][)]_ �ωR1[0] [+ (1][ −] _[ω][)][R]2[0]�−1[�]_ _y2,_ (22b) where ω [0, 1] is found by minimizing J(P ) [12]. The ICI _∈_ gain is given by K = �K1 _K2�_ where _K1 = P_ �(R1[0][)][−][1][ −] _[ω]_ �ωR1[0] [+ (1][ −] _[ω][)][R]2[0]�−1[�]_ _,_ (23a) _K2 = P_ �(R2[0][)][−][1][ −] [(1][ −] _[ω][)]_ �ωR1[0] [+ (1][ −] _[ω][)][R]2[0]�−1[�]_ _,_ (23b) with P according to (22a). Given that the common information structure holds, it is shown in [12] that if (K _[′], P_ _[′]) is any arbitrary CLUE and P_ _[⋆]_ is computed according to (22a), then _P_ _[′]_ _⪯_ _P_ _[⋆]_ =⇒ _P_ _[′]_ = P _[⋆]._ Hence, ICI is a best CLUE under common information. _C. Largest Ellipsoid Method_ Following its first appearance in [9] the LE method has been derived from multiple principles and therefore has been given multiple names: In [55] it is called safe fusion, the authors of [10, 56] suggests the name ellipsoidal intersection, and in [57] it is named internal ellipsoid approximation. It should be noted that there are minor differences between how the estimate is calculated. In this work we use the algorithm proposed in [55]. In the derivations of LE no explicit assumptions on are _A_ made. Below we propose componentwise aligned correlations which is an assumed structure on R12[0] [that is satisfied if there] exists a joint transformation TJ = diag(T, T ) such that _R1 = R1[0][, R]2_ [=][ R]2[0] _∃T, TR1T_ [T] = I _∧_ [TR2T [T]]ij = 0, i ̸= j _∧_ [TR12T [T]]ij = 0, i ̸= j    ��������� = _A_    � _R1 R12_ _R12[T]_ _[R][2]_ � _∈_ **S[m]++** _._ (25) _D = TJ_ _R[0]TJ[T]_ [=] � _I_ _D12_ _D_ _D_ � _,_ (24) Consider the quantities of Algorithm 1. The resulting gain of the LE method is given by _K =_ �K1 _K2�_ = T _[−][1][ �]K1[′]_ _K2[′]_ � _T,_ (26) where K1[′] [and][ K]2[′] [are the gains in the transformed domain, i.e.,] after transformation using T . The matrix K1[′] [is diagonal where] [K1[′] []][ii] [= 1][ if][ [][D]2[′] []][ii] _[≥]_ [1][ and otherwise zero, and][ K]2[′] [=][ I][ −][K]1[′] [14]. **Theorem 4 (Largest Ellipsoid Method—Optimal). If in (24)** _D12 = TR12T_ [T] _is diagonal for T as given in Algorithm 1,_ _then the LE method of Algorithm 1 is a best CLUE._ _Proof. By assumption TJ_ _RTJ[T]_ [=] � _DI12 DD122_ �, where D2 and _D12 are diagonal. The ith component of I is only correlated_ with the ith component of D2. Hence, we only need to consider pairwise correlated scalars. It is then possible to use CI for the merging of scalars correlated to an unknown degree. If P _[′]_ is the covariance in the transformed domain, then [P _[′]]ii = ω[I]ii + (1 −_ _ω)[D2]ii = ω + (1 −_ _ω)[D2]ii,_ which, as a property of CI, is conservative for all ω _∈_ [0, 1]. Minimizing [P _[′]]ii w.r.t. ω is equivalent to [P_ _[′]]ii =_ min (1, [D2]ii), which in particular is the LE solution. In [14, Theorem 4.7] it is shown that LE is a linear unbiased estimator. Hence, LE is a best CLUE under componentwise aligned correlations. VII. THEORY EVALUATION In this section five estimation examples are solved. The covariances P [CI], P [ICI] and P [LE] corresponding to CI, ICI and LE, respectively, the lower bound Pl, and the upper bound Pu are computed wherever applicable. Each example is also solved using the previously proposed RO approach, where the resulting covariance is denoted by P [RO] YALMIP ----- E2a: E2b: E2c: _R1[0]_ _P_ [CI] _P_ [ICI] _P_ [LE] _R2[0]_ Fig. 5. Summary of E2, where P [CI] _≻_ _P_ [ICI] _≻_ _P_ [LE]. If it is possible to exploit more structure in the problem, then it is possible to compute a CLUE having a smaller covariance. The hashed areas in the left part illustrate (H [T]R[−][1]H)[−][1], ∀R ∈A for the different A. implementations in MATLAB[®] for the RO parts are available [from https://gitlab.com/robinforsling/clue. In all examples ex-](https://gitlab.com/robinforsling/clue) cept the last one it is assumed that Hi = I ∈ **R[2][×][2]** for _i = 1, . . ., N and H =_ �H1[T] _. . ._ _HN[T]_ �T. The loss function _J(P_ ) = tr(P ) is chosen which means that the estimator variance is minimized. Example 1 is denoted by E1 and the remaining examples are identified analogously. _A. E1:_ _Is A Finite Set_ _A_ Assume that N = 2 and R[0] = � (RR12[0]1[0][)][T][ R]R12[0]2[0] �, where R1[0] [=] [or[1 0]0 4 −[]][ and]I. Then[ R]2[0] A[= [] =[ 4 0] {0 1Q, S[]][, and where]} ∈ **S[4]++** _[ R][where]12[0]_ _[∈][ Q][R][ =][2][×]�[2][ is either]RI1[0]_ _RI2[0]_ � and[ I] _S =_ � _R1[0]_ _[−][I]_ �. _−I R2[0]_ The BLUE for R[0] = Q is given by _R3[0]_ _R2[0]_ _R1[0]_ _P_ [CI] _P_ [RO] _Pl_ Fig. 6. Summary of E3, where P [CI] _≻_ _P_ [RO] _≻_ _Pl, and hence CI is_ not a best CLUE when N > 2. The gray ellipses in the intersection represent (H [T]R[−][1]H)[−][1] for different R ∈A. |R0 2 R0 3|Col2| |---|---| ||| |PCI PRO|| |P l|| ||E3, where P CI ≻ P RO ≻ P l, and hence CI is en N > 2. The gray ellipses in the intersection )−1 for different R ∈A. case is given by (15). We first look at A = hR 10 G i and S = hR 10 −Gi| _KQ =_ �H [T]Q[−][1]H�−1 H TQ−1 = �1 0 0 0 0 0 0 1 � _,_ (27) _1) E2a: In this case_ is given by (15). We first look at _A_ two elements, Q = � _R1[0]_ _[G]_ � and S = � _R1[0]_ _[−][G]_ � _G R2[0]_ _∈A_ _−G R2[0]_ _∈A_ where G = [ 3.9990 0.9990 ]. Solving (8), but replacing A by A[′] = _S, Q_, yields approximately 1.60I. Since we then _{_ _}_ _A[′]_ _⊂A_ know that Pl ⪰ 1.60I. The matrix C = 2 diag(R1[0][, R]2[0][)][ satisfies][ C][ ⪰] _[R,][ ∀][R][ ∈]_ . This is true since C _R[0]_ = � _R10_ _−R12[0]_ � 0 as a _A_ _−_ _−R21[0]_ _R2[0]_ _≻_ consequence of the assumption R[0] 0. Using (12b) yields _≻_ _Pu = 1.60I which is equivalent to the solution of (8). Using_ Theorem 2 we can hence conclude that a best CLUE is given by (12) with C = 2 diag(R1[0][, R]2[0][)][.] _2) E2b: In this case_ is given by (21). Solving the prob_A_ lem using YALMIP yields P [RO] = 1.18I which is equivalent to the best CLUE solution P _[⋆]_ = P [ICI] = 1.18I computed using ICI. _3) E2c: In this case A is given by (25). KQ according to_ (27) yields KQRKQ[T] [=][ I,] _∀R ∈A. Hence, K_ _[⋆]_ = KQ and _P_ _[⋆]_ = I constitute a best CLUE, where P _[⋆]_ = P [LE] since LE is a best CLUE in case of componentwise aligned correlations. The matrix Rl = � _RI1[0]_ _RI2[0]_ � does not satisfy Rl ⪰ _R, ∀R ∈_ _A, e.g., for B = diag(R1[0][, R]2[0][)][ ∈A][ the difference][ R][l]_ _[−]_ _[B][ is]_ indefinite. However, the matrix C = 2 diag(R1[0][, R]2[0][)][ satisfies] _C_ _R,_ _R_ and is also a minimal bound on . This C yields ⪰ �H ∀ [T]C ∈A[−][1]H�−1 = 1.60I _P ⋆._ _A_ _≻_ _C. E3: Completely Unknown Cross-Correlations N = 3_ Let N = 3 and assume that which yields KQQKQ[T] [=][ K][Q][SK]Q[T] [=][ I][. From][ KQK] [T][ ⪰] _KQQKQ[T]_ _[,][ ∀][K][ subject to][ KH][ =][ I][ and][ P][ ⋆]_ _[⪰]_ _[K][Q][SK]Q[T]_ [it] follows that a best CLUE in this case is given by K _[⋆]_ = KQ and P _[⋆]_ = I. Since (H [T]S[−][1]H)[−][1] = 0.43I (H [T]Q[−][1]H)[−][1] _≺_ we have _Pl_ = _I._ Using a minimal bound _C_ = diag �R1[0] [+][ I, R]2[0] [+][ I]� = diag(2, 5, 5, 2) a strictly upper bound Pu = 1.43I can be computed. RO yields P [RO] = I which is guaranteed to be optimal as is finite. _A_ _B. E2:_ _Is An Infinite Set_ _A_ Assume that N = 2 and R[0] = � (RR12[0]1[0][)][T][ R]R12[0][2] �, with R1[0] [and] _R2[0]_ [defined as in E1. Assume that][ R]12[0] [is unknown and that] now is an infinite set. A best CLUE depends on . We _A_ _A_ will solve this problem for three different assumptions on, _A_ namely: a) completely unknown cross-correlations, b) common information, and c) componentwise aligned correlations. Since R1[0] [and][ R]2[0] [are fixed, CI, ICI and LE yield][ P][ CI][ = 1][.][60][I][,] _P_ [ICI] = 1.18I and P [LE] = I, respectively, in all subcases below. E2 is summarized in Fig 5 with R2[0] [and][ R]3[0] [being generated from rotation of][ R]1[0] [by][ 60][◦] and −60[◦], respectively. Assume that the off-diagonal blocks of R[0] are completely unknown. CI yields P [CI] = 1.88I. In this case YALMIP gives us _P_ [RO] = 1.76I _P_ [CI]. Hence, CI is not a best CLUE under _≺_ completely unknown cross-correlations if N > 2. We have also computed Pl = 1.31I as the smallest ellipse which contains the intersection of the ellipses of R1[0][,][ R]2[0] [and][ R]3[0][,] but we cannot draw any conclusions about whether this Pl is a strictly lower bound on a best CLUE or not �16 0 _R1[0]_ [=] 0 1 � �4.75 6.50 _, R2[0]_ [=] 6.50 12.25 �, R3[0] [=] �−46.75.50 _−126..2550_ � _,_ ----- _R2[0]_ _R1[0]_ _K_ _[⋆]Q(K_ _[⋆])[T]_ _K_ _[⋆]S(K_ _[⋆])[T]_ _Pl_ _P_ _[⋆]_ Fig. 7. Summary of E4, where Pl is a strictly lower bound on P _[⋆]._ Gray ellipses are given by (H [T]R[−][1]H)[−][1] where R ∈A. _D. E4: Lower Bound Is Strict_ In this example it is shown that P _[⋆]_ ≠ _Pl, P_ _[⋆]_ _⪰_ _Pl. Assume_ thatR� 0−2[0].1 05 N −[=].15 = 2��−1which corresponds to1 _− and51_ �, and R[0] _R=_ 12[0]� (R[∈]R12[0]1[0][R][)][T][ R][2] QR[×]12[0][2]2[0][ can either be] and�, where S of R A1[0] =[= [][ [][ 1]0 {.[ 5 1]51 10Q, S1.[]]5[ and][]][ or]}, respectively. Using RO in YALMIP we compute Pl = [ 0[0].[.]45 0[40 0][.].[45]93 []][ and] _P_ [RO] = P _[⋆]_ = [ 0[0].[.]40 0[56 0][.].[40]95 []][. The results are visualized in] Fig. 7, where also K _[⋆]Q(K_ _[⋆])[T]_ and K _[⋆]S(K_ _[⋆])[T]_ are plotted, with K _[⋆]_ being the best CLUE gain. The reason for having _P_ _[⋆]_ ≠ _Pl, P_ _[⋆]_ _⪰_ _Pl is that no K exists such that_ _P_ _KQK_ [T] _P_ _KSK_ [T] _⪰_ _∧_ _⪰_ _∧_ _P_ _[′]_ _⪰_ (H [T]Q[−][1]H)[−][1] _∧_ _P_ _[′]_ _⪰_ (H [T]S[−][1]H)[−][1] and J(P ) = J(P _[′]) hold simultaneously._ _E. E5: Eigenvalue Constrained R12[0]_ In the final example we have N = 2 and assume that the eigenvalues of R12[0] [are constrained. Let][ x][0][ ∈] **[R][2][ and]** tr(P [CI]) tr(P _[′])_ |Col1|tr(PRO)|Col3|Col4|Col5|Col6| |---|---|---|---|---|---| ||||||| ||||||| 0.5 1.0 1.5 2.0 _ρ_ |R0 2|Col2|Col3| |---|---|---| |0 1 K⋆Q(K⋆)T K⋆S(K⋆)T P l P ⋆||Fig. repr| ||RP 1is a −st 1r ic wtl hy rl eo w Re r b Aou .nd −l H ) e ∈|| � 1 1 � _H1 =_ _√2_ _√2_ _,_ _R1[0]_ [= 1][,] _√1_ _√−1_ 2 2 1 0 0 1   _,_ _R2[0]_ [=] 4 0 0 0 4 0 _._   0 0 4 _H2 =_   In this case R12[0] _[∈]_ **[R][1][×][3][ and is assumed to be constrained]** as R12[0] [(][R]12[0] [)][T][ ≤] _[ρ][2][ or equivalently][ ∥][R]12[0]_ _[∥≤]_ _[ρ][ where][ ρ][ ≥]_ [0][.] To have R[0] 0 we require ρ 2 such that ρ [0, 2]. We _⪰_ _≤_ _∈_ now vary ρ ∈ [0, 2] and compute P [RO]. Also P [CI] and P _[′]_ are computed where _P_ _[′]_ = �H1[T][(][R]1[0][)][−][1][H][1] [+][ H]2[T][(][R]2[0][)][−][1][H][2]�−1, such that P _[′]_ is equivalent to the covariance of the BLUE given _∥R12[0]_ _[∥]_ [= 0][. In Fig. 8 the trace of][ P][ RO][,][ P][ CI][ and][ P][ ′][ are plotted.] As ρ increases from 0 to 2, tr(P [RO]) increases from tr(P _[′]) to_ tr(P [CI]). Note, in this example tr(P [RO]) is almost linear in ρ but this is generally not the case Fig. 8. Results for E5, where ∥R12[0] _[∥≤]_ _[ρ][. The green solid line]_ represents tr(P [RO]). _F. Discussion_ The results for E1-E4 are summarized in Table I, where E5 has been excluded since the results of E5 is given on a different format, see Fig. 8. We see that each of CI, ICI and LE yields the same answer for E1 and E2 since R1[0] [and][ R]2[0] are fixed throughout these cases. In E1 and E2 the YALMIP solution is equivalent to a best CLUE, and we further see the benefits of utilizing any extra structure encoded in . _A_ E3 is a counterexample of CI being a best CLUE under completely unknown cross-correlations when N > 2. We do not know if P [RO] is equivalent to P _[⋆]_ since P _[⋆]_ _≻_ _Pl is possible_ even if P [RO] = P _[⋆], cf. Theorem 1. The upper bound is strict_ in this case. In E5 neither of H1 and H2 is identity or even a square matrix, and eigenvalues of R12[0] [(][R]12[0] [)][T][ are constrained to be] smaller than ρ[2]. As ρ is varied from zero to its maximum value, tr(P [RO]) increases from that of the BLUE given R12 = 0 to that of CI. This result is quite specific but nevertheless verifies the generality of the RO methodology. The examples also demonstrate the generality of the CLUE framework and in particular the usability of : (i) it can be _A_ used to select estimation method, e.g., ICI if (21) holds, (ii) it is the basis for deriving and solving general problems using robust optimization, (iii) and it is used to compute lower and upper bounds on a best CLUE. VIII. CONCLUSIONS A framework for conservative linear unbiased estimation was proposed. The backbone of the framework is Definition 2 where a best conservative linear unbiased estimator (best CLUE) is defined. Lower and upper bounds of a best CLUE were derived. TABLE I SUMMARY OF EXAMPLES _Pu_ _Pl_ _P_ [CI] _P_ [ICI] _P_ [LE] _P_ [RO] _P_ _[⋆]_ E1 1.43I **_I_** 1.60I 1.18I **_I_** **_I_** **_I_** E2a **1.60I 1.60I** **1.60I 1.18I** **_I_** **1.60I** **1.60I** E2b - - 1.60I **1.18I** **_I_** **1.18I** **1.18I** E2c 1.60I **_I_** 1.60I 1.18I **_I_** **_I_** **_I_** E3 1.88I **1.31I** 1.88I - - **1.76I** E4 - � 00..45 040 0..9345 � - - - � 00..40 056 0..9540 �� 00..56 040 0..4095 � **black = CLUE, not best CLUE; green = best CLUE; cyan = lower** bound red = not CLUE; yellow = CLUE, might be best CLUE Quantities not computed are marked ” ” ----- Fig. 9. A summary of the main contributions (green boxes) and suggested future directions to take (orange dashed boxes). Current progress (gray hashed boxes) have been included for clarity. The strength of the proposed framework was further demonstrated as best CLUEs were found in more general settings with robust optimization (RO). Using an example we have illustrated that the RO based approach has the potential to perform better than CI if N > 2. Moreover, it was shown that three existing conservative linear estimation methods in fact are a best CLUE under different assumptions about the cross-correlations. This paper suggests two main directions to take for future work. Special cases of a best CLUE: New methods can be derived and connected to a best CLUE by exploiting structures in the set . Conservative estimation using RO: Synthesizing _A_ new theory on RO, particularly in robust semidefinite pro_gramming (SDP), means it is possible to prove tractability_ and optimality for even more general cases than those already stated, and to describe properties of the solution from the RO. We summarize the contributions of this work and suggested future work in Fig. 9. APPENDIX A MATRIX RELATIONS Let V ⊆ **S[n]+** [and][ A, B][ ∈] **[S]+[n]** [. The inequalities][ ⪰] [and][ ≻] [are] defined as _A ⪰_ _B ⇐⇒_ _A −_ _B ⪰_ 0 ⇐⇒ (A − _B) ∈_ **S[n]+[,]** _A ≻_ _B ⇐⇒_ _A −_ _B ≻_ 0 ⇐⇒ (A − _B) ∈_ **S[n]++[.]** A function J : R[n][×][n] **R is matrix nondecreasing if** _→_ _A_ _B =_ _J(A)_ _J(B),_ (28) _⪯_ _⇒_ _≤_ and matrix increasing if _A_ _B, A_ = B = _J(A) < J(B)._ (29) _⪯_ _̸_ _⇒_ The function tr(WA) is matrix nondecreasing if W ∈ **S[n]+** [and] matrix increasing if W ∈ **S[n]++[. In particular][ tr(][A][)][ is matrix]** increasing on S[n]+[. The function][ det(][A][)][ is matrix increasing] on S[n]+ [[40].] An element A is the minimum element of if _∈V_ _V_ _B_ _A,_ _B_ _._ (30) _⪰_ _∀_ _∈V_ An element A is a minimal element of if B and _∈V_ _V_ _∈V_ _B_ _A_ _B_ _A_ (31) _⪯_ _⇒_ Minimal elements of V ⊆ **S[n]++** [are given by] minimize tr(WB), (32) _B∈V_ where W ∈ **S[n]++[. If all][ W][ ∈]** **[S][n]++** [yields the same unique] solution A, then A is a minimum element of . If two or _V_ more minimal elements of a set exist, then does not have _V_ _V_ a minimum element [40]. Assuming J is matrix increasing, the solution to minimize _J(B),_ (33) _B∈V_ yields a minimal element of . This can be shown by contra_V_ diction. Assume that A solves (33) but that A is not a minimal element of . The latter assumption implies that there exists _V_ another element A[′] for which A[′] _A, A[′]_ = A. It follows _∈V_ _⪯_ _̸_ from (29) that J(A[′]) < J(A) which leads to a contradiction since A then cannot be a solution to (33). Hence, the solution to (33) is a minimal element of . _V_ REFERENCES [1] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation _Theory._ Upper Saddle River, NJ, USA: Prentice Hall, 1993. [2] D. Hall, C.-Y. Chong, J. Llinas, and M. Liggins, Distributed Data Fusion _for Network-Centric Operations._ Boca Raton, FL, USA: CRC Press, 2012. [3] L. J. Gleser, Estimation For A Regression Model With An Unknown _Covariance Matrix._ University of California Press, 1972, pp. 541–568. [4] J. Nygårds, V. Deleskog, and G. Hendeby, “Safe fusion compared to established distributed fusion methods,” in Proceedings of the IEEE In_ternational Conference on Multisensor Fusion, Baden-Baden, Germany,_ Sep. 2016. [5] ——, “Decentralized tracking in sensor networks with varying coverage,” in Proceedings of the 21st IEEE International Conference on _Information Fusion, Cambridge, UK, Jul. 2018._ [6] M. A. Bakr and S. Lee, “Distributed multisensor data fusion under unknown correlation and data inconsistency,” Sensors, vol. 17, no. 11, p. 2472, Oct. 2017. [7] S. J. Julier and J. K. Uhlmann, “A non-divergent estimation algorithm in the presence of unknown correlations,” in Proceedings of the 1997 _American Control Conference, Albuquerque, NM, USA, Jun. 1997, pp._ 2369–2373. [8] J. K. Uhlmann, “Covariance consistency methods for fault-tolerant distributed data fusion,” Information Fusion, vol. 4, no. 3, pp. 201–215, Sep. 2003. [9] A. R. Benaskeur, “Consistent fusion of correlated data sources,” in _Proceedings of the 28th Annual Conference of the IEEE Industrial_ _Electronics Society, Sevilla, Spain, Nov. 2002, pp. 2652–2656._ [10] J. Sijs, M. Lazar, and P. P. J. v. d. Bosch, “State fusion with unknown correlation: Ellipsoidal intersection,” in Proceedings of the 2010 American _Control Conference, Baltimore, MD, USA, Jun. 2010, pp. 3992–3997._ [11] B. Noack, J. Sijs, and U. D. Hanebeck, “Algebraic analysis of data fusion with ellipsoidal intersection,” in Proceedings of the IEEE International _Conference on Multisensor Fusion, Baden-Baden, Germany, Sep. 2016,_ pp. 365–370. [12] B. Noack, J. Sijs, M. Reinhardt, and U. D. Hanebeck, “Decentralized data fusion with inverse covariance intersection,” Automatica, vol. 79, pp. 35–41, May 2017. [13] S. S. Blackman and R. Popoli, Design and Analysis of Modern Tracking _Systems._ Norwood, MA, USA: Artech House, 1999. [14] R. Forsling, “Decentralized estimation using conservative information extraction,” Licentiate Thesis, Linköping University, Linköping, Sweden, Dec. 2020. [15] T. Kailath, A. H. Sayed, and B. Hassibi, Linear Estimation. Upper Saddle River, NJ, USA: Prentice Hall, 2000. [16] Y. Bar-Shalom and L. Campo, “The effect of the common process noise on the two-sensor fused-track covariance,” IEEE Transactions on _Aerospace and Electronic Systems, vol. 22, no. 6, pp. 803–805, Nov._ 1986. [17] L. Chen, P. Arambel, and R. Mehra, “Estimation under unknown correlation: Covariance intersection revisited,” IEEE Transactions on _A t_ _ti C_ _t_ _l_ l 47 1879 1882 N 2002 ----- [18] X. Tian and Y. Bar-Shalom, “Exact algorithms for four track-to-track fusion configurations: All you wanted to know but were afraid to ask,” in _Proceedings of the 12th IEEE International Conference on Information_ _Fusion, Seattle, WA, USA, Jul. 2009._ [19] S. H. Grime and H. F. Durrant-Whyte, “Data fusion in decentralized sensor networks,” Control Engineering Practice, vol. 2, no. 5, pp. 849 – 863, Oct. 1994. [20] R. Olfati-Saber and R. Murray, “Consensus problems in networks of agents with switching topology and time-delays,” IEEE Transactions on _Automatic Control, vol. 49, no. 9, pp. 1520–1533, Sep. 2004._ [21] R. Olfati-Saber, “Distributed Kalman filter with embedded consensus filters,” in Proceedings of the 44th IEEE Conference Decision and _Control, Sevilla, Spain, Dec. 2005._ [22] L. Xiao, S. Boyd, and S. Lall, “A scheme for robust distributed sensor fusion based on average consensus,” in Proceedings of the 4th _International Symposium on Information Processing in Sensor Networks_ _(IPSN), Apr. 2005, pp. 63–70._ [23] G. Battistelli, L. Chisci, G. Mugnai, A. Farina, and A. Graziano, “Consensus-based linear and nonlinear filtering,” IEEE Transactions on _Automatic Control, vol. 60, no. 5, pp. 1410–1415, Sep. 2015._ [24] U. A. Khan and J. M. F. Moura, “Distributing the Kalman filter for large-scale systems,” IEEE Transactions on Signal Processing, vol. 56, no. 10, pp. 4919–4935, Oct. 2008. [25] F. Govaers and W. Koch, “Distributed Kalman filter fusion at arbitrary instants of time,” in Proceedings of the 13th IEEE International Con_ference on Information Fusion, Edinburgh, Scotland, Jul. 2010, pp. 1–8._ [26] M. Reinhardt, B. Noack, and U. D. Hanebeck, “The hypothesizing distributed Kalman filter,” in Proceedings of the IEEE International _Conference on Multisensor Fusion, Hamburg, Germany, Sep. 2012._ [27] S. J. Julier and J. K. Uhlmann, “General decentralized data fusion with covariance intersection,” in Handbook of Multisensor Data Fusion: _Theory and Practice, M. Liggins, D. Hall, and J. Llinas, Eds._ Boca Raton, FL, USA: CRC Press, 2009, ch. 14. [28] L. Chen, P. Arambel, and R. Mehra, “Fusion under unknown correlation: Covariance intersection as a special case,” in Proceedings of the 5th _IEEE International Conference on Information Fusion, Annapolis, MD,_ USA, Jul. 2002. [29] M. Reinhardt, B. Noack, P. O. Arambel, and U. D. Hanebeck, “Minimum covariance bounds for the fusion under unknown correlations,” IEEE _Signal Processing Letters, vol. 22, no. 9, pp. 1210–1214, Sep. 2015._ [30] B. Noack, J. Sijs, and U. D. Hanebeck, “Inverse covariance intersection: New insights and properties,” in Proceedings of the 20th IEEE Interna_tional Conference on Information Fusion, Xi’an, China, Jul. 2017._ [31] J. Ajgl and O. Straka, “Inverse covariance intersection fusion of multiple estimates,” in Proceedings of the 23rd IEEE International Conference _on Information Fusion, Virtual Conference, Jul. 2020._ [32] Z. Wu, Q. Cai, and M. Fu, “Covariance intersection for partially correlated random vectors,” IEEE Transactions on Automatic Control, vol. 63, no. 3, pp. 619–629, Mar. 2018. [33] J. Ajgl and O. Straka, “Analysis of partial knowledge of correlations in an estimation fusion problem,” in Proceedings of the 21st IEEE _International Conference on Information Fusion, Cambridge, UK, Jul._ 2018. [34] ——, “Comparison of fusions under unknown and partially known correlations,” IFAC-PapersOnLine, vol. 51, no. 23, pp. 295–300, 2018. [35] Y. Gao, X. R. Li, and E. Song, “Robust linear estimation fusion with allowable unknown cross-covariance,” IEEE Transactions on Systems, _Man, and Cybernetics: Systems, vol. 46, no. 9, pp. 1314–1325, 2016._ [36] A. Sayed, “A framework for state-space estimation with uncertain models,” IEEE Transactions on Automatic Control, vol. 46, no. 7, pp. 998–1013, 2001. [37] E. Delage and Y. Ye, “Distributionally robust optimization under moment uncertainty with application to data-driven problems,” Operations _Research, vol. 58, no. 3, pp. 595–612, 2010._ [38] S. Shafieezadeh-Abadeh, V. A. Nguyen, D. Kuhn, and P. M. Esfahani, “Wasserstein distributionally robust Kalman filtering,” in NeurIPS, Sep. 2018, pp. 8483–8492. [39] S. Wang and Z.-S. Ye, “Distributionally robust state estimation for linear systems subject to uncertainty and outlier,” IEEE Transactions on Signal _Processing, vol. 70, pp. 452–467, Dec. 2022._ [40] S. Boyd and L. Vandenberghe, Convex Optimization. New York, NY, USA: Cambridge University Press, 2004. [41] U. Orguner, “Approximate analytical solutions for the weight optimization problems of CI and ICI,” in 2017 Sensor Data Fusion: Trends, _Solutions, Applications (SDF), 2017, pp. 1–6._ [42] J. H. Taylor, “The Cramér-Rao estimation error lower bound comt ti f d t i i ti li t ” IEEE T _ti_ _Automatic Control, vol. 24, no. 2, pp. 343–344, April 1979._ [43] U. Hanebeck, K. Briechle, and J. Horn, “A tight bound for the joint covariance of two random vectors with unknown but constrained crosscorrelation,” in Proceedings of the IEEE International Conference on _Multisensor Fusion, Baden-Baden, Germany, Aug. 2001, pp. 85–90._ [44] R. Forsling, Z. Sjanic, F. Gustafsson, and G. Hendeby, “Consistent distributed track fusion under communication constraints,” in Proceed_ings of the 22nd IEEE International Conference on Information Fusion,_ Ottawa, Canada, Jul. 2019. [45] J. Ajgl and O. Straka, “Lower bounds in estimation fusion with partial knowledge of correlations,” in Proceedings of the IEEE International _Conference on Multisensor Fusion, Karlsruhe, Germany, Sep. 2021._ [46] A. Ben-Tal, L. E. Ghaoui, and A. Nemirovski, Robust Optimization. Princeton, NJ, USA: Princeton University Press, 2009. [47] G. O. Corrêa and A. Talavera, “Competitive robust estimation for uncertain linear dynamic models,” IEEE Transactions on Signal Processing, vol. 65, no. 18, pp. 4847–4861, Sep. 2017. [48] L. E. Ghaoui, F. Oustry, and H. Lebgret, “Robust solutions to uncertain semidefinite programs,” SIAM J. on Optimization, vol. 9, no. 1, pp. 33– 52, May 1998. [49] J. Löfberg, “YALMIP: A toolbox for modeling and optimization in MATLAB,” in Proceedings of the CACSD Conference, vol. 3, Taipei, Taiwan, Sep. 2004. [50] A. Ben-Tal and A. Nemirovski, “Robust optimization - Methodology and applications,” Math. Program., vol. 92, pp. 453–480, May 2002. [51] G. H. Golub and C. F. van Loan, Matrix Computations, 4th ed. Baltimore, MD, USA: The Johns Hopkins University Press, 2013. [52] A. Ben-Tal and A. Nemirovski, “Robust convex optimization,” Mathe_matics of Operations Research, vol. 23, no. 4, pp. 769–805, Nov. 1998._ [53] J. Löfberg, “Automatic robust convex programming,” Optimization _Methods and Software, vol. 27, no. 1, pp. 115–129, Feb. 2012._ [54] J. Löfberg, “Support for robust conic-conic optimization in YALMIP,” unpublished. [55] F. Gustafsson, Statistical Sensor Fusion. Lund, Sweden: Studentlitteratur, 2018. [56] J. Sijs and M. Lazar, “Empirical case-studies of state fusion via ellipsoidal intersection,” in Proceedings of the 17th IEEE International _Conference on Information Fusion, Salamanca, Spain, Jul. 2014, pp._ 1–8. [57] Y. Zhou and J. Li, “Data fusion of unknown correlations using internal ellipsoidal approximation,” in Proceedings of the 17th Triennial IFAC _World Congress, Seoul, Korea, Jul. 2008, pp. 2856 – 2860._ **Robin Forsling received the M.Sc. degree in en-** gineering physics in 2016 from Umeå University, Umeå, Sweden, and the Lic.Eng. degree in automatic control in 2021 from Linköping University, Linköping, Sweden. He is currently working toward the Ph.D. degree with the Division of Automatic Control, Linköping University. His main research interest is decentralized estimation with a particular focus on conservative estimation methods and communication reducing techniques. Since 2016 he is employed at Saab Aeronautics in Linköping, Sweden, where he has been working as a Systems Engineer with target acquisition systems, decision support systems and sensor f i ----- **Anders Hansson was born in Trelleborg, Sweden,** in 1964. He received the Master of Science degree in Electrical Engineering in 1989, the Degree of Licentiate of Engineering in Automatic Control in 1991, and the PhD in Automatic Control in 1995, all from Lund University, Lund, Sweden. During the academic year 1992-1993 he spent six months at Imperial College in London, UK. From 1995 until 1997 he was a postdoctoral student, and from 1997 until 1998 a research associate at the Information Systems Lab, Department of Electrical Engineering, Stanford University. In 1998 he was appointed assistant professor and in 2000 associate professor at S3-Automatic Control, Royal Institute of Technology, Stockholm, Sweden. In 2001 he was appointed associate professor at the Division of Automatic Control, Linköping University. From 2006 he is full professor at the same department. Anders Hansson is a senior member of the IEEE. During 2006-2007 he was an associate editor of the IEEE Transactions on Automatic Control. He was a member of the EUCA council 20092015. Currently he is a member of the EUCA General Assembly and of the Technical Committee on Systems with Uncertainty of the IEEE Control Systems Society. His research interests are within the fields of optimal control, stochastic control, linear systems, signal processing, applications of control and telecommunications. He got the SAAB-Scania Research Award in 1992. **Fredrik Gustafsson is professor in Sensor In-** formatics at Department of Electrical Engineering, Linköping University, since 2005. He received the M.Sc. degree in electrical engineering 1988 and the Ph.D. degree in Automatic Control, 1992, both from Linköping University. He was an associate editor for IEEE Transactions of Signal Processing 2000-2006, IEEE Transactions on Aerospace and Electronic Systems 2010-2012, and EURASIP Journal on Applied Signal Processing 2007-2012. He was awarded the Arnberg prize by the Royal Swedish Academy of Science (KVA) 2004, elected member of the Royal Academy of Engineering Sciences (IVA) 2007, and elevated to IEEE Fellow 2011. In 2014, he was awarded a Distinguished Professor grant from the Swedish Research Council. He was an adjunct entrepreneurial professor at Twente University 2012-2013. He was awarded the Harry Rowe Mimno Award 2011 for the tutorial "Particle Filter Theory and Practice with Positioning Applications", which was published in the AESS Magazine in July 2010. He was a co-author of "Smoothed state estimates under abrupt changes using sum-of-norms regularization" that received the Automatica paper prize in 2014. He is a co-founder of the companies NIRA Dynamics (automotive safety), Softube (digital music production tools), and Senion (indoor navigation). **Zoran Sjanic received the M.Sc. degree in Com-** puter Science and Engineering in 2002 and the Ph.D. degree in automatic control in 2013, both from Linköping University, Linköping, Sweden. He has been employed by Saab Aeronautics in Linköping, Sweden, since 2001, where he works in the Decision Support Department as a Principal Systems Engineer and Technical Manager for the Image processing and Analysis technical area. Since 2020 he is also Adjunct Associate Professor in the division of Automatic Control, Department of Electrical Engineering, Linköping University. His main research interests are sensor fusion for navigation of manned and unmanned aircraft, radar systems, simultaneous localisation and mapping, distributed estimation and nonlinear estimation methods. **Johan Löfberg received the M.Sc. degree in me-** chanical engineering in 1998 and the Ph.D. degree in automatic control in 2003, both from Linköping University, Linköping, Sweden. After a postdoctoral stay at ETH Zürich from 2003–2006, he now serves as Associate Professor and Docent in the division of Automatic Control, Department of Electrical Engineering, Linköping University. His main research interest is aspects of optimization in control and systems theory, with a particular interest in model predictive control. Driven by applications in control, he is also more generally interested in robust optimization and optimization modelling. He is the author of the MATLAB toolbox YALMIP which has become an important tool for researchers and engineers in many domains. **Gustaf Hendeby (S’04-M’09-SM’17) received the** M.Sc. degree in applied physics and electrical engineering in 2002 and the Ph.D. degree in automatic control in 2008, both from Linköping University, Linköping, Sweden. He is Associate Professor and Docent in the division of Automatic Control, Department of Electrical Engineering, Linköping University. He worked as Senior Researcher at the German Research Center for Artificial Intelligence (DFKI) 2009–2011, and Senior Scientist at Swedish Defense Research Agency (FOI) and held an adjunct Associate Professor position at Linköping University 2011–2015. His main research interests are stochastic signal processing and sensor fusion with applications to nonlinear problems, target tracking, and simultaneous localization and mapping (SLAM), and is the author of several published articles and conference papers in the area. He has experience of both theoretical analysis as well as implementation aspects. Dr. Hendeby is since 2018 an Associate Editor for IEEE Transactions on Aerospace and Electronic Systems in the area of target tracking and multisensor systems. In 2022 he served as general chair for the 25th IEEE International Conference on Information Fusion (FUSION) in Linköping, Sweden. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/TSP.2022.3179841?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/TSP.2022.3179841, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "other-oa", "status": "GREEN", "url": "https://liu.diva-portal.org/smash/get/diva2:1690215/FULLTEXT01" }
2,022
[ "JournalArticle" ]
true
null
[]
25,735
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00f0106e0a9ab64dcd4c0dabe5f1af1e5549f75e
[ "Computer Science" ]
0.909186
Channel Load Aware AP / Extender Selection in Home WiFi Networks Using IEEE 802.11k/v
00f0106e0a9ab64dcd4c0dabe5f1af1e5549f75e
IEEE Access
[ { "authorId": "2393355", "name": "Toni Adame" }, { "authorId": "145167737", "name": "Marc Carrascosa" }, { "authorId": "2825302", "name": "B. Bellalta" }, { "authorId": "2612849", "name": "I. Pretel" }, { "authorId": "83042799", "name": "Iñaki Etxebarria" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "http://ieeexplore.ieee.org/servlet/opac?punumber=6287639" ], "id": "2633f5b2-c15c-49fe-80f5-07523e770c26", "issn": "2169-3536", "name": "IEEE Access", "type": "journal", "url": "http://www.ieee.org/publications_standards/publications/ieee_access.html" }
Next-generation Home WiFi networks have to step forward in terms of performance. New applications such as on-line games, virtual reality or high quality video contents will further demand higher throughput levels, as well as low latency. Beyond physical (PHY) and medium access control (MAC) improvements, deploying multiple access points (APs) in a given area may significantly contribute to achieve those performance goals by simply improving average coverage and data rates. However, it opens a new challenge: to determine the best AP for each given station (STA). This article studies the achievable performance gains of using secondary APs, also called Extenders, in Home WiFi networks in terms of throughput and delay. To do that, we introduce a centralized, easily implementable channel load aware selection mechanism for WiFi networks that takes full advantage of IEEE 802.11k/v capabilities to collect data from STAs, and distribute association decisions accordingly. These decisions are completely computed in the AP (or, alternatively, in an external network controller) based on an AP selection decision metric that, in addition to RSSI, also takes into account the load of both access and backhaul wireless links for each potential STA-AP/Extender connection. Performance evaluation of the proposed channel load aware AP and Extender selection mechanism has been first conducted in a purpose-built simulator, resulting in an overall improvement of the main analyzed metrics (throughput and delay) and the ability to serve, at least, 35% more traffic while keeping the network uncongested when compared to the traditional RSSI-based WiFi association. This trend was confirmed when the channel load aware mechanism was tested in a real deployment, where STAs were associated to the indicated AP/Extender and total throughput was increased by 77.12%.
Received January 29, 2021, accepted February 11, 2021, date of publication February 15, 2021, date of current version February 25, 2021. _Digital Object Identifier 10.1109/ACCESS.2021.3059473_ # Channel Load Aware AP / Extender Selection in Home WiFi Networks Using IEEE 802.11k/v TONI ADAME 1, MARC CARRASCOSA1, BORIS BELLALTA 1, IVÁN PRETEL2, AND IÑAKI ETXEBARRIA[2] 1Department of Information and Communication Technologies, Universitat Pompeu Fabra, 08018 Barcelona, Spain 2FON Labs, 48009 Bilbao, Spain Corresponding author: Toni Adame (toni.adame@upf.edu) This work was supported in part by the Spanish government under Project CDTI IDI-20180274, Project WINDMAL PGC2018-099959-B-100 (MCIU/AEI/FEDER,UE), and Project TEC2016-79510-P; and in part by the Catalan government under Project SGR-2017-1188 and Project SGR-2017-1739. **ABSTRACT Next-generation Home WiFi networks have to step forward in terms of performance. New** applications such as on-line games, virtual reality or high quality video contents will further demand higher throughput levels, as well as low latency. Beyond physical (PHY) and medium access control (MAC) improvements, deploying multiple access points (APs) in a given area may significantly contribute to achieve those performance goals by simply improving average coverage and data rates. However, it opens a new challenge: to determine the best AP for each given station (STA). This article studies the achievable performance gains of using secondary APs, also called Extenders, in Home WiFi networks in terms of throughput and delay. To do that, we introduce a centralized, easily implementable channel load aware selection mechanism for WiFi networks that takes full advantage of IEEE 802.11k/v capabilities to collect data from STAs, and distribute association decisions accordingly. These decisions are completely computed in the AP (or, alternatively, in an external network controller) based on an AP selection decision metric that, in addition to RSSI, also takes into account the load of both access and backhaul wireless links for each potential STA-AP/Extender connection. Performance evaluation of the proposed channel load aware AP and Extender selection mechanism has been first conducted in a purpose-built simulator, resulting in an overall improvement of the main analyzed metrics (throughput and delay) and the ability to serve, at least, 35% more traffic while keeping the network uncongested when compared to the traditional RSSI-based WiFi association. This trend was confirmed when the channel load aware mechanism was tested in a real deployment, where STAs were associated to the indicated AP/Extender and total throughput was increased by 77.12%. **INDEX TERMS Home WiFi, AP selection, extender, load balancing, IEEE 802.11k, IEEE 802.11v.** **I. INTRODUCTION** Since their appearance more than 20 years ago, IEEE 802.11 wireless local area networks (WLANs) have become the worldwide preferred option to provide wireless Internet access to heterogeneous clients in homes, businesses, and public spaces due to their low cost and mobility support. The simplest WLAN contains only a basic service set (BSS), consisting of an access point (AP) connected to a wired infrastructure, and some wireless stations (STAs) associated to the AP. The associate editor coordinating the review of this manuscript and approving it for publication was Kashif Saleem . The increase of devices aiming to use the WLAN technology to access Internet has been accompanied by more demanding user requirements, especially in entertainment contents: on-line games, virtual reality, and high quality video. In consequence, traditional single-AP WLANs deployed in apartments, i.e., Home WiFi networks, may fail to deliver a satisfactory experience due to the existence of areas where the received power from the AP is low, and so the achievable performance [1]. Although IEEE 802.11ac (WiFi 5) [2], IEEE 802.11ax (WiFi 6) [3], [4], and IEEE 802.11be (WiFi 7) [5], [6] amendments provide enhancements on physical (PHY) and medium access control (MAC) protocols that may increase the WLAN efficiency, and also increase the coverage by using ----- beamforming, the best solution is still to deploy more APs to improve the coverage in those areas. In multi-AP deployments, normally only one AP (the main AP) has Internet access, and so the other APs (from now on simply called Extenders) must relay the data to it using a wired or wireless backhaul network. Since presuming the existence of a wired network is not always feasible, Extenders communicate with the main AP wirelessly. In this case, both the main AP and Extenders are equipped with at least two radios, usually operating at different bands. In presence of multiple AP/Extenders, a new challenge appears: how to determine the best AP/Extender for each given STA. According to the default WiFi AP selection mechanism, an STA that receives beacons from several AP/Extenders will initiate the association process with the AP/Extender with the highest received signal strength indicator (RSSI) value. Though simple and easy to implement, this mechanism omits any influence of traffic load and, consequently, can lead to network congestion and low throughput in scenarios with a high number of STAs [7]. Many research activities have already widely tackled the AP selection process in an area commonly referred to as load _balancing, whose goal is to distribute more efficiently STAs_ among the available AP/Extenders in a WLAN. Although multiple effective strategies have been proposed in the literature, most of them lack the prospect of real implementation, as they require changes in the existing IEEE 802.11 standards and/or in STAs’ wireless cards. The channel load aware AP/Extender selection mechanism presented in this article sets out to enhance the overall WLAN performance by including the effect of the channel load into the STA association process. To do it, only already developed IEEE 802.11 amendments are considered: IEEE 802.11k to gather information from AP/Extenders in the WLAN, and IEEE 802.11v to notify each STA of its own prioritized list of AP/Extenders. Particularly, the main contributions of the current work can be summarized into: - Review and classification of multiple existing AP/Extender selection mechanisms, and some background information on the use of IEEE 802.11k/v. - Design of a feasible, practical, and flexible channel load _aware AP/Extender selection mechanism supported by_ IEEE 802.11k/v amendments. - Evaluation of the channel load aware AP/Extender selection mechanism by simulation, studying the performance gains of using Extenders along with the proposed solution. We focus on understanding how the number of Extenders and their position, the fraction of STAs supporting IEEE 802.11k/v, and the load of the access and backhaul links, impact on the system performance in terms of throughput and delay. - Validation of the presented solution in a real testbed, showing the same trends in terms of performance improvements that those obtained by simulation. Lastly, the main lessons that can be learned from this article are listed below : 1) Placement of Extenders: We observe that Extenders must be located at a distance (in RSSI terms) large enough to stimulate the association of farther STAs while maintaining high data rate in its backhaul connection to the AP. Also, we confirm that connecting Extenders through other Extenders not only increases the network coverage, but also the network’s operational range in terms of admitted traffic load. 2) Load of access vs. backhaul links: The relative weight of the load of the access and backhaul(s) link(s) should be generally balanced, without dismissing a proper tuning according to the characteristics of the deploying scenario. 3) STAs supporting IEEE 802.11k/v: We observe that, even for a low fraction of STAs supporting IEEE 802.11k/v, the gains of using the channel load aware AP/Extender selection mechanism are beneficial for the overall network. 4) Throughput and delay improvements: The use of Extenders allows to balance the load of the network, which results in significant gains in throughput and delay for much higher traffic loads. Therefore, the use of Extenders is recommended for high throughput multimedia and delay-constrained applications. The remainder of this article is organized as follows: Section II offers an overview on AP selection in WiFi networks. Section III elaborates on IEEE 802.11k and IEEE 802.11v amendments, paying special attention to the features considered in the proposed channel load aware AP/Extender selection mechanism, which is in turn described in Section IV. Performance results obtained from simulations and real deployments are compiled in Section V and Section VI, respectively. Lastly, Section VII discusses open challenges in future Home WiFi networks and Section VIII presents the obtained conclusions. **II. USE OF EXTENDERS AND AP/EXTENDER SELECTION** **MECHANISMS IN WiFi NETWORKS** The current section reviews the main aspects of the technical framework involving the use of Extenders in next-generation Home WiFi networks, such as the main challenges related to their deployment, the existing options to integrate them into the STA association procedure, and their management through an external platform. _A. MULTI-HOP COMMUNICATION IN WLANs_ The need to expand WLAN coverage to every corner of a targeted area can be satisfied by increasing the AP transmission power or by deploying wired/wireless Extenders. Putting aside the wired option, which is not in scope of the current article, wireless extension of a WLAN can be achieved by means of a wireless mesh network (WMN). ----- In a WMN, multiple deployed APs communicate among them in a multi-hop scheme to relay data from/to STAs. The most representative initiative in this field is IEEE 802.11s, which integrates mesh networking services and protocols with IEEE 802.11 at the MAC layer [8]. Wireless frame forwarding and routing capabilities are managed by the hybrid wireless mesh protocol (HWMP), which combines the flexibility of on-demand route discovery with efficient proactive routing to a mesh portal [9]. As traffic streams in a WMN are mainly oriented towards/from the main AP, they tend to form a tree-based wireless architecture [10]. This architecture strongly relies on the optimal number and position of deployed Extenders, which is determined in [11] as a function of PHY layer parameters with the goal of minimizing latency and maximizing data rate. This analysis is extended in [12], where a model based on PHY and MAC parameters returns those Extender locations that maximize multi-hop throughput. Other approaches such as [13] go far beyond and propose the use of Artifical Intelligence to enable autonomous self-deployment of wireless Extenders. Relaying capabilities of Extenders are also a matter of study, as in [14], where an algorithm is proposed to determine the optimal coding rate and modulation scheme to dynamically control the best band and channel selection. Or in [15], where a low latency relay transmission scheme for WLAN is proposed to simultaneously use multiple frequency bands. All in all, once the number, location and relaying capabilities of Extenders operating in a WLAN are selected, the way in which STAs determine their own parent (i.e., the best AP/Extender located within their coverage area) can impact the overall performance of the network. We will discuss on this issue in the following lines. _B. AP/EXTENDER SELECTION MECHANISMS_ A review of the currently existing AP/Extender selection mechanisms along with the description of the WiFi scanning modes that enable them is offered in the following lines. 1) WiFi SCANNING MODES IEEE 802.11 standard defines two different scanning modes: _passive and active [16]. In passive scanning, for each avail-_ able radio channel, the STA listens to beacons sent by APs for a dwell time. As beacons are usually broadcast by the AP every 100 ms, channel dwell time is typically set to 100-200 ms to guarantee beacon reception [17], [18]. In active scanning, the STA starts broadcasting a probe _request frame on one channel and sets a probe timer. If no_ _probe response is received before the probe timer reaches_ _MinChannelTime, the STA assumes that no AP is working in_ that channel and scans another channel alternatively. Otherwise, if the STA does receive a probe response, it will further wait for responses from other working APs until MaxChan_nelTime is reached by the probe timer. MinChannelTime and_ _MaxChannelTime values are vendor-specific, as they are not_ specified by the IEEE 802.11 standard. Indeed, the obtention of optimum values to minimize the active scanning phase have attracted research attention. In [19], for instance, the author sets these values as low as 6-7 ms and 10-15 ms, respectively. Since passive scanning always has longer latency than _active scanning, wireless cards tend to use the latter to rapidly_ find nearby APs [20]. However, active scanning has three disadvantages: 1) it consumes significant more energy than _passive scanning, 2) it is unable to discover networks that do_ not broadcast their SSID, and 3) it may result in shorter scan ranges because of the lower power level of STAs. It is also usual that mobile STAs periodically perform _active background scanning to discover available APs, and_ then accelerate an eventual roaming operation [21]. In this case, the STA (already associated to an AP and exchanging data) goes periodically off-channel and sends probe requests across other channels. On the other hand, the active on-roam _scanning only occurs after the STA determines a roam is_ necessary. 2) DEFAULT WiFi AP SELECTION MECHANISM Regardless the scanning mode used by an STA to complete its own list of available APs, and the final purpose of this scanning (i.e., the initial association after the STA startup or a roaming operation), the STA executes the default WiFi AP selection mechanism (from now on also named RSSI-based) by choosing the AP of the previous list with the strongest RSSI. This is the approach followed by common APs and available multi-AP commercial solutions, like Google WiFi [22] or Linksys Velop [23], which are especially indicated for homes with coverage problems and few users. In addition, these two solutions also integrate the IEEE 802.11k/v amendments (analyzed later on in Section III), but only to provide faster and seamless roaming. The strongest RSSI might indicate the best channel condition between the STA and the AP. However, only relying on this criteria is not always the best choice, as it can lead to imbalanced loads between APs, inefficient rate selection, and selection of APs with poor throughput, delay, and other performance metrics [24]. 3) ALTERNATIVE AP/EXTENDER SELECTION MECHANISMS The inefficiency of the RSSI-based AP selection mechanism has motivated the emergence of alternative methods that take into account other metrics than solely the RSSI. The most representative examples are compiled in Table 1 and classified according to three different criteria: the AP selection mode, the architecture employed, and the selected decision metric: - AP selection mode: In the active AP selection, the STA considers all potential APs and gathers information regarding one or more performance metrics to make a decision. In [25], the STA scans for all available APs, quickly associates to each, and even runs a set of tests to estimate Internet connection quality. On the contrary, ----- **TABLE 1. Classification of alternative AP/Extender selection mechanisms.** Whereas mechanisms employing to some extent IEEE 802.11k are marked with †, no mechanism employs IEEE 802.11v. (By default, parameters from the decision metric column refer to the STA’s value). the passive AP selection is based on the information that the STA directly extracts from beacon frames or deduces from their physical features, such as the experienced delay in [34]. Lastly, in the hybrid AP selection, the network makes use of the information shared by the STA to give advice on the best possible potential AP. In [39], for instance, clients automatically submit reports on the APs that they use with regard to estimated backhaul capacity, ports blocked, and connectivity failures. - Architecture: This category splits the different mechanisms into decentralized and centralized. Decentral_ized mechanisms are those in which the STA selects its_ AP based on its available information (even combining cross-layer information, as in [38]). On the other hand, _centralized mechanisms imply a certain degree of coor-_ dination between different APs thanks to a central entity (that may well be an SDN controller, as in [30]) intended to balance overall network load. - Decision metric: The AP selection metric can be determined by a single parameter (e.g., AP load in [28]) or a weighted combination of some of them (e.g., throughput and channel occupancy rate in [33]). Apart from RSSI, there exists a vast quantity of available magnitudes for this purpose; however, the most common ones in the reviewed literature are throughput, load, and delay. Furthermore, there exist some novel approaches that have introduced machine learning (ML) techniques into the AP selection process. For instance, in [40] a decentralized cognitive engine based on a neural network trained on past link conditions and throughput performance drives the AP selection process. Likewise, a decentralized approach based on the exploration-exploitation trade-off from Reinforcement Learning algorithms is used in [41], [42]. Under that system, STAs learn the network conditions and associate to the AP that maximizes their throughput. In consequence, STAs stop its exploration, which is only resumed when there is a change in network’s topology. Another decentralized ML-based approach is proposed in [43], where the AP selection mechanism is formulated as a non-cooperative game in which each STA tries to maximize its throughput. Then, an adaptive algorithm based on no-regret learning makes the system converge to an equilibrium state. _C. COMMERCIAL WLAN MANAGEMENT PLATFORMS_ Centralized network management platforms are commonly used in commercial solutions, as they give full control of the network to the operator. These management platforms focus not only on the AP selection, but also cover several network performance enhancements such as channel and band selection, and transmit power adjustment. Nighthawk Mesh WiFi 6 System [44] intelligently selects the fastest WiFi band for every connected STA, and Insight Management Solution [45] recalculates the optimum channel and transmit power for all the APs every 24 hours. Based on signal strength and channel utilization metrics, ArubaOS network operating system has components (i.e. AirMatch [46] and ClientMatch [47]) which dynamically balance STAs across channels and encourage dual-band capable STAs to stay on the 5GHz band on dual-band APs. Lastly, Cognitive Hotspot Technology (CHT) [48] is a multi-platform software that can be installed on a wide range of APs. It brings distributed intelligence to any WiFi network to control the radio resources including AP automatic channel selection, load balancing, as well as client and band steering for STAs. The channel load aware AP/Extender selection mechanism presented in this work could be easily integrated in these centralized platforms and even be further enhanced by exploiting the know-how gathered from different Home WiFi networks. **III. IEEE 802.11k/V AMENDMENTS** The constant evolution of the IEEE 802.11 standard has been fostered by the incremental incorporation of technical amendments addressing different challenges in the context of WLANs. In particular, the optimization of the AP selection process and the minimization of the roaming interruption time are tackled in two different amendments: IEEE 802.11k and IEEE 802.11v [49]. _A. IEEE 802.11k: RADIO RESOURCE MEASUREMENT_ The IEEE 802.11k amendment on radio resource measurement [50] defines methods for information exchange about the radio environment between APs and STAs. This information may be thus used for radio resource management ----- strategies, making devices more likely to properly adapt to the dynamic radio environment. Radio environment information exchange between two devices running IEEE 802.11k occurs through a two-part frame request/report exchange carried within radio measurement report frames (i.e., a purpose-specific category of action frames). Despite the wide set of possible measurements, the AP/Extender selection mechanism presented in this work will only consider beacon and channel load reports. The beacon request/report pair enables an AP to ask an STA for the list of APs it is able to listen effectively to on a specified channel or channels. The request also includes the measurement mode that should be performed by the targeted STA: active scanning (i.e., information comes from _probe responses), passive scanning (i.e., information comes_ from beacons), or beacon table (i.e., use of previously stored beacon information). Whenever an STA receives a beacon request, it creates a new beacon report containing the BSSID, operating frequency, channel number, and RSSI (among other parameters) of each detected AP within its range during the measurement duration specified in the beacon request. At the end of the measurement duration, the STA will send a beacon report with all the aforementioned gathered information. Similarly, the channel load request/report exchange allows an AP to receive information on the channel condition of a targeted network device. The channel load report contains the channel number, actual measurement start time, measurement duration, and channel busy fraction [51]. _B. IEEE 802.11v: WIRELESS NETWORK MANAGEMENT_ The IEEE 802.11v amendment [52] on wireless network management uses network information to influence client roaming decisions. Whereas IEEE 802.11k only targets the radio environment, IEEE 802.11v includes broader operational data regarding network conditions, thus allowing STAs to acquire better knowledge on the topology and state of the network. In fact, there are a multitude of new services powered by IEEE 802.11v, including power saving mechanisms, interference avoidance mechanisms, fast roaming, or an improved location system, among others. In all cases, the exchange of data among network devices takes place through several action frame formats defined for wireless network management purposes. The BSS transition management service is of special interest to our current work, as it enables to suggest a set of preferred candidate APs to an STA according to a pre-established policy. IEEE 802.11v defines 3 types of BSS transition man_agement frames: query, request, and response._ - A query is sent by an STA asking for a BSS transition _candidate list to its corresponding AP._ - An AP responds to a query frame with the BSS transition _candidate list; that is, a request frame containing a pri-_ oritized list of preferred APs, their operating frequency, and their channel number, among other information. In fact, the AP may also send a BSS transition candidate _list to a compatible IEEE 802.11v STA at any time to_ accelerate any eventual roaming process. - A response frame is sent by the STA back to the AP, informing whether it accepts or denies the transition. Once received a BSS transition candidate list and accepted its proposed transition, the STA will follow the provided APs candidate list in order of priority, trying to reassociate to such a network. As operating frequency and channel number of each candidate AP is also provided, total scan process time in the reassociation operation can be minimized. **IV. CHANNEL LOAD AWARE AP/EXTENDER SELECTION** We introduce in this section the proposed channel load aware AP/Extender selection mechanism. We aim to define a general approach that allows us to study the trade-off between received power and channel load-based metrics to make the AP/Extender selection decision. The proposed AP/Extender selection mechanism is intended to be applied on a WLAN topology like the one from Figure 1, consisting of an AP, several Extenders wirelessly connected to the AP, and multiple STAs willing to associate to the network.[1] It is fully based on the existing IEEE 802.11k/v amendments, which enables its real implementation, and can be executed as part of the association process of an STA in any of the following circumstances: - An STA has just associated to the network through the AP/Extender selected by using the default RSSI-based criteria. - An STA is performing a roaming procedure between different AP/Extenders from the same WLAN. - The AP initiates an operation to reassociate all previously associated STAs in case network topology has changed (e.g., a new Extender is connected), or an overall load balance operation is executed (e.g., as consequence of new traffic demands coming from STAs). In a real implementation, all computation associated with this mechanism would be executed in the AP, as it is the single, centralized element in the architecture with a global vision of the network. Alternatively, computation tasks could be assumed by an external network controller run into a server, either directly connected to the AP or placed in a remote, cloud-based location. _A. OPERATION OF THE AP/EXTENDER SELECTION_ _MECHANISM_ The channel load aware AP/Extender selection mechanism splits the selection process into four differentiated stages. Figure 2 shows the sequence of their main tasks, which are described in the following lines: 1) Initial association (IEEE 802.11) 1If Extenders were connected to the AP by means of wired links, the proposed channel load aware mechanism would be likewise applicable. ----- **FIGURE 1. WLAN topology with Extenders. Note that M(ai** **_,j ) corresponds to the access link metric from STA_** _i to AP/Extender j. As for M(bj ), it corresponds to the backhaul link metric of Extender j_ . 2) Collection and exchange of information (IEEE 802.11k) - The AP (or the network controller) initiates a new information collection stage by sending (directly or through the corresponding Extenders) a beacon _request to the STA._ - Depending on the type of the beacon request received, the STA initiates an active scanning, a _passive scanning, or simply consults its own bea-_ _con table._ - The surrounding AP/Extenders respond to an _active scanning with a probe response or simply_ emit their own beacon frames. - The STA transmits the resulting beacon report to its corresponding AP/Extender. - The AP/Extender, in turn, retransmits this beacon _report to the AP (or the network controller)._ - Lastly, the AP emits channel load requests to the network Extenders. - The Extenders measure the observed channel occupation and send a channel load report to the AP. 3) Computation and transmission of decision (IEEE 802.11v) **FIGURE 2. Sequence diagram of the channel load aware AP/Extender** selection mechanism, using active scanning as measurement mode and explicit channel load request/report exchange. - After an active or passive scanning, the STA sends an association request to the AP/Extender with the best observed RSSI value. - The AP/Extender registers the new STA and confirms its association. Moreover, it checks if the STA supports IEEE 802.11k and IEEE 802.11v modes, which are indispensable to properly perform the next steps of the mechanism. - The AP/Extender notifies the AP (or the network controller) of the new associated STA and its capabilities. - The AP (or the network controller) computes the _Yi,j decision metric (defined in the next subsection)_ for each AP/Extender detected by the STA. - The AP sends the STA the resulting BSS transition _candidate list._ 4) Reassociation (IEEE 802.11) - The STA starts a new association process with the first AP/Extender recommended in the BSS _transition candidate list. If it fails, the STA tries_ to associate to the next AP/Extender in the list. - The new AP/Extender registers the new STA and confirms its association. - The new AP/Extender notifies the AP of the new associated STA. ----- **FIGURE 4. RSSI weighting applied in the channel load aware** AP/Extender selection mechanism. metric by combining parameters from both access and backhaul links. More specifically, Yi,j is the decision metric employed in our proposal per each pair formed by STA i and AP/Extender _j. Then, the best AP/Extender for STA i will be the one with_ the minimum Yi,j value according to _Yi,j = α · M_ (ai,j) + (1 − _α) · M_ (bj) � � � = α · RSSI[∗]i,j [+][ C][a]i,j + (1 − _α) ·_ _Cbj(k),_ (1) _k∈Nj_ **FIGURE 3. Example of a WLAN before and after applying the channel** _load aware AP/Extender selection mechanism._ - Every reassociation to a new AP/Extender within the WLAN would require a complete authentication process, unless the fast BSS transition feature from IEEE 802.11r is employed [53]. According to the classification criteria from Table 1, the AP selection mode in this new AP/Extender selection mechanism is hybrid, because STAs share with the AP information about the network state, the architecture is centralized, as the AP (or the network controller) computes the best AP/Extender for each STA, and the parameters of the decision metric are: the RSSI observed by the STA and the channel load observed by the different AP/Extenders. As a matter of example, Figure 3 offers a graphical view of a complete WLAN before and after applying the channel load aware AP/Extender selection mechanism. _B. AP/EXTENDER SELECTION METRIC_ The decision metric used in the proposed approach combines parameters observed both in the access link M (ai,j) (i.e., from STA i to AP/Extender j) and in the backhaul link(s) M (bj) (i.e., those in the route from Extender j to the AP) [36]. When using the RSSI-based AP selection mechanism, STAs simply choose the AP/Extender with the strongest RSSI value in the access link. Differently, our AP/Extender selection mechanism takes advantage of the capabilities offered by IEEE 802.11k and IEEE 802.11v to create a new decision where α is a configurable factor that weights the influence of access and backhaul links (0 ≤ _α ≤_ 1) and Cai,j is the channel load of the access link observed by AP/Extender j. Considering Nj as the set of backhaul links in the path between Extender j and the AP, Cbj(k) is the channel load of backhaul link k. Note that when j corresponds to the AP, there are no backhaul links (i.e., Nj = ∅). Channel load C is here considered as the fraction of time during which the wireless channel is sensed busy, as indicated by either the physical or virtual carrier sense mechanism, with 0 _C_ 1 [54]. The AP (or the network controller) can ≤ ≤ obtain this information explicitly (by means of the channel _load request/report exchange) or implicitly (from the BSS_ _load element contained in both beacon frames and probe_ _responses emitted by AP/Extenders) [50]._ In fact, unlike other parameters employed in alternative decision metrics, the channel load is able to provide information not only from the targeted WLAN, but also from the influence of other external networks. In consequence, the WLAN is more able to balance the traffic load of newly associated STAs to the less congested AP/Extenders, thus increasing the adaptability degree to the state of the frequency channel. For its part, RSSI[∗]i,j [corresponds to an inverse weighting] of the signal strength received by STA i from AP/Extender j, which is computed as RSSIi,j − _Ptj_ RSSI[∗]i,j [=] _Si −_ _Ptj_ _,_ (2) where RSSIi,j is the signal strength received by STA i from AP/Extender j in dBm, Ptj is the transmission power level of AP/Extender j in dBm, and Si is the carrier sense threshold (i.e., sensitivity level) of STA i in dBm. As shown in Figure 4, the weighting of possible input values of RSSIi,j ∈ [Si, Ptj] from (2) applied in the AP/Extender ----- selection mechanism results in output values of RSSI[∗]i,j [∈] [0, 1]. Consequently, low RSSI values (i.e., those close to the sensitivity level Si) are highly penalized. **V. PERFORMANCE EVALUATION** This section is first intended to understand the benefits of adding Extenders to a WLAN, and determine their optimal number and location for a given area. Then, the very concept of a WLAN with Extenders is applied to a typical Home WiFi scenario aiming to evaluate the impact of the main parameters involved in the AP/Extender selection mechanism on network’s performance. _A. SIMULATION FRAMEWORK_ MATLAB was the selected tool to develop a simulator that enables the deployment, setting, testing, and performance evaluation of a WLAN. Specifically, our simulator focused on the AP/Extender selection mechanism contained in the STA association process, the transmission of uplink (UL) data packets (i.e., those from STAs to the AP), and the computation of metrics in the AP with respect to the received traffic. As for the PHY layer, it was assumed that, once the network topology was established, all devices adjusted their data rate according to the link condition. Specifically, simulations used the ITU-R indoor site-general path loss model according to PLITU(di,j) = 20 · log10(fc) + N · log10(di,j) + Lf − 28, (3) where PLITU is the path loss value (in dB), di,j is the distance between transmitter i and receiver j (in m), fc is the employed frequency (in MHz), N is the distance power loss coefficient (in our particular case and according to the model guidelines, _N = 31), and Lf is the floor penetration loss factor (which_ was removed as a single floor was always considered) [55]. The distributed coordination function (DCF) was used by all AP/Extenders and STAs. We assumed that all AP/Extenders and STAs were within the coverage area of the others, given they operated in the same channel. Therefore, an STA was able to associate to any AP/Extender in the area of interest. Only UL transmissions were considered in simulations, as they represent the worst case in a WLAN; that is, when multiple non-coordinated devices compete for the same wireless spectrum. Though excluded from the current study, downlink (DL) communications could either follow the same topology resulting from the STA association process or, as it is already conceived by designers of future WiFi 7, establish their own paths by means of the multi-link operation capability (in our particular case, according to an alternative decision metric) [6]. WLAN performance metrics (throughput, delay, and congestion) were obtained using the IEEE 802.11 DCF model presented and validated in [56], which supports heterogeneous finite-load traffic flows as required in this work. Details from two different wireless standards were implemented in the simulator: IEEE 802.11n and IEEE 802.11ac. Due to the **TABLE 2. List of common simulation parameters.** higher penetration of 2.4 GHz compatible devices in real deployments, all tests employed IEEE 802.11n at 2.4 GHz in access links (with up to 3 available orthogonal channels) and IEEE 802.11ac at 5 GHz in backhaul links (with a single channel).[2] Nonetheless, the simulator supports any combination of standards over the aforementioned network links. A wide set of tests was conducted on several predefined scenarios to evaluate the impact of different WLAN topologies, configurations, and AP/Extender selection mechanisms on the network’s performance. The definition of the scenarios together with their corresponding tests is provided in the following subsections. Lastly, a comprehensive list of common simulation parameters is offered in Table 2, whose values were applied to all subsequent tests, if not otherwise specified. As for test-specific simulation parameters, we refer the reader to Table 3. _B. SCENARIO #1: CIRCULAR AREA_ A circular area was defined by the maximum coverage range of the AP at 2.4 GHz (Dmax); i.e., the distance in which an STA would receive a signal with the same strength as its sensitivity level. Three different network topologies were there considered: only a single AP, an AP and 2 Extenders, and an AP and 4 Extenders forming a cross (see Figure 5). Position of Extenders was in turn limited by the maximum coverage range of the AP at 5 GHz (dmax). 2Data rates were computed from the observed RSSI and according to the corresponding modulation and coding scheme (MCS) table. ----- **TABLE 3. List of test-specific simulation parameters.** **FIGURE 5. Network topologies of Scenario #1.** 1) TEST 1.1: AP-EXTENDER DISTANCE The goal of this test was to evaluate the effect of the distance between the AP and any Extender (dAP,E) on network’s performance. To keep symmetry, the topology from Figure 5c was used, moving all Extenders far from the AP, with RSSI values at any Extender (RSSIAP,E) ranging from −50 dBm to −90 dBm (i.e., being the latter the RSSIAP,E value at dmax), in intervals of 1 dB. The case without Extenders was also included for comparative purposes. A number of NSTA = 10 STAs with a common traffic load of BSTA = 2.4 Mbps were uniformly and randomly deployed _k_ 1000 times on the AP coverage area. Both the RSSI= _based and the channel load aware AP/Extender selection_ mechanisms were used in each deployment. In the latter case, _α was set to 0.5 to give the same importance to access and_ backhaul links when selecting an AP/Extender. As shown in Figure 6, the use of Extenders almost always improved the network’s performance in terms of throughput, delay, and congestion regardless RSSIAP,E. In general, the best range to place Extenders was RSSIAP,E ∈ [−50, −72] dBm, as throughput was maintained over 99% in multi-channel cases when using any of the analyzed AP/Extender selection mechanisms.[3] More specifically, the channel load aware mechanism was able to ensure 100% of throughput and keep delay below 10 ms regardless RSSIAP,E. This was not the case when using a single communication channel, because almost all STAs were directly connected to the AP (thus resembling the case without Extenders, where furthest STAs hindered the operation of the rest due to their higher channel occupancy), unless they were really close to an alternative Extender. As for the RSSI-based mechanism, it always behaved worse than the _channel_ _load_ _aware_ mechanism in multi-channel cases, but provided better performance in single channel ones. In fact, although the number of STAs connected to Extenders decayed as we moved Extenders far away from the AP, that value was still much higher than in the _channel load aware mechanism. However, the adoption by_ Extenders of MCS 1 from RSSIAP,E = −77 dBm on, severely impacted on network’s performance, as they were not able to appropriately transmit all packets gathered from STAs. As a result of this test, dAP,E was set in following tests to the value that made RSSIAP,E = −70 dBm. 2) TEST 1.2: NETWORK’s RANGE EXTENSION To prove the benefit of using Extenders to increase the network coverage, the same topologies of Scenario #1 were 3In this test, but also as generalized practice in the rest of tests from this article, results of each network configuration were obtained as the mean of values from all k deployments, whether the network got congested or not. ----- **FIGURE 6. Test 1.1. AP-Extender distance.** **FIGURE 7. Test 1.3. Number of Extenders.** used. However, in this case, STAs were placed uniformly at random over a circular area of radius 1.2 - _Dmax. Again, RSSI-_ _based and channel load aware (with α = 0.5) AP/Extender_ selection mechanisms were employed. A number of NSTA = 10 STAs were randomly deployed _k_ 10000 times on the predefined area, with the result= ing average rate of successful associations from Table 4. As expected, the higher the number of Extenders, the higher the total percentage of STAs that found an AP/Extender within their coverage area and got associated. In fact, both AP/Extender selection mechanisms achieved the same STA association rates, because they only depended on whether there were available AP/Extenders within each STA coverage area. 3) TEST 1.3: NUMBER OF EXTENDERS In all three topologies from Scenario #1 were placed a number of NSTA = 10 STAs, each one with the same traffic load ranging from BSTA = 12 kbps to BSTA = 3.6 Mbps (i.e., a total network traffic, BT = NSTA · BSTA, from BT = 0.12 **TABLE 4. Test 1.2. Network’s range extension.** Mbps to BT = 36 Mbps). STA deployments were randomly selected k 1000 times and the whole network operated = under both the RSSI-based and the channel load aware (with _α = 0.5) AP/Extender selection mechanisms. In this test, only_ the multi-channel case was considered. Results from Figure 7 justify the use of Extenders to increase the range in which the network operates without congestion, going up to BT ≈ 13 Mbps without Extenders, up to BT ≈ 16 Mbps in the RSSI-based mechanism, and up to BT ≈ 25 Mbps in the channel load aware one. Furthermore, the channel load aware mechanism guaranteed the minimum observed delay for any considered value of BT > 5 Mbps. ----- **TABLE 5. Test 1.3. Number of Extenders (network’s operational range expressed in terms of BT ).** The influence of the number of Extenders on performance was different in function of the AP/Extender selection mechanism. Whereas it was barely relevant in the channel _load aware mechanism due to the effective load balancing_ among Extenders and AP, it provided heterogeneous results when using the RSSI-based mechanism. Particularly, the use of 4 Extenders left the AP with a very low number of directly connected STAs, thus overloading backhaul links with respect to the case with only 2 Extenders. Lastly, further details on network’s operational range are detailed in Table 5 according to three different metrics based on throughput, delay, and congestion. _C. SCENARIO #2: HOME WiFi_ In this case, STAs were deployed within a rectangular area emulating a typical Home WiFi scenario defined according to a set of RSSI values (see Figure 8). Three network topologies were there considered: only a single AP, an AP connected to a single Extender, and an AP connected to two linked Extenders. 1) TEST 2.1: USE OF LINKED EXTENDERS To evaluate the effect of linking two Extenders in the backhaul, a set of NSTA = 10 STAs were randomly placed k = 1000 times on all topologies from Figure 8, with BSTA ranging from 12 kbps to 6 Mbps (i.e., BT took values from 0.12 Mbps to 60 Mbps). Both the RSSI-based and the channel load _aware AP/Extender selection mechanisms were considered_ (the latter with α = 0.5 to balance access and backhaul links). This test was first performed in a multi-channel case, where the channel load aware mechanism was able to avoid network congestion until almost BT = 40 Mbps and improve the performance offered by the RSSI-based mechanism, as seen in Figure 9. Furthermore, the use of a second Extender linked to the first one was justified to increase the network’s operational range, as shown in Table 6. As for the single channel case, the use of a second Extender (whether under the RSSI-based or the channel load aware mechanism) here did not result in a significant improvement of any analyzed performance metric. The fact that all STAs (even some of them with low transmission rates) ended up competing for the same channel resources increased the overall occupation and led to congestion for BT _< 25 Mbps_ regardless the number of Extenders. **FIGURE 8. Network topologies of Scenario #2.** 2) TEST 2.2: IMPACT OF ACCESS AND BACKHAUL LINKS Assuming the network topology from Figure 8c with 2 linked Extenders, the effect of α parameter on the channel load _aware AP/Extender selection mechanism was studied for α =_ {0, 0.25, 0.5, 0.75, 1} and BSTA = {1.8, 3, 4.2, 5.4} Mbps (i.e., a total network traffic of BT = {18, 30, 42, 54} Mbps, respectively). As shown in Figure 10, values of α ∈ [0.5, 0.75] in the multi-channel case were able to guarantee the best network performance in terms of throughput (> 95%) and delay (< 50 ms) for any considered BT value. In fact, to give all the weight in (1) either to the access link (α = 1) or to the backhaul links (α = 0) never resulted in the best exploitation of network resources. On the other hand, the best performance in the single channel case was achieved when α = 1; that is, when the ----- **FIGURE 9. Test 2.1. Use of linked Extenders (multi-channel case).** **FIGURE 10. Test 2.2. Impact of access and backhaul links.** _channel load aware mechanism behaved as the RSSI-based_ one and therefore only the RSSI value was taken into account to compute the best AP/Extender for each STA.[4] 3) TEST 2.3: SHARE OF IEEE 802.11k/V CAPABLE STAs The channel load aware AP/Extender selection mechanism can be executed by IEEE 802.11k/v capable STAs without detriment to the rest of STAs, which would continue using the RSSI-based mechanism as usual. This test intended to evaluate this effect on overall network’s performance. Assuming again the network topology from Figure 8c with 2 linked Extenders, the effect of the share of IEEE 802.11k/v capable STAs (here noted as β) on the channel _load aware mechanism was studied for α_ = 0.5, β = {0, 25, 50, 75, 100} %, and BSTA = {1.8, 3, 4.2, 5.4} Mbps (i.e., a total network traffic of BT = {18, 30, 42, 54} Mbps, respectively). As shown in Figure 11, there was a clear trend in the multi-channel case that made network’s performance grew 4In the single channel case, Cai,j element in (1) is the same for any access link. Then, if α = 1 (i.e., all the weight is given to the access link), the decisive factor is RSSIi,j. together with the share of IEEE 802.11k/v capable STAs, even ensuring more than 95% of throughput for any considered BT value when half or more of STAs were IEEE 802.11k/v capable. On the contrary, in the single channel case the best results were achieved when β = 0 or, in other words, when none STA had IEEE 802.11k/v capabilities and therefore all of them applied the traditional RSSI-based mechanism. 4) TEST 2.4: INTERFERENCE FROM EXTERNAL NETWORKS We aimed to evaluate the potential negative effect that the presence of neighboring WLANs could have on the channel _load aware AP/Extender selection mechanism, and verify if_ that mechanism continued outperforming the RSSI-based one in terms of total throughput and average delay. A particular scenario with an AP, an Extender and 10 STAs was considered following the deployment shown in Figure 12, where the Extender shared its access link channel at 2.4 GHz band with an external network. Whereas the traffic load of each STA was set to BSTA = 4.32 Mbps, the load of the external network ranged from BEXT = 0 Mbps to _BEXT = 12 Mbps._ ----- **TABLE 6. Test 2.1. Use of linked Extenders (network’s operational range expressed in terms of BT ).** **FIGURE 11. Test 2.3. Share of IEEE 802.11k/v capable STAs.** Figure 13a shows that, for any considered α value, the _channel load aware mechanism was able to deliver 100%_ of throughput for higher BEXT values than the RSSI-based configuration, having the highest α values the best performance. The topology without Extenders, here maintained as a reference, again demonstrates the utility of Extenders in such Home WiFi scenarios. The average delay of STAs followed the same trend (see Figure 13b), having again the channel load aware mechanism the best performance, maintaining it below 5 ms in any configuration given BEXT < 5 Mbps. Observing the delay, it is worth noting the difference between the gradual delay increase in the RSSI-based mechanism (due to the progressive saturation of the access link to the Extender when BEXT ∈ [1.5, 3.5] Mbps) in comparison with its abrupt change in the channel load aware one. This was due to a different AP/Extender selection of one or more STAs from a given _BEXT value on._ **VI. PERFORMANCE OF THE AP/EXTENDER SELECTION** **MECHANISM IN A REAL DEPLOYMENT** A testbed was deployed at Universitat Pompeu Fabra (UPF) to emulate a Home WiFi network and, therefore, further study the benefits of using Extenders and the performance of the _channel load aware AP/Extender selection mechanism._ The hardware employed consisted of an AP, an Extender, and 5 laptops acting as traffic generation STAs. A sixth **FIGURE 12. Network topology and STA deployment of Test 2.4.** laptop was connected to the AP through Ethernet to act as the traffic sink. The AP and the Extender were placed at a distance that guaranteed RSSIAP,E = −70 dBm at 5 GHz, as in the previous simulated scenarios. As for the 2.4 GHz band, non-overlapping communications were ensured by using orthogonal channels. STAs were deployed in 2 different sets of positions (see Figure 14). Then, using the RSSI and load parameters from each STA, all network links were obtained according to the ----- **FIGURE 13. Test 2.4. Interference from external networks.** appropriate AP/Extender selection mechanism. These links were then set in the real deployment to get the performance results. Tests were performed using iPerf[5] version 2.09 or higher, which allowed the use of enhanced reports that included both the average throughput and the delay of the different network links. The clocks of the STAs needed to be synchronized for the delay calculation, and this was achieved using the network time protocol (NTP).[6] UDP traffic was used in all iPerf tests. Several traffic loads were used in each test, and 5 trials were performed for each traffic load. Each trial lasted 60 seconds. Clocks were re-synchronized before every new load was tested (i.e., every 5 trials), leading to an average clock offset of +/ − 0.154 ms. All trials were performed during non-working ours, and there were no other WiFi users at UPF during the tests. _A. EXPERIMENT 1: ON THE BENEFITS OF USING_ _EXTENDERS_ Testbed #1 was designed to analyze the performance of a network that consisted of one AP and one Extender, considering only the RSSI-based association mechanism. The device placement for this experiment can be found in Figure 14a. Two cases were considered: the first one was the deployment without the Extender, meaning that all STAs were forced to associate to the AP. The second case did consider the Extender, allowing STAs to associate to either the AP or the Extender. The association for each case can be found in Table 7, as well as the RSSI of each STA for both the AP and the Extender. In the first case, where all STAs associated to the AP, we can observe that the RSSI was very low for STAs #4 and #5, as expected. Once we added the Extender in the second case, STAs #4 and #5 were associated to it, and so they improved their RSSI. Specifically, STA #4 got an increase of 30.51%, and STA #5 experienced an increase of 46.15%, [5iPerf main website: https://iperf.fr/](https://iperf.fr/) [6NTP main website: http://www.ntp.org/](http://www.ntp.org/) **FIGURE 14. Plan map of testbeds performed at UPF and placement of** network devices. respectively. The average RSSI of the different links was also increased, going from -47.20 dBm to -37.60 dBm (i.e., 20.34% higher). Three different total network traffic loads (BT ), as a result of the corresponding traffic load per STA (BSTA), were tested in each case, starting with BSTA = 1 Mbps (i.e., BT = 5 Mbps), then BSTA = 3 Mbps (i.e., BT = 15 Mbps), and lastly BSTA = 7.5 Mbps (i.e., BT = 37.5 Mbps). Figure 15 shows the throughput achieved for each load, as well as the average delay for the network. Regardless the presence of the Extender, 100% of throughput was achieved for BT = 5 Mbps. Higher differences appeared for BT = 15 Mbps and BT = 37.5 Mbps, as without the Extender the network was saturated, whereas 100% of the desired throughput was achieved when using the Extender. ----- **TABLE 7. RSSI values received by STAs from AP/Extender and selected next hop in Testbed #1 and #2.** **FIGURE 15. Throughput and delay achieved in Testbed #1.** **FIGURE 16. Average delay by STA in Testbed #1.** The use of an Extender is also beneficial for the average delay, as even in the worst case, when BT = 37.5 Mbps, this value was reduced from 6633.84 ms to 4.10 ms. The reason of selection mechanisms. The resulting association for all STAs, such huge delays when not using Extenders can be observed as well as their traffic loads can be found in Table 7, where in Figure 16, where the delay breakdown per STA shows how we can observe that at least one STA was always associated to STA #4 and STA #5 influenced the overall average values. the Extender when using the channel load aware mechanism, In this experiment we have shown that the use of Extenders thus resulting in better use of network resources. in a Home WiFi network can be beneficial beyond the exten- Figure 17 shows the results obtained for each AP/Extender sion of the coverage area, increasing both the minimum and selection mechanism. For BT = 5 Mbps, BT = 37.5 Mbps the average RSSI for the whole network, as well as achieving and BT = 50 Mbps, both the RSSI-based and the channel higher throughput capacity and lower delays. These results _load aware mechanisms achieved 100% of desired through-_ therefore support our previous simulations, whose results are put. However, only the channel load aware mechanism was compiled in Table 4, Table 5, and Figure 7. capable of reaching 100% for BT = 75 Mbps, with the RSSI _based mechanism reaching only 66.9 Mbps. Finally, although_ _B. EXPERIMENT 2: VALIDATION OF THE CHANNEL LOAD_ the network was always congested for BT = 100 Mbps, _AWARE AP/EXTENDER SELECTION MECHANISM_ the channel load aware mechanism managed to boost the Testbed #2 was deployed following Figure 14b to evaluate throughput from 49.22 Mbps to 87.18 Mbps. the performance of the channel load aware AP/Extender In terms of delay, the channel load aware mechanism selection mechanism and compare it to the RSSI-based mech- always had the minimum values. As a matter of example, anism. The AP and the Extender were always active and in in the worst case, with BT = 100 Mbps, the delay was equal non-overlapping channels. All STAs were inside the office to 130.24 ms and 37.34 ms for the RSSI-based and the channel that contained the AP, and we applied both selection mecha- _load aware mechanisms, respectively._ nisms to every STA. For the channel load aware mechanism, In this experiment, we have shown that the channel the α used was 0.5; i.e., the influence of the access and the _load aware AP/Extender selection mechanism outperforms_ backhaul links was the same when selecting an AP/Extender. the network performance in Home WiFi scenarios of the Five increasing loads were used to compare the per- _RSSI-based one in terms of throughput and delay. Further-_ formance of the RSSI-based and the channel load aware more, results also corroborate those obtained in previous ----- **FIGURE 17. Throughput and delay achieved in Testbed #2.** simulations (compiled in Table 6), in which the channel _load aware mechanism is shown to keep more deployments_ uncongested. **VII. THE FUTURE OF HOME WiFi NETWORKS WITH** **MULTIPLE AP/EXTENDERS** In the last years, the emergence of a plethora of new applications and services in addition to the necessity of ubiquitous communication have made Home WiFi networks be more densely populated with wireless devices. Consequently, WiFi traditional spectrum at 2.4 GHz band has become scarce, and it has been necessary to extend the WiFi paradigm into new bands operating at 5 GHz and 6 GHz, with much higher resources availability. Next generation WiFi amendments such as IEEE 802.11ax and IEEE 802.11be are taking advantage of these new bands of free license-exempt spectrum to develop physical PHY/MAC enhancements that provide Home networks with higher capacity, lower delay, and higher reliability, thus expanding WiFi into next-generation applications from the audiovisual, health care, industrial, transport, and financial sector, among others. Nonetheless, regardless the operating band, the increasing demand of wireless resources in terms of throughput, bandwidth, and for longer connection periods makes crucial to take into consideration the interplay not only with other devices from the same Home WiFi network, but also with overlapping networks when accessing to the shared medium, including other AP/Extenders belonging to the same WLAN. In this last case, the proliferation of WLAN management platforms as discussed in Section II may facilitate the coordination of the network, as well as with the help of some new features coming in IEEE 802.11ax and IEEE 802.11be amendments, such as spatial reuse, OFDMA, and target wake time (TWT) solutions [57], including their cooperative multiAP/Extender counterparts. For WLANs with multiple AP/Extenders, there are still many open challenges to properly design and implement real-time load balancing schemes among AP/Extenders when considering STA (and AP) mobility and traffic heterogeneity, including UL and DL traffic. Particularly, to create a potentially effective AP/Extender selection mechanism adapted to the aforementioned conditions, its decision metric(s) should be enriched with new parameters describing the instantaneous state of available AP/Extenders such as the number of hops to the AP, the packet latency, the available rate(s), the bit error rate (BER), or even the distance to the targeted STA. In this last regard, the IEEE 802.11az Task Group (TGaz) aims at providing improved absolute and relative location, tracking, and positioning of STAs by using fine timing measurement (FTM) instead of signal-strength techniques [58]. Specifically, FTM protocol enables a pair of WiFi cards to estimate distance between them from round-trip timing measurement of a given transmitted signal. Lastly, and in line with what was stated in Section II, there is wide scope for the introduction of ML techniques into the AP/Extender selection mechanism. Particularly, the weight(s) of the decision metric(s) could be determined through ML, either dynamically according to a real-time observation and feedback process on the network state, or by applying the values corresponding to the most similar case from a set of predetermined patterns and scenarios. **VIII. CONCLUSION** The RSSI-based AP selection mechanism, used by default in IEEE 802.11 WLANs, only relies on the signal strength received from available APs. Therefore, in spite of its simplicity, it may result in an unbalanced load distribution between AP/Extenders and, consequently, in a degradation of the overall WLAN performance. Though several alternatives can be found in the literature addressing this issue, the channel load aware AP/Extender selection mechanism presented in this article stands out by its feasibility, as it is fully based on the already existing IEEE 802.11k/v amendments, without requiring to modify the firmware of end devices to facilitate real implementation. The potential of the channel load aware mechanism is shown through simulations and real testbed results. It is able to outperform the traditional RSSI-based mechanism in multi-channel scenarios consisting of multiple AP/Extenders in terms of throughput, delay, and number of situations that are satisfactorily solved, thus extending the WLAN operational range in, at least, 35%. Furthermore, results from a real testbed show that the throughput is boosted up to 77.12% with respect to the traditional RSSI-based mechanism in the considered setup. As for the measured delay, it is consistently lower with the channel _load aware mechanism, with differences ranging from 1.398_ to 92.895 ms. **REFERENCES** [1] T. Høiland-Jørgensen, M. Kazior, D. Täht, P. Hurtig, and A. Brunstrom, ‘‘Ending the anomaly: Achieving low latency and airtime fairness in wifi,’’ in Proc. USENIX Annu. Tech. Conf. (USENIX ATC), 2017, pp. 139–151. ----- [2] IEEE Standard for Information Technology—Telecommunications and _Information Exchange Between Systems Local and Metropolitan Area_ _Network—Specific Requirements—Part 11: Wireless LAN Medium Access_ _Control (MAC) and Physical Layer (PHY) Specifications—Amendment 4:_ _Enhancements for Very High Throughput for Operation in Bands Below 6_ _ghz, document 802.11ac-2013 (Amendment to IEEE Std 802.11-2012, as_ amended by IEEE Std 802.11ae-2012, IEEE Std 802.11aa-2012, and IEEE Std 802.11ad-2012), Dec. 2013, pp. 1–425. [3] B. Bellalta, ‘‘IEEE 802.11ax: High-efficiency WLANS,’’ IEEE Wireless _Commun., vol. 23, no. 1, pp. 38–46, Feb. 2016._ [4] E. Khorov, A. Kiryanov, A. Lyakhov, and G. Bianchi, ‘‘A tutorial on IEEE 802.11ax high efficiency WLANs,’’ IEEE Commun. Surveys Tuts., vol. 21, no. 1, pp. 197–216, 1st Quart., 2019. [5] D. Lopez-Perez, A. Garcia-Rodriguez, L. Galati-Giordano, M. Kasslin, and K. Doppler, ‘‘IEEE 802.11be extremely high throughput: The next generation of Wi-Fi technology beyond 802.11ax,’’ IEEE Commun. Mag., vol. 57, no. 9, pp. 113–119, Sep. 2019. [6] T. Adame, M. Carrascosa, and B. Bellalta, ‘‘Time-sensitive networking in IEEE 802.11be: On the way to low-latency WiFi 7,’’ 2019, _arXiv:1912.06086. [Online]. Available: http://arxiv.org/abs/1912.06086_ [7] L.-H. Yen, T.-T. Yeh, and K.-H. Chi, ‘‘Load balancing in IEEE 802.11 networks,’’ IEEE Internet Comput., vol. 13, no. 1, pp. 56–64, Jan. 2009. [8] G. R. Hiertz, D. Denteneer, S. Max, R. Taori, J. Cardona, L. Berlemann, and B. Walke, ‘‘IEEE 802.11 s: The WLAN mesh standard,’’ IEEE Wireless _Commun., vol. 17, no. 1, pp. 104–111, 2010._ [9] S. M. S. Bari, F. Anwar, and M. H. Masud, ‘‘Performance study of hybrid wireless mesh protocol (HWMP) for IEEE 802.11s WLAN mesh networks,’’ in Proc. Int. Conf. Comput. Commun. Eng. (ICCCE), Jul. 2012, pp. 712–716. [10] S. Waharte and R. Boutaba, ‘‘Tree-based wireless mesh network architecture: Topology analysis,’’ in Proc. 1st Int. Workshop Wireless Mesh Netw. _(MeshNets), Budapest, Hungary, 2005, pp. 1–11._ [11] M. Herlich and S. Yamada, ‘‘Optimal distance of multi-hop 802.11 WiFi relays,’’ in Proc. IEICE Soc. Conf., Sep. 2014, pp. 1–2. [12] S. A. Hassan, ‘‘Optimal throughput analysis for indoor multi-hop wireless networks in IEEE 802.11n,’’ in Proc. WAMICON, Apr. 2013, pp. 1–5. [13] R. Atawia and H. Gacanin, ‘‘Self-deployment of future indoor Wi-Fi networks: An artificial intelligence approach,’’ in Proc. GLOBECOM IEEE _Global Commun. Conf., Dec. 2017, pp. 1–6._ [14] Z. M. Fadlullah, Y. Kawamoto, H. Nishiyama, N. Kato, N. Egashira, K. Yano, and T. Kumagai, ‘‘Multi-hop wireless transmission in multiband WLAN systems: Proposal and future perspective,’’ IEEE Wireless _Commun., vol. 26, no. 1, pp. 108–113, Feb. 2019._ [15] N. Egashira, K. Yano, S. Tsukamoto, J. Webber, and T. Kumagai, ‘‘Low latency relay processing scheme for WLAN systems employing multiband simultaneous transmission,’’ in Proc. IEEE Wireless Commun. Netw. Conf. _(WCNC), Mar. 2017, pp. 1–6._ [16] IEEE Standard for Information Technology—Telecommunications and _Information Exchange Between Systems Local and Metropolitan Area_ _Networks—Specific Requirements—Part 11: Wireless LAN Medium Access_ _Control (MAC) and Physical Layer (PHY) Specifications, Standard 802.11-_ 2016 (Revision of IEEE Std 802.11-2012), Dec. 2016, pp. 1–3534. [17] X. Chen and D. Qiao, ‘‘HaND: Fast handoff with null dwell time for IEEE 802.11 networks,’’ in Proc. IEEE INFOCOM, Mar. 2010, pp. 1–9. [18] T. Choi, Y. Chon, and H. Cha, ‘‘Energy-efficient WiFi scanning for localization,’’ Pervas. Mobile Comput., vol. 37, pp. 124–138, Jun. 2017. [19] A. Mishra, M. Shin, and W. Arbaugh, ‘‘An empirical analysis of the IEEE 802.11 MAC layer handoff process,’’ ACM SIGCOMM Comput. Commun. _Rev., vol. 33, no. 2, pp. 93–102, Apr. 2003._ [20] H. Velayos and G. Karlsson, ‘‘Techniques to reduce the IEEE 802.11b handoff time,’’ in Proc. IEEE Int. Conf. Commun., Jun. 2004, pp. 3844–3848. [21] S. Lee, M. Kim, S. Kang, K. Lee, and I. Jung, ‘‘Smart scanning for mobile devices in WLANs,’’ in Proc. IEEE Int. Conf. Commun. (ICC), Jun. 2012, pp. 4960–4964. [22] Google LLC. (2019). Google WiFi Main Website. Accessed: Jan. 29, 2021. [Online]. Available: https://store.google.com/product/google_wifi [23] Linksys. (2019). Linksys Velop Main Website. Accessed: Jan. 29, 2019. [Online]. Available: https://www.linksys.com/us/velop/ [24] J. B. Ernst, S. Kremer, and J. J. P. C. Rodrigues, ‘‘A utility based access point selection method for IEEE 802.11 wireless networks with enhanced quality of experience,’’ in Proc. IEEE Int. Conf. Commun. (ICC), Jun. 2014, pp. 2363–2368. [25] A. J. Nicholson, Y. Chawathe, M. Y. Chen, B. D. Noble, and D. Wetherall, ‘‘Improved access point selection,’’ in Proc. 4th Int. Conf. Mobile Syst., _Appl. Services - MobiSys, 2006, pp. 233–245._ [26] L.-H. Yen, J.-J. Li, and C.-M. Lin, ‘‘Stability and fairness of AP selection games in IEEE 802.11 access networks,’’ IEEE Trans. Veh. Technol., vol. 60, no. 3, pp. 1150–1160, Mar. 2011. [27] Y. Fukuda and Y. Oie, ‘‘Decentralized access point selection architecture for wireless LANs,’’ IEICE Trans. Commun., vol. E90-B, no. 9, pp. 2513–2523, Sep. 2007. [28] H. Gong, K. Nahm, and J. Kim, ‘‘Distributed fair access point selection for multi-rate IEEE 802.11 WLANs,’’ in Proc. 5th IEEE Consum. Commun. _Netw. Conf., Jan. 2008, pp. 528–532._ [29] F. Xu, C. C. Tan, Q. Li, G. Yan, and J. Wu, ‘‘Designing a practical access point association protocol,’’ in Proc. IEEE INFOCOM, Mar. 2010, pp. 1–9. [30] A. Raschella, F. Bouhafs, M. Seyedebrahimi, M. Mackay, and Q. Shi, ‘‘A centralized framework for smart access point selection based on the fittingness factor,’’ in Proc. 23rd Int. Conf. Telecommun. (ICT), May 2016, pp. 1–5. [31] M. Abusubaih and A. Wolisz, ‘‘An optimal station association policy for multi-rate ieee 802.11 wireless lans,’’ in Proc. 10th ACM Symp. Modeling, _Anal., Simulation Wireless Mobile Syst. - MSWiM, 2007, pp. 117–123._ [32] L. Du, Y. Bai, and L. Chen, ‘‘Access point selection strategy for large-scale wireless local area networks,’’ in Proc. IEEE Wireless Commun. Netw. _Conf., Mar. 2007, pp. 2161–2166._ [33] M. Abusubaih, J. Gross, S. Wiethoelter, and A. Wolisz, ‘‘On access point selection in IEEE 802.11 wireless local area networks,’’ in Proc. 31st IEEE _Conf. Local Comput. Netw., Nov. 2006, pp. 879–886._ [34] S. Vasudevan, K. Papagiannaki, C. Diot, J. Kurose, and D. Towsley, ‘‘Facilitating access point selection in IEEE 802.11 wireless networks,’’ in Proc. 5th ACM SIGCOMM Conf. Internet Meas. IMC, 2005, p. 26. [35] K. Mittal, E. M. Belding, and S. Suri, ‘‘A game-theoretic analysis of wireless access point selection by mobile users,’’ Comput. Commun., vol. 31, no. 10, pp. 2049–2062, Jun. 2008. [36] L. Luo, D. Raychaudhuri, H. Liu, M. Wu, and D. Li, ‘‘Improving end-to-end performance of wireless mesh networks through smart association,’’ in Proc. IEEE Wireless Commun. Netw. Conf., Mar. 2008, pp. 2087–2092. [37] M. Abusubaih and A. Wolisz, ‘‘Interference-aware decentralized access point selection policy for multi-rate IEEE 802.11 wireless LANs,’’ in Proc. _IEEE 19th Int. Symp. Pers., Indoor Mobile Radio Commun., Sep. 2008,_ pp. 1–6. [38] K. Sundaresan and K. Papagiannaki, ‘‘The need for cross-layer information in access point selection algorithms,’’ in Proc. 6th ACM SIGCOMM _Internet Meas. - IMC, 2006, pp. 257–262._ [39] J. Pang, B. Greenstein, M. Kaminsky, D. McCoy, and S. Seshan, ‘‘WiFireports: Improving wireless network selection with collaboration,’’ IEEE _Trans. Mobile Comput., vol. 9, no. 12, pp. 1713–1731, Dec. 2010._ [40] B. Bojovic, N. Baldo, and P. Dini, ‘‘A neural network based cognitive engine for IEEE 802.11 WLAN access point selection,’’ in Proc. IEEE _Consum. Commun. Netw. Conf. (CCNC), Jan. 2012, pp. 864–868._ [41] M. Carrascosa and B. Bellalta, ‘‘Decentralized AP selection using multiarmed bandits: Opportunistic ε-Greedy with stickiness,’’ in Proc. IEEE _Symp. Comput. Commun. (ISCC), Jun. 2019, pp. 1–7._ [42] M. Carrascosa and B. Bellalta, ‘‘Multi-armed bandits for decentralized AP selection in enterprise WLANs,’’ 2020, arXiv:2001.00392. [Online]. Available: http://arxiv.org/abs/2001.00392 [43] L. Chen, ‘‘A distributed access point selection algorithm based on no-regret learning for wireless access networks,’’ in Proc. IEEE 71st Veh. Technol. _Conf., May 2010, pp. 1–5._ [44] Nighthawk Mesh WiFi 6 System, Netgear, Inc, San Jose, CA, USA, 2019. [45] Netgear Support. (2018). How do I use Smart WiFi or Auto-Radio Resource _Management (RRM) in Insight. Accessed: Jan. 29, 2021. [Online]. Avail-_ able: https://kb.netgear.com/000053883 [46] Aruba AirMatch Technology, Aruba Networks, Inc., Santa Clara, CA, USA, 2018. [47] What is Aruba ClientMatch,, Aruba Networks, Inc., Santa Clara, CA, USA, 2018. [48] J. A. G. Garrido and J. D. Alonso, ‘‘System and method for decentralized control of wireless networks,’’ U.S. Patent 10 397 932, Aug. 27, 2019. [49] M. I. Sanchez and A. Boukerche, ‘‘On IEEE 802.11K/R/V amendments: Do they have a real impact?’’ IEEE Wireless Commun., vol. 23, no. 1, pp. 48–55, Feb. 2016. ----- [50] IEEE Standard for Information technology—Local and metropolitan area _networks—Specific requirements—Part 11: Wireless LAN Medium Access_ _Control (MAC) and Physical Layer (PHY) Specifications Amendment 1:_ _Radio Resource Measurement of Wireless LANs, Standard 802.11k-2008,_ (Amendment to IEEE Std 802.11-2007), Jun. 2008, pp. 1–244. [51] E. A. Panaousis, P. A. Frangoudis, C. N. Ververidis, and G. C. Polyzos, ‘‘Optimizing the channel load reporting process in IEEE 802.11kenabled WLANs,’’ in Proc. 16th IEEE Workshop Local Metrop. Area _Netw., Sep. 2008, pp. 37–42._ [52] IEEE Standard for Information technology—Local and metropolitan _area networks—Specific requirements—Part 11: Wireless LAN Medium_ _Access Control (MAC) and Physical Layer (PHY) Specifications Amend-_ _ment 8: IEEE 802.11 Wireless Network Management, Standard 802.11v-_ 2011, (Amendment to IEEE Std 802.11-2007 as amended by IEEE Std 802.11k-2008, IEEE Std 802.11r-2008, IEEE Std 802.11y-2008, IEEE Std 802.11w-2009, IEEE Std 802.11n-2009, IEEE Std 802.11p-2010, and IEEE Std 802.11z-2010), Feb. 2011, pp. 1–433. [53] S. Bangolae, C. Bell, and E. Qi, ‘‘Performance study of fast BSS transition using IEEE 802.11r,’’ in Proc. Int. Conf. Commun. Mobile Comput. _IWCMC, Jul. 2006, pp. 737–742._ [54] C. Thorpe, S. Murphy, and L. Murphy, ‘‘Analysis of variation in IEEE 802.11k channel load measurements for neighbouring WLAN systems,’’ in Proc. 17th ICT Mobile Wireless Commun. Summit (ICT-MobileSummit), 2008. [Online]. Available: https://e-archivo.uc3m.es/handle/10016/14054 [55] Propagation Data and Prediction Methods for the Planning of Indoor _Radio Communication Systems and Radio Local Area Networks in the Fre-_ _quency Range 900 MHz to 100 GHz, document ITU-R, Recommendation_ P.1238-7, P Series. Radiowave propagation, 2012. [56] B. Bellalta, M. Oliver, M. Meo, and M. Guerrero, ‘‘A simple model of the IEEE 802.11 MAC protocol with heterogeneous traffic flows,’’ in Proc. _EUROCON Int. Conf. Comput. Tool, Nov. 2005, pp. 1830–1833._ [57] M. Nurchis and B. Bellalta, ‘‘Target wake time: Scheduled access in IEEE 802.11ax WLANs,’’ IEEE Wireless Commun., vol. 26, no. 2, pp. 142–150, Apr. 2019. [58] C.-C. Wang. IEEE P802.11, Wireless LANs, Specification Framework for _TGaz. Accessed: Jan. 29, 2021. https://mentor.ieee.org/802.11/dcn/17/11-_ 17-0462-16-00az-11-az-tg-sfd.doc TONI ADAME received the M.Sc. degree in telecommunications engineering from the Universitat Politecnica de Catalunya (UPC), in 2009. He is currently a Senior Researcher with the Department of Information and Communication Technologies (DTIC), Universitat Pompeu Fabra (UPF), responsible for the design of technical solutions in research and development projects-based on heterogeneous wireless technologies. He also collaborates as an Associate Lecturer in several IT degrees with UPF and Universitat Oberta de Catalunya (UOC). MARC CARRASCOSA received the B.Sc. degree in telematics engineering and the M.Sc. degree in intelligent and interactive systems from the Universitat Pompeu Fabra (UPF), in 2018 and 2019, respectively. He is currently pursuing the Ph.D. degree with the Wireless Networking Research Group, Department of Information and Communication Technologies (DTIC), UPF. His research interest includes performance optimization in wireless networks. BORIS BELLALTA is currently an Associate Professor with the Department of Information and Communication Technologies (DTIC), Universitat Pompeu Fabra (UPF), where he is also the Head of the Wireless Networking Research Group. IVÁN PRETEL received the M.Sc. degree in development and integration of software solutions from the University of Deusto, in 2010, and the Ph.D. degree in computer engineering and telecommunications, in 2015. He is currently a Research Engineer with Fon Labs. In 2008, he began his research career with the MORElab Research Group, Deusto Foundation, where he started as a Research Intern in the mobile services area participating in more than 20 international and national research projects related to system architectures, human–computer interaction, and societal challenges. He is also involved in research projects related to data science and 5G technologies, such as the 5GENESIS H2020 project. He also collaborates in several master degrees as an Associate Lecturer with the University of Deusto, giving several courses on mobile platforms, big data, and business intelligence. His research interests include data science and advanced mobile services. IÑAKI ETXEBARRIA received the degree in telecommunications engineering from the Escuela Superior de Ingeniería de Bilbao (ETSI). He has developed his professional career in private business, before Fon he worked with Erictel M2M working on IoT, embedded equipment development, and fleet management software solutions. Since 2015, he has been working with Fon Labs, where he has specialized in communication networks, specifically WiFi, developing innovation projects on product and technology. He is currently a Research and Development Engineer with Fon Labs. He has worked on several projects in international consortiums integrating WiFi in 5G networks. He also combines engineering work with the management of the Fon Labs team. -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2004.08110, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09354613.pdf" }
2,020
[ "JournalArticle" ]
true
2020-04-17T00:00:00
[ { "paperId": "2d14884fc362e2e18321372aae44f8575e4bd4b8", "title": "IEEE Standard for Information Technology--Telecommunications and Information Exchange between Systems - Local and Metropolitan Area Networks--Specific Requirements - Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications - Redline" }, { "paperId": "31e33e50ac50e3e273571b980e24e646d4b80432", "title": "Multi-Armed Bandits for Decentralized AP selection in Enterprise WLANs" }, { "paperId": "397de1bd830d10c80bc4edec0b6d2de6df1f693e", "title": "Time-Sensitive Networking in IEEE 802.11be: On the Way to Low-Latency WiFi 7" }, { "paperId": "8ca96b9fc3ed781ac47fd99dbd04e4a593b01e16", "title": "Decentralized AP selection using Multi-Armed Bandits: Opportunistic ε-Greedy with Stickiness" }, { "paperId": "d7a6c480e4364145e7e328875efe2fac3832c574", "title": "IEEE 802.11be Extremely High Throughput: The Next Generation of Wi-Fi Technology Beyond 802.11ax" }, { "paperId": "c21d4bda27eda671604b5093632af9bc8b25c318", "title": "Multi-Hop Wireless Transmission in Multi-Band WLAN Systems: Proposal and Future Perspective" }, { "paperId": "bc021725fdf50de15fc0a91cfb44bd9e1c28de81", "title": "Target Wake Time: Scheduled Access in IEEE 802.11ax WLANs" }, { "paperId": "db3ae1dac8a7b45f7e6bb7e9ae7fca4d9d6a7c11", "title": "Self-Deployment of Future Indoor Wi-Fi Networks: An Artificial Intelligence Approach" }, { "paperId": "05d11d5850b5b1f0a5dc6b1518914e49dd1f014b", "title": "Energy-efficient WiFi scanning for localization" }, { "paperId": "b69f834743a0d8b0544054385869a69cdc8d5ff0", "title": "Low Latency Relay Processing Scheme for WLAN Systems Employing Multiband Simultaneous Transmission" }, { "paperId": "ba7f5b3f80159579b7b050380a8ff958bf392801", "title": "Ending the Anomaly: Achieving Low Latency and Airtime Fairness in WiFi" }, { "paperId": "1d06c514ab7844f083ac899f1f477de2698c5b1a", "title": "A centralized framework for smart access point selection based on the Fittingness Factor" }, { "paperId": "629b5649cff12a388a7772f49b020050696caffa", "title": "On IEEE 802.11K/R/V amendments: Do they have a real impact?" }, { "paperId": "5ea0c7762e62a928564e239cb99e837de0777e76", "title": "IEEE 802.11ax: High-efficiency WLANS" }, { "paperId": "03e8f41b833ab570d1ffc292605c5cd4d34ee897", "title": "A utility based access point selection method for IEEE 802.11 wireless networks with enhanced quality of experience" }, { "paperId": "6096c4bc3e6335e1b55c49f2e8133e86079b6296", "title": "Optimal throughput analysis for indoor multi-hop wireless networks in IEEE 802.11n" }, { "paperId": "f2ca20d38a8013a36b4d2ddcedc68e78320c9c38", "title": "Performance study of hybrid Wireless Mesh Protocol (HWMP) for IEEE 802.11s WLAN mesh networks" }, { "paperId": "5395e8bdf03de0a5ea2320fec012263c2acd2c4e", "title": "Smart scanning for mobile devices in WLANs" }, { "paperId": "805d987b6e1de715e227bea633efb7fe97f1a1b2", "title": "A Neural Network based cognitive engine for IEEE 802.11 WLAN Access Point selection" }, { "paperId": "cfb820f627b4e3186716fc1a85898608a6a1c3a2", "title": "Stability and Fairness of AP Selection Games in IEEE 802.11 Access Networks" }, { "paperId": "f861380b479d292e00629db16057867857cf36b6", "title": "A Distributed Access Point Selection Algorithm Based on No-Regret Learning for Wireless Access Networks" }, { "paperId": "badd21a2238020776ba3946957e01de3b0f995cb", "title": "Designing a Practical Access Point Association Protocol" }, { "paperId": "f5109e138a0a6cb0dc28e7932b1327a85fd4e7ab", "title": "HaND: Fast Handoff with Null Dwell Time for IEEE 802.11 Networks" }, { "paperId": "265cffeefbe3188b9acfb4eafc5965e2408bd5c3", "title": "IEEE 802.11s: The WLAN Mesh Standard" }, { "paperId": "0d868769b3e8e49220c45cc0c069b6fb240ad365", "title": "Wifi-Reports: Improving Wireless Network Selection with Collaboration" }, { "paperId": "abda94e2d53bad912ca23516d4f9108761432471", "title": "Interference-aware decentralized access point selection policy for multi-rate IEEE 802.11 Wireless LANs" }, { "paperId": "667423cfc8625a7cd17dea6a9157ad14c07d2c50", "title": "Optimizing the channel load reporting process in IEEE 802.11k-enabled WLANs" }, { "paperId": "c9ccaa8a21215caf240ddf8d527a920101ee16d8", "title": "A game-theoretic analysis of wireless access point selection by mobile users" }, { "paperId": "63314ce11513f43cca0a0067e73fb7c1bf4463a5", "title": "Improving End-to-End Performance of Wireless Mesh Networks through Smart Association" }, { "paperId": "3b67435a67dcd692e8b006389dae14d1951d9d6e", "title": "Distributed Fair Access Point Selection for Multi-Rate IEEE 802.11 WLANs" }, { "paperId": "e48926854eecdf2840894167bd28a5411b9abfd9", "title": "An optimal station association policy for multi-rate ieee 802.11 wireless lans" }, { "paperId": "09fd9b208a7df88371a98c8fdd828bf2a075def7", "title": "Decentralized access point selection architecture for wireless LANs" }, { "paperId": "f31d55c51a8d4e36bd57e6ccc8d60f08d9c93b55", "title": "Access Point Selection Strategy for Large-Scale Wireless Local Area Networks" }, { "paperId": "7688c418ce8984087b3141a1272c6fb8be88d641", "title": "On Access Point Selection in IEEE 802.11 Wireless Local Area Networks" }, { "paperId": "7a180be3f8d994789c45d5b283d7d8bd4c01d74e", "title": "The need for cross-layer information in access point selection algorithms" }, { "paperId": "c4639cf655a4ef6a761c2a35ad39ac7e8ec68521", "title": "Performance study of fast BSS transition using IEEE 802.11r" }, { "paperId": "063ca005b4714603f907bbc8b89b27e3af535caf", "title": "Improved access point selection" }, { "paperId": "3665b1dbb30a285d067ed10ff6e901ecd19865e5", "title": "Facilitating access point selection in IEEE 802.11 wireless networks" }, { "paperId": "c7e4583c59fb2f9fa8d4aa0b3e477be5cdd1d4f6", "title": "Techniques to reduce the IEEE 802.11b handoff time" }, { "paperId": "dfeb0437c05df795cdfd0c7f944fd55c37b6db8b", "title": "An empirical analysis of the IEEE 802.11 MAC layer handoff process" }, { "paperId": null, "title": "IEEE P802.11, Wireless LANs, Specification Framework for TGaz, https://mentor.ieee.org/802.11/dcn/17/11-17-0462-16-00az-11-az-tg-sfd.doc" }, { "paperId": "f9f8a0cc797db5fbc8924541b973968747f928e1", "title": "A Tutorial on IEEE 802.11ax High Efficiency WLANs" }, { "paperId": null, "title": "System and method for decentralized con-trol of wireless networks" }, { "paperId": null, "title": "Nighthawk Mesh WiFi 6 System" }, { "paperId": null, "title": "What is Aruba ClientMatch?" }, { "paperId": null, "title": "Aruba AirMatch Technology" }, { "paperId": "7fc0131a272f7b80c4a40f2b1f76008f084121ff", "title": "Optimal Distance of Multi-hop 802 . 11 WiFi Relays" }, { "paperId": "2106f250967871c41b9457b92a386d8ee1d02aee", "title": "Load Balancing in IEEE 802.11 Networks" }, { "paperId": "b4ca9499d24f41f84eb96c2aef7422b78bc82d9d", "title": "Analysis of Variation in IEEE 802 . 11 k Channel Load Measurements for Neighbouring WLAN Systems" }, { "paperId": "99364dbd96fd1db90894bc3cec100d895b2da0b8", "title": "A Simple Model of the IEEE 802.11 MAC Protocol with Heterogeneous Traffic Flows" }, { "paperId": "3fc99e90563e52461a510ca291c29ddeb97a04b4", "title": "Tree-based Wireless Mesh Network Architecture : Topology Analysis" }, { "paperId": "e454a4f94495c8593c9f8d654f1627c8e54e78e8", "title": "Wireless lan medium access control (mac) and physical layer (phy) specifications" }, { "paperId": null, "title": "Enhancements for Very High Throughput for Operation in Bands Below 6 ghz, document 802" }, { "paperId": null, "title": "How do I use Smart WiFi or Auto-Radio Resource Management (RRM) in Insight" }, { "paperId": null, "title": "Wireless LANs, Specification Framework for TGaz" }, { "paperId": null, "title": "- 7 : Propagation data and prediction methods for the planning of indoor radio communication systems and radio local area networks in the frequency range 900 MHz to 100 GHz , P Series" }, { "paperId": null, "title": "Linksys Velop Main Website" } ]
19,542
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00f030043a1ca3a957a37cdf03060a15fefeb932
[]
0.918561
Paradigm shift in the music industry: Adaptation of blockchain technology and its transformative effects
00f030043a1ca3a957a37cdf03060a15fefeb932
JOURNAL OF ARTS
[ { "authorId": "2264798292", "name": "Betül Yarar Koçer" } ]
{ "alternate_issns": [ "2426-8356" ], "alternate_names": [ "J ART", "J art", "Journal des arts" ], "alternate_urls": [ "https://gallica.bnf.fr/ark:/12148/cb32799197c/date" ], "id": "3a539138-8634-48b5-af70-11b63711ab77", "issn": "2636-7718", "name": "JOURNAL OF ARTS", "type": "journal", "url": "http://www.ratingacademy.com.tr/ojs/index.php/arts" }
The music industry is undergoing a profound transformation thanks to blockchain technology. This article extensively examines how the core components of music – production, distribution, performance, and presentation – are undergoing radical changes through the integration of blockchain technology. The traditional music industry faces significant challenges, particularly in vital areas like copyright management, music distribution, and artist compensation. These challenges have become even more complex with the digitization of music and the rise of online platforms. However, blockchain technology, with its decentralized and transparent structure, has the potential to overcome these obstacles. This technology takes important steps in addressing disputes related to copyright by enhancing the traceability and verifiability of music works throughout their lifecycle, thereby contributing to fairer compensation for artists. Moreover, this article also delves into other intersecting domains related to the music industry, focusing on safeguarding intellectual property in music and presenting innovative solutions to the intricate music economy. Relevant data gathered through qualitative research methods is systematically presented to comprehensively explore the potential role of blockchain technology in the music industry’s future. This exploratory analysis also investigates blockchain-supported platforms, providing an in-depth examination of their current development status and business models. The article places special emphasis on fundamental concepts such as copyright, ownership of artistic works, cultural heritage, and the role of blockchain technology in shaping the music industry, artists, and the ongoing digital transformation. In this rapidly evolving dynamic process, the transformative role of blockchain technology in the music industry and its potential must be continuously monitored, serving as a foundation for future-oriented initiatives. This comprehensive approach reflects the concerted effort to understand the effects of blockchain technology, which is shaping the trajectory of the music industry’s future, from a broader perspective.
Volume/Cilt: 6, Issue/Sayı: 4 Year/Yıl: 2023, pp. 243-253 E-ISSN: 2636-7718 [URL: https://journals.gen.tr/arts](https://journals.gen.tr/arts) DOI: https://doi.org/10.31566/arts.2163 Received / Geliş: 08/08/2023 Acccepted / Kabul: 27/09/2023 RESEARCH ARTICLE / ARAŞTIRMA MAKALESİ # Paradigm shift in the music industry: Adaptation of blockchain technology and its transformative effects ## Betül Yarar Koçer Assist. Prof. Dr., Mersin University, State Conservatory, Department of Music,Türkiye, e-mail: betulyarar@mersin.edu.tr **Abstract** The music industry is undergoing a profound transformation thanks to blockchain technology. This article extensively examines how the core components of music – production, distribution, performance, and presentation – are undergoing radical changes through the integration of blockchain technology. The traditional music industry faces significant challenges, particularly in vital areas like copyright management, music distribution, and artist compensation. These challenges have become even more complex with the digitization of music and the rise of online platforms. However, blockchain technology, with its decentralized and transparent structure, has the potential to overcome these obstacles. This technology takes important steps in addressing disputes related to copyright by enhancing the traceability and verifiability of music works throughout their lifecycle, thereby contributing to fairer compensation for artists. Moreover, this article also delves into other intersecting domains related to the music industry, focusing on safeguarding intellectual property in music and presenting innovative solutions to the intricate music economy. Relevant data gathered through qualitative research methods is systematically presented to comprehensively explore the potential role of blockchain technology in the music industry’s future. This exploratory analysis also investigates blockchain-supported platforms, providing an in-depth examination of their current development status and business models. The article places special emphasis on fundamental concepts such as copyright, ownership of artistic works, cultural heritage, and the role of blockchain technology in shaping the music industry, artists, and the ongoing digital transformation. In this rapidly evolving dynamic process, the transformative role of blockchain technology in the music industry and its potential must be continuously monitored, serving as a foundation for future-oriented initiatives. This comprehensive approach reflects the concerted effort to understand the effects of blockchain technology, which is shaping the trajectory of the music industry’s future, from a broader perspective. **Keywords: Blockchain, Music Industry, Digital Music Distribution, Licensing, Copyright** **Citation/Atıf: YARAR KOÇER, B. (2023). Paradigm shift in the music industry: Adaptation of blockchain technology and its transformative** effects. Journal of Arts. 6(4): 243-253, DOI: 10.31566/arts.2163 **Corresponding Author/ Sorumlu Yazar:** Betül Yarar Koçer E-mail: betulyarar@mersin.edu.tr Bu çalışma, Creative Commons Atıf 4.0 Uluslararası Lisansı ile lisanslanmıştır. This work is licensed under a Creative Commons Attribution 4.0 International License. ----- ## 1. INTRODUCTION The advancement of technology has brought about a radical transformation in the distribution and recording methods of music. Recorded sounds have enabled musical compositions to reach wider audiences, expanding the boundaries of cultural sharing. However, this transformation has also given rise to concepts such as copyright and intellectual property. Traditionally, ownership of music compositions, copyright regulations, and protection methods have been shaped by legal frameworks aimed at controlling the usage of musicians’ and artists’ works. Technological progress, especially since the mid-20th century, has significantly widened the scope of cultural sharing through the use of recorded sounds, making music accessible to broader audiences. Music, previously limited to live performances, has become easily accessible in recorded formats. Cassettes, records and then digital formats made it easy for everyone to listen to music in the venue of their choice. Yet, this technological shift has also brought forth issues related to copyright and intellectual property. The rise of digital formats has increased the copyability and shareability of music compositions, leading to unauthorized use, duplication, and distribution of artists’ works. Copyright and intellectual property laws have been developed to protect the control and income of musicians by adapting to the changing dynamics of the music industry and aiming to safeguard artists’ creative efforts. While the music industry is rapidly adapting to the effects of digital transformation, the traditional processes of music distribution and copyright management remain complex and contentious. Lack of trust can arise in areas such as revenue sharing and copyright tracking among artists, producers, songwriters, and other stakeholders. The challenge of generating reasonable income from music production has become increasingly difficult, driven by the surge in inter-stakeholder content sharing and the distribution of intellectual property rights. Currently, the involvement of numerous intermediaries in the distribution stage has led to a chaotic process, contributing to a significant reduction in artists’ income due to low sales and inadequate royalty payments. “New technologies can radically simplify methods of identifying and compensating music rights owners, enabling sustainable business models for artists, entrepreneurs, and music enterprises.” (Panay, 2016). In this context, blockchain technology emerges as a potential solution. Mougayar (2016) asserts that blockchain technology is as crucial as the World Wide Web. The potential impacts of blockchain technology on the music industry encompass a wide range. In the transformation of music, it is observed that blockchain technology could play a significant role in music production, distribution, and consumption. Particularly, by reducing the number of intermediaries and enabling instant payments, it can address complex payment issues within the music industry. Additionally, blockchain technology has the capacity to enhance copyright management and traceability of music compositions by providing a decentralized structure. It can assist artists in gaining more control and transparency by enabling digital ownership, tracking, and sharing of music compositions. The historical and cultural evolution of music, combined with technological advancements and new business models, is shaping the future of the music industry. The impact of this technology on the future of the music industry is of great importance in terms of preserving cultural heritage and valuing artists’ efforts. This article aims to delve deeper into understanding the potential impact of blockchain technology in the field of music and to discuss the transformation in the industry. To achieve this goal, after examining the current state and dynamics of the music industry, a comparative analysis of global blockchain music companies will be conducted. The analysis results will provide a discussion on the potential contribution of blockchain technology to the music industry and its possible effects. ----- ## 2. METHOD This article employs a qualitative research method to examine the potential and transformative effects of blockchain technology on the music industry. Specifically, the focus has been on how blockchain can impact the music industry and how new business models can be defined through technology. “Qualitative research can be defined as a series of interpretive techniques that attempt to explain, analyze, and translate concepts and phenomena rather than record their frequency in society” (Van Maanen, 1983). For this purpose, a qualitative approach has been adopted since it deals with “how” questions. The research initiated with a comprehensive literature review. The literature review laid the foundation for data collection and analysis by providing guiding frameworks for the research (Vom Brocke et al., 2015). Relevant sources were selected from platforms such as Scopus and Google Scholar, and an overview was obtained by skimming through identified texts. Online materials like social media content, blockchain platforms, and industry reports were also utilized to gain a comprehensive understanding. Additionally, the snowball sampling method was employed as an efficient way to find relevant literature in terms of time. In the initial phase, the supply chain processes and relationships of the traditional music industry were examined in detail through an exploratory analysis. This analytical approach was utilized to comprehend how the chain operates and to identify challenges within these processes. The same analytical method was applied in the exploration of new music platforms supported by blockchain technology. Prominent blockchain-based platforms like Resonate, Opus, Musicoin, and Audius were investigated at this point. The functionality, purpose of use, adopted practices, and how they are used were systematically explored using a comprehensive content analysis method. Content analysis proved to be a critical tool in shedding light on the unique features and usage patterns of each platform. Through this analytical approach, the advantages, challenges, functionality, and purpose of use of each platform were discussed in detail. In this context, this study comprehensively addresses the impact of blockchain technology on the digital transformation of the music industry. Comparing the traditional music industry’s supply chain model with the potential offered by blockchain technology provides valuable perspectives for the industry’s future evolution. This analytical framework aims to contribute to a broader understanding of the transformation of the music industry within a larger context. ## 3. DEVELOPMENT OF DIGITAL MUSIC AND CHALLENGES IN THE EVOLUTION OF THE INTERNET The evolution of the internet has revolutionized the music industry and brought significant changes to how music is created, distributed, and consumed. Digital music, defined in its fundamental sense, is a visual-auditory medium stored in digital format that can be transmitted over the internet and wireless networks. When compared to traditional music, digital music is not only low-cost, highly efficient, and personalized, but also caters to the consumption needs of consumers in the era of new technologies. The internet has completely transformed music distribution today by enabling easier and faster access to music. Instead of traditional physical formats like vinyl, cassette, or CDs, the internet allows music to be downloaded digitally or streamed online through streaming services, making music more accessible. Digital technologies and the internet have democratized the process of music creation and recording. Digital audio workstations (DAWs) and various software tools allow musicians to create professional-quality music from their homes. Additionally, online collaboration platforms facilitate musicians’ collaboration from around the world. Moreover, new avenues have been provided for artists and record labels to promote and market their music. Through social media, music videos, online radio, podcasts, music platforms like Spotify and Apple Music, and other digital platforms, artists can reach a global audience. Music listeners can now interact with their favorite artists through social media and ----- listen to artists live through online concerts and live streams. However, this transition has come with various challenges and impacts. The launch of the iTunes Music Store by Apple in April 2003 is considered a significant milestone in the digital music transformation. This platform reduced the cost of downloading a single song to $0.99 and an album to $9.99 through iTunes 4.0, providing a 33% discount compared to traditional CD formats (Dutra et al., 2018). This move encouraged the consumption of digital music and marked a significant transformation in the music industry. The price reduction made music more accessible on digital platforms and influenced music consumers’ habits. Digital music and the internet have created new challenges and opportunities in terms of copyright and licensing. Artists and rights holders are required to change how they manage their music’s online use and revenue generation. This transformation has brought both new opportunities and challenges. Digital music and the internet have created new challenges and opportunities in the realms of copyright and licensing, necessitating a change in how artists and rights holders manage the use of their music online and how they earn income from it. This process of transformation has brought forth numerous new prospects online streaming, while download rates are alongside its challenges. For instance, issues like piracy and copyright infringements have emerged as significant problems affecting both the music industry and artists. In recent years, the rise in popularity of music streaming services has somewhat mitigated piracy, as these services often offer users access to a vast music library at a low cost or for free. Nevertheless, piracy continues to pose a significant challenge for the music industry. To address these and other issues, the music industry and technology developers continually explore new solutions and models. Furthermore, many artists contend that the revenue derived from music streaming platforms is unfair. Notably, musicians, including influential figures like David Bowie, have been at the forefront of advocating for and actively engaging in discussions on this transformation (For more detailed information, refer to Pareles, 2002). However, during the complex transitional period spanning from 2000 to 2015, public discourse paid limited attention to how musicians would generate income in this emerging digital age, the funding sources available to them, and the means by which they could sustain their music careers. Discussions during this period primarily revolved around speculations regarding new opportunities and changes, with relatively little focus on the income-generation declining. Table 1. Global Recorded Music Industry Revenues 1999 - 2022 (Billion US Dollars) (ifpi.org) **Table 1. Global Recorded Music Industry Revenues 1999 - 2022 (Billion US Dollars) (ifpi.org)** ----- challenges faced by musicians (Hesmondhalgh, 2021, p. 3594). The global music industry continues to grow in recent years. According to the International Federation of the Phonographic Industry (IFPI), which measures the music industry’s growth, it reported a total revenue of $26.2 billion in 2022 based on data from record companies. IFPI notes that while revenue from physical formats (such as vinyl and CD revenue) has decreased over the past decade, digital revenue has increased. Furthermore, there is an observed increase in online streaming, while download rates are declining. The revenue from digitally sold music has been unequally distributed among stakeholders in the music industry. In broad terms, the revenue from music streams is divided as follows: 30% to on-demand streaming services, 60% to record companies and publishers, and 10% to songwriters, artists, and music groups. According to analyses, Apple Music has paid unsigned artists $0.0064 and signed artists $0.0073, while Spotify has paid $0.007 and $0.0044 respectively. In 2017, in the United States, for an artist to earn the minimum wage of $1,472, their songs would need to be streamed around 230,000 times on Apple and 380,000 times on Spotify. For YouTube, considering an artist receives only $0.003 per stream, their content would need to be streamed around 4.2 million times (Sanchez, 2017). Ensuring fair and equitable distribution of revenues among rights holders, particularly among stakeholders, is critical for the sustainability of the industry and to support artists. At this point, collaboration among all stakeholders in the industry is necessary to explore appropriate solutions. ## 4. BLOCKCHAIN TECHNOLOGY “Blockchain is a shared, trusted, public ledger that everyone can inspect, but which no single user controls. It operates by consensus, and once recorded, the data in any given block cannot be altered retroactively.” (BlockchainHub, 2023). Since its introduction through the Bitcoin whitepaper published by an anonymous individual or group using the pseudonym Satoshi Nakamoto in 2008, blockchain technology has come a long way (Nakamoto, 2008). A blockchain consists of a virtual chain of blocks, each with a unique identifier (referred to as a hash) and containing information such as financial transactions, contracts, or other documents. A blockchain operates on a decentralized network of computers (referred to as nodes) collectively verifying the information entering a block. Reaching a consensus on what information should be included in a block is necessary to minimize the chances of accepting incorrect information, as nodes mostly reject a block without the need for a central entity (Peters and Panayi, 2016). The database is distributed based on the principle that each copy of new data is sent to not just one computer but to all users in the chain or system. To change any bit of the database, hackers would need to change the copies of inputs in the system by 51%, and each copy would need to include all previous interactions with that data (Nguyen & Dang, 2018, pp. 483-484). In essence, no singular entity owns a blockchain, making it immutable and devoid of a single point of vulnerability for those attempting to hack or otherwise tamper with the data in the blockchain ledger. For this reason, blockchain is the first technology to enable the transfer of digital ownership in a decentralized and trustless manner (Iinuma, 2018). Creating a blockchain transaction involves the following steps: defining the transaction and providing access to the sender network, including the recipient’s address, transaction value, and digital signature. Nodes verify the user’s digital signature through encryption. The verified transaction is added to a pool. Pending transactions are combined into a block, creating an updated record maintained by a node. The block is accepted by the network’s verification nodes and added to the blockchain. This process is typically completed within 2 to 10 seconds (Gheorghe et al., 2017, p. 218). Blockchain provides high security and flexibility through high interaction, successfully eliminating third parties and rendering processes more transparent, democratic, decentralized, costeffective, and secure. This technology has various applications, including smart contracts, supply ----- chain traceability, digital identity verification, and many more. Blockchain technology offers transformation potential across numerous industries through these and other applications. Due to its decentralized and transparent nature, blockchain can offer a reliable framework for copyright management. Through smart contracts, automatic and transparent revenue sharing can occur between copyright holders and licensees. Additionally, copyright tracking and monitoring processes can be automated, reducing copyright infringements and disputes. In the field of music distribution, blockchain technology can enable artists to directly reach listeners and eliminate the costs of traditional intermediaries. This could create a fairer and more sustainable revenue model, particularly for independent artists. ## 5. USAGE AND PRACTICE OF BLOCKCHAIN IN MUSIC In the digital age, music is considered data, and metadata is the data about that data, containing information about the music itself. Metadata embedded in each recorded music track can include usage conditions and contact details of copyright holders, making it easier to locate owners of a recorded music piece and acquire licenses. The concept is to attribute a purpose to music, allowing it to act as if it were alive. Gradually placing copyright data onto the blockchain could eventually lead to the creation of a comprehensive copyright database for music (LO’Dair, 2016). In the contemporary music landscape, the fusion of blockchain technology, smart contracts, and cryptocurrency is forming the foundation of a new music ecosystem that reflects inclusivity, integrity, transparency, and fair compensation ethics. Producers and consumers of digital music content are deciding how to share their content in the online world. On these new-generation platforms, artists can easily upload their music and associated content to a centralized online location, making it accessible to everyone. Rights, ownership, and usage of the content shift the focus from traditional music company or distributor policies to a technically artist centric model built on blockchain architecture. This model enables artists to offer their work for listening, sharing, remixing, or purchase directly to audiences. (Tapscott & Tapscott, 2016, pp. 287290). The music industry ecosystem is a centralized database network. These databases connect rights and licensing flows while providing a revenue stream. The DotBlockchain architecture is designed for Blockchain technology, aiming to develop the future music ecosystem by utilizing a balanced ring architecture. This architecture encompasses all participants from traditional labels and publishers to performing rights organizations and composition editors. Collaborating partners can store their data in a metadata chain by combining their individual databases. This chain resides within a public data block. The DotBlockchain architecture works compatibly with existing media formats, maintaining data safety and accuracy (Gheorge, 2017, pp. 2022-2024). However, integrating blockchain technology into the music industry could face challenges such as standardizing copyright management and licensing processes and creating a legal framework. Additionally, collaboration and data sharing among all stakeholders need to be encouraged. The various roles and applications of blockchain technology in the music industry include: **Copyrights and Licensing:** Blockchain can be used to verify and track ownership of a song or album. This facilitates the verification of copyright and licensing information for each track, leading to more accurate revenue distribution. **Music Distribution: Artists and groups can** distribute music directly to consumers using blockchain technology. This bypasses traditional distribution channels, giving them more control and potential revenue. **Micro Payments: Blockchain facilitates artists** receiving micro payments for their tracks. This allows listeners to directly purchase specific songs or albums. ----- **NFTs (Non-Fungible Tokens): Artists can use** NFTs to create unique digital products. This provides fans with the opportunity to own unique pieces and offers artists new revenue streams. **Interaction** **with** **Fans:** Some artists use blockchain technology to engage more with their fans. They can provide exclusive access and experiences using tokenized rewards. Nevertheless, the full impact of blockchain technology on the music industry is still unfolding and evolving. However, this technology has significant potential to fundamentally change how musicians create, distribute, and earn from their music. ## 6. INNOVATIVE BLOCKCHAIN- BASED PLATFORMS IN MUSIC DISTRIBUTION AND A LOOK INTO THE FUTURE The music industry is undergoing profound changes due to the impact of digital transformation. Challenges such as traditional distribution models, copyright issues, and the lack of fair compensation for artists necessitate new and innovative solutions for music to adapt to the digital age. At this juncture, blockchain technology comes into play, enabling data to be stored transparently, securely, and in a decentralized manner. However, the widespread adoption and acceptance of these platforms by the general public can give rise to significant challenges, considering factors such as technological capabilities, user behavior, and industry standards. Blockchain-based music platforms are offering a new perspective to the music industry, reshaping the interaction between artists and listeners. However, how these platforms will be embraced as alternatives to traditional music distribution models and how they will impact the music industry will be better understood through future studies and adoption processes. Since its early days, blockchain technology has garnered significant interest across various industries. Platforms like Bittunes, Ujo Music, Voise, Musicoin, and Resonate are standout examples of blockchain-based streaming platforms that have emerged in recent years. These platforms promise to employ smart contracts to reward artists and pledge fair compensation. They also provide the capability for users to directly tip artists. However, the acquisition of the necessary cryptocurrency (such as Bitcoin, Ethereum, or Musicoin) for these platforms might not be as user-friendly as the payment processes of traditional music platforms, potentially slowing down the adoption process (Sciaky, 2019). Some of the innovative platforms in the music industry are as follows: **Audius[1]: Built on the Ethereum blockchain,** Audius allows artists to independently release their music and interact directly with listeners. This enables artists to overcome the limitations of traditional music distribution channels and reach broader audiences, effectively marketing their music (Audius, 2023). Audius boasts several important features that set it apart from other blockchain-based music platforms. These features enable the platform to provide a more effective and appealing experience for users. The user-friendly interface facilitates the rapid adoption of Audius. Both artists and listeners find navigation and content uploading on the platform hassle-free, ensuring a more comfortable and enjoyable user experience. Audius’ ability to provide wider access is also a noteworthy feature. When artists upload their music to the platform, they can reach listeners from different cultures and geographies, allowing their music to reach broader audiences. Audius’ innovative business models distinguish it from other platforms. Artists can choose to make their music available for free listening or license it for a certain fee. Additionally, adjusting usage rights for works based on different regions or platforms is also possible. These features differentiate Audius from other blockchainbased music platforms. The platform presents an innovative approach aimed at providing a more sustainable, fair, and enriching music experience for both artists and listeners. **Musicoin[2]:** Operating on micro-payments between artists and listeners, Musicoin offers a ----- fair payment model. As listeners enjoy music, they can make payments to artists using cryptocurrency. Artists, in turn, are rewarded with the Musicoin cryptocurrency as they share their content. This approach enables artists to better determine the value of their music and manage their copyrights more directly. Artists can establish closer connections with their listeners, receive feedback, and even offer exclusive content or experiences for a certain amount of Musicoin. This not only allows for listening to music but also enriches the experience by forming a more personal connection with artists. Being free and ad-free, this platform utilizes the Universal Basic Income (UBI) model, ensuring that each contribution is fairly rewarded (Musicoin, 2023). **Resonate3: It presents an alternative approach** to subscription-based models like Spotify and Apple Music. It offers music to listeners at affordable prices while committing to providing artists with higher payments compared to their competitors. Operating on a blockchain-based democratic governance system, it ensures that artists receive their earnings through a perlistener payment model, while listeners gain access to music through a fair subscription model. One of Resonate’s most striking features is the “Stream2Own” model. In this model, users make micro-payments for each streaming session, and the amount they pay is instantly transferred to the artist’s wallet. This enables artists to earn instant income from every play, fostering a more equitable distribution of revenue compared to traditional music streaming platforms. Additionally, the Resonate platform grants artists more control over how they license their music. Artists can determine the usage terms for their works, which are automatically enforced through smart contracts. This enhances the protection of copyright and empowers artists to manage their music. Another area where Resonate stands out is user experience. The platform allows users not only to listen to music but also to get closer insights into artists’ stories and music creation processes. This approach transforms music into not just sound but also a story and experience, fostering a deeper connection. Resonate’s unique features offer a fresh perspective on the digitalization of the music industry, with functions like fair revenue sharing, licensing control, and enhanced music experience. Serving as a robust and effective bridge between artists and music consumers, this platform positions itself as a contender to shape the future of music by innovating in the realms of fair revenue distribution, licensing control, and music experience. **Opus[4]: It stands out as a blockchain-based** platform built to store and distribute high-quality audio files. It specifically enables the storage of high-resolution audio files in the FLAC (Free Lossless Audio Codec) format. The FLAC format maintains audio quality while incorporating compression capabilities. This provides music artists and producers with the capacity to preserve their creations at the highest level. Opus’s primary goal is to enhance audio quality within the music industry. Traditional digital music platforms often use compressed audio formats, which may lead to quality degradation. Opus, on the other hand, aims to deliver a superior listening experience by offering highresolution audio through the FLAC format. Opus is built on a blockchain technology that ensures fair revenue sharing. Artists receive direct payments as listeners engage with their music on the platform. Smart contracts facilitate revenue sharing based on predefined ratios. Furthermore, Opus enables accurate management of copyright. Artists can establish usage terms for their works, and smart contracts automatically enforce these conditions. The platform supports various licensing models. Artists can make their works available for free streaming or license them for a specified fee. Additionally, they can customize the usage rights based on geographical regions or platforms. With a global vision, Opus provides access to listeners and artists worldwide. When artists upload their works to the platform, listeners from different geographies and cultural backgrounds can access these creations. Opus’s fundamental aim is to deliver a high-quality audio experience while ensuring fair revenue sharing and copyright management. With its innovative approach, Opus contributes to a more transparent and accessible future for the music industry. ----- These platforms in the music industry carry the potential to offer a more equitable and transparent experience to artists and listeners. They exemplify instances of the digital transformation within the music industry. While each platform shares a similar core purpose and functionality, their unique features and intended uses exhibit noticeable differences. Notably, Resonate’s innovative distribution model, Opus’s high-quality audio storage concept, Musicoin’s copyright management, and Audius’s decentralized music streaming platform all stand out for their distinctive attributes. The adoption process of these platforms and their impacts on the music industry will become clearer based on the outcomes of future research and studies. ## 7. DISADVANTAGES OF BLOCK- CHAIN-BASED MUSIC PLATFORMS While blockchain platforms offer several advantages, they also come with certain disadvantages. The requirement to transact with cryptocurrencies and the complexity of payment processes can be significant factors limiting user acceptance. Blockchain-based music platforms are often designed to facilitate payments with cryptocurrencies and manage copyright, which may demand users to possess cryptocurrencies or purchase them. Despite the prevalence of cryptocurrencies, many individuals may still have limited knowledge or desire to engage with them. This can lead to hesitation among music enthusiasts or artists to use such platforms. The complexity of payment processes can also pose a barrier. Payments on blockchain-based platforms are typically conducted through smart contracts, which can be different and more technical compared to traditional payment methods. Users may need to understand and navigate these processes correctly. Additionally, factors such as the volatility of cryptocurrency values and the verification process of payment transactions can complicate the payment experience for users. These disadvantages, especially for users who are less familiar with technology or have limited exposure to cryptocurrencies, can reduce their willingness to adopt the platforms. User tendencies to prioritize security and simplicity can influence the adoption rate of these platforms. Therefore, platform providers might reach a broader audience by offering user-friendly interfaces, simplifying payment processes, and supporting traditional payment methods instead of cryptocurrencies. Taking these issues into account, the future success of blockchain-based music platforms will depend on how effectively they can make the user experience simple and secure. Another disadvantage is scalability issues. Blockchain infrastructure might struggle to handle high-volume transactions, limiting the platforms’ growth and reach to a wider user base. Scalability issues might become even more pronounced if the platform gains popularity. Energy consumption is another concern. Some blockchain protocols can require high energy consumption for transaction validation, raising environmental and sustainability concerns. Transaction speed and duration could also pose a disadvantage. During peak periods or increased network traffic, transaction speeds could slow down, making it unsuitable for scenarios requiring instant payments or quick transactions. Data storage concerns should also be considered. Since blockchain records every transaction on a public ledger, safeguarding personal or sensitive data might be challenging. This becomes a risk, especially when situations demand the storage of confidential information. Lastly, the legal and regulatory aspects of blockchain technology are still uncertain. Matters like the legal status and taxation of cryptocurrencies can vary from country to country. These uncertainties might affect the operational processes of platforms. These disadvantages stand out as factors that could limit the widespread acceptance and usage of blockchain-based music platforms. However, with developers’ efforts to address these challenges and the evolution of technology, these disadvantages could be overcome over time, allowing platforms to reach a broader user base. ----- ## 8. CONCLUSION The “blockchain” technology has rapidly advanced and is widely used in various fields. Today, music works have transitioned from the recording era to digitalization. Digitized music works can be accessed widely via the “Internet,” but they also bring along contentious copyright issues. Effective strategies using blockchain technology in the future will make it possible to protect digital music copyrights on the internet, which will greatly enhance the development of the digital music industry. Blockchain has the potential to simplify many complex processes by offering benefits such as reduced transaction costs, payment speed, elimination of intermediaries, and sharing of copyright fees through smart contracts, compared to traditional methods. However, there are disadvantages associated with blockchain technology, including the lack of recognition among potential users, indifference towards music that can be accessed in this way, and the high volatility of cryptocurrency values. Thus, there is a prevalence of speculative and misinformation-laden discourse surrounding blockchain technology. In the music sector, technology benefits both in the online sale of concert tickets and in delivering live music performances to listeners, benefiting amateur musicians and small groups as well. The nature of open-source software offered by technology allows all these processes to be decentralized, enabling users and existing institutions in the music industry to set up their own web stores and reach listeners directly. Additionally, by creating synchronous streams that suit the nature of music, it provides brand new performance models and economic gain systems. Blockchain technology can play a significant role in areas such as copyright management, music distribution, and artist compensation in the music industry. However, given the technical, legal, and collaboration challenges, the adoption process for this technology will be lengthy and require careful planning. In the future, it is expected that blockchain technology will bring more innovation and transformation to the music industry. While blockchain platforms offer several advantages, they also come with certain disadvantages. Blockchain-based music platforms face disadvantages such as the requirement to transact with cryptocurrencies and the complexity of payment processes. These factors can limit user acceptance and particularly deter users unfamiliar with cryptocurrencies. Additionally, challenges like scalability issues, high energy consumption, transaction speed, and data storage concerns can restrict the expansion and usage of these platforms. Although these issues can be overcome with solutions that enhance user experience, it’s anticipated that as technology evolves and regulatory clarity is achieved, blockchain-based music platforms will become more widespread. **Endnotes** https://audius.co/ https://musicoin.org/ https://resonate.coop/ https://opus.audio/ **REFERENCES** AUDIUS. 2023. Audius. Accessed on June 9, 2023. https://audius.co/ BLOCKCHAINHUB (2017). Blockchain Explained - Intro - _Beginners Guide to Blockchain. Available at:_ https://blockchainhub.net/blockchain-intro/ (Accessed on May 17, 2023). DUTRA, A., TUMASJAN, A., & WELPE, I. M. (2018). Blockchain is Changing How Media and Entertainment Companies Compete. _MIT Sloan Management Review,_ Fall 2018. GHEORGHE, P. N. M., TIGĂNOAIA, B., & NICULESCU, A. (2017, October). Blockchain and smart contracts in the music industry–streaming vs. downloading. _International Conference on Management_ _and Industrial Engineering (pp. 215-228)._ _Niculescu_ _Publishing House._ HESMONDHALGH, D. (2021). Is music streaming bad for musicians? Problems of evidence and argument. _New Media & Society,_ _23(12), 3593–_ ----- [3615. https://doi.org/10.1177/1461444820953541](https://doi.org/10.1177/1461444820953541) IINUMA, A. (2018, April 5). What Is Blockchain And What Can Businesses Benefit From It? Forbes. https://www. forbes.com/sites/forbesagencycouncil/2018/04/05/ what-is-blockchain-and-what-can-businesses-benefitfrom-it/?sh=3aef0a6f675f (Accessed on July 5, 2023). LO’DAIR, M. (2016), Music on The Blockchain. [online]. Blockchain For Creative Industries Research Cluster Middlesex University, Report No. 1, https://www. mdx.ac.uk/__data/assets/pdf_file/0026/230696/MusicOn-The-Blockchain.pdf (Accessed on July 10, 2023). MOUGAYAR, W. (2016). _The business blockchain:_ _promise, practice, and application of the next Internet_ _technology. John Wiley & Sons._ MUSICOIN. 2023. Musicoin. Accessed on June 9, 2023. https://musicoin.org/. NAKAMOTO, S. (2008). Bitcoin: A peer-to-peer electronic _cash system. https://bitcoin.org/bitcoin.pdf (Accessed_ on July 10, 2023). NGUYEN, Q. K., & DANG, Q. V. (2018, November). Blockchain Technology for the Advancement of the Future. _2018 4th international conference on green_ _technology and sustainable development (GTSD) (pp. 483-_ 486). IEEE. OPUS. (2023). Accessed on June 9, 2023. https://opus. audio/ PANAYİ, P. (2016), Why Us, Why Now: Convening the Open Music Initiative. _BerkleeICE, https://www._ berklee.edu/panos-panay-open-music-initiative (Accessed on July 3, 2023) PARELES, J. (2002). David Bowie, 21st-century entrepreneur. The New York Times, 9(06), 2002. PETERS, G.W., PANAYİ, E. (2016). Understanding modern banking ledgers through blockchain technologies: Future of transaction processing and smart contracts on the internet of money. _Banking_ _Beyond Banks and Money. Springer, pp. 239–278._ RESONATE. (2023). Resonate Mission. Accessed on June 9, 2023. https://resonate.is/. SANCHEZ, D. (2017). What Streaming Music Services Pay. _Digital Music News, https://www._ digitalmusicnews.com/2017/07/24/what-streamingmusic-services-pay-updated-for-2017/. (Accessed on July 1, 2023). SCIAKY, D. (2019). The digital transformation of the music industry through applications of blockchain technology. https://kth.diva-portal.org/smash/record. jsf?pid=diva2%3A1375894&dswid=984 (Accessed on June 1, 2023). TAPSCOTT, D., & TAPSCOTT, A. (2016). _Blockchain_ _revolution: how the technology behind bitcoin is changing_ _money, business, and the world. Penguin._ VAN MAANEN, J. (1983). _Qualitative methodology._ Sage. VOM BROCKE, J., SIMONS, A., RIEMER, K., NIEHAVES, B., PLATTFAUT, R., CLEVEN, A. (2015). Standing on the Shoulders of Giants: Challenges and Recommendations of Literature Search in Information Systems Research. Communications of the Association of _Information Systems, 37(9)._ -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.31566/arts.2163?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.31566/arts.2163, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://journals.gen.tr/index.php/arts/article/download/2163/1441" }
2,023
[ "JournalArticle" ]
true
2023-10-23T00:00:00
[]
9,408
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Education", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00f24857803eefcd0900689dda52c32505105132
[ "Computer Science" ]
0.905613
THUNDER: helping underfunded NPO’s distribute electronic resources
00f24857803eefcd0900689dda52c32505105132
Journal of Cloud Computing: Advances, Systems and Applications
[ { "authorId": "145729001", "name": "Gabriel Loewen" }, { "authorId": "2068756", "name": "Jeffrey M. Galloway" }, { "authorId": "2115206220", "name": "Jeffrey A. Robinson" }, { "authorId": "1696277", "name": "X. Hong" }, { "authorId": "1770322", "name": "Susan V. Vrbsky" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
As federal funding in many public non-profit organizations (NPO’s) seems to be dwindling, it is of the utmost importance that efforts are focused on reducing operating costs of needy organizations, such as public schools. Our approach for reducing organizational costs is through the combined benefits of a high performance cloud architecture and low-power, thin-client devices. However, general-purpose private cloud architectures are not easily deployable by average users, or even those with some computing knowledge. For this reason, we propose a new vertical cloud architecture, which is focused on ease of deployment and management, as well as providing organizations with cost-efficient virtualization and storage, and other organization-specific utilities. We postulate that if organizations are provided with on-demand access to electronic resources in a way that is cost-efficient, then the operating costs may be reduced, such that the user experience and organizational efficiency may be increased. In this paper we discuss our private vertical cloud architecture called THUNDER. Additionally, we introduce a number of methodologies that could enable needy non-profit organizations to decrease costs and also provide many additional benefits for the users. Specifically, this paper introduces our current implementation of THUNDER, details about the architecture, and the software system that we have designed to specifically target the needs of underfunded organizations.
http://www.journalofcloudcomputing.com/content/2/1/24 ## RESEARCH Open Access # THUNDER: helping underfunded NPO’s distribute electronic resources ### Gabriel Loewen[*], Jeffrey Galloway, Jeffrey Robinson, Xiaoyan Hong and Susan Vrbsky **Abstract** As federal funding in many public non-profit organizations (NPO’s) seems to be dwindling, it is of the utmost importance that efforts are focused on reducing operating costs of needy organizations, such as public schools. Our approach for reducing organizational costs is through the combined benefits of a high performance cloud architecture and low-power, thin-client devices. However, general-purpose private cloud architectures are not easily deployable by average users, or even those with some computing knowledge. For this reason, we propose a new vertical cloud architecture, which is focused on ease of deployment and management, as well as providing organizations with cost-efficient virtualization and storage, and other organization-specific utilities. We postulate that if organizations are provided with on-demand access to electronic resources in a way that is cost-efficient, then the operating costs may be reduced, such that the user experience and organizational efficiency may be increased. In this paper we discuss our private vertical cloud architecture called THUNDER. Additionally, we introduce a number of methodologies that could enable needy non-profit organizations to decrease costs and also provide many additional benefits for the users. Specifically, this paper introduces our current implementation of THUNDER, details about the architecture, and the software system that we have designed to specifically target the needs of underfunded organizations. **Introduction** Within the past several years there has been a lot of work in the area of cloud computing. Some may see this as a trend, whereas the term “cloud” is used simply as a buzzword. However, if viewed as a serious contender for managing services offered within an organization, or a specific market, cloud computing is a conglomerate of several very desirable qualities. Cloud computing is known for being scalable, which means that resource availability scales up or down based on need. Additionally, cloud computing represents highly available and ondemand services, which allow users to easily satisfy their computational needs, as well as access any other required services, such as storage and even complete software systems. Although there is no formal definition for cloud computing, we define cloud computing as a set of serviceoriented architectures, which allow users to access a number of resources in a way that is elastic, cost-efficient, and on-demand. General cloud computing can be separated into three categories: Infrastructure-as-a-Service (IaaS), [*Correspondence: gloewen@crimson.ua.edu](mailto:gloewen@crimson.ua.edu) Department of Computer Science, The University of Alabama, Tuscaloosa, AL, USA Platform-as-a-Service (PaaS), and Software-as-a-Service (SaaS). Infrastructure-as-a-Service provides access to virtual hardware and is considered the lowest service layer in the typical cloud stack. An example of Infrastructureas-a-Service is the highly regarded Amazon EC2, which is subsystem of Amazon Web Services [1]. At the highest layer is Software-as-a-Service, which provides complete software solutions. An example software solution, which exists as a cloud service is Google Docs. Google Docs is a SaaS which gives users access to document editing tools, which may be used from a web browser. In between SaaS and IaaS is Platform-as-a-Service, which allows users to access programming tools and complete API’s for development. An example of a PaaS is Google AppEngine, which gives developers access to robust API’s and tools for software development in a number of different languages. We are beginning to see many software services being offered by a number of public cloud providers, including image editing software, email clients, development tools, and even language translation tools. However, these tools are all offered by different providers and are not necessarily free for general use. © 2013 Loewen et al.; licensee Springer. This is an Open Access article distributed under the terms of the Creative Commons [Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction](http://creativecommons.org/licenses/by/2.0) in any medium, provided the original work is properly cited. ----- http://www.journalofcloudcomputing.com/content/2/1/24 Considering that non-profit organizations cannot always afford to purchase access to software, we propose that these organizations should simply maintain their own private cloud, which could decrease the costs associated with software licensing. There are several freely available cloud architectures that may be considered. However, general-purpose cloud architectures are not suitable for organization that do not have highly trained professionals to manage such a system. This downfall of most generalpurpose architectures is due to the lack of an easy to use user interface and somewhat complicated deployment process. Many architectures, such as Eucalyptus [2] and OpenStack [3], rely heavily on the command line for interfacing with the system, which isn’t desirable for markets that do not have experts readily available for troubleshooting. A cloud architecture designed for these specific markets must have the following attributes: ease of deployment, user friendly interface, energy efficiency, and cost effectiveness. In consideration of these qualities we have designed a new IaaS cloud architecture, which we call THUNDER (THUNDER Helps Underfunded NPO’s Distribute Electronic Resources). THUNDER utilizes the notion of simplicity at all levels in order to ensure that all users, regardless of their technical experience, will be able to use the system or redeploy the architecture if necessary. Most IaaS cloud architectures rely upon the general case model. In the general case, an IaaS cloud architecture supports low-level aspects of the cloud stack, such as hardware virtualization, load balancing of virtual machine instances, elastic storage, and modularity of physical hardware. Vertical clouds, on the other hand, are defined by a specific market, and therefore, are able to abstract the general case IaaS cloud model to provide features that are tailored for a specific set of uses. We see vertical clouds predominantly in the healthcare sector with the ehealth cloud architecture. The THUNDER architecture is an abstraction of the general case model by taking care of the low-level details of hardware virtualization, load balancing, and storage in a way that is considerate of the technical maturity of the users, as well as the level of expertise expected from the administrators. This abstraction is possible in a vertical cloud designed for the non-profit sector because we can make an assumption about the maximum number of virtual machines, the type of software required, and the expected level of experience of the users. We assume the number of virtual machine instances is congruent to the number of client devices in an office or computer lab. Additionally, the software available on the cloud is defined by a set of use cases specific to the organization. For example, THUNDER deployed to a school may be used in conjunction with a mathematics course, which would be associated with a virtual machine image containing mathematics software, such as Matlab or Maple. Additionally, we assume that the technical experience of administrators and instructors in a school setting is low. Therefore, by deviating from the general case model of an IaaS cloud architecture, and by considering the special needs of the market, we can minimize the complexity of deployment by removing the necessity of a fine-tuned configuration. In the following sections we discuss related background work in private vertical cloud architectures, our proposed architecture, future work, and we end with a summary and conclusion. **Background and motivation** There has been much discussion on the topic of cloud computing for various administrative purposes at educational institutions. However, cloud computing is a topic that until recently has not been widely considered for the high school grade bracket. Due to the nature of cloud computing, being a service oriented architecture, there is a lot of potential in adopting a cloud architecture that can be used in a classroom [4]. Cloud computing in the classroom could be used to provide valuable educational tools and resources in a way that is scalable, and supportive of the ever-changing environment of the classroom. Production of knowledgeable students is not a trivial task. Researchers in education are focused on providing young students with the tools necessary to be productive members of society [4]. The past decade has seen, in some cases, a dramatic decrease in state and local funding for public secondary education. This reduction in funding indicates that a paradigm shift in how technology is utilized in the classroom is necessary in order to continue to provide high quality education. The authors of [4-7] believe that cloud computing may be a viable solution to recapture students’ interests and improve student success. **Education** Researchers at North Carolina State University (NCSU) have developed a cloud architecture, which is designed to provide young students with tools that help to engage students in the field of mathematics [4]. This cloud architecture, known as “Virtual Computing Lab” or “VCL”, has been provided as a public service to rural North Carolina 9th and 10th grade algebra and geometry classes. The goal of this study is to broaden the education of STEM related topics using the VCL in these schools, and two applications were selected to be used in the course curriculums: Geometer’s Sketchpad 5, and Fathom 2. The authors describe a set of key challenges that were encountered during the study, including: diversity of software, software licensing, security, network availability, life expectancy of hardware, affordability, as well as technical barriers. Software availability is a prime concern when it comes to provisioning educational tools for academic use. ----- http://www.journalofcloudcomputing.com/content/2/1/24 The specific needs of the classroom, in many cases, require specific software packages. When deploying software to a cloud architecture, it is not always possible to provide certain software packages as cloud services. For this reason, it is common to bundle software with virtual machine images, which are spawned on an IaaS cloud. A virtual machine image is a single file that contains a filesystem along with a guest operating system and software packages. Additionally, software packages may have some conflicts with one another that can create an issue with the logistics of the system [4]. Another software concern is related to software-specific licensing, and how it affects the cloud. Many software packages require licensing fees to be paid per user of the system, or as a volume license, which may or may not impose a maximum number of users allowed access to the software. Therefore, depending on the specific requirements of the school and course, software licensing fees must be paid for accordingly. For example, when geometers sketchpad was deployed to the VCL, the authors made sure that the software licensing fees were paid for in accordance with the software publishers’ license agreement. The necessity for licensing does affect the cost effectiveness of using a cloud in this setting, however it is no different than licensing software for traditional workstations [4]. The authors of [8] have created a private cloud architecture, called CloudIA, which supports e-Learning services at every layer of the cloud stack. At the IaaS layer, the CloudIA architecture supports an automated virtual machine image generator, which utilizes a web interface for creating custom virtual machine images with predefined software packages installed. At the PaaS layer, the CloudIA architecture supports computer science students with a robust API for writing software that utilizes cloud services. At the SaaS layer, the CloudIA architecture supports collaborative software for students to utilize for projects and discussion. The authors of [9] describe the benefits of cloud computing for education. The main point that the authors make is that cloud computing provides a flexible and cost effective way to utilize hardware for improving the way information is presented to students. Additionally, the authors describe details about the ability of cloud computing to shift the traditional expenses from a distributed IT infrastructure model to a more pay-as-you-go model, where services are paid for based on specific needs. Authors of [5] discuss “Seattle”, which is a cloud application framework and architecture, enabling users to interact with the cloud using a robust API. By using this platform students can execute experiments for learning about cloud computing, networking, and other STEM topics. The authors also describe a complimentary programming language built upon Python, which gives students easy access to the Seattle platform. The authors of [10] discuss a new model for SaaS, which they have named ESaaS. ESaaS is defined as a Softwareas-a-Service cloud architecture with a focus on providing educational resources. The authors discuss the need for a managed digital library and a global repository for educational content, which is easily accessible through a web interface. The proposed architecture is meant to integrate into existing secondary and post-secondary institutions as a supplementary resource to their existing programs. **LTSP** One approach is the use of thin client devices, which have been used in other educational endeavors, such as the Linux Terminal Server Project (LTSP) [11]. Thin client solutions, when paired with an IaaS cloud, offer low power alternatives to traditional computing infrastructures. The authors of [12] analyze energy savings opportunities in the thin-client computing paradigm. Authors of [13] discuss design considerations for a low power and modular cloud solution. In this study the LTSP [11] architecture is reviewed and compared to the authors cloud architecture design. LTSP is a popular low power thin client solution for accessing free and open source Linux environments using a cluster of server machines and thin client devices. The LTSP architecture provides services, which are very similar to an IaaS cloud architecture with a few notable limitations. Firstly, LTSP only offers Linux environments, which differs from an IaaS cloud in that the cloud can host Linux, Windows, and in some instances Apple OSX virtual machine instances. Additionally, LTSP does not utilize virtualization technology, rather it provides several minimal Linux and X windows environments on the same host computer. Interfacing with an LTSP instance also differs from an IaaS cloud in that an LTSP terminal will boot directly from the host machine using PXE or NetBoot, which is a remote booting protocol. A client connected to an IaaS cloud will typically rely upon the Remote Desktop Protocol (RDP) for accessing Windows instances, or the Virtual Network Computing (VNC) protocol for Linux instances. **Other work** All of the previous work relate to educational resources and services in the cloud. However, most of the related work is integrated using public cloud vendors and is specific towards one particular subject, as is presented in [4] and [5]. The authors of [14] present their solution, SQRT-C, which is a light-weight and scalable resource monitoring and dissemination solution using the publisher/subscribe model, similar to what is described in this manuscript. The approach considers three major design implementations as top priority: Accessing physical ----- http://www.journalofcloudcomputing.com/content/2/1/24 resource usage in a virtualized environment, managing data distribution service (DDS) entities, and shielding cloud users from complex DDS QoS configurations. SQRT-C can be deployed seamlessly in other cloud platforms, such as Eucalyptus and OpenStack, since it relies on the libvirt library for information on resource management in the IaaS cloud. In [15], the authors propose a middleware for enterprise cloud computing architectures that can automatically manage the resource allocation of services, platforms, and infrastructures. The middleware API set used in their cloud is built around servicing specific entities using the cloud resources. For end users, the API toolkit provides interaction for requesting services. These requests are submitted through a web interface. Internal interface APIs communicate between physical and virtual cloud resources to construct interfaces for users and determine resource allocation. A service directory API is provided for users based on user privileges. A monitoring API is used to monitor and calculate the use of cloud system resources. This relates to the middleware introduced in this manuscript; however, it addresses architectures more suitable for large enterprises. The authors of [16] propose a resource manager that handles user requests for virtual machines in a cloud environment. Their architecture deploys a resource manager and a policy enforcer module. First, the resource manager decides if the user has the rights to request a certain virtual machine. If the decision is made to deploy the virtual machine, the policy enforcer module communicates with the cloud front-end and executes an RPC procedure for creating the virtual machine. Authors of [17] describe how cloud platforms should provide services on-demand that helps the user complete their job quickly. Also mentioned is the cloud’s responsibility of hiding low-level technical issues, such as hardware configuration, network management, and maintenance of guest and host operating systems. The cloud should also reduce costs by using dynamic provisioning of resources, consuming less power to complete jobs (within the job constraints), and by keeping human interaction to cloud maintenance to a minimum. Development of cloud APIs is discussed in [18]. The author mentions three goals of a good cloud API: Consistency, Performance, and Dependencies. Consistency implies the guarantees that the cloud API can provide. Performance is relatively considered in forms of decreasing latency while performing actions. Cloud dependencies are other processes that must be handled, other than spawning virtual machines and querying cloud resource and user states. These three issues are considered in the development process of our own IaaS cloud architecture. **Proposed architecture** Our focus is to provide underfunded non-profit organizations with the means to facilitate the computing needs of their users in a cost-effective manner. The THUNDER architecture is composed of a special purpose private cloud stack, and an array of low power embedded systems, such as Raspberry Pi’s [19] or other low-power devices. The THUNDER stack differs from the general-purpose private cloud model in a number of ways. General purpose cloud stacks, such as Eucalyptus [2] and OpenStack [3], are focused on providing users with many different options as to how the cloud can be configured. These general-purpose solutions are great for large organizations because the architecture is flexible enough to be useful for diverse markets. However, non-profit organizations do not typically have the resources to construct a generalpurpose cloud architecture. Therefore, a special-purpose or vertical cloud architecture is desirable because it circumvents the typical cloud deployment process by making assumptions about the use of the architecture. THUNDER may be utilized by various NPO’s and for various purposes, but a secondary focus of THUNDER is focused on the education market. Research in cloud computing for education has shown that educational services in high school settings are successful in motivating students to learn and achieve greater success in the classroom [4]. The THUNDER cloud stack utilizes a number of commodity compute nodes, in addition to persistent storage nodes with a redundant backup, as well as a custom DHCP, MySQL, and system administration server. Each compute node is capable of accommodating four Windows virtual machines or twelve Linux virtual machines. The lab consists of low-power client devices with a keyboard, mouse, and monitor connected to a gigabit network. A custom web-based interface allows users to login, select their desired virtual machine from a list of predefined images, and then launch the virtual machine image. For example, students taking a course in Python programming might be required to use a GNU/Linux based computer for development. However, a receptionist in an office setting might be required to use a Microsoft Windows system. Therefore, regardless of the user requirements, THUNDER will be able to provide all necessary software components to each user independently. Figure 1 illustrates the THUNDER network topology. The THUNDER network topology resembles a typical cloud topology, where the compute cluster is connected to a single shared LAN switch, and support nodes share a separate LAN switch. Additionally, the topology shows the client devices and how they interface with the rest of the system. Table 1 shows a power cost comparison between THUNDER and a typical 20 PC lab, and shows a possible savings of 50% when compared to a traditional computer lab. ----- http://www.journalofcloudcomputing.com/content/2/1/24 **Figure 1 Networking topology for the THUNDER cloud architecture.** **Network topology** The THUNDER network topology in Figure 1 is most cost efficient when combined with low-power or thi– client devices, but can also be paired with regular desktop and laptop computers. One of the common advancements in wired Ethernet technology is the use of switches. **Table 1 Power cost comparison between THUNDER and a** **typical lab** **Typical lab** **THUNDER** **#** **Hardware** **Watts** **#** **Hardware** **Watts** 20 PC Desktops 6,000 20 Thin client 60 20 Display 2,000 20 Display 2,000 3 Compute node 1,200 2 Storage node 500 1 Admin/Web 250 Total 8,000 Total 4,010 Monthly bill: $152 Monthly bill: $77 Ethernet switches allow for adjacent nodes connected to the switch to communicate simultaneously without causing collisions. The network interface cards used in all of the devices of THUNDER support full duplex operation, which further allows nodes to send and receive data over the network at the same time. Ethernet is a LinkLayer protocol, which determines how physical devices on the network communicate. The clients communicate with THUNDER through simple socket commands and a virtual desktop viewing client, such as VNC or RDP viewers. **Compute and store resources** THUNDER compute and storage resources will consume a considerable amount of network bandwidth. The compute nodes are responsible for hosting virtual machines that are accessed by the clients. These compute nodes will mount the user’s persistent data as the virtual machine is booting. Each compute node will communicate with the THUNDER cloud resources and the client devices using a 1 Gbps network interface. The client devices should ----- http://www.journalofcloudcomputing.com/content/2/1/24 be equipped with a 10/100/1000 Mbps network adapter, and considering the limited number of cloud servers, it is unlikely that the 1 Gbps network switch will become completely saturated with traffic. Userspace storage nodes are connected to the same switch as the compute nodes, which allows for tighter coupling of storage nodes and compute nodes, decreasing the potential delay for persistent file access. Backup storage nodes are connected to a separate network switch, and are used to backup the cloud system in case of a system failure. **_Administrative resources_** The THUNDER administrative resources include a MySQL database, Web interface, and Networking services. These services are hosted on a single physical machine, with a backup machine isolated to the same network switch. There will be no need for a high amount of resources in the administrative node since the numbers of compute and storage nodes determine the amount of clients that can be connected to THUNDER. The THUNDER cloud, when accessed from devices external to the organization’s private network, can be routed to a secondary administrative node, such that the on-site users will not experience any quality of service issues. **Network provisioning for low latency** Using a top down approach, the amount of bandwidth (maximum) needed for a twenty node THUNDER lab can be described. If we assume that each client device requires a sustained 1 Mbps network throughput, we would need to accommodate for sustaining 20 Mbps within the network switch used by the clients. Since the network switch is isolated to communicating with the resources of THUNDER, this throughput needs to be sustainable on the uplink port. This is relatively easy, given the costs of gigabit switches on the market today. The specification that needs close attention is the total bandwidth of the switch backplane. Making the assumption that this bandwidth is the number of ports multiplied by the switch speed is not always true. In our case, the bandwidth needed, 20 Mbps, is much lower than the maximum throughput of a twenty-four port gigabit switch. There is little to no communication between THUNDER compute node resources. These devices are used to host virtual machines that are interfaced to the clients directly. Given a THUNDER lab size of twenty clients, the network bandwidth needed on the isolated network containing the compute nodes should be above 20 Mbps, assuming each client consumes 1 Mbps of bandwidth. The THUNDER storage node resources are also isolated to the same gigabit network switch as the compute node resources. When the user logs into THUNDER and requests a virtual machine, their persistent storage is mounted inside the virtual machine for them to use. The data created by the users has to be accessed while they are using a virtual machine. **Middleware design and implementation** One of the core components in building a cloud architecture is the development of a middleware solution, allowing for ease in resource management. Additionally, in order to improve the quality of service (QOS) an emphasis on minimizing resource utilization and increasing system reliability is desirable. Our reasoning for developing a new cloud middleware API is to address issues that we have encountered in current cloud middleware solutions, which are centered upon ease of deployment and ease of interfacing with the system. Additionally, we have utilized our API to build a novel cloud middleware solution for use in THUNDER. Specifically, this middleware solution is designed for management of compute resources, including instantiation of virtual machine images, construction and mounting of storage volumes, metadata aggregation, and other management tasks. We present the design and implementation for our cloud middleware solution and we introduce preliminary results from our study into the construction of THUNDER, which is our lightweight private vertical IaaS cloud architecture. Management of resources is a key challenge in the development of a cloud architecture. Moreover, there is a necessity for minimizing the complexity and overhead in management solutions in addition to facilitating attributes of cloud computing, such as scalability and elasticity. Another desirable quality of a cloud management solution is modularity. We define modularity as the ability to painlessly add or remove components on-the-fly without the necessity to reconfigure any services or systems. The field of cloud management exists within several overlapping domains, which include service management, system deployment, access control management, and others. We address the requirements of a cloud management middleware API, which is intended to support the implementation of the private cloud architecture currently in development. Additionally, we compare our cloud management solution to solutions provided by freely available private IaaS cloud architectures. When examining the current state of the art in cloud management, there are few options. We are confined to free and open source (FOSS) cloud implementations, such as Eucalyptus [2] and Openstack [3]. Cloud management solutions used in closed-source, and often more popular cloud architectures, such as Amazon EC2, are out of reach from an academic and research perspective due to their closed nature. However, there has been an effort to make Eucalyptus and Openstack compatible with Amazon EC2 by implementing a compatible API and command line tools, such as eucatools [20] and Nova [21], respectively. The compatibility of API’s makes it easy to form a basis ----- http://www.journalofcloudcomputing.com/content/2/1/24 of comparison between different architectures. Although, this compatibility may also serve as a downfall because if one API suffers from a bug, it may also be present in other API’s. **Eucalyptus discussion** The methodology for management of resources in Eucalyptus is predominantly reliant upon establishing a control structure between nodes, such that one cluster is managed by one second-tier controller, which is managed by a centralized cloud controller. In the case of Eucalyptus, there are five controller types: cloud controller, cluster controller, block-based storage controller (EBS), bucketbased storage controller (S3), and node controller. The cloud controller is responsible for managing attributes of the cloud, such as the registration of controllers, access control management, as well as facilitating user interaction through command-line and, in some cases, webbased interfacing. The cluster controller is responsible for managing a cluster of node controllers, which entails transmission of control messages for instantiation of virtual machine images and other necessities required for compute nodes. Block-based storage controllers provide an abstract interface for creation of storage blocks, which are dynamically allocated virtual storage devices that can be utilized as persistent storage. Bucket-based storage controllers are not allocated as block-level devices, but instead are treated as containers by which files, namely virtual machine images, may be stored. Node controllers are responsible for hosting virtual machine instances and for facilitating remote access via RDP [22], SSH [23], VNC [24], and other remote access protocols. **OpenStack discussion** Similar to the methodology used by Eucalyptus, OpenStack also maintains a control structure based on the elements present in the Amazon EC2 cloud. OpenStack maintains five controllers: compute controller (Nova), object-level storage (Swift), block-level storage (Cinder), networking controller (Quantum), and dashboard (Horizon). There are many parallels between the controller of OpenStack and the controllers of Eucalyptus. The Nova controller of OpenStack is similar to the node controller of Eucalyptus. Similarly we see parallels between Swift in OpenStack with the bucket-based controller in Eucalyptus, and Cinder in Openstack with the block-based storage of Eucalyptus. There seems to be a discretion in implementation between the highestlevel controller in each architecture. OpenStack maintains different controllers for interfacing and network management, while Eucalyptus maintains a single cloud controller combining these functionalities. Additionally, OpenStack does not maintain a higher-level control structure for managing compute components, which is a deviation from the cluster controller mechanism present in Eucalyptus. **Middleware interfacing, communication, and** **authentication** In developing our middleware solution we encountered challenges regarding the method by which it would interface with the various resources in the cloud. Many different methodologies were considered. However, we decided to use an event-driven mechanism, which is similar to remote procedure calls (RPC). One of the prime differences in the way Eucalyptus and OpenStack perform management tasks is in the means of communication. Eucalyptus utilizes non-persistent SSH connections between controllers and nodes in order to remotely execute tasks. OpenStack, on the other hand utilizes remote procedure call, or RPC’s. In keeping with the methodology introduced by OpenStack and its current momentum in the open source cloud computing community, we utilize an event driven model, which presents a very similar mechanism to that of RPC. However, these two architectures share a common component. They both utilize the libvirt [25] library, which is the same library that we utilize in our architecture. Additionally, authentication was a challenge because in reducing the complexity of authentication we introduce new possible security threats. Although, we believe the security threats posed by our authentication model are minimal, additional threats could be uncovered during system testing. We believe that this solution is important because we address concerns regarding the overall usage of the cloud architecture, and our initial performance results in Figure 2 show that our middleware performs well when compared to Eucalyptus [2]. **Node-to-node communication scheme** In contrast to the methodologies used by Eucalyptus, OpenStack, and presumably Amazon, our cloud middleware API addresses resource management in a simplified and more direct manner. The hierarchy of controllers used in Eucalyptus introduces extra complexity that we have deemed unnecessary. For this reason, our solution utilizes a simple publisher/subscriber model by which compute, storage, and image repository nodes may construct a closed network. The publisher/subscriber system operates in conjunction with event driven programming, which allows events to be triggered over the private network to groups of nodes subscribed to the controller node. Figure 3 shows the logical topology and lines of communication constructed using this model. In constructing the communication in this manner we are able to broadcast messages to logical groups in order to gather metadata about the nodes subscribed to that group. Message passing is useful for retrieving the status ----- http://www.journalofcloudcomputing.com/content/2/1/24 **Figure 2 Performance comparison of NetEvent and SSH authentication protocols.** of nodes, including virtual machine utilization, CPU and memory utilization, and other details pertaining to each logical group. Additionally, we are able to transmit messages to individual nodes in order to facilitate virtual machine instantiation, storage allocation, image transfer, and other functions that pertain to individual nodes. **_Registration of nodes_** Communication between nodes utilizes non-persistent socket connections, such that the controller node maintains a static pre-determined port for receiving messages, while other nodes may use any available port on the system. Thus, each node in the cloud, excluding the controller node, automatically selects an available port at boot time. Initial communication between nodes is done during boot time to establish a connection to the controller node. We utilize a methodology for automatically finding and connecting to the controller node via linear search over the fourth octet of the private IP range (xxx.xxx.xxx.0 to xxx.xxx.xxx.255). Our assumption in this case is that the controller node will exist on a predefined subnet that allows us to easily establish lines of communication without having to manually register nodes. Additionally, we can guarantee sequential ordering of IP addresses with our privately managed DHCP server. Once a communication link is established between a node and the controller node, the node will request membership within a specific logical group, after which communication between the controller node and that logical group will contain the node in question. The registration methodology used in our middleware solution differs from the methodology used by Eucalyptus and OpenStack. For example, Eucalyptus relies upon command line tools to perform RSA keysharing and for **Figure 3 Logical topology - logical groups represent group-wise membership in publisher/subscriber model.** ----- http://www.journalofcloudcomputing.com/content/2/1/24 establishing membership with a particular controller. We do not perform key sharing, and instead rely upon a pre-shared secret key and generated nonce values. This approach is commonly known as challenge-response [26], and it ensures that nodes requesting admission into the cluster are authentic before communication is allowed. When a node wishes to be registered as a valid and authentic node within a cluster, a nonce value is sent to the originating node. The node will then encrypt the nonce with the pre-shared key and transmit the value back to the controller. We validate the message by comparing the decrypted nonce produced by the receiver and the nonce produced by the sender. Thus, we do not rely upon manual sharing of RSA keys beforehand, and instead we eliminate the need for RSA keys altogether and utilize a more dynamic approach for validation of communication during the registration process. Figure 4 presents the registration protocol. **Middleware API** As stated in the introduction, our methodology for constructing a middleware API for cloud resource management centers around the decreasing overhead when compared to general-purpose solutions. In order to facilitate a simple middleware solution, our API was designed to provide a powerful interface for cloud management while not introducing excessive code overhead. We have titled our API “NetEvent”, which is indicative of its intended purpose as an API for triggering events over a network. This API is utilized within our private IaaS cloud architecture as a means for communication, management of resources, and interaction with our cloud interface. Figures 5 and 6 illustrate the manner in which the API is accessed. Although, the code examples presented here are incomplete, they illustrate the simplicity of creating events to be triggered by the system for management of resources. In Figure 5 we present sample code for the creation of a controller node, which is responsible for relaying commands from the web interface to the cloud servers. In Figure 6 we present a skeleton for the creation of a compute node with events written for instantiation of virtual machine images and for retrieving the status of the node. Although, we do not present code for the implementation of storage or image repository nodes, the implementations are similar to that of the compute node. In addition, the code examples presented in this paper show only a subset of the functionality contained within the production code. The API presented here provides a powerful interface for implementing private cloud architectures. By means of event triggering over a private network we are able to instantiate virtual machine images, mount storage volumes, retrieve node status data, transfer virtual machine images, monitor activity, and more. The implementation of the system is completely dependent upon the developer’s needs and may even be used in distributed systems, which may or may not be implemented as a **Figure 4 Node registration protocol.** ----- http://www.journalofcloudcomputing.com/content/2/1/24 **Figure 5 Example controller node service written in pseudocode.** cloud architecture. This approach is different than the more traditional approach of remote execution of tasks by means of SSH tunneling. **User interfacing** In the previous section we introduced our middleware API for managing cloud resources. However, another important component is a reasonable way to interface with the middelware solution. Although, the middleware API solution is completely independent from the interface, we have chosen to use a message passing approach that is different from that of general-purpose architectures. In this approach our web interface, which is written in PHP, connects to the controller node in order to trigger the “INVOKE” event. By interfacing with the controller node we are able to pass messages to groups or individual nodes in order to manage the resources of that node and receive responses. The ability to interface in this manner allows our interface to remain decoupled from the logical implementation, while allowing for flexibility in the interface and user experience. Figure 7 shows an example PHP script for interfacing with the resources in the manner described in this section. **Figure 6 Skeleton for compute node service written in pseudocode.** ----- http://www.journalofcloudcomputing.com/content/2/1/24 **Figure 7 Example communication interface in PHP.** The PHP interface presented in Figure 7 illustrates the methodology behind how we may capture and display data about the nodes, as well as provide a means for user interaction in resource allocation and management. Although, we do not present the full source code in this paper, additional functions could be written. For example, a function could be written that instructs compute nodes to instantiate a particular virtual machine image. One important aspect of this system is that the mode of communication remains consistent at every level of the cloud stack. Every message sent is implemented via nonpersistent socket connections. This allows for greater data consistency without modifying the semantics of messages between the different systems. Figure 8 shows an example interface for metadata aggregation of a logical compute group. Figure 9 presents a sequence diagram for the VM selection interface. **Supporting storage services** Pinnacle to the development of a complete cloud architecture, and a pre-requisite to supporting compute services is the ability for a cloud middleware to support the mounting and construction of persistent storage volumes. Storage service support is a pre-requisite of compute services because it is common for virtual machine images to reside on a separate image repository or network attached storage device. Therefore, before compute services can be fully realized it is necessary to be able to mount the image repository, such that the local hypervisor may have access to the virtual machine images. We can support storage services using the storage driver provided by libvirt. Figure 10 shows the XML specification provided to libvirt, which is required by the storage driver. Once the storage pool has been mounted, then the user of the cloud may be provided access to storage space, if it is persistent userspace. Alternatively, if the share is a image repository, then the compute node will be given access to the virtual machine images provided by the storage pool. **Supporting compute services** The NetEvent API allows for services to be written and distributed to nodes within a private cluster. These services utilize the NetEvent API as a means for triggering events remotely. Within cloud architectures there are a few important events that must be supported. Firstly, the instantiation of virtual machine images must be supported by all cloud architectures. Compute services may be supported by combining the flexibility of the NetEvent API and a hypervisor, such as KVM. A proper compute service should maintain an image instantiation event which invokes the hypervisor and instructs it to instantiate a specific virtual machine image. The steps involved in supporting compute services start with mounting the storage share containing the virtual machine images. This is made possible with the function, _mountVMPool, which constructs a storage pool located in_ the directory “/var/lib/iibvirt/images”, and is the default location by which libvirt may locate the available domains or virtual machine images available to the system. Once the virtual machine pool is mounted then a specific virtual machine may be instantiated, which is made possible with the function, instantiateVM. This function looks up the virtual machine, and if it exists in the storage pool, it will be instantiated. Once the VM is instantiated, a domain object will be returned to the node, which provides the methods for managing the virtual machine instantiation. ----- http://www.journalofcloudcomputing.com/content/2/1/24 **Figure 8 Example interface for metadata aggregation with two nodes being polled for data.** **Figure 9 Sequence diagram for virtual machine image selection process.** ----- http://www.journalofcloudcomputing.com/content/2/1/24 **Figure 10 Storage pool XML specification required by libvirt.** **Supporting metadata aggregation** Metadata aggregation refers to the ability to retrieve data about each node within a specific group. This data may be used for informative purposes, or for more complex management tasks. Example metadata includes the nodes IP address, operating system, and kernel. Additionally, dynamic data may be aggregated as well, including RAM availability and CPU load. We can support metadata aggregation in each service by introducing events that retrieve the data and transmit it to the controller node. **Performance results** One of the many reasons for not using SSH, which seems to be the industry standard approach for inter-node communication in general-purpose cloud architectures, is that SSH produces excessive overhead. The communication approach used by NetEvent is very simplified and does not introduce data encryption or a lengthy handshake protocol. The downfall of simplifying the communication structure is that the system becomes at risk for loss of sensitive data being transmitted between nodes. However, in the case of this system no sensitive data is ever transmitted, and instead only simple commands are ever sent between nodes. For this reason encryption is unnecessary. However, authentication is still required in order to determine if nodes are legitimate. In testing the performance of NetEvent we compared the elapsed time for authenticating a node with the controller and establishing a connection with the elapsed time for SSH to authenticate and establish a connection. We gathered data over five trials, which is presented in Table 2. Additionally, Figure 2 presents the average latencies between SSH and NetEvent. From the performance comparison we draw the conclu sion that general-purpose cloud architectures that utilize SSH connections, such as Eucalyptus, sacrifice up to a 99% loss in performance when compared to traditional sockets. However, this comparison is being made at optimal conditions, because the servers are under minimal load. More data needs to be gathered to determine how much the performance is affected when the servers are overloaded. **Supporting software services** The preceding sections discussed our implementation of the software system necessary for supporting IaaS cloud services, namely hardware virtualization and persistent storage. Building upon virtualization of hardware, we are able to provide software services as custom virtual machine instances. The approach that THUNDER takes is instantiation of server virtual machine images, which deploy web services for user collaboration, research, and other tools and utilities. By implementing services in this fashion, no modifications are required to the infrastructure of the cloud, and administrators may easily start services by allocating hardware resources and stop services by deallocating resources. This approach differs from typical SaaS architectures in that no additional configuration is necessary outside of what is required for regular instantiation of virtual machines. The only difference is that regular users do not have access to connecting to service instances directly using VNC. **Future work and conclusion** We would like to introduce this architecture in a select number of organizations in order to determine the effectiveness and usability of the architecture from both the user’s and administrator’s perspectives. Based on the results of the study, we will alleviate any possible concerns from users or administrators. We plan to form an incremental process, such that various aspects of the system are studied in different organizational environments, then small changes will be made to the system **Table 2 Performance results comparing NetEvent to** **traditional SSH-based authentication** **Trial #** **SSH (ms)** **NetEvent (ms)** 1 298 3.1 2 301 3.2 3 302 9.2 4 298 3.1 5 299 3.0 ----- http://www.journalofcloudcomputing.com/content/2/1/24 before running another study. In this fashion, we will have more control over which features are beneficial to organizations, and which features are least significant. Currently, the THUNDER cloud is comprised of a set of standalone servers which are not organized in a shared structure, such as a rack-style chassis. We would like to build a complete prototype that is presentable and able to be easily taken to organizations for demonstration purposes. One of the key challenges in building a cloud infrastructure is the development of a middleware solution, which allows for ease in resource management. The work presented in this paper demonstrates that a middleware solution does not have to be as complex as those found in the popular cloud architectures, Eucalyptus and Openstack. We also introduced the model by which our middleware API offers communication between nodes, namely utilizing event driven programming and socket communication. We have developed our API to be efficient, light weight, and easily adaptable for the development of vertical cloud architectures. Additionally, we showed the manner in which a web interface may interact with the middleware API in order to send messages and receive responses from nodes within the cloud. For future work we would like to investigate approaches for fault tolerance in this architecture. Additionally, we would like to perform an overall system performance benchmark and make comparisons between other cloud architectures. We would also like to implement a method for obfuscation of management traffic such that the system may not be as susceptible to malicious users. We have presented our work in designing and implementing a new private cloud architecture, called THUNDER. This architecture is implemented as a vertical cloud, which is designed for use in non-profit organizations, such as publicly funded schools. We leverage a number of technologies, such as Apache2 web server, and MySQL for implementation of the architecture. Additionally, we introduce socket programming and RPC as a viable alternative to the more common SSH based solution for inter-node communication. We have established that the primary goal of THUNDER is not to replace traditional private cloud architectures, but to serve as an alternative, which is custom tailored for reducing complexity, costs, and overhead in underprivileged and underfunded markets. We also demonstrate that if an organization were to adopt the THUNDER architecture they could benefit by reducing up to 50% of their power bill due to the low power usage when compared to a traditional computer lab. We believe that the power and cost savings, when combined with the features and qualities of THUNDER as presented in this paper, make THUNDER a desirable architecture for school computer labs and other organizations. Further studies and analysis will validate the effectiveness of the architecture. **Competing interests** The authors declare that they have no competing interests. **Authors’ contributions** GL performed the research, design, and development of the THUNDER architecture. JG and JR revised the manuscript and contributed to the background work. XH provided insight and guidance in developing the networking model for THUNDER. SV edited and revised the final manuscript. All authors read and approved the final manuscript. Received: 11 August 2013 Accepted: 14 December 2013 Published: 21 December 2013 **References** 1. [Amazon Web Services. [http://aws.amazon.com/]. Accessed: 12/18/2012](http://aws.amazon.com/) 2. [Eucalyptus Enterprise Cloud. [http://eucalyptus.com/]](http://eucalyptus.com/) 3. [OpenStack. [http://openstack.org/]](http://openstack.org/) 4. Stein S, Ware J, Laboy J, Schaffer HE (2012) Improving K-12 pedagogy via a Cloud designed for education. Int J Informat Manage. [[http://linkinghub.elsevier.com/retrieve/pii/S0268401212000977]](http://linkinghub.elsevier.com/retrieve/pii/S0268401212000977) 5. Cappos J, Beschastnikh I (2009) Seattle: a platform for educational cloud [computing. ACM SIGCSE 111–115. [http://dl.acm.org/citation.cfm?id=](http://dl.acm.org/citation.cfm?id=1508905) [1508905]](http://dl.acm.org/citation.cfm?id=1508905) 6. Donathan K, Ericson B (2011) Successful K-12 outreach strategies. In: Proceedings of the 42nd ACM technical symposium on Computer [science education, pp 159–160. [http://dl.acm.org/citation.cfm?id=](http://dl.acm.org/citation.cfm?id=1953211) [1953211]](http://dl.acm.org/citation.cfm?id=1953211) 7. Ercan T (2010) Effective use of cloud computing in educational institutions. Procedia - Social and Behavioral Sci 2(2):938–942. [[http://linkinghub.elsevier.com/retrieve/pii/S1877042810001709]](http://linkinghub.elsevier.com/retrieve/pii/S1877042810001709) 8. Doelitzscher F, Sulistio A, Reich C, Kuijs H, Wolf D (2010) Private cloud for collaboration and e-Learning services: from IaaS to SaaS. Comput 91: [23–42. [http://www.springerlink.com/index/10.1007/s00607-010-0106-z]](http://www.springerlink.com/index/10.1007/s00607-010-0106-z) 9. Sultan N (2010) Cloud computing for education: A new dawn? Int J [Inform Manage 30(2):109–116. [http://linkinghub.elsevier.com/retrieve/](http://linkinghub.elsevier.com/retrieve/pii/S0268401209001170) [pii/S0268401209001170]](http://linkinghub.elsevier.com/retrieve/pii/S0268401209001170) 10. Masud M, Huang X (2011) ESaaS: A new education software model in [E-learning systems. Inform Manage Eng 468–475. [http://www.](http://www.springerlink.com/index/H5547506220H73K1.pdf) [springerlink.com/index/H5547506220H73K1.pdf]](http://www.springerlink.com/index/H5547506220H73K1.pdf) [11. Linux Terminal Server Project. http://ltsp.org/. [Accessed: 12/23/2012]](http://ltsp.org/) 12. Willem Vereecken LD (2010) Energy efficiency in thin client solutions. GridNets 25:109–116 13. Cardellini V, Iannucci S (2012) Designing a flexible and modular architecture for a private cloud: a case study. In: Proceedings of the 6th international workshop on Virtualization Technologies in Distributed Computing Date, VTDC ’12. ACM, New York, NY, USA, pp 37–44. [[http://doi.acm.org/10.1145/2287056.2287067]](http://doi.acm.org/10.1145/2287056.2287067) 14. An K, Pradhan S, Caglar F (2012) Gokhale AA publish/subscribe middleware for dependable and real-time resource monitoring in the cloud. In: Proceedings of the Workshop on Secure and Dependable Middleware for Cloud Monitoring and Management, SDMCMM ’12. ACM, [New York, NY, USA, pp 1–3:6. [http://doi.acm.org/10.1145/2405186.](http://doi.acm.org/10.1145/2405186.2405189) [2405189]](http://doi.acm.org/10.1145/2405186.2405189) 15. Lee SY, Tang D, Chen T, Chu WC (2012) A QoS Assurance middleware model for enterprise cloud computing. In: IEEE 36th Annual Computer Software and Applications Conference Workshops (COMPSACW), 2012, pp 322–327 16. Apostol E, Baluta I, Gorgoi A, Cristea V (2011) Efficient manager for virtualized resource provisioning in cloud systems. In: IEEE International Conference on Intelligent Computer Communication and Processing (ICCP), 2011, pp 511–517 17. Khalidi Y (2011) Building a cloud computing platform for new possibilities. Computer 44(3):29–34 18. Pallis G (2010) Cloud computing: the new frontier of internet computing. IEEE Int Comput 14(5):70–74 [19. Raspberry Pi Foundation (2013). http://www.raspberrypi.org/](http://www.raspberrypi.org/) [Accessed: 11/15/2012] [20. EC2 Tools (2013). [http://www.eucalyptus.com/eucalyptus-cloud/tools/](http://www.eucalyptus.com/eucalyptus-cloud/tools/ec2) [ec2]](http://www.eucalyptus.com/eucalyptus-cloud/tools/ec2) [21. OpenStack Nova (2013). [http://nova.openstack.org/]](http://nova.openstack.org/) ----- http://www.journalofcloudcomputing.com/content/2/1/24 22. Surhone L, Timpledon M, Marseken S (2010) Remote desktop protocol. VDM Verlag Dr. Mueller AG & Company Kg, Saarbruecken, Germany 23. Barrett DJ, Silverman RE, Byrnes RG (2005) SSH, The secure shell: the definitive guide. O’Reilly Media, Sebastopol, CA, USA [24. VNC - Virtual network computing (2013). [http://www.hep.phy.cam.ac.uk/](http://www.hep.phy.cam.ac.uk/vnc_docs/index.html) [vnc_docs/index.html]. [Accessed: 12/18/2012]](http://www.hep.phy.cam.ac.uk/vnc_docs/index.html) [25. libvirt - The virtualization API (2013). http://libvirt.org/.](http://libvirt.org/) [Accessed: 7/21/2013] 26. M’Raihi D, Rydell J, Bajaj S, Machani S, Naccache D “OCRA: OATH Challenge-Response Algorithm”, RFC 6287. June 2011 doi:10.1186/2192-113X-2-24 **Cite this article as: Loewen et al.: THUNDER: helping underfunded NPO’s** distribute electronic resources. Journal of Cloud Computing: Advances, Sys_tems and Applications 2013 2:24._ -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1186/2192-113X-2-24?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1186/2192-113X-2-24, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://journalofcloudcomputing.springeropen.com/track/pdf/10.1186/2192-113X-2-24" }
2,013
[ "JournalArticle" ]
true
2013-12-01T00:00:00
[ { "paperId": "9ec00129eb540f667e750b1d1f6c1c3c10513d15", "title": "Improving K-12 pedagogy via a Cloud designed for education" }, { "paperId": "a39a58733def7147f0e8188676628098d7e92c9c", "title": "A publish/subscribe middleware for dependable and real-time resource monitoring in the cloud" }, { "paperId": "dc751f2c80660080741b22c73d4174af5e49a983", "title": "A QoS Assurance Middleware Model for Enterprise Cloud Computing" }, { "paperId": "f45fda5f62283d02d7a3c10e14d585fbab47ad5e", "title": "Designing a flexible and modular architecture for a private cloud: a case study" }, { "paperId": "1cf578c7693a804865d58353be48ff22ba2f85b5", "title": "Efficient manager for virtualized resource provisioning in Cloud Systems" }, { "paperId": "eeb1b41918b7e298d5b9c1511c8e6678a5ea013b", "title": "ESaaS: A New Education Software Model in E-learning Systems" }, { "paperId": "538f9486d42432ee85890e6d402c631e891f2311", "title": "OCRA: OATH Challenge-Response Algorithm" }, { "paperId": "edc1097dfb9bef7df0c0e185904e0ff7c2eab6d1", "title": "Successful K-12 outreach strategies" }, { "paperId": "77f35465e74d10349931f50cb7dcbab8648be885", "title": "Proceedings of the 42nd ACM technical symposium on Computer science education" }, { "paperId": "551e5e3adcc7bd7059bea60ac0c284ce63044261", "title": "Building a Cloud Computing Platform for New Possibilities" }, { "paperId": "d404b2193a62c0f99e37c33ed323fa69fb76a449", "title": "Cloud Computing: The New Frontier of Internet Computing" }, { "paperId": "1db3b7d018e2981d4436535a32afccf3741e8866", "title": "Cloud computing for education: A new dawn?" }, { "paperId": "8eb8251ec42f2a47501d72098000dad88700852c", "title": "Energy Efficiency in Thin Client Solutions" }, { "paperId": "634148c772d4affa2d8ed4235e14125b4a823276", "title": "Seattle: a platform for educational cloud computing" }, { "paperId": "54f67af65db0ae1b4a7d7a1b85022494d2679959", "title": "SSH, The Secure Shell: The Definitive Guide" }, { "paperId": null, "title": "Eucalyptus Enterprise Cloud" }, { "paperId": null, "title": "http://www.raspberrypi.org/ [Accessed: 11/15" }, { "paperId": "2686eeb5ae98872d0ab7418b244ef9765310700c", "title": "Private cloud for collaboration and e-Learning services: from IaaS to SaaS" }, { "paperId": "fdd3d480e784dddc7dc9ce0be48bfcc1b667bc06", "title": "Effective use of cloud computing in educational institutions" }, { "paperId": null, "title": "Remote desktop protocol. VDM Verlag Dr SSH, The secure shell: the definitive guide. O'Reilly Media" } ]
12,964
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Chemistry", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00f2f02c7694836eb4950f4395aae455571d91c1
[]
0.875635
Blockchain technology in quantum chemistry: A tutorial review for running simulations on a blockchain
00f2f02c7694836eb4950f4395aae455571d91c1
International Journal of Quantum Chemistry
[ { "authorId": "1395946625", "name": "M. W. Hanson-Heine" }, { "authorId": "1729356540", "name": "Alexander P. Ashmore" } ]
{ "alternate_issns": null, "alternate_names": [ "Int J Quantum Chem" ], "alternate_urls": [ "https://onlinelibrary.wiley.com/journal/1097461X", "http://onlinelibrary.wiley.com/journal/10.1002/(ISSN)1097-461X" ], "id": "b0b4cd85-08f5-4199-bd28-5512a3b0a0d9", "issn": "0020-7608", "name": "International Journal of Quantum Chemistry", "type": "journal", "url": "http://www3.interscience.wiley.com/cgi-bin/jtoc?ID=29830" }
null
DOI: 10.1002/qua.27035 #### R E V I E W # Blockchain technology in quantum chemistry: A tutorial review for running simulations on a blockchain ## Magnus W. D. Hanson-Heine [1] | Alexander P. Ashmore [2] 1School of Chemistry, University of Nottingham, Nottingham, UK 2Oxinet, Bath, UK Correspondence Magnus W. D. Hanson-Heine, School of Chemistry, University of Nottingham, University Park, Nottingham NG7 2RD, UK. [Email: magnus.hansonheine@nottingham.ac.](mailto:magnus.hansonheine@nottingham.ac.uk) [uk, magnus.hansonheine@gmail.com](mailto:magnus.hansonheine@nottingham.ac.uk) ### 1 | INTRODUCTION ### Abstract #### Simulations of molecules have recently been performed directly on a blockchain vir tual computer at atomic resolution. This tutorial review covers the current applica tions of blockchain technology for molecular modeling in physics, chemistry, and biology, and provides a step-by-step tutorial for computational scientists looking to use blockchain computers to simulate physical and scientific processes in general. Simulations of carbon monoxide have been carried out using molecular dynamics #### software on the Ethereum blockchain in order to facilitate the tutorial. K E Y W O R D S blockchain, computational science, distributed ledger, quantum chemistry Blockchain and distributed ledger technology is widely used in areas ranging from finance [1], to medicine [2], energy markets [3], and transparent and censorship resistant computers [4]. Blockchains have been integrated into chemical databases for genomics [5–21], electron microscopy [22], and automated experimental chemistry [23], and these databases aim to improve both privacy and openness when storing, accessing, and sharing chemical and molecular data. Blockchains have been integrated with machine learning techniques for computational modeling in order to facilitate data sharing and collaborative or federated model construction [24–41]. Distributed ledgers have driven novel algorithm developments in quan tum computing [42–52], blockchains have been used to create non-fungible tokens (NFTs) from scientific data [53], and blockchains have seen extensive use in the chemical industry [54–59]. However, the development of blockchains that are capable of performing generalized Turing complete computation now means that, in prin ciple, the full range of computational science, quantum chemistry, and quantum physics simulations that are possible using conventional com puters can also be implemented directly on blockchain computers. Quantum chemistry applications that use blockchain architectures to solve the Schrödinger equation have not yet been implemented; however, the implicit solutions for electronic structures that are contained within empirical parameterization of molecular mechanics and molecular dynamics force fields are available using existing software. Calculations of this type have been performed directly on a blockchain virtual computer to model the relatively simple physical process of a vibrating carbon monoxide molecule [60, 61]. The ability to perform computational and quantum science using blockchain computers is the focus of the current tutorial review. ### 2 | BLOCKCHAIN COMPUTERS—AN OVERVIEW Blockchains are a series of algorithms that enable peer-to-peer consensus on the state and history of a decentralized ledger in an open network [1]. Open public blockchains aim to store a provably tamperproof record of data or computational operations. This is done through the use of cryptographic hashing functions that map variable length input data to a fixed-length output called a hash. Any change in the input results in an [This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium,](http://creativecommons.org/licenses/by/4.0/) provided the original work is properly cited. ----- unpredictable but deterministic change to the hash, and consensus on the state of the blockchain is maintained using algorithms such as proof-of work [1, 62]. Proof-of-work algorithmic details and the details of blockchain virtual computer operations have been discussed in detail elsewhere [1, 4]. However, in basic terms, a blockchain virtual computer could be described using the following protocol: a set of desired computations, ter med “transactions,” are broadcast and executed by the computer nodes in the blockchain network including “miner” nodes. The outputs of these computations are hashed along with supplemental data by a miner node. The supplemental data includes an update to the state of the ledger that creates new tokens, or “coins,” that are under the miner node's control. The supplemental data is modified by a miner node until the numerical hash of the combined input is below a threshold (network difficulty), which is set in order to ensure that on average a certain amount of computa tional hashing work needs to be performed by a given miner in order for a valid combination of inputs to be found. Once a sufficiently small hash is found, the miner node that found the corresponding set of inputs broadcasts this new “block” of data to the rest of the network. Other nodes in the network then perform a much smaller number of hashing operations on these broadcast inputs in order to check that the hash is valid, and check that the other rules that the blockchain operates by have been followed. If the new block of data is accepted by the rest of the nodes in the network, then the miner receives control of the new tokens in addition to any fees that were sent to them as part of the computational work. This remunerates them for the costs associated with doing the hashing and broadcasting work. If the rules are not followed, then the new block is rejected by the other nodes, and the newly created tokens are not received in the copies of the blockchain held by the other nodes in the net work. This creates an incentive for accurate and honest computational work, and disincentivizes fraudulent or erroneous work. Newly appended blocks of data contain a hash of the previous block as part of their supplemental data, which then causes revisions to old blocks of data to invali date the hashes of the subsequent blocks, leading to them being rejected. The information in each “block” is therefore mathematically “chained” to the other blocks in the series through the sequential cryptographic hashes. All computer nodes in the network can therefore reliably maintain consensus on the state and history of the blockchain database, and on the computational work used to modify that state, creating a decentralized virtual computer. At the time of writing, the Ethereum blockchain and associated Ethereum virtual machine (EVM) is the oldest and most widely used blockchain that performs Turing complete general computation [4]. ### 3 | BLOCKCHAIN MOLECULAR DYNAMICS The first computational chemistry experiments performed entirely within a blockchain virtual computer have been reported in the scientific litera ture. These simulations involved modeling the vibrational motion of the carbon monoxide molecule on the Ethereum blockchain [60, 61]. In this case, atomic resolution molecular dynamics simulations were performed for the carbon monoxide molecule with a harmonic potential representing the energetics of the carbon–oxygen chemical bond over a 1 and 40 fs timeframe, respectively. Computer software compatible with the EVM was written in the specialized Solidity programming language, and was used to calculate a diatomic molecular dynamics trajectory with a variation FIGURE 1 Molecular dynamics trajectories showing carbon–oxygen bond length variations for the carbon monoxide molecule when (A) coded in C# and run on a local computer, and (B) coded in solidity and run on the Ethereum blockchain. Figure reproduced from HansonHeine and Ashmore [60] TABLE 1 The Ethereum blockchain transaction identification numbers (hexadecimal) for the simulations of carbon monoxide Operation Transaction identification number Software upload 0x1c98dbb671dbe76b7cb4188d7585296e5adbe0f64fe8144546b4a82081089152 Preliminary molecular dynamics 0x13eb9c343f380334262706975c676aa5006ac2103b2132dfba0e1667c7dfddc6 Production molecular dynamics 0x36b510f3bdb2fa67a0d4f749899ab8630b54000ee5a211172d699d750f94f94f ----- FIGURE 2 Abstract representation of computational state replication in blockchain virtual computers. Image credit: Giovanni Ciatto, University of Bologna of the velocity Verlet algorithm by integrating Newton's equations of motion in atomic units [63]. These simulations were executed for 10 and 400 time steps of 0.1 fs each, respectively, with an initial carbon–oxygen bond length of 120 pm. This model used an equilibrium bond length parameter of 112.8 pm and a force constant of 1855 N/m, together with the masses of [12]C and [16]O assigned to the atoms. An equivalent simula tion was also written using the C# programming language, and was executed on a local machine for comparison (see Figure 1). The outputs of these calculations were recorded on the Ethereum blockchain in blocks 9360161 and 9360178, respectively, and are available for the public to review. These specific transactions can be identified in the listed blocks using the hexadecimal identification numbers in Table 1. ### 4 | SMART CONTRACTS Ethereum, and many other blockchains, allow for software to be deployed and executed through transactions on the blockchain. These pieces of software are known as “smart contracts,” and since they are deployed as part of blocks, any contract added to the blockchain cannot be changed and any transactions interacting with the contract are also recorded permanently. This means that any deployed contract is always accessible regardless of the source code being lost, made closed source, or any hardware changes that might otherwise limit the accessibility. Interactions with the blockchain are broadcast as messages by the nodes that form part of the blockchain network and are packaged into blocks by miner nodes as part of the proof-of-work algorithm. These messages can include instructions to deploy or execute smart contracts, and are termed “transactions.” The term transaction is used in blockchain nomenclature as a holdover from the first blockchain, Bitcoin, which was ini tially used exclusively for financial payments, and this term is still used today when referring to the corresponding operations on blockchain net works that are capable of Turing complete computation. For example, the molecular dynamics simulation described here was carried out by executing a smart contract uploaded in Ethereum block 9360156 with the transaction identification number shown in Table 1. ### 5 | BLOCKCHAIN TIMESTAMPS When similar scientific discoveries are made independently, the original discovery is often attributed to the person considered to have made the relevant observations or calculations first. One of the more famous examples of this is the controversy between Leibniz and Newton over who invented calculus; however, this can be a regular occurrence when researchers share a common knowledge base and goal set. Knowing the order in which experiments were carried out is also useful when analyzing methods of hypothesis testing and conclusion formation that can differ depending on the order in which observations happen. An important property of many blockchains is therefore the creation of an internal chro nology. The regularity and sequential way in which new blocks of data are added and immutably chained to the existing blocks cryptographically, means that the position which calculations have in the blockchain acts as an automatic timestamp that can be used to verify the order in which the calculations were performed. In the case of the two carbon monoxide simulations discussed previously, the 1 fs simulation took place chrono logically prior to the 10 fs simulation. Under normal circumstances this statement would be hard to prove. However, in this case the two trajecto ries are recorded in Ethereum blocks 9360161 and 9360178 respectively which creates an effective timestamp proving the order in which the ----- FIGURE 3 Screen capture showing the default workspace in the Remix IDE FIGURE 4 Screen capture showing the provided contract in the text editor of the Remix IDE simulations were performed for as long as the Ethereum blockchain continues to operate with a coherent state and history. Similar uses of this time stamp data have been proposed in the blockchain integrated automatic experiment platform (BiaeP) protocol [23]. ### 6 | SCIENTIFIC REPRODUCTION Scientific reproduction, and the replication of findings, including computational work, is of critical importance for error checking and establishing consensus among scientists. The internal state of blockchain computers are reproduced by multiple independent local computer nodes (Figure 2). This makes them an ideal target for running scientific simulations in an openly verifiable and repeatable manner which can enable improved peer review and replication in the computational sciences. Any simulation written and deployed to the blockchain can always be found in the same place on that blockchain and the simulation can be verified by anyone running a node and can be re-executed by anyone using the blockchain ----- FIGURE 5 Screen capture showing the Remix IDE Solidity compiler tab FIGURE 6 Screen capture showing the Remix IDE “Deploy and Run Transactions” tab and willing to pay the transaction fees. Previous executions of smart contracts can also often be identified, and all of the parameters used to exe cute them can be easily verified and retrieved, ensuring that any simulation that was run can also be repeated with the computational certainty that executing the same code with the same parameters will result in the same output ----- FIGURE 7 Screen capture of the manual interface for the DiatomicMD contract in Remix IDE FIGURE 8 Screen capture of the Remix IDE transaction viewer, displaying the example simulation results ----- While this does not prevent a malicious contract, for example, one that would check the number of times it has been run and then produce a different result dependent on that number, the compiled bytecode of the contract can be retrieved from the blockchain and decompiled into code that can be manually checked for any errors or attempts at deception. This can be done regardless of whether the original source code is available, and the bytecode can also be used to confirm that the source code, if provided, is the same as the code that was deployed on the blockchain. FIGURE 9 Screen capture showing the manual interface for the DiatomicMD contract with the getSimOutput call expanded FIGURE 10 Screen capture showing the transaction viewer displaying results retrieved with a call to getSimOutput ----- FIGURE 11 Screen capture of the “Deploy and Run Transactions” tab after compiling the contract FIGURE 12 Screen capture of the MyEtherWallet website homepage ----- FIGURE 13 Screen capture of the MyEtherWallet website wallet access menu Smart contracts can make records and retrieve data from blocks, which can serve as an optional additional record of a simulation. While it is generally more computationally expensive to record data on the blockchain itself, if that data is to be accessed later multiple times, this could prove more cost effective, as well as making the simulation easier to verify, since, any output stored will have been confirmed and recorded in a way that makes tampering almost impossible when the blockchain is operating as intended. ### 7 | CENSORSHIP RESISTANT COMPUTATION Calculations on open public blockchains can be made available to anyone for review. Previous blockchain executions can be identified, verified, and executed again using the same code to produce the same output, and new pieces of software can be uploaded and exe cuted with pseudo-anonymity for the researchers involved. Once the transactions are broadcast and the correct fees are paid, a given computation cannot be stopped by any central authority so long as the blockchain operates normally. Blockchains can therefore be useful for scientists working under conditions where concerns about fraud and censorship become meaningful considerations. Blockchain consensus algorithms can create permanent and practically immutable or tamperproof records of computational data. While relatively rare, in extreme cases, state actors have been known to block or edit data contained in peer-reviewed journals [64, 65]. ----- FIGURE 14 Screen capture showing the MyEtherWallet “Deploy Contract” menu option ### 8 | BLOCKCHAIN LIMITATIONS The requirement that multiple computer nodes reproduce the computational work in each block in order to reach and maintain internal consensus currently makes the computational rate of blockchains significantly slower than conventional computers. More efficient architectures are cur rently under development; however, this remains a significant limitation. For example, the production molecular dynamics simulation of carbon monoxide previously mentioned in this tutorial review was included in Ethereum block 9360178, and this block was mined in 11 s, compared to � a 135 ms execution time for an equivalent C# simulation executed on a standalone desktop machine running with a i7-4790k CPU and 32GB of � 1333 MHz DDR3 RAM. The relative inefficiency of blockchain computation compared with trusted and centralized computational work, and the electricity needed to perform the hashing operations in proof-of-work, has led to open questions over the ecological impact of blockchain computation [66]. Consensus algorithms such as proof-of-stake and even green-proof-of-work have been proposed as energy efficient alternatives [67]. The mining competi tion in proof-of-work means that miners to use energy to reach consensus even when they fail to find the correct hash before their competitors, which could be considered wasteful when compared with alternative methods. However, this process also incentivizes miners to use energy in an efficient manner in order to maximize their chance of being the first to generate a new block by using electricity in areas and at times when it is cheaper, and would otherwise be wasted or used less efficiently. This makes direct watt-for-watt comparisons between blockchain proof-of-work mining and other types of computational energy usage potentially misleading as a direct measure of their relative environmental impact. Blockchain computers are currently many orders of magnitude less efficient than centralized or trusted alternatives; however, they also provide certain computational advantages that these other systems cannot replicate ----- FIGURE 15 Screen capture of MyEtherWallet, showing the contract deployment section for the tutorial program Ethereum also has a limit to the maximum amount of computation that can be performed as part of generating a single block. This is measured in units known as “gas,” and exists in part to make sure that individual programs do not halt the blockchain and prevent it from regularly forming new blocks. A detailed description of the relationship between gas and specific computational operations can be found in the Ethereum yellow paper “Ethereum: A [Secure Decentralized Generalized Transaction Ledger EIP-150 Revision” by Gavin Wood and currently hosted on the website gavwood.com/paper.pdf at](http://gavwood.com/paper.pdf) the time of writing. The hard limit on the maximum amount of computation that can be performed as part of generating a single block in the case of the Ethereum blockchain is currently set to 30 000 000 gas per-block, with a target of 15 000 000. The example molecular dynamics transaction carried out in this work made use of 545 720 gas. However, if the entire block limit were used for a simulation it would then be stopped by the Ethereum network at that point even if instructions and additional funds were available for a longer run time. In principle both the results and the current state of a simulation after one transaction can be stored and used to send a second transaction in a future block in order to continue the simulation from where it left off, but this is not currently possible without action from outside the Ethereum blockchain in order to trigger the continuation of the simulation. Furthermore, since each node needs to re-execute every transaction and reproduce the same results, the programming languages used for smart contracts are either entirely deterministic, or, in the cases where they use pre-existing languages, are limited to ensure that certain features are not used. As an example, this means that smart contracts cannot use random number generators that generate (pseudo)random numbers within a contract in a way which might lead to different computers generating different values, they also cannot contain any kind of variable that involves an estimation which can vary across different programming languages and hardware, and they cannot contain any calculation involving the current date, time, or location specific data that is determined by the local computer node at the point that the code is executed. It is possible to implement pseudo-random number generators and provide seed data as an input parameter to generate pseudo-randomness in a deterministic manner that allows for consensus on the output between nodes in the blockchain. ### 9 | BLOCKCHAIN SIMULATION TUTORIAL The smart contract used in this tutorial is the source code and set of execution instructions for a piece of molecular dynamics software that has been designed to simulate the general vibration of diatomic molecules using a variation of the velocity Verlet algorithm [63] with a harmonic potential ----- FIGURE 16 Screen capture showing the MyEtherWallet smart contract deployment transaction confirmation representing the energetics of chemical bonding. This contract is similar to the contract used to carry out the original simulations of the carbon monoxide molecule mentioned previously, but it has been written to allow for a greater number of user-defined input parameters in order to facilitate the general simulation of any diatomic molecule and to improve its usefulness as a potential teaching aid. The source code for this smart contract can be found in the Supporting Information and on the Ethereum blockchain (vide infra), and will be referenced explicitly here. This molecular dynamics software has two modes of execution. The first mode reproduces the original molecular dynamics of carbon monoxide using the parameters used during the simulations carried out in Ethereum blocks 9 360 161 and 9 360 178, with a user specified output precision and number of time steps. The second mode also allows for user defined input values to be set for the masses of the two atoms individually, the harmonic bond force constant, the time step size, the initial bond length, and the equilibrium bond length. The details of these functions can be found in the smart contract and for simplicity the first mode will be used for the demonstration in this tutorial. However, any custom built simulation software that is written in Solidity can also be substituted for the contract we have used here in order to run any other physically meaningful simulation or any other computational science application of the users choice. ### 10 | SUPPORTING SOFTWARE AND TOOLS To follow this tutorial, some third-party software will be required. To simplify this process, minimize hardware requirements, and to reduce the chances of encountering version differences, web applications have been primarily chosen here. There are however, many different options for each of these categories and this is only one of many ways to deploy a smart contract. ### 10.1 | Ethereum wallet A wallet is a piece of software (or hardware) that stores the private key of a cryptographic private and public key pair that is used to access and sign transactions for an account on a blockchain Wallets will often contain further features ranging from running ----- FIGURE 17 Screen capture showing the smart contract deployment on the Ethereum blockchain using the Etherscan blockchain block explorer website entire nodes or miner nodes, to lite-wallets that interact with third-party nodes in order to broadcast transactions and query the blockchain. For this tutorial, a browser based lite-wallet was chosen. These are wallets that are integrated into a web browser and allow for more conve nient access to web applications that utilize blockchains. While this tutorial is written with these wallets in mind, any wallet that is supported by MyEtherWallet (MEW) can be used here, which includes a variety of physical hardware wallets and mobile applications. To complete the deployment and execution process, a sum of the Ether token (ETH), or the native token associated with the blockchain being used to compute the simulation, will be required in order to pay for any associated transaction fees (Appendix). This sum can vary greatly depending on the relative exchange value of the native token, and the congestion of the network. Two notable Ethereum browser wallets that function very similarly are the “Brave Wallet” and “Metamask” software. The former is a wallet integrated with the Brave browser, while the latter is a plug-in that can be installed on most mainstream browsers such as Chrome and Firefox. [Brave can be downloaded from brave.com and Metamask can be downloaded from metamask.io.](http://brave.com) ### 10.2 | Integrated development environment An integrated development environment (IDE) is an application designed to fulfill many or all of the requirements of software development. This typically incorporates a text editor for working with source code, alongside other features such as file management, a code compiler, and/or debugging tools. For this tutorial, an IDE for Solidity, the native programming language of Ethereum, is required. The Remix IDE is the one used here, and it [has the capability to run smart contracts locally for debugging purposes Remix can be accessed through the website remix ethereum org](http://remix.ethereum.org) ----- FIGURE 18 Screen capture showing the “interact with contract” option for the deployment smart contract in MyEtherWallet FIGURE 19 Screen capture of the contract listed in MyEtherWallet (MEW)'s search after deployment ----- FIGURE 20 Screen capture of the contract once selected, with address and ABI fields populated ### 10.3 | MyEtherWallet MEW is a website that allows users to view and manage their account on the Ethereum blockchain using a variety of different wallets. It offers a simple graphical user interface (GUI) for both deploying and interacting with smart contracts, and it supports both of the browser wallets suggested for this tutorial alongside numerous other options. While it is possible to deploy a contract in many other ways, MEW's GUI has been [deemed to be a particularly user friendly option. MEW can be accessed through myetherwallet.com.](http://myetherwallet.com) ### 10.4 | Off-chain smart contract testing Before deploying a contract, it first needs to be compiled into bytecode that can be understood by the EVM. During this stage, it is also important to test the contract to ensure that it functions as intended, since, once deployed, the contract cannot be altered or removed, and each new deployment incurs an additional cost in transaction fees. To test the contract, the first step is to open Remix IDE. By default, Remix will provide an example workspace with sample code. A new workspace is not required and the existing sample code can be safely ignored as it will not be used. To begin, a new file must be created by clicking the document icon located below the workspace name as shown in Figure 3. In this instance the file is named “diatomicmd.sol.” Once created, the new diatomicmd.sol file must be opened and the contents of the provided contract pasted in, ensuring that the entire con tract is present and that no lines have been missed. An example of this is shown in Figure 4. The contract begins with a “pragma solidity” state ment denoting the version number of the compiler to use and ends with a closing curly brace “}” on the final line. Once the contract code has been entered into Remix's text editor, it must be compiled before it can be tested or deployed. On the toolbar to the left, clicking the third item listed will open the Solidity Compiler tab. Once opened, a button will be displayed with the text “Compile diatomicmd.sol,” shown in Figure 5, which after being clicked, will have prepared the bytecode and application binary interface (ABI) for this con tract. The ABI is an interface that describes the functions of a contract, which makes it possible for other applications to understand how to inter act with the deployed contract itself. A single compiler warning should also appear below this panel, but can be safely ignored for the purposes of this tutorial ----- FIGURE 21 Screen capture showing MyEtherWallet's contract interaction panel with the RunMD function selected With the contract compiled, it needs to then be deployed to an environment for testing. Remix IDE supports running an EVM inside the local web browser using JavaScript so that contracts can be tested without being deployed to a live or public environment. Clicking the fourth item in the toolbar on the left will open the “Deploy & Run Transactions” tab. The environment can then be set to the JavaScript VM option, and with the contract set to the DiatomicMD contract that was compiled, the “Deploy” button shown in Figure 6 can then be clicked to deploy the contract for testing outside of the blockchain. The contract functions will be presented at the bottom of the tab and can be run in the browser using the JavaScript VM. There will be two functions named “runMD,” one accepting two parameters and one accepting eight. The two-parameter version will, when called, run the simula tion using pre-configured parameters taken from the first simulation of carbon monoxide run on the EVM [60, 61]. The eight-parameter version can be used to run a custom simulation, however, each input variable will first need to be converted to bytes using the “getValueBytes” function, providing an integer representation of the input value and an integer indicating the number of decimal places, the output of which will then be used as the input for calling the “runMd” function. For example, if an input value was meant to be 1.25, you would call “getValueBytes” using 125 as the “val” parameter and 2 as the “precision” parameter. Expand the second listed two-parameter “runMd” function and provide a number of steps and precision. In this case it has been carried out using five steps, with a precision of ten decimal places, shown in Figure 7. When ready, click “transact” to run the test simulation. Due to limita tions running a copy of the EVM in a web browser with JavaScript, and the fact that software has not yet been designed with these types of simu lation in mind, it is likely that longer running functions will cause the browser to stop responding or crash, while this is not an issue on a full Ethereum node it is recommended to use a smaller number of steps to mitigate this during testing ----- FIGURE 22 Screen capture showing the Etherscan block explorer data associated with this new molecular dynamics trajectory of carbon monoxide At the bottom of the debug window, two transactions will now be listed (shown in Figure 8) and can be expanded. The first transaction is the deployment of the contract itself, while the second is the test execution of the simulation. Expanding the second transaction and looking at the “decoded input” line will list the parameters that were entered previously, while the “decoded output” line will list the results of the simulation as a comma separated list. The first value in this list is the run number, which can be used later to retrieve these results, with the following values being the raw simula tion output which must then be divided by 10[precision] to produce the diatomic internuclear separation distances for each time step in atomic units. With the simulation having been run, it is now possible to confirm that the results have been recorded by using the “getSimOutput” function. Expanding the “getSimOutput” menu in Figure 9, entering the run number of the test, and clicking “call,” will add a further item below the previ ous two transactions shown in Figure 10. The “decoded input” line will list the run number that was entered and the “decoded output” will display the same results as the previous transaction. ### 10.5 | Blockchain contract deployment Once satisfied that the contract is working as intended, the next step is to obtain the bytecode and ABI for the contract. The bytecode is the compiled contract in a form that the EVM can understand that will actually be deployed on the blockchain network. Switching back to the “Deploy and Run Transactions” tab there should be two buttons to copy both the bytecode and ABI. These buttons are located at the bottom of Figure 11. These will be needed in the next steps, and should each be copied and recorded where they can be easily retrieved during later steps. ----- FIGURE 23 Screen capture showing the results of this molecular dynamics trajectory With the ABI and bytecode prepared, the next step is to switch to MEW and access MetaMask or the Brave wallet from within the Brave browser. In order to do this, navigate to the MEW website and click the “Access My Wallet” button (Figure 12), followed by the Browser exten sion button (Figure 13). If using a different wallet from those suggested but still using MEW, select the appropriate option on the MEW site and follow the instruc tions provided. The remaining steps will be identical to using the suggested wallets with the only exception being the process of signing transactions. When the wallet is connected, expand the “Contract” section and navigate to the “Deploy Contract” page in Figure 14 using the menu on the left side of the webpage. Once the webpage loads, paste the bytecode and ABI into the two text areas, and enter a name for the smart contract. Only the bytecode is deployed to the blockchain, while the ABI and smart contract name are recorded by the website to simplify accessing the contract in the future. While the contract name is unimportant, the ABI should be retained as there is no guarantee that MEW will maintain a record of it and without a recognized standard for contracts of this type, interacting with the contract without the ABI may be difficult ----- TABLE 2 The carbon monoxide bond lengths calculated using the Ethereum virtual computer in blocks 14710078 and 9360178 [60, 61] Time (fs) RCO (a0) Block 14710078 RCO (a0) Block 9360178 0.1 2.2676 2.2676 0.2 2.2672 2.2672 0.3 2.2667 2.2667 0.4 2.2659 2.2659 0.5 2.2649 2.2649 TABLE 3 Decimal block numbers, hexadecimal block header hashes, and hexadecimal transaction identification numbers for the simulation in this work Contract deployment Block number 14709931 Transaction ID 0xa3f600317f39a6e065299b506c2e950676befc3080dce1d4d68eafdec7c5bdbc Block header hash 0x03eefc07fcefc9d06f43c877986d7394deb065e445a5a5234e039bb659b262fb Molecular dynamics simulation Block number 14710078 Transaction ID 0x7b8bdb43dadc827d80de190f5df2753baecfb56ed1b2536ea0141de4a1f97e32 Block header hash 0x440c0aec7ed4ceff351575fd9f72693e4c61f232e68fddf70fd9155d11c14ff7 With the contract data filled in, the “Sign Transaction” button in Figure 15 can be clicked to deploy the contract onto the blockchain. Once a smart contact is deployed it can no longer be changed, and modifying any software contained in the smart contract will therefore require that an updated contract be deployed separately, so a user may wish to take a moment to ensure that everything is filled out correctly at this point. After clicking the “Confirm & Send” button shown in Figure 16, the wallet will prompt for the transaction to be cryptographically signed, at which point it will be broadcast to the Ethereum network. MEW will also update to confirm that the transaction has been initiated. The transac tion can be viewed, and progress of the broadcast can be monitored, by selecting the “View on Etherscan” option. Recommended practice is to copy and save the contract address from Etherscan for future reference. This may not be needed on MEW ini tially, but can save time in the future when calling the contract and looking up transactions. The contract address is listed in the “To:” section on the Etherscan block explorer (see Figure 17). Should you lose or forget to record the contract address, it can always be retrieved from the transaction history of the account used to deploy the contract. With the contract now deployed, in this case in block 14709931 of the Ethereum blockchain with the transaction identification number 0xa3f600317f39a6e065299b506c2e950676befc3080dce1d4d68eafdec7c5bdbc, return to MEW and switch to the “Interact with Contract” section in the menu (Figure 18). Select the contract in the contract dropdown menu (Figure 19). If the contract is not listed by name, copy the contract address from earlier into the appropriate field and add the ABI (see Figure 20). MEW does not, at time of writing, support calling overloaded functions (having two functions with the same name but a different set of parameters). When running the simulation with custom inputs this will not be an issue, however, it will not be possible to call the simplified func tion used in testing, that is, to recreate the original simulations of carbon monoxide, without an altered ABI on MEW. For simplicity, this tutorial uses a slightly altered ABI which is provided in the Supporting Information. This is a limitation of MEW specifically, and not a limitation of Ethereum itself. With the address and ABI filled in, click “Interact” and then the “RunMD” function from the list provided. For demonstration purposes we have opted for five steps with a precision of five (Figure 21). Click “Call” and sign the transaction just as before when deploying the contract, this will likely take longer to complete and can again be monitored using Etherscan or an equivalent block explorer (see Figure 22). The simulation has now taken place on the Ethereum blockchain. In this case the simulation of carbon monoxide was recorded in Ethereum block 14710078, with a transaction ID of 0x7b8bdb43dadc827d80de190f5df2753baecfb56ed1b2536ea0141de4a1f97e32 (Table 3). Once com pleted, return to MEW and select the “GetSimOutput” function, enter “1” as the “runNum” parameter and click “Call.” This function is call-only and has no cost, nor will it be recorded on the blockchain. If everything has worked correctly, the output of the simulation will be retrieved and displayed below in the “Results” section, starting with the run number and followed by an array of values. These values will need to be multiplied by 10[�][precision] in order to produce the final intended output for this contract and should again be quoted to one fewer decimal places than the ----- precision specification. The results for this simulation of carbon monoxide are shown in unprocessed form in Figure 23, and compared with the original outputs from simulation of carbon monoxide in block 9360178 in Table 2. Finally, the block header hashes for the specific blocks that contained the simulation data can be taken either directly from the blockchain or using block explorer software like Etherscan to provide extra information about the blocks reproducibility. The block header hashes for the blocks in which the software deployment and molecular dynamics run took place are given with the block numbers and transaction ID numbers in Table 3. ### 10.6 | Conclusions and perspective This tutorial provides a template for researchers looking to perform scientific computation using blockchain computers. Blockchains can provide clear advantages over conventional computers in terms of error recording, computational reproducibility, provenance and determining calculation order, and censorship resistance in scientific work. The use of blockchain cryptography can also aid in federated calculations that involve data sharing. The high costs relative to centralized architectures currently make blockchains less useful for routine calculations where these properties carry less of a premium. However, it is our view that blockchain calculations are likely to have an increasing impact as blockchain throughputs con tinue to scale. Hybrid calculations that use blockchains for certain computational steps and conventional off-chain computers for other steps may also allow more common use of these methods at a significantly reduced cost. In practical terms, while speculative, more developed blockchain calculations could prevent future uncertainties like the “war over supercooled water” between the research groups of Chandler and Debenedetti [68]. We look forward to seeing how this field develops over time. AUTHOR CONTRIBUTIONS Magnus W. D. Hanson-Heine: Conceptualization; data curation; project administration; supervision; writing – original draft. Alexander P. Ashmore: Software; writing – original draft. CONFLICT OF INTEREST The authors declare no competing financial interest. It has come to our attention that the Ethereum blockchain has begun a hard fork between the proof-of-work and proof-of-stake consensus algorithms while the present article was under peer review. DATA AVAILABILITY STATEMENT The data generated during this study is available in this article, and is also available on the Ethereum blockchain in the referenced blocks. ORCID Magnus W. D. Hanson-Heine [https://orcid.org/0000-0002-6709-297X](https://orcid.org/0000-0002-6709-297X) Alexander P. Ashmore [https://orcid.org/0000-0001-7498-8873](https://orcid.org/0000-0001-7498-8873) REFERENCES [[1] S. Nakamoto, Bitcoin: A Peer-to-Peer Electronic Cash System, 2008 Bitcoin.org, https://bitcoin.org/bitcoin.pdf (accessed: June 22 2019).](http://bitcoin.org) [2] D. R. Wong, S. Bhattacharya, A. J. Butte, Nat. Commun. 2019, 10(1), 917. [3] M. Andoni, V. Robu, D. Flynn, S. Abram, D. Geach, D. Jenkins, P. McCallum, A. Peacock, Renew. Sustain. Energy Rev. 2019, 100, 143. [[4] V. Buterin, A Next-Generation Smart Contract and Decentralized Application Platform, 2014 (Cryptorating.eu) https://cryptorating.eu/whitepapers/](https://cryptorating.eu/whitepapers/Ethereum/Ethereum_white_paper.pdf) [Ethereum/Ethereum_white_paper.pdf. (accessed November 13 2019).](https://cryptorating.eu/whitepapers/Ethereum/Ethereum_white_paper.pdf) [5] T.-T. Kuo, X. Jiang, H. Tang, X. F. Wang, T. Bath, D. Bu, L. Wang, A. Harmanci, S. Zhang, D. Zhi, H. J. Sofia, L. Ohno-Machado, BMC Med. Genom. 2020, 13(7), 98. [6] D. Grishin, K. Obbad, G. M. Church, Nat. Biotechnol. 2019, 37(10), 1115. [7] G. Gürsoy, C. M. Brannon, M. Gerstein, BMC Med. Genom. 2020, 13(1), 74. [8] N. D. Pattengale, C. M. Hudson, BMC Med. Genom. 2020, 13(7), 102. [9] T.-T. Kuo, T. Bath, S. Ma, N. Pattengale, M. Yang, Y. Cao, C. M. Hudson, J. Kim, K. Post, L. Xiong, L. Ohno-Machado, Int. J. Med. Inform. 2021, 154, 104559. [10] S. Ma, Y. Cao, L. Xiong, BMC Med. Genom. 2020, 13(7), 91. [11] H. I. Ozercan, A. M. Ileri, E. Ayday, C. Alkan, Genome Res. 2018, 28(9), 1255. [12] M. S. Ozdayi, M. Kantarcioglu, B. Malin, BMC Med. Genom. 2020, 13(7), 82. [13] G. Gürsoy, R. Bjornson, M. E. Green, M. Gerstein, BMC Med. Genom. 2020, 13(7), 78. [14] T.-T. Kuo, R. A. Gabriel, L. Ohno-Machado, J. Am. Med. Inform. Assoc. 2019, 26(5), 392. [15] B. Adanur Dedeturk, A. Soran, B. Bakir-Gungor, PeerJ 2021, 9, e12130. [16] F. Carlini, R. Carlini, S. Dalla Palma, R. Pareschi, F. Zappone, Front. Blockchain 2020, 3(57), 1. [17] V. Racine, Sci. Eng. Ethics 2021, 27(3), 35. [18] X.-L. Jin, M. Zhang, Z. Zhou, X. Yu, J. Med. Internet Res. 2019, 21(9), e13587. [19] B S Glicksberg S Burns R Currie A Griffin Z J Wang D Haussler T Goldstein E Collisson J Med Internet Res 2020 22(3) e16810 ----- [20] M. Shabani, J. Am. Med. Inform. Assoc. 2019, 26(1), 76. [21] M. Kang, V. Lemieux, Ledger 2021, 6, 126. [22] D. R. Ortega, C. M. Oikonomou, H. J. Ding, P. Rees-Lee, Alexandria, G. J. Jensen, PLoS One 2019, 14(4), e0215531. [23] Y. Xu, R. Liu, J. Li, Y. Xu, X. Zhu, J. Phys. Chem. Lett. 2020, 11(23), 9995. [24] H. Kim, S. Kim, J. Y. Hwang, C. Seo, IEEE Access 2019, 7, 136481. [25] F. Chen, H. Wan, H. Cai, G. Cheng, Can. J. Stat. 2021, 49(4), 1364. [26] D. Li, D. Han, T.-H. Weng, Z. Zheng, H. Li, H. Liu, A. Castiglione, K.-C. Li, Soft Comput. 2021, 26, 4423. [27] M. Imran, U. Zaman, Imran, J. Imtiaz, M. Fayaz, J. Gwak, Electronics 2021, 10(20), 2501. [28] F. Zerka, V. Urovi, A. Vaidyanathan, S. Barakat, R. T. H. Leijenaar, S. Walsh, H. Gabrani-Juma, B. Miraglio, H. C. Woodruff, M. Dumontier, P. Lambin, IEEE Access 2020, 8, 183939. [29] V. Drungilas, E. Vaiciukynas, M. Jurgelaitis, R. Butkienˇ ė, L. Ceponien[ˇ] ė, Appl. Sci. 2021, 11(3), 1010. [30] E. A. Mantey, C. Zhou, J. H. Anajemba, I. M. Okpalaoguchi, O. D.-M. Chiadika, Front. Public Health. 2021, 9(1256), 737269. [31] Y. Liu, Y. Qu, C. Xu, Z. Hao, B. Gu, Sensors 2021, 21(10), 3335. [32] R. Zhang, M. Song, T. Li, Z. Yu, Y. Dai, X. Liu, G. Wang, J. Syst. Archit. 2021, 118, 102205. [33] J. Weng, J. Zhang, M. Li, Y. Zhang, W. Luo, IEEE Trans. Dependable Secure Comput. 2021, 18(5), 2438. [34] H. Jin, X. Dai, J. Xiao, B. Li, H. Li, Y. Zhang, IEEE Internet Things J. 2021, 8(21), 15776. [35] M. Ali, H. Karimipour, M. Tariq, Comput. Secur. 2021, 108, 102355. [36] D. Unal, M. Hammoudeh, M. A. Khan, A. Abuarqoub, G. Epiphaniou, R. Hamila, Comput. Secur. 2021, 109, 102393. [37] M. Shen, H. Wang, B. Zhang, L. Zhu, K. Xu, Q. Li, X. du, IEEE Internet Things J. 2021, 8(4), 2265. [38] Z. Mahmood, V. Jusas, Symmetry 2021, 13(7), 1116. [39] C. Hickman, H. Alshubbar, J. Chambost, C. Jacques, C.-A. Pena, A. Drakeley, T. Freour, Fertil. Steril. 2020, 114(5), 927. [40] U. Majeed, L. U. Khan, A. Yousafzai, Z. Han, B. J. Park, C. S. Hong, IEEE Access 2021, 9, 155634. [41] C. Jiang, C. Xu, Y. Zhang, Inf. Sci. 2021, 576, 288. [42] E. O. Kiktenko, N. O. Pozhar, M. N. Anufriev, A. S. Trushechkin, R. R. Yunusov, Y. V. Kurochkin, A. I. Lvovsky, A. K. Fedorov, Quantum Sci. Technol. 2018, 3(3), 035004. [43] I. Stewart, D. Ilie, A. Zamyatin, S. Werner, M. F. Torshizi, W. J. Knottenbelt, R. Soc. Open Sci. 2018, 5(6), 180410. [44] M. Edwards, A. Mashatan, S. Ghose, Quantum Inf. Process. 2020, 19(6), 184. [45] Y.-L. Gao, X. B. Chen, G. Xu, K. G. Yuan, W. Liu, Y. X. Yang, Quantum Inf. Process. 2020, 19(12), 420. [46] S. Singh, N. K. Rajput, V. K. Rathi, H. M. Pandey, A. K. Jaiswal, P. Tiwari, Neural Process. Lett 2020, 1. [47] C. Li, X. Chen, Y. Chen, Y. Hou, J. Li, IEEE Access 2019, 7, 2026. [48] G. Iovane, IEEE Access 2021, 9, 39827. [49] S. Banerjee, A. Mukherjee, P. K. Panigrahi, Phys. Rev. Res. 2020, 2(1), 013322. [50] Q. Li, M. Luo, C. Hsu, L. Wang, D. He, IEEE Syst. J. 2022, 16, 4816. [51] P. Zhang, L. Wang, W. Wang, K. Fu, J. Wang, Secur. Commun. Netw. 2021, 2021, 6671648. [52] Z. Cai, J. Qu, P. Liu, J. Yu, IEEE Access 2019, 7, 138657. [53] J. Nicola, Nature 2021, 594(7864), 481. [54] Z. C. Kennedy, D. E. Stephenson, J. F. Christ, T. R. Pope, B. W. Arey, C. A. Barrett, M. G. Warner, J. Mater. Chem. C 2017, 5(37), 9570. [55] J. J. Sikorski, J. Haughton, M. Kraft, Appl. Energy 2017, 195, 234. [56] J. Leng, P. Jiang, K. Xu, Q. Liu, J. L. Zhao, Y. Bian, R. Shi, J. Cleaner Prod. 2019, 234, 767. [57] S. S. Takhar, K. Liyanage, Int. J. Internet Technol. Secur. Trans. 2021, 11(1), 75. [58] K. Bhubalan, A. M. Tamothran, S. H. Kee, S. Y. Foong, S. S. Lam, K. Ganeson, S. Vigneswari, A.-A. Amirul, S. Ramakrishna, Environ. Res. 2022, 213, 113631. [59] H. Lu, K. Huang, M. Azimi, L. Guo, IEEE Access 2019, 7, 41426. [60] M. W. D. Hanson-Heine, A. P. Ashmore, Chem. Sci. 2020, 11(18), 4644. [61] M. W. D. Hanson-Heine, A. P. Ashmore, J. Phys. Chem. Lett. 2020, 11(16), 6618. [62] C. Dwork, M. Naor, Pricing via processing or combatting junk mail. in Advances in Cryptology—CRYPTO '92 (Ed: E. F. Brickell), Springer, Berlin Heidelberg 1993, p. 139. [63] W. C. Swope, H. C. Andersen, P. H. Berens, K. R. Wilson, J. Chem. Phys. 1982, 76(1), 637. [64] W. D. Carey, Science 1983, 219(4587), 911. [65] U. Hoßfeld, L. Olsson, Nature 2006, 443(7109), 271. [66] C. Schinckus, Renew. Sustain. Energy Rev. 2021, 152, 111682. [67] N. Lasla, L. Al-Sahan, M. Abdallah, M. Younis, Comput. Netw. 2022, 214, 109118. [[68] G. S. Ashley, Phys. Today 2018, 1945. https://doi.org/10.1063/PT.6.1.20180822a](https://doi.org/10.1063/PT.6.1.20180822a) AUTHOR BIOGRAPHIES ----- SUPPORTING INFORMATION Additional supporting information can be found online in the Supporting Information section at the end of this article. APPENDIX This appendix details the steps of how to setup a blockchain wallet on the Ethereum blockchain and purchase Ether tokens to run a simulation of the kind outlined in this publication. The appendix assumes the use of the Chrome web browser and MetaMask wallet plug-in for this purpose at the time of writing. 1. Go to the Chrome web store and install the MetaMask plug-in. A copy of this plug-in can also be found through the MetaMask website [(which is currently metamask.io).](http://metamask.io) 2. Click the “Add to Chrome” button to install the plug-in. 3. Once installed, a new tab titled “Welcome to MetaMask” will open. Click the “Get started” button. 4. Click either the “No thanks” or “I agree” button to share diagnostic data with MetaMask based on personal preferences. Neither will directly affect any further steps in this process. 5. Click the “Create a wallet” button. 6. Enter a password of your choice and click “Create” to generate a new “wallet.” 7. The next page will allow you to watch a video explaining what a “recovery phrase” is in the context of cryptographic key pairs, as well as pre senting some basic information on how to keep this phrase secure. 8. The next page will allow you to click to view your recovery phrase. You should record this somewhere secure, however, if you do not, it is possible to retrieve it from the MetaMask plug-in at a later date. Should this phrase be lost however, there is no method to retrieve it, and any tokens associated with the wallet will be lost. 9. Click either “Remind me later” or “Next.” If you click “Next,” you will be prompted to enter the phrase to confirm you have backed it up. Once done, click “Confirm.” 10. The wallet is now set up and ready to use. To obtain some Ether (ETH), which is the native token of the Ethereum blockchain, click the “Buy” button. 11. The option to select multiple payment gateways will then be presented where you can choose to purchase ETH with fiat currencies such as Dollars ($), Euros (€), or Pounds Sterling (£), and have it directly deposited into the wallet you have just set up. The payment gateway to select will depend on the preferred method to pay for the ETH (e.g., debit card, credit card, bank transfer, apple pay, etc). 12. Once installed, MetaMask is available from the top-right plug-in bar in the Chrome browser. 13. Should this wallet need to be used on a separate machine, repeat the same steps until the “Create a wallet” option is presented and instead click “Recover a wallet,” at which point you should enter the recovery phrase from the original setup to restore that wallet instead. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1002/qua.27035?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1002/qua.27035, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GREEN", "url": "https://nottingham-repository.worktribe.com/preview/13449034/Int%20J%20of%20Quantum%20Chemistry%20-%202022%20-%20Hanson___Heine%20-%20Blockchain%20technology%20in%20quantum%20chemistry%20%20A%20tutorial%20review%20for.pdf" }
2,022
[ "Review" ]
true
2022-10-30T00:00:00
[ { "paperId": "de67654589823c196f81521726d78fc35ebac509", "title": "Leveraging blockchain concepts as watermarkers of plastics for sustainable waste management in progressing circular economy." }, { "paperId": "7db7e190025906c1b3741413eb5e2c8668fbd54a", "title": "Proof-of-work based blockchain technology and Anthropocene: An undermined situation?" }, { "paperId": "29717b25795cd599543ccbdfc05e80e3e73d5a38", "title": "A Decentralized Identity-Based Blockchain Solution for Privacy-Preserving Licensing of Individual-Controlled Data to Prevent Unauthorized Secondary Data Usage" }, { "paperId": "2becfa0f1f61441e2ba5dd0721cfc44ad5ffdfd0", "title": "Blockchain for federated learning toward secure distributed machine learning systems: a systemic survey" }, { "paperId": "b62ddfa76a90b48fa2096814f98572ce2f0187eb", "title": "Comprehensive Survey of IoT, Machine Learning, and Blockchain for Health Care Applications: A Topical Assessment for Pandemic Preparedness, Challenges, and Solutions" }, { "paperId": "87a54749baeb5c1642b5d7a1cad1bf401860991b", "title": "Integration of federated machine learning and blockchain for the provision of secure big data analytics for Internet of Things" }, { "paperId": "bb0c90fd90ca9481919d1fa14e20044709e3cf5e", "title": "Blockchain for genomics and healthcare: a literature review, current status, classification and open issues" }, { "paperId": "82921c5f3622cb54aba641e7e12f5bb0befe8764", "title": "Blockchain-Secured Recommender System for Special Need Patients Using Deep Learning" }, { "paperId": "1f467e7d2805b945ebca678be3240c023d524cbc", "title": "Democratic learning: hardware/software co-design for lightweight blockchain-secured on-device machine learning" }, { "paperId": "887f6f6a4ebe631f2332e7f2c7be847fa07b1c43", "title": "Integration of blockchain and federated learning for Internet of Things: Recent advances and future challenges" }, { "paperId": "b277f77fef0b5d117765140359927fe45c6c6fa6", "title": "Benchmarking blockchain-based gene-drug interaction data sharing methods: A case study from the iDASH 2019 secure genome analysis competition blockchain track" }, { "paperId": "b2c4e325c49611a2886134dafc2fe4271264a08f", "title": "Implementation Framework for a Blockchain-Based Federated Learning Model for Classification Problems" }, { "paperId": "1ddc151446a22f4da7263d410a1ee31ac41aad83", "title": "How scientists are embracing NFTs" }, { "paperId": "696133fc007432f68045379e0618fb4bf011fa94", "title": "PFLM: Privacy-preserving federated learning with membership proof" }, { "paperId": "5477382231adec3319c0a2089eb63c3d4edb71e9", "title": "Can Blockchain Solve the Dilemma in the Ethics of Genomic Biobanks?" }, { "paperId": "166d0012a6711aab6be7ba957d0249f6260a0a4b", "title": "Blockchain-Enabled Asynchronous Federated Learning in Edge Computing" }, { "paperId": "b64c941afa33573a441cc512e68f38b7b7766af8", "title": "Towards Blockchain-Based Federated Machine Learning: Smart Contract for Model Inference" }, { "paperId": "eeabfd10ec53acbd0a9cdf4a47730ede06853e46", "title": "A novel quantum blockchain scheme base on quantum entanglement and DPoS" }, { "paperId": "629d2985d0a72fbf0ba69097c787c80b20072ca3", "title": "The Blockchain Integrated Automatic Experiment Platform (BiaeP)." }, { "paperId": "7493e0e7fe9be5de44ebc6cbd09b3693c79b8a68", "title": "Data sharing: using blockchain and decentralized data technologies to unlock the potential of artificial intelligence: What can assisted reproduction learn from other areas of medicine?" }, { "paperId": "c60c4667076bc0cfe40870212b745eb785c86386", "title": "Calculating with Permanent Marker: How Blockchains Record Immutable Mistakes in Computational Chemistry." }, { "paperId": "b656b2c65dc0c3b22a08f7f1aa1a1f5a9f01f2fc", "title": "Green-PoW: An Energy-Efficient Blockchain Proof-of-Work Consensus Algorithm" }, { "paperId": "386c4d9e439132ec06ff7bdc9ae14e0c9843a37c", "title": "Decentralized genomics audit logging via permissioned blockchain ledgering" }, { "paperId": "a4c0b3e005cc8ff86ea36fd89b640a9b36faa019", "title": "iDASH secure genome analysis competition 2018: blockchain genomic data access logging, homomorphic encryption on GWAS, and DNA segment searching" }, { "paperId": "6c8660f101d6fb712217d6e036f68a3785a130a0", "title": "Using blockchain to log genome dataset access: efficient storage and query" }, { "paperId": "66776d20e43f25b5f8b5cc303e02caa10271cd07", "title": "Blockchain-Authenticated Sharing of Genomic and Clinical Outcomes Data of Patients With Cancer: A Prospective Cohort Study" }, { "paperId": "07734c769838f696fd2cebae792d8944dbdc2db8", "title": "Leveraging blockchain for immutable logging and querying across multiple sites" }, { "paperId": "47cea5a55d5f36f0c3db63c4b7ba6798bff0584b", "title": "A review of quantum and hybrid quantum/classical blockchain protocols" }, { "paperId": "8bda3f00fd1db58cdcd3198c4cc14707b9773b34", "title": "Using Ethereum blockchain to store and query pharmacogenomics data via smart contracts" }, { "paperId": "42bcfcd7a1372643f34ab500c1fe0d676e254765", "title": "Makerchain: A blockchain with chemical signature for self-organizing process in social manufacturing" }, { "paperId": "080a16243c559464798895642e6ee616a496fbc4", "title": "Data privacy in the age of personal genomics" }, { "paperId": "47b5f5fce2ba7799e3750deaf762c604ba037e1d", "title": "Machine learning in/for blockchain: Future and challenges" }, { "paperId": "b779cdd9e02232eff538f8ce6f8ae6e7399df425", "title": "Efficient logging and querying for blockchain-based cross-site genomic dataset access audit" }, { "paperId": "9b510736a3cbdff7a1417142b8e89b38b0702553", "title": "Fair compute loads enabled by blockchain: sharing models by alternating client and server roles" }, { "paperId": "3f48b08829cd419e2898fbeb14beab2d07f9744d", "title": "Prototype of running clinical trials in an untrustworthy environment using blockchain" }, { "paperId": "7d31850f128e4bf854f6dc6d2eb4ab91568927e6", "title": "Application of a Blockchain Platform to Manage and Secure Personal Genomic Data: A Case Study of LifeCODE.ai in China" }, { "paperId": "60be2610dba19761d6458bbac27527b744b0109e", "title": "Blockchain technology in the energy sector: A systematic review of challenges and opportunities" }, { "paperId": "c5f70ee65eabde057f0ecc37c12e333182c5f94d", "title": "Blockchain-based platforms for genomic data sharing: a de-centralized approach in response to the governance problems?" }, { "paperId": "fa24095e91e253064000da94c63e14562dc578b6", "title": "ETDB-Caltech: A blockchain-based distributed public database for electron tomography" }, { "paperId": "bb6716d6b37ebba651e882d5077b1f9bf3f88408", "title": "Realizing the potential of blockchain technologies in genomics" }, { "paperId": "32fd79744afaa92970a3a57305497121af55785b", "title": "Committing to quantum resistance: a slow defence for Bitcoin against a fast quantum computing attack" }, { "paperId": "eabb07994b757d329a434a082e72b81fca9f6237", "title": "Blockchain technology in the chemical industry: Machine-to-machine electricity market" }, { "paperId": "8f7579327850b5719547287f540b88c2efd4f10b", "title": "Quantum-secured blockchain" }, { "paperId": "efcf9e4952a6904241975b95eaf45a645a59cd08", "title": "Freedom of the mind got Nature banned by the Nazis" }, { "paperId": "0ffa8cd252fbf280c088296a94728cb6ede16e31", "title": "Censorship, soviet style." }, { "paperId": "ac81ede712252992b364566fc2f44637e3b0eab3", "title": "A computer simulation method for the calculation of equilibrium constants for the formation of physi" } ]
13,199
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00f3a470a4e9d926afcc53cfcd5deeb000595e12
[]
0.885048
A Secure Land Asset Transfer System using Blockchain
00f3a470a4e9d926afcc53cfcd5deeb000595e12
International Journal of Information &amp; Computation Technology
[ { "authorId": "2105991621", "name": "P. Ramya" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
International Journal of Information & Computation Technology. ISSN 0974-2239 Volume 13, Number 1 (2023), pp. 1-6 © International Research Publications House https://dx.doi.org/10.37624/IJICT/13.1.2023.1-6 # A Secure Land Asset Transfer System using Blockchain **P.N. Ramya[1], Geetha Krishna Guruju[2], Ayushi Verma[3],** **Achintya Prathikantam[4] and Priyanka Myaragalla[5 ]** _12,3,4,5G. Narayanamma Institute of Technology and Science (for Women),_ _Hyderabad TS 500084, India_ _[1e-mail: ramyapn1@gmail.com](mailto:ramyapn1@gmail.com)_ **Abstract** Property transfer is one of the use cases that involves a lot of intermediaries to put trust in the system. In the present scenario, property transactions are carried out on paper, giving rise to countless conflicts. Maintaining accurate records of land ownership and transfers is a very difficult task, made even more challenging by fraudulent or incomplete registries that can be extremely hard to trace back through history. The integrity of these records is crucial, but ensuring their accuracy is a complex undertaking. Blockchain can be utilized to overcome these predicaments faced in land dealings. The transparent nature of blockchain makes it possible to securely track the transfer of ownership from one individual to another reliably. Blockchain’s immutable, auditable, and traceable features makes it a suitable solution for this use case. IPFS is a decentralized protocol and peer-to-peer network that facilitates the storage and sharing of data in a distributed file system. It's designed to enable efficient and secure sharing of files across a network of computers without relying on a central server. A solution of decentralized application or DAPP on Ethereum Blockchain is proposed through this work, which will be a one stop platform for buying, selling, or registering land. A systematic approach is used, right from the registration of the land inspector/buyer/seller to the registration of lands, making it available to sell, etc. **Keywords: Blockchain, Smart Contracts, Ganache, Ethereum, IPFS, Decentralized** Application **1** **Introduction** In our country, property ownership is a contentious issue because of the lack of proper documentation and legal conflicts. The system's weaknesses lie in legacy paper trails and poorly maintained centralized systems, which can be easily manipulated by ----- 2 _P.N. Ramya et al_ fraudulent users. To address these issues, an Ethereum blockchain-based Decentralized application is being proposed. Blockchain technology is created by combining a blockchain system or network with a data structure. The data is stored in blocks which are interconnected and also has hash references to the previous block. The hash references are also used for storing transaction data. A hash function is a tool that takes in data of any size and converts it into a specific and unchanging string of bits called a hash value or hash reference. While these hash values can be quickly calculated, it's extremely challenging to reverse the process and turn the hash value back into the original data, according to computational theory. The proposed system uses the IPFS, which is a distributed file system for storing the confidential land and user identity documents. The proposed system aims to create a secure, decentralized, and tamperproof platform for buying, selling, and registering land. By enabling direct communication between buyers and sellers and eliminating the need for intermediaries, the system will increase transparency and efficiency. The goal is to create a one-stop decentralized application that can maintain immutable and tamper-proof records of transactions. **2** **Related work** In 2021, Suganthe et al. [1] proposed a system that provides the precise details of land records and ownership. The major drawback of this system is that it can only get the land details and store them within the blockchain and hence it doesn’t enable users to buy/sell the land or transfer the ownership of land. The proposed work of Mohammed Moazzam Zahuruddin et al. [2] is implemented on Ethereum blockchain with solidity. The system being proposed employs a double consensus mechanism for transactions, where the landowner initiates the transaction and the buyer completes it. This approach addresses situations where the landowner is unavailable, and assigns ownership to the government. The major drawback of this system is offline land details verification. This paper [3] proposes a land document registration system based on Ethereum and IPFS. This method ensures that user papers kept in the IPFS garage are secure. They expanded an information garage software to illustrate the process. The log files are saved on the IPFS network, which also provides the Hash. The major drawback of the system is that it simply secures the documents stored on IPFS and doesn’t provide any provision for communication between the land buyers and sellers. **3** **Proposed system** The system we propose is a decentralized application which provides a user-friendly interface for direct communication between the buyers and sellers without any middle man. We aim to implement this system using Solidity and Flutter. The DAPP will be a one stop platform for buying, selling and transferring land assets. The system has user/land authentication and verification to prevent fraudulent activity. The system facilitates the users to buy/sell land assets conveniently and every transaction is recorded on the blockchain to maintain transparency and immutability. IPFS is used to ----- _A Secure Land Asset Transfer System using Blockchain_ 3 store the land and identity documents in a decentralized way. The GIS mapping software is used to draw and display the layout of the land. All the transactions take place in Ethers using Metamask wallet. As a proof of ownership transfer, a land sale deed document is generated and stored in IPFS. **4** **Implementation** 4.1 **Module Description** **Authentication Module.** In this module, the verification of the user and the lands is done. The user first registers with the help of his private key and identity documents. After the user has registered, the Land Inspector verifies his identity documents and authenticates him. If the user adds a land to his profile, the Land Inspector should verify the land documents in order to enable the user to take further actions. **User Processes Module.** In this module, the user can add lands to his dashboard, make lands available for sale or buy the available lands from the land gallery. Each of these actions are followed by verification in every step. Finally, the transfer of ownership takes place and a digital document is generated and stored. **Transactions Module.** Metamask is used for making transactions in our system. Right from the buyer’s payment to recording the transfer of ownership, all the transactions are tamper-proof and unchangeable. **4.2** **System Design** There are three stakeholders namely, contract owner, land inspector and the user interact with the decentralized web application. It is built using flutter and solidity. Web3.js is used as API support for communication between the DAPP and blockchain. An Ethereum wallet is required for performing all the blockchain transactions. The documents are stored on IPFS which is a decentralized file system. **Figure 1. Proposed System Design** ----- 4 _P.N. Ramya et al_ **4.3** **Working** The contract owner initiates the whole process by deploying the smart contract on the Ethereum blockchain and acts as the Admin of the system. The contract owner can add/remove the Land Inspectors. First-time users have to register into the system by using a private key and by providing personal information. The system verifies the user and the land inspector verifies the identity document produced at the time of registration. Once the user is authenticated, the option to add lands is enabled. The land inspector verifies the land documents before approving the addition of land to the user’s profile. Once the land is verified, it gets added to the land gallery of all the users. The user can make it available for sale or can buy an already available land. If the seller accepts the request, payment is made and transaction begins. The land inspector verifies the transaction and transfers the ownership of the land. A transfer of ownership document proof is generated and stored in IPFS securely. **5** **Results** **Figure 2: Land Inspector Dashboard** **Figure 3: User Dashboard-Land Gallery** ----- _A Secure Land Asset Transfer System using Blockchain_ 5 **Figure 4: Land Details Page showing Land Map drawn using GIS** **Figure 5: Payment Confirmation Page** **Figure 6: Payment in transit using Metamask wallet** ----- 6 _P.N. Ramya et al_ **Figur 7: Land sale deed document generated and stored using IPFS** **6** **Conclusion and Future Scope** The proposed system is a single platform for the users to buy/sell/register a land. By providing the details of the user, identity documents and land documents which are verified by a Land Inspector, the system ensures the credibility of the data. On the contrary, if these were provided in the traditional system, the data could be altered easily. With the help of blockchain and by storing the data on a decentralized file system, we have ensured that no fraudulent activity takes place and the data remains tamper-proof. As we know that there’s always a scope for improvement, there are certain aspects that could be added to our system to increase its overall efficacy. The system can be further enhanced by automating the user and the land verification process. We can also predict the approximate price of land and suggest the users about the current land price trends. We can also include land splitting or gifting options. **References** [1] Suganthe, R. C., Shanthi N., Latha R. S., Gowtham K., Deepakkumar S., and Elango R.: Blockchain enabled Digitization of Land Registration. In: 2021 International Conference on Computer Communication and Informatics (ICCCI)(2021) [2] Mohammed M.Z., Gupta S., Shaik W.A.: Land Registration using Blockchain Technology. In: International Journal of Emerging Technologies and Innovative Research (2021) [3] Kumar K.V.R., Gokul A.R., Kumar.V.N.: Blockchain and Smart Contract for Land Registration using Ethereum Network. In: International Journal of Engineering Research & Technology (IJERT)(2022) [4] Roopa.C,, Suganthe, R. C., Shanthi N.: Blockchain Based Certificate Verification Using Ethereum And Smart Contract. In: Journal of Critical Reviews(2020) -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.37624/ijict/13.1.2023.1-6?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.37624/ijict/13.1.2023.1-6, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://doi.org/10.37624/ijict/13.1.2023.1-6" }
2,023
[ "JournalArticle" ]
true
2023-07-06T00:00:00
[]
2,482
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00f658832a6f1d5d53f20b9bc361594b72a8ddf8
[ "Computer Science" ]
0.85086
A Review of Dynamic Wireless Power Transfer for In‐Motion Electric Vehicles
00f658832a6f1d5d53f20b9bc361594b72a8ddf8
[ { "authorId": "2093280131", "name": "Kai Song" }, { "authorId": "39474042", "name": "K. Koh" }, { "authorId": "3333859", "name": "Chunbo Zhu" }, { "authorId": "31139085", "name": "Jinhai Jiang" }, { "authorId": "2144447138", "name": "Chao Wang" }, { "authorId": "47932866", "name": "Xiaoliang Huang" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
## We are IntechOpen, the world’s leading publisher of Open Access books Built by scientists, for scientists # 6,700 Open access books available # 180,000 195M International authors and editors Our authors are among the Downloads # 12.2% # 154 TOP 1% Countries delivered to most cited scientists Contributors from top 500 universities Selection of our books indexed in the Book Citation Index in Web of Science™ Core Collection (BKCI) ### Interested in publishing with us? Contact book.department@intechopen.com Numbers displayed above are based on latest data collected. For more information visit www.intechopen.com ----- ##### Chapter 6 #### A Review of Dynamic Wireless Power Transfer for In‐ Motion Electric Vehicles ##### Kai Song, Kim Ean Koh, Chunbo Zhu, Jinhai Jiang, Chao Wang and Xiaoliang Huang Additional information is available at the end of the chapter http://dx.doi.org/10.5772/64331 **Abstract** Dynamic wireless power transfer system (DWPT) in urban area ensures an uninterrupt‐ ed power supply for electric vehicles (EVs), extending or even providing an infinite driving range with significantly reduced battery capacity. The underground power supply network also saves more space and hence is important in urban areas. It must be noted that the railways have become an indispensable form of public transportation to reduce pollution and traffic congestion. In recent years, there has been a consistent increase in the number of high‐speed railways in major cities of China, thereby improving accessi‐ bility. Wireless power transfer for train is safer and more robust when compared with |Col1|Col2| |---|---| ||S| conductive power transfer through pantograph mounted on the trains. Direct contact is subject to wear and tear; in particular, the average speed of modern trains has been increasing. When the pressure of pantograph is not sufficient, arcs, variations of the current, and even interruption in power supply may occur. This chapter provides a review of the latest research and development of dynamic wireless power transfer for urban EV and electric train (ET). The following key technology issues have been discussed: (1) power rails and pickups, (2) segmentations and power supply schemes, (3) circuit topologies and dynamic impedance matching, (4) control strategies, and (5) electromagnetic interference. **Keywords:** dynamic wireless power transfer, magnetic coupler, circuit topologies, control strategies, electromagnetic interference © 2016 The Author(s). Licensee InTech. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. ----- 110 Wireless Power Transfer - Fundamentals and Technologies ##### 1. Introduction In recent years, studies on DWPT have gained traction especially from The University of Auckland, Korea Advanced Institute of Science and Technology (KAIST), The University of Tokyo, Oak Ridge National Laboratory (ORNL), and many other international institutions. The topics discussed include system modeling, control theories, converter topologies, magnetic coupling optimization, and electromagnetic shielding technologies for DWPT. The University of Auckland and Conductix‐Wampfler manufactured the world's first WPT bus with 30 kW power. A demo ET with 100 kW WPT capability and a 400 m long track without any on‐board battery was also constructed [1] as shown in Figure 1. **Figure 1. WPT for EV and ET.** KAIST constructed electric buses powered by an online electric vehicle (OLEV) system. The buses are deployed in Gumi city for public transportation, running on two fixed routes covering a total distance of 24 km as shown in Figure 2. The OLEV system on these routes is able to supply 100 kW power with 85% of transfer efficiency [2]. **Figure 2. KAIST OLEV.** The research in Oak Ridge National Laboratory focuses on coupling configuration, transfer characteristics, medium loss, and magnetic shielding. The dynamic charging system as shown in Figure 3 constructed by ORNL consists of a full bridge inverter powering two transmitters simultaneously through a series connection. The experimental results show that the positions of the electric vehicle significantly affect the transferred power and efficiency [3]. ----- A Review of Dynamic Wireless Power Transfer for In‐Motion Electric Vehicles 111 http://dx.doi.org/10.5772/64331 **Figure 3. DWPT system of ORNL.** Researchers in The University of Tokyo proposed using the combination of a feedforward controller and a feedback controller to adjust the duty cycle of the power converters in the DWPT system to achieve optimum efficiency. With the advanced control method, a wireless in‐wheel motor is developed as shown in Figure 4. The current WPT is from the car body to the in‐wheel motor. In future, the wireless in‐wheel motor can be powered directly from the ground using a dynamic charging system [4]. **Figure 4. Wireless in‐wheel motor.** On the other hand, the Korea Railroad Corporation (KRRI) designed a WPT system for the implementation in railway track. A 1 MW, 128‐m‐long railway track was developed to demonstrate the dynamic charging technology for EV. The coupling mechanism consists of a long transmitter track and two small U‐shaped magnetic ferrites to increase the coupling strength. As a long transmitter track has high inductance, high voltage drop will occur when the current flows through it. In order to reduce this voltage stress, the compensation capacitors are distributed along the track as shown in Figure 5 [5]. **Figure 5. Wireless power rail developed by KRRI.** ----- 112 Wireless Power Transfer - Fundamentals and Technologies The researchers from the Japan Railway Technical Research Institute proposed a different design of coupling mechanism for the ET. The transmitters are long bipolar coils, and “figure‐ 8” coils are used as the matching pickups as shown in Figure 6. The system is able to transfer 50 kW of power across a 7.5‐mm gap with 10‐kHz frequency [6]. **Figure 6. The non‐contact power supply system for railway vehicle.** Bombardier Primove from Germany is currently leading in WPT technology for EV and ET. Studies have been primarily conducted for better exploitation of the technology. Apparently, the technical information of the WPT system developed by Bombardier Primove has not been published. In 2013, the company proposed a design shown in Figure 7 to ensure high reliability when powering the EV. The main DC bus is supplied by k‐number of AC/DC substations connected in parallel. This configuration is used to increase the robustness of the system. If one of the AC/DC substations breaks down, that particular substation will be disconnected from the system and other neighboring substations can continue functioning normally, thus avoiding power interruption. Each transmitter cluster is supplied by multiple high‐frequency DC/AC inverters in parallel. Similar to the DC bus, the power supply at the AC bus will not be interrupted if an inverter breaks down. At the receiver side, the train contains a DC bus as shown in Figure 7. Multiple receivers are supplying to the DC bus simultaneously via AC/DC rectification. The DC bus powers the motor through a controller. If any of the rectifiers is damaged, other receivers can continue providing sufficient power to the DC bus [7]. **Figure 7. DWPT system for railway vehicle.** ----- A Review of Dynamic Wireless Power Transfer for In‐Motion Electric Vehicles 113 http://dx.doi.org/10.5772/64331 The Harbin Institute of Technology demonstrated dynamic charging using segmented transmitters with parallel connections to the inverter [8]. At the receiver side, two layers of flat coils wounded in the same direction are stacked against each other to cancel the points, where transferred power is zero, thereby increasing the overall efficiency. Using the decoupling principle to design the size and position of the two‐phase coil, the cross‐coupling is cancelled and high efficiency is then achieved at any position [9]. Although several studies have been conducted all over the world yielding exceptional results, factors such as power transfer performance, construction cost, and maintenance cost still require improvement. Other important considerations for practical DWPT implementation include high‐power rail, robust control strategies, and EMC. ##### 2. Power rails and pickups Core‐less rectangular coils and bipolar coils are the two general types of coils used in WPT. The University of Auckland proposed using long rectangular rails to transfer power. A larger surface area for road construction necessitates less amount of power to be transferred per surface area. The design is also sensitive to lateral displacement of the electric vehicles. Moreover, a high level of magnetic field leakage occurs at both sides of the rail [10]. KAIST proposed an improved version by adding a magnetic core with an optimized design. Com‐ pared to the transmitter rail proposed by the University of Auckland, the transfer efficiency and transfer distance are increased. However, the construction cost is also higher. KAIST presented an advanced coupling mechanism design and optimization technology in their past research. In 2009, the first‐generation OLEV was successfully produced. An E‐shaped magnetic core is used as the power transmission rail. The air gap is only 1 cm and the transfer efficiency 80% [2]. A U‐shaped transmission rail was also proposed in the same year by significantly increasing the transmission gap to 17 cm with an efficiency of 72%. In 2010, a skeleton‐type W‐shaped magnetic core is proposed, thus further increasing the transfer distance to 20 cm and efficiency to 83% [2]. From 2011 to 2015, researchers from KAIST designed fourth‐generation I‐shaped bipolar rails and fifth‐generation S‐shaped bipolar rails with even larger transfer gap, narrower frame, and higher efficiency [2]. With bipolar rails, the magnetic field path is parallel to the moving direction of the vehicle instead of being orthogonal to the moving direction. The new design is well suited for DWPT due to its advantages such as high power density, narrow frame, and therefore lower construction complexity, robust to lateral displacement, and lower magnetic field exposure on both sides of the rail [10–12] (Tables 1 and 2). In 2015, KAIST proposed using a dq‐two‐phase transmitter rail for cancelling the zero coupling points along the moving direction [13] using the control method which is relatively complex. A double loop control is implemented by detecting the phase of the primary current. The amplitudes and phases of the d‐q currents are controlled using a phase‐locked loop and DC chopper according to the position of the receiver. ----- 114 Wireless Power Transfer - Fundamentals and Technologies **Type** **Coreless long coil** **Bipolar rail** Merits Even magnetic field distribution, stable power transfer, coreless, and low manufacturing cost Demerits Low power density, sensitive to lateral displacement, large surface area is needed for High power density, narrow design, robust to lateral displacement, low construction complexity, and low level of magnetic field exposure Uneven magnetic field distribution, zero coupling point. High cost due to the usage of ferrite core **Table 1. Advantages and disadvantages of commonly used powering rail.** **Table 2. Wireless power rails and receiving pickups developed by KAIST (From generation 1 to 6).** ##### 3. Segment and power supply scheme In order to overcome the issues of low transfer efficiency and high sensitivity to the changing parameters in a centralized power supply system, a new segmented scheme is proposed [14]. The voltage at the 50 Hz AC bus is first stepped up to reduce transmission loss. Then, before the segmented transmitters, the voltage is stepped down via the inverter. Constant current is also used at the transmitters. Efficient converter topologies are also reviewed for implementing a centralized power supply system. (1) Centralized power supply scheme (Figure 8) |Col1|Col2|displacement, large surface area is neede construction, and high level of magnetic field exposure|d for point. High cost due to the usage of ferrite core|Col5| |---|---|---|---|---| |Table 1. Advantages and disadvantages of commonly used powering rail.||||| ----- A Review of Dynamic Wireless Power Transfer for In‐Motion Electric Vehicles 115 http://dx.doi.org/10.5772/64331 With the increasing length of the transmitter rail, the bandwidth of the primary side channel becomes narrower. Therefore, the system is more sensitive to the variations of parameters, and the robustness is decreased. The controller for the centralized power supply is relatively **a.** High requirements of the components due to a single module supporting large power. **b.** The whole rail is activated and causes high loss. **c.** Low reliability due to any breakdown will affect the whole rail. **d.** The efficiency is low when the load is small. **e.** High self‐inductance and therefore high voltage across capacitor. **f.** Highly sensitive toward the variations in parameters, causing low stability. **Figure 8. Centralized power supply scheme.** **Figure 9. Power frequency scheme—segmented rail mode.** (2) Power frequency scheme—segmented rail mode (Figure 9) The advantages of segmented rails are as follows: **a.** Different segments can be turned on at different time periods, decreasing the power loss; **b.** Smaller‐sized power converters; **c.** Higher reliability, when one of the segments breaks down, other segments will still be functioning normally; ----- 116 Wireless Power Transfer - Fundamentals and Technologies **d.** Lower self‐inductance, less sensitive to variations in parameters, and therefore increasing the system stability. However, segmented rails also have the following disadvantages: **a.** High number converters, difficult to control and high maintenance and construction cost; **b.** High number of components is required and therefore low reliability of the whole system. (3) High frequency scheme—segmented rail mode (Figure 10) With segmented rails and centralized power supply, the advantages of this design are as follows: **a.** Lesser power converter units, easier to maintain; **b.** Different segments can be activated at different time periods, lesser power loss; **c.** Lower self‐inductance, less sensitive to variations in parameters, increases the system stability. **Figure 10. High frequency scheme—segmented rail mode.** However, this design has the following disadvantages: **a.** When the power supply breaks down, all of the segmented rails will stop functioning, thus lowering the system reliability; **b.** High loss in the cable connecting the power supply to the segmented rails; **c.** High capacity power supply and therefore large requirements of the components; (4) High frequency and high voltage scheme and low voltage and constant current rail mode (Figure 11). ----- A Review of Dynamic Wireless Power Transfer for In‐Motion Electric Vehicles 117 http://dx.doi.org/10.5772/64331 **Figure 11. High frequency and high voltage scheme—low voltage and constant current rail mode.** (5) Combination scheme (Figure 12) This type of rails combines the advantages of abovementioned rails; however, the system is complex and only suitable for a large‐scale dynamic charging system. **Figure 12. Combined type rail scheme.** ##### 4. Circuit topologies and impedance matching In the DWPT system, the gap between the receiver and transmitter is always changing. Different cars have different heights with respect to the ground and the coupling coefficient will varies significantly. Coupling coefficient is an important parameter in WPT. If the value is too low, the efficiency may drop considerably. Contrarily, frequency splitting phenomena may occur if the coupling coefficient is too high, and the system functions in the unstable ----- 118 Wireless Power Transfer - Fundamentals and Technologies region. Therefore, the circuit topology should be designed to be insensitive to coupling changes. In order to achieve a steady power supply with variations in coupling and to increase the system stability in the light‐load region, an LCLC topology can be used. The current at the primary is kept constant and stress on switches is reduced during on‐off. At the receiver side, a parallel‐T configuration can increase the tolerance of the system toward coupling variation. The proposed topology is shown in Figure 13. **Figure 13. Circuit topology of double LCLC.** The transmitter current is written as follows: _ip_ = (U _i_ - _U_ _r_ 0 ) / (w0Lp ) (1) With λ = L s / L 2 <1 as the load coefficient, the receiver output voltage is as follows: _U_ _o_ = _U_ _oc_ l= w0k _L L Ip_ _s_ _p_ l (2) The output voltage is 1/λ times the receiver voltage. A step‐up voltage converter is used to provide sufficient power when coupling is low, therefore increasing the tolerance of the system against lateral displacement. The voltage ratio and efficiency are given as follows: ìG = _MRl_ _L R0_ ( l + _rs_ ) + _r C0_ _p_ (M w0 + _r Rp_ ( l + _rs_ )) ïíïîh= w02M R2 l2L0 (w02M 2 + _r Rp_ ( l2 + _rs_ ))(L R0 ( l2 + _rs_ ) + _C_ _pr0_ (w02M 2 + _r Rp_ ( l2 + (3) where r0 is the internal resistance of the inverter circuit, rp is the resistance of the transmitter, and rp is the resistance of the receiver. The power and efficiency curves are given in **Figure 14. The efficiency is high at the low‐** coupling region which is particularly important for the DWPT application. As shown by the curves in Figure 15, the efficiency and power are significantly improved for different loads and coupling coefficient compared to series topology. ----- A Review of Dynamic Wireless Power Transfer for In‐Motion Electric Vehicles 119 http://dx.doi.org/10.5772/64331 **Figure 14. Efficiency and voltage gain vs. coupling coefficient.** **Figure 15. Power and efficiency of the two kinds of structure vs. coupling coefficient.** While designing the circuit of WPT, the compensation is performed under no‐load condition. In normal operating condition, frequency tracking is used to ensure resonance by keeping the same phase between primary voltage and primary current [12]. Besides, to ensure the EMC and system stability, control is used to achieve constant current. The magnetic field from the transmitter is in steady state. For example, in the WPT system developed by KAIST, the input voltage of the inverter is adjusted using a three‐phase thyristor converter shown in **Fig‐** **ure 16 to achieve constant current at the transmitter.** ----- 120 Wireless Power Transfer - Fundamentals and Technologies **Figure 16. Diagram of the KAIST IPTS showing a power inverter, a power supply rail, and a pickup.** For the secondary side, in order to realize constant current, constant voltage, or constant power, a DC/DC converter is usually implemented. Figures 17 and 18 show the DC/DC converters used in the WPT systems of the University of Auckland and KAIST [15, 16]. **Figure 17. Secondary DC/DC converter.** **Figure 18. Functional diagram of OLEV power receiver system.** ----- A Review of Dynamic Wireless Power Transfer for In‐Motion Electric Vehicles 121 http://dx.doi.org/10.5772/64331 **Figure 19 shows a secondary‐side circuit which consists of both controllable rectifier and DC/** DC converter. SPWM synchronous rectification is employed at the controllable rectifier. The duty cycle of the rectifier is regulated through SPWM; the effective resistance can be adjusted in the range of Rload ~ _∞. While for a boost converter, the effective resistance can be in the range_ of 0∼ _∞. Therefore, any desired values of the effective resistance can be realized to improve_ the system overall efficiency. **Figure 19. Dynamic impedance adjustment for secondary side pickups.** ##### 5. Control strategies Three types of control were proposed for DWPT: primary control, secondary control, and double‐side control. The University of Auckland proposed adjusting the duty cycle of the inverter to control primary resonant current, simplifying the system configuration [17]. KAIST designed constant current control at the primary. A DC/DC converter is added before the inverter, and the DC voltage from the main line is adjusted to achieve constant current for different loads [13]. The main objective of primary control is to produce constant magnetic field, then robust power control can be implemented. The University of Tokyo utilizes secondary control strategy. A buck converter is added after the rectifier [4]. General state space averaging (GSSA) is used to construct the small‐signal model. Constant power or maximum efficiency is then realized using PI pole placement [18]. In addition, controllable rectifier and hysteresis comparator are also proposed for implementation at the secondary side to control the output power or maximum efficiency [19]. Double‐side control can be with or without communication. ORNL combines the control of both sides, using a closed loop control and frequency adjustment with communication to realize wireless charging [3]. The Hong Kong University proposed simultaneous control of both power and maximum efficiency without communication. The smallest input power is searched to realize constant output power of the inverter [20] (Table 3). ----- 122 Wireless Power Transfer - Fundamentals and Technologies **Control** **strategy** **Primary control** **Secondary control** **Both side control** **Without close‐loop** **communication** Both desired power and maximum efficiency are achievable simultaneously Conflict control between Merits Constant current in transmitter, steady magnetic field, no need to **With close‐loop** **communication** Both desired power and maximum efficiency are achievable simultaneously Additional wireless communication is required, lower the system reliability and real‐time performance consider reflected impedance Demerits Unable to control for maximum efficiency, primary side and limited control of output load, and constant current charging is not realizable Constant charging current, constant charging voltage, or maximum efficiency Adjustable range of the secondary side is limited, and accurate model is required secondary side **Table 3. Comparison of advantages and disadvantages of various control strategies.** The DWPT system is subject to disturbances such as variation of mutual inductance caused by movement of the vehicles. New robust control strategies, which are more superior to PID controllers [4,18,19] in disturbance suppression, are currently being studied. ##### 6. Electromagnetic interference The DWPT uses a high‐frequency, strong magnetic field to transfer power wirelessly. The EMC is an important consideration as the DPWT system is surrounded by many sensitive electronic circuits. The requirements include shielding design, frequency allocation, and grounding design. According to the standard set by the International Commission on Non‐Ionizing Radiation Protection (ICNIRP), the current density exposed to the public is 200 mA/m[2], when the frequency is 100 kHz. The values may affect the nerve system of human body. The limit of specific absorption rate (SAR) is 2 W/kg and power density is 10 W/m[2]; if the exposure to the human body is higher than these limits, heating of the human tissues may occur (Table 4). **Shielding** **Metal conductor** **Magnetic material** **Active shielding** **Resonant reactive shielding** **method** |spec hum|ifi a|c absorption rate (SAR) is 2 W/kg a n body is higher than these limits, h|nd power density is 10 W/m2; if the exposure to the eating of the human tissues may occur (Table 4).|Col5|Col6| |---|---|---|---|---|---| |Shielding Metal conductor Magnetic material Active shield||||in|g Resonant reactive shielding| Does not consume power from the system, controllable Difficult to design, complex configuration Flexible placement, good shielding effect Merits Fully enclosed metal conductor housing provide excellent shielding effect Demerits Eddy loss affecting the system efficiency Magnetic field shaping, increasing coupling coefficient and therefore low loss Limited shielding effect Additional coil lower the system efficiency **Table 4. Comparison of merit and demerit of various magnetic shielding methods.** ----- A Review of Dynamic Wireless Power Transfer for In‐Motion Electric Vehicles 123 http://dx.doi.org/10.5772/64331 The suppression of the leakage field can be divided into active shielding and passive shielding. In passive shielding, a magnetic path is created using magnetic material or canceling field using a low magnetic permeability metallic conductor [21–23]. The self‐inductance and mutual inductance are increased when using magnetic material. The magnetic flux distribution is improved due to higher coupling coefficient, and transfer loss is decreased. However, the shielding effect is limited. Metallic shield is widely used in a high‐frequency magnetic field to suppress electromagnetic interference. Both KAIST and ORNL utilize this kind of shielding method. The advantages include simple design and easy to use. However, metallic shielding cannot cover the transmitter and receiver completely. The exposed conductor is subject to friction and eddy current which will increase the heat loss. KAIST proposed a new active shielding method in 2015. A conventional ferrite plate is embedded in multiple metallic sheets as shown in Figure 20. Experimental results show that the magnetic interference is effectively reduced [24]. **Figure 20. Ferrite shielding structure using an embedded metal sheet.** Regarding active shielding, additional coils with or without power supply are implemented at the WPT system to create a cancelling field as shown in Figure 21. Compared to metallic shielding, the space required is smaller. ----- 124 Wireless Power Transfer - Fundamentals and Technologies **Figure 21. Magnetic field cancellation using a resonant coil.** KAIST published a paper in 2013, proposing an active shielding method using a resonant coil. A switching array is used to change the values of compensated capacitors, thereby controlling the amplitude and phase of the cancelling field. An experiment was performed using green public transportation [25]. In 2015, an improved version using double loop and phase adjust‐ ment to achieve resonance was proposed to achieve an active shielding without power supply. The shielding coils are placed at the side of the coupling mechanism as shown in Figure 22. The current induced by leakage field is then sensed. Magnetic field with the same amplitude but opposite polarity with the leakage is then created for field cancellation [26]. **Figure 22. Resonant reactive power shielding with double coils and four capacitors.** In 2013, ORNL proposed using an aluminum board to reduce electromagnetic interference [27]. As shown in Figure 23, a 1‐mm‐thick aluminum shield is placed above the cables. The magnetic field measured at the passenger‐side front tire is reduced from 18.72 μT to 3.22 μT. **Figure 23. Suppression of magnetic field after adding aluminum plate and its effect.** ----- A Review of Dynamic Wireless Power Transfer for In‐Motion Electric Vehicles 125 http://dx.doi.org/10.5772/64331 ##### 7. Conclusions With the advancement of EV and ET, the significance of DWPT has been consistently growing. Recent developments in DWPT for EV and ET have been presented throughout this chapter. Five different aspects of this technology, such as power rail and pickup design, power supply schemes, circuit topologies and impedance matching, control strategies, and EMC, are reviewed. Despite obtaining significant results post study in this field, some issues of concern are yet to be resolved. Previous results as well as the challenges in deployment of DWPT in real application have been highlighted in this chapter. ##### Author details Kai Song[1*], Kim Ean Koh[2], Chunbo Zhu[1], Jinhai Jiang[1], Chao Wang[1] and Xiaoliang Huang[2] *Address all correspondence to: kaisong@hit.edu.cn 1 School of Electrical Engineering, Harbin Institute of Technology, Harbin, China 2 Department of Electrical Engineering, The University of Tokyo, Tokyo, Japan ##### References [1] Chen L, Nagendra G.R, Boys J.T, Covic G.A. Double‐Coupled Systems for IPT Roadway Applications. IEEE Journal of Emerging and Selected Topics in Power Electronics. 2015;3(1):37–49. DOI: 10.1109/JESTPE.2014.2325943 [2] Choi S.Y, Gu B.W, Jeong S.Y, Rim C.T. Advances in Wireless Power Transfer Systems for Roadway‐Powered Electric Vehicles. IEEE Journal of Emerging and Selected Topics in Power Electronics. 2015;3(1):18–36. DOI: 10.1109/JESTPE.2014.2343674 [3] Miller J.M, Onar O.C, Chinthavali M. Primary‐Side Power Flow Control of Wireless Power Transfer for Electric Vehicle Charging. IEEE Journal of Emerging and Selected Topics in Power Electronics. 2015;3(1):147–162. DOI: 10.1109/JESTPE.2014.2382569 [4] Kobayashi D, Imra T, Hori Y. Real‐time coupling coefficient estimation and maximum efficiency control on dynamic wireless power transfer for electric vehicles. In: IEEE WoW 2015 ‐ IEEE PELS Workshop on Emerging Technologies: Wireless Power, Proceedings ; June 5, 2015 ‐ June 6, 2015; Daejeon, Korea. Piscataway, United States: Institute of Electrical and Electronics Engineers Inc.; 2015. p. 1–6. DOI: 10.1109/WoW. 2015.7132799 ----- 126 Wireless Power Transfer - Fundamentals and Technologies [5] Kim J.H, Lee B.S, Lee J.H, Lee S.H, Park C.B, Jung S.M. Development of 1‐MW Inductive Power Transfer System for a High‐Speed Train. IEEE Transactions on Industrial Electronics. 2015;62(10):6242–6250. DOI: 10.1109/TIE.2015.2417122 [6] Ukita K, Kashiwagi T, Sakamoto Y, Sasakawa T. Evaluation of a non‐contact power supply system with a figure‐of‐eight coil for railway vehicles. In: IEEE WoW 2015 ‐ IEEE PELS Workshop on Emerging Technologies: Wireless Power, Proceedings; June 5, 2015 ‐ June 6, 2015; Daejeon, Korea. Piscataway, United States: Institute of Electrical and Electronics Engineers Inc.; 2015. p. 1–6. DOI: 10.1109/WoW.2015.7132807 [7] Winter J, Mayer S, Kaimer S, Seitz P, Pagenkopf J, Streit S. Inductive power supply for heavy rail vehicles. In: 2013 3rd International Electric Drives Production Conference, EDPC 2013 ‐ Proceedings; October 29, 2013 ‐ October 30, 2013; Nuremberg, Germa‐ ny.Piscataway, United States: IEEE Computer Society; 2013. p. 1–9. DOI: 10.1109/EDPC. 2013.6689749 [8] Song K, Zhu C, Li Y, Guo Y, Jiang J, Zhang J. Wireless power transfer technology for electric vehicle dynamic charging using multi‐parallel primary coils. Zhongguo Dianji Gongcheng Xuebao/Proceedings of the Chinese Society of Electrical Engineering. 2015;35(17):4445–4453. DOI: 10.13334/j.0258‐8013.pcsee.2015.17.020 [9] Zhu C, Song K, Wei G, Zhang Q. Novel power receiver for dynamic wireless power transfer system. In: Industrial Electronics Society, IECON 2015 ‐ 41st Annual Confer‐ ence of the IEEE; 9 November 2015‐12 November 2015; Yokohama.Piscataway, United States: Institute of Electrical and Electronics Engineers Inc.; 2015. p. 002247– 002251. DOI: 10.1109/IECON.2015.7392436 [10] Chen L, Nagendra G.R, Boys J.T, Covic G.A. Ultraslim S‐Type Power Supply Rails for Roadway‐Powered Electric Vehicles. IEEE Transactions on Power Electronics. 2015;30(11):6456–6468. DOI: 10.1109/TPEL.2015.2444894 [11] Lee W.Y, Huh J, Choi S.Y, Thai X.V, Kim J.H, Al‐Ammar E.A. Finite‐Width Magnetic Mirror Models of Mono and Dual Coils for Wireless Electric Vehicles. IEEE Transac‐ tions on Power Electronics. 2013;28(3):1413–1428. DOI: 10.1109/TPEL.2012.2206404 [12] Huh J, Lee S.W, Lee W.Y, Cho G.H, Rim C.T. Narrow‐Width Inductive Power Transfer System for Online Electrical Vehicles. IEEE Transactions on Power Electronics. 2011;26(12):3666–3679. DOI: 10.1109/TPEL.2011.2160972 [13] Park C, Lee S, Jeong S.Y, Cho G.H, Rim C.T. Uniform Power I‐Type Inductive Power Transfer System With DQ Power Supply Rails for On‐Line Electric Vehicles. IEEE Transactions on Power Electronics. 2015;30(11):6446–6455. DOI: 10.1109/TPEL. 2015.2420372 [14] Tian Y. Research on Key Issues of Sectional Track‐Based Wireless Power Supply Technology for Electric Vehicles [thesis]. Chongqing, China: Chongqing University; 2012. ----- A Review of Dynamic Wireless Power Transfer for In‐Motion Electric Vehicles 127 http://dx.doi.org/10.5772/64331 [15] Keeling N.A, Covic G.A, Boys J.T. A Unity‐Power‐Factor IPT Pickup for High‐Power Applications. IEEE Transactions on Industrial Electronics. 2010;57(2):744–751. DOI: 10.1109/TIE.2009.2027255 [16] Shin J, Shin S, Kim Y, Ahn S, Lee S, Jung G, et al. Design and Implementation of Shaped Magnetic‐Resonance‐Based Wireless Power Transfer System for Roadway‐Powered Moving Electric Vehicles. IEEE Transactions on Industrial Electronics. 2014;61(3):1179– 1192. DOI: 10.1109/TIE.2013.2258294 [17] Abdolkhani A, Hu A.P. Improved autonomous current‐fed push‐pull resonant inverter. Iet Power Electronics. 2014;7(8):2103–2110. DOI: 10.1049/iet‐pel.2013.0749 [18] Hata K, Imura T, Hori Y. Maximum efficiency control of wireless power transfer via magnetic resonant coupling considering dynamics of DC‐DC converter for moving electric vehicles. In: Conference Proceedings ‐ IEEE Applied Power Electronics Conference and Exposition ‐ APEC; March 15, 2015 ‐ March 19, 2015; Charlotte, NC, United states. Piscataway, United States: Institute of Electrical and Electronics Engi‐ neers Inc.; 2015. p. 3301–3306. DOI: 10.1109/APEC.2015.7104826 [19] Gunji D, Imura T, Fujimoto H. Envelope model of load voltage on series‐series compensated wireless power transfer via magnetic resonance coupling. In: IEEE WoW 2015 ‐ IEEE PELS Workshop on Emerging Technologies: Wireless Power, Proceedings; June 5, 2015 ‐ June 6, 2015; Daejeon, Korea. Piscataway, United States: Institute of Electrical and Electronics Engineers Inc.; 2015. p. 1–6. DOI: 10.1109/WoW.2015.7132853 [20] Zhong W.X, Hui S.Y.R. Maximum Energy Efficiency Tracking for Wireless Power Transfer Systems. IEEE Transactions on Power Electronics. 2015;30(7):4025‐4034. DOI: 10.1109/TPEL.2014.2351496 [21] Kim J, Kim H, Kim M, Ahn S, Kim J, Kim J. Analysis of EMF noise from the receiving coil topologies for wireless power transfer. In: cccc2012 Asia‐Pacific Symposium on Electromagnetic Compatibility, APEMC 2012 ‐ Proceedings; May 21, 2012 ‐ May 24, 2012; Singapore, Singapore. Piscataway, United States: IEEE Computer Society; 2012. p. 645–648. DOI: 10.1109/APEMC.2012.6237964 [22] Kim H, Song C, Kim J, Kim J. Shielded Coil Structure Suppressing Leakage Magnetic Field from 100W‐Class Wireless Power Transfer System with Higher Efficiency. International Microwave Workshop Series on Innovative Wireless Power Transmis‐ sion: Technologies. 2012;16(3):83–86. DOI: 10.1109/IMWS.2012.6215825 [23] Ahn S, Park H.H, Choi C.S, Kim J, Song E, Paek H.B, et al. Reduction of electromagnetic field (EMF) of wireless power transfer system using quadruple coil for laptop appli‐ cations. In: 2012 IEEE MTT‐S International Microwave Workshop Series on Innovative Wireless Power Transmission: Technologies, Systems, and Applications, IMWS‐IWPT 2012 ‐ Proceedings; May 10, 2012 ‐ May 11, 2012; Kyoto, Japan. Piscataway, United States: IEEE Computer Society; 2012. p. 65–68. DOI: 10.1109/IMWS.2012.6215821 ----- 128 Wireless Power Transfer - Fundamentals and Technologies [24] Park H.H, Lwon J.H, Kwak S.I, Ahn S. Magnetic Shielding Analysis of a Ferrite Plate with a Periodic Metal Strip. IEEE Transactions on Magnetics. 2015;51(8):1–8. DOI: 10.1109/TMAG.2015.2425796 [25] Kim J, Kim J, Kong S, Kim H, Suh I.S, Suh N.P, et al. Coil Design and Shielding Methods for a Magnetic Resonant Wireless Power Transfer System. Proceedings of the IEEE. 2013;101(6):1332–1342. DOI: 10.1109/JPROC.2013.2247551 [26] Moon H, Kim S, Park H.H, Ahn S. Design of a Resonant Reactive Shield With Double Coils and a Phase Shifter for Wireless Charging of Electric Vehicles. IEEE Transactions on Magnetics. 2015;51(3):1–4. DOI: 10.1109/TMAG.2014.2360701 [27] Ahn S, Park J, Song T, Lee H, Byun J, Kang D, et al. Low frequency electromagnetic field reduction techniques for the On‐Line Electric Vehicle (OLEV). In: IEEE Interna‐ tional Symposium on Electromagnetic Compatibility; July 25, 2010 ‐ July 30, 2010; Fort Lauderdale, FL, United states. Piscataway, United States: Institute of Electrical and Electronics Engineers Inc.; 2010. p. 625–630. DOI: 10.1109/ISEMC.2010.5711349 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.5772/64331?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5772/64331, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://www.intechopen.com/citation-pdf-url/51247" }
2,016
[ "Review" ]
true
2016-06-29T00:00:00
[]
9,236
en
[ { "category": "Physics", "source": "external" }, { "category": "Computer Science", "source": "external" }, { "category": "Mathematics", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Physics", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00f6cf73c647f116d7a23d17e8a4c45a0555dd81
[ "Physics", "Computer Science", "Mathematics" ]
0.859475
Phase transitions in distributed control systems with multiplicative noise
00f6cf73c647f116d7a23d17e8a4c45a0555dd81
arXiv.org
[ { "authorId": "8054025", "name": "N. Allegra" }, { "authorId": "1713790", "name": "Bassam Bamieh" }, { "authorId": "48850725", "name": "P. Mitra" }, { "authorId": "5225213", "name": "C. Sire" } ]
{ "alternate_issns": null, "alternate_names": [ "ArXiv" ], "alternate_urls": null, "id": "1901e811-ee72-4b20-8f7e-de08cd395a10", "issn": "2331-8422", "name": "arXiv.org", "type": null, "url": "https://arxiv.org" }
Contemporary technological challenges often involve many degrees of freedom in a distributed or networked setting. Three aspects are notable: the variables are usually associated with the nodes of a graph with limited communication resources, hindering centralized control; the communication is subject to noise; and the number of variables can be very large. These three aspects make tools and techniques from statistical physics particularly suitable for the performance analysis of such networked systems in the limit of many variables (analogous to the thermodynamic limit in statistical physics). Perhaps not surprisingly, phase-transition like phenomena appear in these systems, where a sharp change in performance can be observed with a smooth parameter variation, with the change becoming discontinuous or singular in the limit of infinite system size. In this paper, we analyze the so called network consensus problem, prototypical of the above considerations, that has previously been analyzed mostly in the context of additive noise. We show that qualitatively new phase-transition like phenomena appear for this problem in the presence of multiplicative noise. Depending on dimensions, and on the presence or absence of a conservation law, the system performance shows a discontinuous change at a threshold value of the multiplicative noise strength. In the absence of the conservation law, and for graph spectral dimension less than two, the multiplicative noise threshold (the stability margin of the control problem) is zero. This is reminiscent of the absence of robust controllers for certain classes of centralized control problems. Although our study involves a ‘toy’ model, we believe that the qualitative features are generic, with implications for the robust stability of distributed control systems, as well as the effect of roundoff errors and communication noise on distributed algorithms.
## Phase transitions in distributed control systems with multiplicative noise **Nicolas Allegra[1][,][2], Bassam Bamieh[1], Partha Mitra[3]** **and Cl´ement Sire[4]** 1 Department of Mechanical Engineering, University of California, Santa Barbara, USA 3 Cold Spring Harbor Laboratory, Cold Spring Harbor, NY, USA 4 Laboratoire de Physique Th´eorique, Universit´e de Toulouse, UPS, CNRS, F-31062 Toulouse, France **Abstract.** Contemporary technological challenges often involve many degrees of freedom in a distributed or networked setting. Three aspects are notable: the variables are usually associated with the nodes of a graph with limited communication resources, hindering centralized control; the communication is subjected to noise; and the number of variables can be very large. These three aspects make tools and techniques from statistical physics particularly suitable for the performance analysis of such networked systems in the limit of many variables (analogous to the thermodynamic limit in statistical physics). Perhaps not surprisingly, phase-transition like phenomena appear in these systems, where a sharp change in performance can be observed with a smooth parameter variation, with the change becoming discontinuous or singular in the limit of infinite system size. In this paper we analyze the so called network consensus problem, prototypical of the above considerations, that has been previously analyzed mostly in the context of additive noise. We show that qualitatively new phase-transition like phenomena appear for this problem in the presence of multiplicative noise. Depending on dimensions and on the presence or absence of a conservation law, the system performance shows a discontinuous change at a threshold value of the multiplicative noise strength. In the absence of the conservation law, and for graph spectral dimension less than two, the multiplicative noise threshold (the stability margin of the control problem) is zero. This is reminiscent of the absence of robust controllers for certain classes of centralized control problems. Although our study involves a ”toy” model we believe that the qualitative features are generic, with implication for the robust stability of distributed control systems, as well as the effect of roundoff errors and communication noise on distributed algorithms. PACS numbers: 05.10.-a, 02.30.Yy, 89.75.Fb Submitted to: JSM ----- **Contents** **1** **Introduction** **3** 1.1 Example of a simple Laplacian consensus algorithm . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2 Plan of the article . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 **2** **Consensus algorithms in a random environment** **6** 2.1 Definition of the system and conservation laws . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2.2 Correlations induced by exact conservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.2.1 Asymmetric links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.2.2 Symmetric links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2.3 Isotropic links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.2.4 Gain noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.3 Average conservation law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 **3** **Threshold behavior of the network coherence** **10** 3.1 Phase diagram of the exactly conserved model . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 3.1.1 Noise threshold in any dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.1.2 Finite size dependence of the coherence . . . . . . . . . . . . . . . . . . . . . . . . . . 12 3.1.3 Exponential regime above the noise threshold . . . . . . . . . . . . . . . . . . . . . . 13 3.2 Phase diagram of the average conserved model . . . . . . . . . . . . . . . . . . . . . . . . . . 14 3.2.1 Noise threshold in high dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 3.2.2 Robustness of the coherence and finite-size dependance . . . . . . . . . . . . . . . . . 16 3.3 A word on the other forms of symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 **4** **Summary and conclusions** **17** 2 ----- **1. Introduction** Phase transition phenomena have played a central role in twentieth century theoretical physics, ranging from condensed matter physics to particle physics and cosmology. In recent decades, phase transition phenomena, corresponding to non-analytic behavior arising from large system-size limit in systems of many interacting variables, have increasingly appeared in the engineering disciplines, in communications and computation [1], robotics [2, 3], control theory [4, 5] and machine learning. Related phenomena have also been noted in behavioral biology [6], social sciences [7] and economics [8, 9]. It is not as widely appreciated that phase transition-like behavior may also be observed in distributed control systems (or distributed algorithms) with many variables, for similar mathematical reasons that they appear in statistical physics, namely the presence of many interacting degrees of freedom. Problems in a variety of engineering areas that involve interconnections of dynamic systems are closely related to consensus problems for multi-agent systems [10]. The problem of synchronization of coupled oscillators [11] has attracted numerous scientists from diverse fields ([12] for a review). Flocks of mobile agents equipped with sensing and communication devices can serve as mobile sensor networks for massive distributed sensing [13, 2, 4]. In recent years, network design problems for achieving faster consensus algorithms has attracted considerable attention from a number of researchers [14]. Another common form of consensus problems is rendezvous in space [15, 16]. This is equivalent to reaching a consensus in position by a number of agents with an interaction topology that is position induced. Multi-vehicle systems are also an important category of networked systems due to their technological applications [17]. Recently, consensus algorithms had been generalized to quantum systems [18] and opening new research directions towards distributed quantum information applications [19]. In contrast to the above mentioned work, a relatively less studied problem is that of robustness or resilience of the algorithms to uncertainties in models, environments or interaction signals. This is the central question in the area of Robust Control, but it has been relatively less studied in the context distributed control systems. This issue is quite significant since algorithms that work well for a small number of interacting subsystems may become arbitrarily fragile in the large-scale limit. Thus, of particular interest to us in this paper is how large-scale distributed control algorithms behave in various uncertainty scenarios. Additive noise models have been studied in the context of networked consensus algorithms [20, 21], including scaling limits for large networks [22, 23]. More relevant to the present work are uncertainty models where link or node failures, message distortions, or randomized algorithms [24, 25, 26, 27, 28] are modeled by multiplicative noise. In this latter case, the phenomenology is much richer than the case of only additive noise. The basic building block of multi-agent and distributed control systems is the so-called consensus algorithm. In this paper we develop a rather general model of consensus algorithms in random environments that covers the above mentioned uncertainty scenarios, and we study its large-size scaling limits for d-dimensional lattices. We now motivate the problem formulation of this paper using a simple version of the consensus algorithm. _1.1. Example of a simple Laplacian consensus algorithm_ We consider a local, linear first-order consensus algorithms over an undirected, connected network modeled by an undirected graph G with N nodes and M edges. We denote the adjacency matrix of G by A, and D is the degree matrix. The Laplacian matrix of the graph G is denoted by L and is defined as L = D _A. In_ _−_ the first-order consensus problem, each node i has a single state ui(t). The state of the entire system at time _t is given by the vector u(t) ∈_ R[N] . Each node state is subject to stochastic disturbances, and the objective is for the nodes to maintain consensus at the average of their current states. The dynamics of this system in continuous time is given by **u˙** (t) = _αLu(t) + n(t),_ (1) _−_ where α is the gain on the communication links, and n(t) ∈ R[N] is a white noise. This model covers a large number of systems. A few examples are _• In flocking problems ui(t) is the current heading angle of a vehicle or agent [6]. L is the Laplacian of_ the agents’ connectivity network determined by e.g. a distance criterion. The disturbance ni(t) models random forcing on the i’th agent. 3 ----- _• In load balancing algorithms for a distributed computation network ui(t) is the current load on a_ computing node. L is the Laplacian of the connectivity graph [29, 30] . The disturbance ni(t) models arriving (when positive) or completed (when negative) jobs. _• In distributed sensing networks [13, 2, 4], the i’th sensor makes a local measurement ui(0) of a global_ quantity. The aim is for the sensors to communicate and agree on the value of the sensed quantity without having a central authority, thus they communicate locally based on a connectivity graph with _L as its Laplacian. The dynamics (3) represents averaging of each sensor with their neighbors, while_ the disturbances ni(t) can represent the effective noise in communicating with neighbors. The aim is for all sensors to reach the same estimate (which would be the initial network mean) asymptotically, i.e. for each i, limt→∞ _ui(t) =_ _N1_ �i _[u][i][(0).]_ We should note that in much of the literature, the algorithm Eq. (3) has been studied for finite systems and in the absence of disturbances n(t). In the large-scale limit however, there are significant differences in phenomenology between the disturbance-free versus the uncertain scenarios. In the absence of disturbances and for a connected graph, the system Eq. (3) converges asymptotically to the average of the initial states. With the additive noise term, the nodes do not converge to consensus, but instead, node values fluctuate around the average of the current node states. Let us denote the spatial average across the network (”network mean”) of the sites by m(t), _m(t) = [1]_ _N_ _N_ � _uk(t)._ (2) _k=1_ In the absence of noise, m(t) = m(0) for a finite graph. Although m(t) shows a diffusive behavior for finite sized networks in the presence of additive noise, the diffusion coefficient is proportional to 1/N so that in the infinite lattice size limit one still obtains m(t) = m(0). This can be seen by spatially averaging the consensus equation. The noise-free dynamics preserves the spatial average (equivalently, the Graph Laplacian is chosen to be left stochastic), so in the absence of noise ˙m(t) = 0. In the presence of the additive noise, _m˙N_ (t) = [1] _N_ � _ni(t),_ (3) _i_ Assuming the noise is uncorrelated between sites, this means that var(mN (t)) ∝ _Nt_ [, so that] limN _→∞_ _var(mN_ (t)) = 0. Thus, in the infinite lattice size limit, m(t) = m(0). Note that this continues to be the case for the multiplicative noise model discussed in the paper when there is a conservation law that preserves the network mean in the absence of the additive noise component. The spatial variance across the lattice sites has been called the ”network coherence” in previous work [23]. 1 _CN[∞]_ [:= lim]t→∞ _N_ _N_ � � � var _ui(t) −_ _m(t)_ _,_ (4) _i=1_ Note that in the infinite size limit for the consensus problem with additive noise m(t) = m(0) so the spatial variance (”network coherence”) also characterizes the variance of the individual site variables from the desired mean m(0) at the initial time. If this variance is finite in the limit of large N, then under the consensus dynamics the individual lattice sites settle down to stationary fluctuations around the desired initial spatial mean, and this quantity can be recovered from a single site at long times, by taking a time average. It has been shown that CN[∞] [is completely determined by the spectrum of the matrix][ L][ (see [][10][] for a] review). Let the eigenvalues of L denoted by 0 = λ1 < ... < λN . The network coherence is then equal to 1 _._ (5) _λi_ 4 1 _CN[∞]_ [:=] 2αN _N_ � _i=2_ ----- A classical result [23] shows that, if one considers the network to be Z[d], then the large-scale (N →∞) properties of the coherence show the following behavior _CN[∞]_ _[∼]_    _N_ in d = 1 log N in d = 2 1 in d > 2. (6) Compared to a perfect Laplacian algorithm (without additive noise), the addition of disturbances makes the system unable to reach the global average in low dimensions. In higher dimensions, the network coherence becomes finite and the algorithm performs a statistical consensus at large time. This result shows that there are fundamental limitations to the efficiency of the algorithm for a large-scale system in presence of additive noise in low dimensions. Let us mention that this algorithm can be extended to a second-order dynamic where each nodes has two state variables u(t) and v(t). These system dynamics arise in the problem of autonomous vehicle formation control (the state of a node is given by the position and velocity of the vehicle). The vehicles attempt to maintain a specified formation traveling at a fixed velocity while subject to stochastic external perturbations and a similar coherence definition can be introduced, and the scaling of the coherence with N has been computed [23] and similar dimensional limitations has been unraveled. _Robustness and Multiplicative Noise_ In engineering systems, N may be large but perhaps not as large as might be considered in condensed matter problems. In the distributed control setting, one can give a robustness interpretation of asymptotic limits like Eq. (6). For example, in the d = 1 case (which is relevant to the so-called automated vehicular platoons problem [23]) one can say that as the formation size increases, it becomes susceptible to arbitrarily small forcing disturbances. A more proper robustness interpretation however requires considering scenarios with _multiplicative noise as a model for uncertain system dynamics, and not just uncertain forcing or measurement_ noise (which is typically modeled with additive noise). Consider the following more general version of Eq. (3) **u˙** (t) = _αL(t) u(t) + n(t),_ (7) _−_ where the Laplacian L(t) is now a time-varying matrix with some “structured” randomness. This is now a multiplicative noise model since the randomness multiplies the state u(t). This is used to model effects of random environments such as networked algorithms in which links or nodes may fail randomly, randomly varying system parameters, round-off error in floating-point arithmetic, or more generally any random errors that are proportional to the current state [24, 25, 26, 27, 28]. If arbitrarily small probabilities of error in _L(t) can produce large effects in the dynamics of Eq. (7) then the algorithms are fragile and lack robustness._ The main question in this paper is to study a fairly general model of networked algorithms in random environments like Eq. (7). In particular, we study scaling limits in d-dimensional lattices, and characterize how the system responds to multiplicative noise and whether (or not) the algorithm can remain stable and perform its task once some arbitrary small multiplicative noise in incorporated. It is highly unlikely that taking account of multiplicative noise (with variance σ[2]) is going to improve the scaling in low dimensions, but the challenge is to characterize the robustness of the algorithm in high dimensions. This measure of robustness can be quantified by the amount of noise variance σ[2] that the system can sustain before hitting a threshold σc[2] [above which the algorithm becomes unstable and the coherence grows unboundedly. The range] of noise 0 ⩽ _σ[2]_ _< σc[2][,]_ (8) in which the coherence is bounded is often called margin of stability in the control literature [31]. The main goal of this article is to quantify this margin, and to explicitly give its behavior in a wide range of models. _1.2. Plan of the article_ The plan of the article is the following, after the introduction of the general system and its main properties, we shall discuss the several conservation laws that one might consider and the effects on the correlations 5 ----- in the system. Furthermore, different cases of the form of the random environment, which is modeled by multiplicative random variables will be analyzed, leading to various types of correlations between the different degrees of freedom of our model. The main part of the article will be the study of the time-behavior of the network coherence Eq. (14) via the 2-point correlation function. We shall see that, depending on the conservation law that we are imposing in our system, different behaviors of the coherence is found. The stability margin Eq. (8), introduced in the previous section, shall be fully characterized and its explicit value will be given for each considered cases. The complete phase diagram of the network-coherence will be explored in any dimension, with an emphasis on the presence of a noise- threshold between various large-scale behaviors. The phase diagrams shall be explained in a more general point of view by analyzing the various universality classes of the system and making contact with the well-known phenomenology of disordered systems. Although our results are derived for first-order consensus algorithms Eq. (3), we expect qualitatively similar behaviors to occur in higher-order consensus algorithms such as those used in vehicular formation control [23]. **2. Consensus algorithms in a random environment** _2.1. Definition of the system and conservation laws_ In this section, one introduces a class of algorithms generalizing the noisy Laplacian algorithm Eq. (3). The generalized algorithm is a discrete-time stochastic evolution for the quantity ui(t) ∈ R with both additive and multiplicative noises. Let us define the system with a general coupling kernel K on a square-lattice of size L and dimension d � _ui(t + 1) =_ _ξij(t)K(ri, rj)uj(t) + Vi(t)ui(t) + ni(t),_ (9) _j_ where ξij(t), Vi(t) are random variables with mean ⟨ξ⟩ and ⟨V ⟩, and Vi being the neighborhood of the site _i. Here we choose the convention that ξii(t) = 0 such that the off-diagonal randomness is only in Vi. The_ link ξij(t) and onsite Vi(t) variables are Gaussian variables with correlations that will be detailed in the next sections. The additive noise is centered and uncorrelated in space and time with variance σn[2] _⟨ni(t)⟩_ = 0, (10) _⟨ni(t)nj(t[′])⟩_ = σn[2] _[δ][ij][δ][tt][′]_ _[,]_ (11) where δab is the Kronecker symbol. The system is translationally invariant so that K(ri, rj) = K(ri **rj). In** _−_ the following, we shall use the convenient notation Ki−j := K(ri − **rj). We assume also that the distributions** do not vary with time. As a consequence, all the moments of the different variables do not depend of the space-time coordinates ( e.g. _⟨ξij(t)[p]⟩_ = ⟨ξ[p]⟩∀i, j). Let us define the variances σξ[2] [=][ ⟨][ξ][2][⟩−⟨][ξ][⟩][2][ and] _σV[2]_ [=][ ⟨][V][ 2][⟩−⟨][V][ ⟩][2][. In the spirit of Eq. (][3][), the system can be written in a matrix form] **u(t + 1) = M(t)u(t) + n(t),** (12) where the matrix M(t) is a time-dependent random matrix with matrix elements Mij(t) = Ki−jξij(t)+δijVj(t) and n(t) is the additive noise vector. This algorithm is a generalization of Eq. (3) in which the Laplacian matrix is replaced by a time-dependent random matrix. The problem is then related to the asymptotic properties of product of random matrices (see [32] or [33, 34] for example in the context of consensus and wireless communications). The exponential growth rate of the matrix powers M[t] as t is controlled _→∞_ by the eigenvalue of M with the largest absolute value. While the stationary distribution was guaranteed in the additive model Eq. (3), the multiplicative model Eq. (9) may or may not have a stationary solution depending on the parameters and the dimension d. The mean value can be written _⟨ui(t + 1)⟩_ = �∥K∥1⟨ξ⟩ + ⟨V ⟩�⟨ui(t)⟩, (13) where ∥K∥1 = [�]i [K][i][. Then obviously][ ⟨][u][i][(][t][)][⟩] [= 0 when][ ⟨][V][ ⟩] [=][ −∥][K][∥][1][⟨][ξ][⟩] [and][ ⟨][u][(][t][ + 1)][⟩] [=][ ⟨][u][i][(][t][)][⟩] [when] _⟨V ⟩_ = 1 −∥K∥1⟨ξ⟩. In general, the sequence converges if |(∥K∥1⟨ξ⟩ + ⟨V ⟩| ⩽ 1 and diverges otherwise for a 6 ----- positive initial condition _u(0)_ _> 0. One way of studying the margin will be to compute the time-behavior_ _⟨_ _⟩_ of the coherence _L[d]_ � Gr(t) −⟨m(t)⟩[2], (14) _r=1_ _CL(t) = [1]_ _L[d]_ _L[d]_ � _⟨(ur(t) −_ _m(t))(u0(t) −_ _m(t)))⟩_ = [1] _L[d]_ _r=1_ where Gr(t) is the local 2-point correlation function ⟨ur(t)u0(t)⟩ that captures correlations between an arbitrary node 0 and a node r and m(t) = L[−][d][ �]i _[u][i][(][t][) being the space average of the value][ u][i][(][t][). The]_ term ⟨m(t)⟩[2] is stationary if |(∥K∥1⟨ξ⟩ + ⟨V ⟩| ⩽ 1. Therefore the time-evolution of CL(t) is simply governed by the time-evolution of Gr(t). This function will be the main quantity that one will be interested in to characterize the scaling behavior of the performance of the algorithm. In the following, we shall impose the time conservation of the average m(t) either exactly or in average at the thermodynamic limit L . _→∞_ _2.2. Correlations induced by exact conservation_ In this section, we summarize all the particular cases of Eq. (12), while preserving conservation of m(t) exactly (i.e [�]i [M][ij][ = 1) at the thermodynamic limit. The randomness is fully defined by specifying the] form of the correlations ⟨Mij(t)Mpq(t)⟩. The exact conservation is defined by m(t + 1) = m(t) . As a consequence the matrix M satisfies column-stochasticity [�]i [M][ij][ = 1. Hence, it is straitforward to show that] we need to impose � _Vj(t) = 1 −_ _ξij(t)Ki−j,_ (15) _i_ for that conservation law to be satisfied Let us notice that the summation is over the first index, while it is on the second in the evolution equation Eq. (9). In this case, the multiplicative quantities are linearly dependent and one can write everything in terms of the variance of one or the other. Additionally one can assume many different form for the link variables ξij(t). The simplest case is to assume that ξij(t) and ξji(t) are independent, this assumption does not add any new correlation between the onsite variables Vi(t). The opposite scenario is to consider ξij(t) = ξji(t) which enforces correlations between onsite variables as we shall see in the following. We will also consider both cases where the link variables ξij(t), depend only on one index, i.e. ξij(t) = Jj(t) and ξij(t) = gi(t) which we call respectively isotropic and gain noise cases. For each considered possibilities, we shall explicitly compute the relevant correlations that we will need later on for the calculation of the coherence and in particular for the calculation of Gr(t). _2.2.1. Asymmetric links_ Now let us consider the model with asymmetric links _ξij(t) ̸= ξji(t)._ (16) The model can be seen as acting on a directed graph where the incoming and outgoing links are two different independent random variables. As we saw in the previous section, exact conservation enforces � _Vi(t) = 1 −_ _ξki(t)Kk−i_ then _⟨V ⟩_ = 1 −⟨ξ⟩∥K∥1. (17) _k_ Let us write down an explicit example (ignoring the additive noise) for a line with three sites with closed boundary conditions and nearest-neighbors interactions. The evolution equation Eq. (9), written in the matrix form Eq. (12), reads  _ui−1(t + 1)_   _Vi−1_ _ξi−1,i_ 0   _ui−1(t)_   _ui(t + 1)_  =  _ξi,i−1_ _Vi_ _ξi,i+1_   _ui(t)_  _,_ (18) _ui+1(t + 1)_ 0 _ξi+1,i_ _Vi+1_ _ui+1(t)_ where the column-stochastic condition is _Vi−1(t) = 1 −_ _ξi,i−1(t),_ (19) _Vi(t)) = 1 −_ _ξi−1,i −_ _ξi+1,i(t),_ _Vi+1(t) = 1 −_ _ξi,i+1._ 7 ----- ### ξi,i−1 ξi−1,i ### ξi+1,i ξi,i+1 **Figure 1. Diagram corresponding to the asymmetric case. It describes the simplest example defined by** Eq. (18) and Eq. (19). The link variables are four different uncorrelated random variables (represented by different colors). The rules are the following, the arrows exiting site i correspond to the quantity which is distributed to the neighbors. And the arrows coming in i represent the quantity that is given from the neighbors. ### ξi,i−1 ξi,i−1 ### ξi,i+1 ξi,i+1 **Figure 2. Diagram corresponding to the symmetric case for the example Eq. (18). The link variables are** now the same ξij (t) = ξji(t) between two neighbors. We can check that ui−1(t + 1) + ui+1(t + 1) + ui(t + 1) = ui−1(t) + ui(t) + ui+1(t) ‡. In that case, the matrix M is not correlated, all the entries are independent. It can be useful to write Eq. (18) and Eq. (19) as a diagram (see Fig .(1) ). The rules are the following, the arrows exiting site i correspond to the quantity which is distributed to the neighbors. And the arrows coming in i represent the quantity that is given from the neighbors. With exact conservation law, the variances of the link and on-site noises are related by _σV[2]_ [=][ ∥][K][∥][2][σ]ξ[2][, where][ ∥][K][∥] [=] ��i [K]i[2][. Consequently, the correlations between the link variables and the] on-site variables take the form _⟨ξrp(t)V0(t)⟩−⟨ξ⟩⟨V ⟩_ = −σξ[2][K][r][δ][p,][0][.] (20) The exact conservation law spatially correlates the link and on-site variables together. Because of Eq. (17), there is no cross-correlation between onsite variables at different sites _⟨Vr(t)V0(t)⟩−⟨V ⟩[2]_ = σV[2] _[∥][K][∥][2][δ][r,][0][.]_ (21) That shows that the onsite variables are simply uncorrelated in space. This asymmetric link assumption will be fully detailed in the next sections, but it will be useful to explore the other cases for a better understanding of the differences. The asymmetric case is the only one which does not create correlations inside the matrix _M, as we shall see next._ _2.2.2. Symmetric links_ Now let us consider the same model but with symmetric links _ξij(t) = ξji(t),_ (22) and where one still have the deterministic relation Eq. (15) between the variables. The example of the past section can be written as well and the diagram becomes the following. In that situation, the matrix M is now doubly-stochastic because of the symmetry, but not uncorrelated anymore. The corresponding diagram _‡ Let us remark that ⟨V ⟩_ = 1 −⟨ξ⟩∥K∥1 is not satisfied here because our example is not invariant by translation. 8 ----- can be seen on (see Fig .2 ) Indeed the symmetry induced some counter-diagonal correlations in the matrix M. Here we still have the relation σV[2] [=][ ∥][K][∥][2][σ]ξ[2] [and] _⟨ξrp(t)V0(t)⟩−⟨ξ⟩⟨V ⟩_ = σξ[2][K][r][δ][p,][0][,] (23) like in the previous model. The main difference comes from the fact that the V ’s are now spatially correlated, and using Eq. (15) one finds _⟨Vr(t)V0(t)⟩−⟨V ⟩[2]_ = σξ[2] �∥K∥[2]δr,0 + K[2]r� _._ (24) We can observe, that the onsite variables are not delta-correlated anymore but are short-range correlated with a correlation kernel K[2]r[. The symmetry between links induces correlations between the onsite variables] at different positions. _§_ _2.2.3. Isotropic links_ Let us consider the case where, the link variable does not depend on the first index _ξij(t) := Jj(t),_ (25) of variance σJ[2] [, then all the out-going links have the same value, but the incoming links are independent.] The diagram of Eq. (9) in that case is Fig .3. In that case, one has � _Vi(t) = 1 −_ _Ji(t)_ Kk−i = 1 − _Ji∥K∥1._ (26) _k_ and σV[2] [=][ σ]J[2] [(][∥][K][∥][1][)][2][ where][ ∥][K][∥][1][ =][ �]i [K][i][. The exact conservation condition implies row-stochasticity of] the matrix M. In that case, the correlations between the link variables and the on-site variables are ### Ji−1 Ji ### Ji Ji+1 **Figure 3. Diagram corresponding to the isotropic case ξij** (t) = Jj (t) for the example Eq. (18). The outgoing links are equal (red) and the incoming are different uncorrelated random variables (blue and green). _⟨Jr(t)V0(t)⟩−⟨J⟩⟨V ⟩_ = −σJ[2] _[∥][K][∥][1][δ][r,][0][.]_ (27) Because of Eq. (26), there is no correlations between Vr(t) and V0(t) at different sites _⟨Vr(t)V0(t)⟩−⟨V ⟩[2]_ = σJ[2] [(][∥][K][∥][1][)][2][ δ][r,][0][.] (28) Contrary to previous symmetric case, the isotropic assumption does not couple the on-site variables. This case is actually very interesting because of the form of the matrix M, although it will not be discussed in this article. _2.2.4. Gain noise_ The last case that one might be interested in here is _ξij(t) := gi(t),_ (29) of variance σg[2][. This assumption can be seen as the opposite of the previous isotropic situation. Here we] have � _Vi(t) = 1 −_ _gk(t)Kk−i._ (30) _k_ The diagram is now Fig .4 (notice that the diagram is reversed compared to the previous isotropic case). The relation between the variances is now σV[2] [=][ ∥][K][∥][2][σ]g[2][. Here the exact conservation condition implies] 9 ----- ### gi gi−1 ### gi+1 gi **Figure 4.** Diagram corresponding to the gain noise case ξij (t) = gi(t) for the example Eq. (18). The incoming links are equal (red) and the outgoing are different independent random variables (blue and green). column-stochasticity of the matrix M. The correlations between link variables and on-site variables take the form _⟨gr(t)V0(t)⟩−⟨g⟩⟨V ⟩_ = −σg[2][K][r][.] (31) The correlations between the onsite variables are also different, one can show that _⟨Vr(t)V0(t)⟩−⟨V ⟩[2]_ = σg[2][∥][K][∥][1][K][r][.] (32) In this case, the V _[′]s are spatially correlated by the kernel Kr du to the form of the links and are also_ correlated with the links with the same kernel. In a sense, this specific form is closer to the symmetric case, but the expressions of the correlations are different. _2.3. Average conservation law_ The other scenario is to impose conservation in average at the thermodynamic limit _m(t+1)_ = _m(t)_ . This _⟨_ _⟩_ _⟨_ _⟩_ condition is a less restrictive constraint on the time-evolution of the system. This condition only imposes a relation in average between the onsite and link variables _⟨V ⟩_ = 1 −⟨ξ⟩∥K∥1, (33) where ∥K∥1 = [�]i [K(][i][). The diagonal and off-diagonal noises are now independent random quantities while] there were linearly dependent when exact conservation was enforced. We are not repeating the analysis of correlations between noises in every cases here, let us just focus on the asymmetric model that one will be interested in in the rest of the article. Of course, when only average conservation is enforced, no relation between the variances σV[2] [and][ σ]ξ[2] [exists, and there is no cross-correlation between onsite and link variables] _⟨ξrp(t)V0(t)⟩−⟨ξ⟩⟨V ⟩_ = 0. (34) Furthermore, the onsite variables V are simply delta-correlated _⟨Vr(t)V0(t)⟩−⟨V ⟩[2]_ = σV[2] _[δ][r,][0][.]_ (35) In the formulation Eq. (12), the matrix M is not stochastic anymore [�]i [M][ij][ ̸][= 1 but is average-stochastic] �i[⟨][M][ij][⟩] [= 1.] We shall see that those two different conservation conditions (exact or average) lead to different results for the behavior of the coherence, that can be understood in terms of their corresponding continuum space-time evolution equation. **3. Threshold behavior of the network coherence** _3.1. Phase diagram of the exactly conserved model_ As explained earlier, the behavior of the coherence of the system can be analyzed by computing the 2-point correlation function Gr(t) = ⟨ur(t)u0(t)⟩ between ur(t) and u0(t) at same time t. The calculation is shown here for the asymmetric case with exact conservation law, but the steps are the same for all the different _§ If now we impose average conservation and Mij_ ’s are Gaussian distributed, then M belongs to the Gaussian Orthonormal Ensemble (GOE) [35]. 10 ----- particular cases. In the asymmetric case with exact conservation, we have Eq. (20) and Eq. (21), therefore, it is straitforward to write down the time-evolution of the correlator Gr(t) as � Gr(t + 1) = ⟨ξ⟩[2][ �] Kr−kKlGk−l(t) + 2⟨V ⟩⟨ξ⟩ Kr−kGk(t) _k,l_ _k_ + ⟨V ⟩[2]Gr(t) + δr,0G0(t) �σξ[2][∥][K][∥][2][ +][ σ]V[2] _[−]_ [2][σ]ξ[2][K][2]r� + σn[2] _[δ][r,][0][.]_ (36) From now on, we will focus on the analysis of this evolution equation for a short-range kernel but let us mention that this equation can be easily solved on the complete graph where the kernel takes the form Kr = 1 − _δr,0, the details of this calculation will be presented elsewhere._ _3.1.1._ _Noise threshold in any dimension One could notice that Eq. (36) can be written down as a_ convolution, then in Fourier space the result is a simple product plus a term that contains Gr(t) in r = 0 � �2 � � G(� **q, t + 1) =** _⟨ξ⟩K([�]_ **q) + ⟨V ⟩** �G(q, t) + G0(t) _σξ[2][∥][K][∥][2][ +][ σ]V[2]_ _[−]_ [2][σ]ξ[2]K[�][2]r + σn[2] _[,]_ (37) with the following definition of the Fourier transform ˆuq(t) = [�]ri∈Z[d][ u][r]i [(][t][)][e][i][qr][i] [. As we saw earlier, the] exact conservation law implies σV[2] [=][ σ]ξ[2][∥][K][∥][2][ then Eq. (][37][) reads] � G(� **q, t + 1) = λ(q)[2]G(�** **q, t) + 2G0(t)σξ[2]** _∥K∥[2]_ _−_ K[�][2]r � + σn[2] = λ(q)[2][ �]G(q, t) + 2G0(t)σξ[2][(]K[�][2](0) − K[�][2](q)) + σn[2] _[,]_ (38) where we have _λ(q) =_ _ξ_ K(q) + _V_ _⟨_ _⟩_ [�] _⟨_ _⟩_ � � = 1 −⟨ξ⟩ _∥K∥1 −_ K([�] **q)** � � = 1 −⟨ξ⟩ K(0)� _−_ K(� **q)** _._ (39) Now let us consider the continuum space limit of this problem, ur∈Zd (t) → _u(r ∈_ R[d], t). Let us notice that the case _ξ_ = 0 is therefore rather particular, indeed it implies _V_ = 1 _λ(q) = 1, and the evolution_ _⟨_ _⟩_ _⟨_ _⟩_ _→_ equation is trivial to solve, hence the correlator is always exponential for any value of the variances. For _∥_ _⟨ξ⟩̸= 0, the stationary solution_ G[�] st(q) is given by solving G([�] **q, t + 1) −** G([�] **q, t) = 0, it follows** � � 2Gst(0)σξ[2] K�[2](0) − K[�][2](q) + σn[2] G� st(q) = _,_ (40) 1 _λ(q)[2]_ _−_ where we used the notation Gst(0) := Gst(r = 0). The self-consistent equation for Gst(0) gives Gst(0) = _cdσn[2]_ 1−2σξ[2][f][d][ where, in the continuum-lattice limit, we have] � _∞_ _−∞_ � _∞_ K�[2](0) − K[�][2](q) d[d]q _._ (42) 1 _λ(q)[2]_ _−∞_ _−_ d[d]q (41) 1 _λ(q)[2][,]_ _−_ and 1 _cd =_ (2π)[d] 1 _fd =_ (2π)[d] Until now, the calculation holds for any integrable kernel, and we shall now focus on a local kernel which scales as a power-law in Fourier space K(q) **q[2][θ], the simplest being the Laplacian corresponding to θ = 1.** [�] _∼_ For a local kernel, one has λ(q)[2] _∼_ 1+ _p⟨ξ⟩q[2][θ]_ where p is a constant. The integral cd converges in d > 2θ and _∥_ For symmetric links, this case is not trivial anymore. 11 ----- _fd is convergent in any dimension. The integrals cd and fd can be defined, by dimensional regularization,_ for any real d ¶. The stationary solution exists if and only if 1 − 2fdσξ[2] _[>][ 0. The critical value]_ 1 _σc[2]_ [=] _,_ (43) 2fd where fd is given by Eq. (42), is the maximum value of the variance σξ[2] [such that the stationary state is] reachable and such that the system is stable. The constant fd depends only on the explicit form of the kernel, the mean of the link variables _ξ_ and the dimension d of the space. The value of this threshold can _⟨_ _⟩_ be tuned by changing the value of _ξ_ . So a first result shows that, in that situation, the stability margin _⟨_ _⟩_ Eq. (8) is always positive in any dimension, therefore the system can support some amount of multiplicative noise and reach its stationary state. The full explicit form of the stationary correlator can be computed. For _σξ[2]_ _[< σ]c[2]_ [and for][ d >][ 2][θ][ the stationary solution can be written] Gst(r) = 2σξ[2] _cdσn[2]_ (2π)[d] 1 − 2σξ[2][f][d] � K�[2](0) − K[�][2](q) d[d]q _e[−][i][qr]._ (44) 1 _λ(q)[2]_ _−_ The stationary solution converges to a finite value in d > 2θ. As usual, the divergence is logarithmic at dc. The behavior can be easily understood in a renormalization group (RG) language. In the phase σξ[2] _[< σ]c[2][, we]_ can define the quantity A[ex]σ[2] 1 _A[ex]σ[2][ =]_ _._ (45) 1 − 2σξ[2][f][d] This quantity, which goes to 1 when σξ[2] _[→]_ [0 and goes to][ ∞] [when][ σ]ξ[2] _[→]_ _[σ]c[2][, can be seen as a coefficient]_ which renormalizes the additive noise. Indeed we have Gst(0) = A[ex]σ[2] _[σ]n[2]_ _[f][d][, which is, up to the constant][ A][ex]σ[2]_ [,] the stationary solution of a pure additive system (σξ[2] [= 0), as we shall see in the next paragraph. What] it means, is that below the noise threshold, the randomness of the link and onsite variables are irrelevant in a RG sense and the system renormalizes to a pure additive model with a renormalized additive noise _n˜i(t) =_ [�]A[ex]σ[2] _[n][i][(][t][) where][ n][i][(][t][) is the original additive noise. The correlator equation becomes]_ G(� **q, t + 1) = λ(q)[2]G(�** **q, t) + A[ex]σ[2]** _[σ]n[2]_ _[.]_ (46) Since σξ[2] [is the only relevant parameter dictating the large-scale behavior of the system, the effective] description Eq. (46) of the system below σc[2] [is still valid in the non-stationary regime][ d][ ⩽] [2][θ][. The full] space-time-dependent solution of Eq. (46), for a kernel of the form K(q) **q[2][θ]** verifies the following scaling [�] _∼_ form � _t_ � G(r, t) = A[ex]σ[2] _[σ]n[2]_ **[r][2][θ][−][d][Ψ]** _._ (47) **r[2][θ]** The form of the correlator is valid for any dimension below the threshold. Here Ψ(y) is a scaling function with properties that Ψ(y) _const as y_ and Ψ(y) _y[(2][θ][−][d][)][/][2][θ]_ as y 0. This scaling form Eq. (47) is _→_ _→∞_ _→_ _→_ the so-called Family-Viczek (FV) scaling [36], well-known in statistical physics of interfaces. The case θ = 1 and θ = 2 are respectively the Edwards-Wilkinson (EW) and Mullins-Herring (MH) universality classes [37]. The upper-critical dimension of this system is dc = 2θ, above that dimension, the correlator converges to a finite value. For σξ[2] _[< σ]c[2]_ [and for][ d][ ⩽] [2][θ][ the behavior is a power-law following Eq. (][47][). The correlator grows] as t[(2][θ][−][d][)][/][2][θ] and the stationary solution is never reached for an infinite system. This exponent is named 2β in the context of growing interfaces and is equal to β = 1/4 (resp β = 3/8) for the EW class (resp. MH). The exponent increases with θ. _3.1.2. Finite size dependence of the coherence_ Now we can extend the result on the large-time coherence scaling for the Laplacian algorithm (θ = 1) sketched in the introduction. From the result on the calculation of G(r, t) below the noise threshold, one can translate the informations in terms of the time-dependance of CN (t) and the stationary value CN[∞] [of the network coherence defined (see Eq. (][4][) and Eq. (][14][)) in] the introduction for a system of N agents. The finite-size behavior of the coherence can be extracted by _¶ then the calculation also holds for a fractal graph of non-integer dimension d_ 12 ----- introducing a cut-off in Fourier space in Eq. (41) and Eq. (42). For a system of size L with N = L[d] nodes, the coherence is behaving for short times (t _L[2])_ _≪_ � _CN_ (t) ∝ [1] d[d]rG(r, t) ∼ _t[(2][−][d][)][/][2]._ (48) _N_ So for short times, the coherence grows as a power-law, independently of system-size. The coherence reaches then a stationary value which scales with system-size as _CN[∞]_ _[∝]_ _N[1]_ � 2−d d[d]rGst(r) ∼ _N_ _d,_ (49) for any d. This value is reached within a time-scale tc ∼ _L[2]_ = N [2][/d]. The value of the infinite coherence CN[∞] grows unboundedly in d < 3 and converges to a finite value in d > 2, in agreement with Eq. (6). In terms of performance, in d > 2 the algorithm is still capable to perform its averaging task even with the presence of multiplicative noise, as long as the variance is below the threshold σc[2][.] _3.1.3. Exponential regime above the noise threshold_ Now that we have understood the behavior of the correlator below the threshold, where the large-scale behavior was dictated by the additive noise, one can study the behavior in the other regime, where the behavior will be controlled by the parameter σξ[2][. For] _σξ[2]_ [⩾] _[σ]c[2][, the additive noise is irrelevant since we expect an exponential growth of the correlator, and the]_ equation Eq. (38) can be written as G(� **q, t + 1) = D(G(�** **q, t)).** (50) The large-time behavior of this equation is dominated by the largest eigenvalue λmax of the operator D. If _λmax = 1, the FV scaling Eq. 47 holds. If λmax > 1 then the correlator behaves exponentially as_ G(� **q, t) ∼** _λ[t]max[g][�][(][q][)][,]_ (51) where _g(q) has to determined self-consistently. Inserting Eq. (51) inside Eq. (37), one ends up with_ � 2σξ[2][(]K[�][2](0) − K[�][2](q) _g(q) =_ _._ (52) � _λmax −_ _λ(q)[2]_ Now let us write a condition on the form of the largest eigenvalue λmax of the operator D. We have K�[2](0) − K[�][2](q) G(� **q, t) = 2σξ[2]** (53) _λmax −_ _λ(q)[2][ G(0][, t][) =][ �][g][(][q][)G(0][, t][)][.]_ Now using (2π)[−][d][ �] d[d]qG([�] **q, t) = G(0, t) we end up with a condition on the largest eigenvalue λmax** � d[d]q (54) (2π)[d][ �][g][(][q][) = 1][,] where K�[2](0) − K[�][2](q) _g(q) = [d][d][q]_ (55) � (2π)[d] _λmax −_ _λ(q)[2][,]_ and λ(q)[2] _≈_ 1 + p⟨ξ⟩q[2]. The integral Eq. (54) when λmax → 1, converges in any d by dimensional regularization, meaning that the FV scaling holds in any d below the threshold. The behavior of the coherence C(t) in the regime σξ[2] [⩾] _[σ]c[2]_ [is always exponential in any dimension and for any system-size. Those] different regimes can be understood by looking at the corresponding continuum equation describing the large-scale fluctuations of Eq. (9). Without loss of generalities, let us focus on the Laplacian kernel from now on. We showed previously that, when exact conservation is enforced, it exists a deterministic relation between the onsite and link noises Eq. (17). Therefore the asymptotic behavior of our model is governed by a continuum space-time equation with only one multiplicative source of randomness and additive noise _∂tu(x, t) = ∂x [ξ(x, t)∂xu(x, t)] + n(x, t)._ (56) 13 ----- _d_ Finite Exponential _CN[∞]_ [=] 1−cdσ2σξ [2]n[2][fd] _CN_ (t) ∼ _λ[t]max_ _dc = 2_ Algebraic and stationary _CCNN((tt ≫ ≪_ _ttcc)) ∼ ∼_ _tN[2][β]2−d_ _d_ _CExponentialN_ (t) ∼ _λ[t]max_ 0 _σc[2]_ [=] 2fd1 _σξ[2]_ **Figure 5. Phase diagram in the exact conservation case on a d-dimensional lattice for the Laplacian kernel** _θ = 1 with N sites. On the left of the critical line (red line) σξ[2]_ _[< σ]c[2][, the behavior of the coherence follows]_ the FV scaling Eq. (47), i.e. algebraic (with exponent β = (2 − _d)/4) for short time then stationary at_ large time, in low d (below the blue dashed line) and finite for higher d. On the other side of the line, the coherence grows exponentially for any d. In that situation, the algorithm remains resilient to multiplicative noise below the noise threshold, showing that the stability margin is positive in any dimension. From our analysis, one shows that there are physically two different regimes. The first regime, is when _ξ(x, t) is irrelevant at large distance and one ends up with the large-scale behavior of the Edwards-Wilkinson_ equation _∂tu(x, t) = ∂x[2][u][(][x][, t][) + ˜][n][(][x][, t][)][,]_ (57) where the additive noise ˜n(x, t) is renormalized by the value of the variance σξ[2] [of the multiplicative noise] as ˜n(x, t) = [�]A[ex]σ[2] _[n][(][x][, t][), where][ n][(][x][, t][) is the original additive noise. The other regime is exponential (see]_ Eq. (51)), where the large-scale behavior is dominated by the multiplicative noise ξ(x, t) and where the additive noise does not change the asymptotic behavior. This two regimes are separated by a threshold σc[2] that is finite in any dimension and equal to σc[2] [=] 2f1d [where][ f][d][ is the integral given by Eq. (][42][).] _3.2. Phase diagram of the average conserved model_ Now let us look at the our system when the time-evolution of m(t) is conserved in average. We will see that the behavior changes drastically when one imposes average conservation, especially in high dimensions. In the aymmetric case with average conservation we have no correlation between the onsite and link variables, therefore the evolution of the correlator Gr(t) takes the following form � Gr(t + 1) = ⟨ξ⟩[2][ �] Kr−kKlGk−l(t) + 2⟨V ⟩⟨ξ⟩ Kr−kGk(t) _k,l_ _k_ + ⟨V ⟩[2]Gr(t) + δr,0G0(t) �σξ[2][∥][K][∥][2][ +][ σ]V[2] � + σn[2] _[δ][r,][0][.]_ (58) Let us focus on the analysis of the evolution equation in the next section. The steps of calculation are very similar to the exactly conserved system, therefore one jumps immediately to the analysis of the solution. _3.2.1._ _Noise threshold in high dimensions_ Let us again focus on the Laplacian kernel without loss of generality. The self-consistent equation for Gst(0) gives Gst(0) = _cdσn[2]_ (59) 1 − _cdσ[2][,]_ where cd is given by the same integral Eq. (41) as in the previous section and σ[2] = σV[2] [+][ σ]ξ[2][∥][K][∥][2][. The] stationary solution exists if and only if 1 − _cdσ[2]_ _> 0. The critical value of the variance being_ _σc[2]_ [= 1] _._ (60) _cd_ 14 ----- _σσξξ[2]_ Exponential Phase Finite Phase Finite Phase _σσVV[2]_ **Figure 6.** There is a critical line f (σV[2] _[, σ]ξ[2][) in the plane (][σ]V[2]_ _[, σ]ξ[2][) which is parametrized by the ellipse]_ equation _[σ]σV[2]c[2]_ [+] _σσξc[2]_ _[∥][K][∥][2][ = 1. In the white domain defined by][ f]_ [(][σ]V[2] _[, σ]ξ[2][)][ ⩾]_ [1, the algorithm becomes unstable] and the coherence is exponential. In the blue domain defined by f (σV[2] _[, σ]ξ[2][)][ <][ 1, it is finite and the algorithm]_ reaches a global consensus. Let us mention that the critical line exists only in high dimensions d > 2, the size of the ellipse decreases (the dashed ellipses) as d → _dc and eventually shrinks down to a single point at_ _dc. The system is then completely unstable for any non-zero variances._ There is a critical line f (σV[2] _[, σ]ξ[2][) in the plane (][σ]V[2]_ _[, σ]ξ[2][) which is parametrized by the ellipse (see Fig .(][6][))]_ equation _σV[2]_ + _σξ[2]_ K = 1. (61) _∥_ _∥[2]_ _σc[2]_ _σc[2]_ The constant cd depends only of the explicit form of the kernel, the mean of the link variables ⟨ξ⟩ and the dimension d of the space. Because of the divergence of cd in d ⩽ 2, there is no threshold in low dimensions and the system is always controlled by the variables ξ’s and V ’s as we will see later on. In d ⩽ 2, the stability margin Eq. (8) becomes infinitesimal and any non-zero amount of multiplicative noise makes the algorithm unstable. The full space-dependance of the stationary solution below σc[2] [and in][ d >][ 2 takes the following] form �� _∞_ _e[−][i][qr]_ d[d]q (62) 1 _λ(q)[2][ .]_ _−∞_ _−_ Gst(r) = _σn[2]_ (2π)[d] � 1 1 − _cdσ[2]_ The picture is almost the same as in the exactly conserved system where below the threshold, the system behaves as a pure additive model Eq. (46). In this regime, the renormalization factor becomes 1 _A[av]σ[2][ =]_ (63) 1 − _cdσ[2][ .]_ Below σc[2][, the system is also described by the pure additive model Eq. (][46][) with ˜][n][i][(][t][) =][ �][A][av]σ[2] _[n][i][(][t][) and]_ the FV scaling Eq. (47)) holds with the same set of exponents. The main difference is the FV scaling holds only in d > 2, where the exponent β goes to zero, and where the stationary solution Eq. (62) converges to a finite value Eq. (59). We have actually a critical line (see Fig .(6)) in the plane (σV[2] _[, σ]ξ[2][). The behavior is]_ the same anywhere below the line where the large-scale dynamic is described by Eq. (46). Because there is no relation between the variances σV[2] [and][ σ]ξ[2][, we can send independently][ σ]V[2] [or][ σ]ξ[2] [to zero. The large-scale] dynamic above the threshold can then be described by � _ui(t + 1) = ⟨ξ⟩_ Ki−juj(t) + V[˜]i(t)ui(t), (64) _j_ with V[˜]i(t) has a renormalized variance σV[2] [+] _[σ]ξ[2][∥][K][∥][2][. The behavior of our system when average conservation]_ is imposed, varies greatly from the exactly conserved algorithm, where the system could not be described by 15 ----- _d_ _σc[2]_ [=] _cd1_ Finite Exponential _CN[∞]_ [=] 1−cdσ2σ[2]n[2]cd _CN_ (t) ∼ _λ[t]max_ _dc = 2_ Exponential _CN_ (t) ∼ _λ[t]max_ 0 _σ[2]_ **Figure 7. Phase diagram in the average conservation case for the Laplacian kernel θ = 1. Here the scenario** is rather different than the exact conserved case in dimension d. In low d, the behavior of the coherence is always exponential for any system-size. In higher d there is a transition between a finite phase and an exponential regime at the critical value σc[2][.] Eq. (64) due to the relation between the variances σV[2] [and][ σ]ξ[2][. The eigenvalue equation for this model can] also be written as Eq. (50). If λmax > 1 then the correlator behaves as G([�] **q, t) ∼** _λ[t]max[g][�][(][q][), where]_ _σV[2]_ [+][ σ]ξ[2][∥][K][∥][2] _g(q) =_ (65) � _λmax −_ _λ(q)[2][ .]_ The condition on λmax is now � d[d]q (66) (2π)[d][ �][g][(][q][) = 1][.] The convergence properties of this integral are fairly different than the one found in the previous section. The integral Eq. (66) when λmax → 1 converges in d > 2. Thus in d ⩽ 2 there is no stationary state and the correlator grows exponentially. In d > 2 there is a noise threshold σc[2][, below this value the behavior is] algebraic, and above it is exponential for any σ[2]. In that case, the phase transition occurs at d > 2, indeed in d ⩽ 2 the system remains exponential for any value of σ[2]. The low d phase is the strong coupling regime and in higher d, there is two phases depending on the value of σ[2]. Let us finally notice that, contrary to the exact conservation case, here the phase transition may happen at an infinitesimal value of σ, just above the critical dimension dc. The complete phase diagram is shown in Fig .(7). Two regimes are present, just as before, the EW limit where the variance σ[2] is smaller than the critical value and where the multiplicative noise is irrelevant. This happens in low dimensions. The other regime can be shown to be described by the stochastic heat equation (SHE) _∂tu(x, t) = ∂x[2][u][(][x][, t][) + ˜][V][ (][x][, t][)][u][(][x][, t][)][.]_ (67) This equation is known to describe the partition function u(x, t) of a directed polymer in a quenched random potential V[˜] (x, t) and the height h(x, t) log u(x, t) of a random interface verifying the Kardar-Parisi-Zhang _∝_ equation (see [38] for details or [39] for even more details). _3.2.2. Robustness of the coherence and finite-size dependance_ The analysis of the correlator G(r, t) tells us that in low dimensions, the consensus algorithm is extremely sensitive to multiplicative noise. The network coherence becomes essentially exponential as soon as a finite value of multiplicative noise is introduced in the system, making the algorithm unstable and unable to perform its task. The stability margin, is essentially zero in d < 3. In higher dimensions d > 2, the stability margin becomes is non-zero, and the system remains stable as long as the noise variance remains below the threshold. The higher the dimension and the bigger the stability margin is. From that result on the calculation of G(r, t) below the noise threshold, one can translate the information in terms of the stationary value CN[∞] [of the network coherence defined (see Eq. (][4][)). The] 16 ----- finite-size behavior of the coherence can be extracted by introducing a cut-off in Fourier space in Eq. (41). For a system of size L with N = L[d] nodes, the coherence is reaching a stationary value which scales as system-size as _CN[∞]_ _[∝]_ _N[1]_ � 2−d d[d]rGst(r) ∼ _N_ _d ._ (68) In that context, in d > 2 the algorithm is still capable to perform its task, even with the presence of multiplicative noise, as long as the variance is below the threshold σc[2][. In lower dimensions, the stability] margin is inexistent and there is no stationary state. An infinitesimal amount of multiplicative noise makes the system exponentially unstable, the system is thus infinitely fragile. _3.3. A word on the other forms of symmetry_ As we said earlier in the article, the asymmetric link case, is the simplest case where the link and onsite variables are uncorrelated in space, see Eq. (21)), regardless the type of conservation condition that we consider. Therefore, this case leads to a rather simple equation for the time-evolution of the correlator, and the calculation can be done for any kernel. The other cases might be more or less involved depending on the form of the links and the conservation condition, indeed when one enforces symmetry between link variables for example, the onsite random variables Vi(t) become spatially correlated (see Eq. (24)) leading to different asymptotic behaviors. Nevertheless, the method that we have proposed here can be applied directly to those situations, leading to the explicit form of the noise threshold. The details of those models will be detailed elsewhere although let us finally mention, that the crucial ingredient in the different universal behaviors that we have observed is the presence of some form of a conservation law, and we believe that the specific details of the system will not change the dimensional behaviors and phase diagrams Fig .(5) and Fig .(7), but could have some relevance for more specific applications to real consensus algorithms and communication networks. **4. Summary and conclusions** In this paper, we have quantified the performance of a consensus algorithm or a distributed system by studying the network coherence Eq. (4) of a system of N agents _CN_ (t) = [1] _N_ � d[d]r �G(r, t) _m(t)_ _,_ (69) _−⟨_ _⟩[2][�]_ in the large-scale limit N and where G(r, t) is the local 2-point correlation _u(r, t)u(0, t)_ . We were _→∞_ _⟨_ _⟩_ interested in the behavior of this quantity for a system with diverse sources of uncorrelated multiplicative and additive noise (see Eq. (9)) in a lattice of dimension d in the continuum-space limit. We showed how the quantity G(r, t) behaves in both cases where one imposes conservation of m(t) either exactly or in average. Let us summarize the results of the continuum-space calculation of the time-behavior of the network coherence CN (t), for a system of N nodes, when varying the strength of the link and onsite random variables. The general behavior in the exact conservation case, is that in any dimension, below a critical value _σc[2][, the network coherence grows algebraically then reaches the stationary state at large time, and above]_ _σc[2][, the network coherence grows exponentially, see Fig .(][5][). The stability margin is always non-zero in any]_ dimension, and the system can remain stable as long as the variance of the noise is below the threshold. The other case is when average conservation of m(t) is enforced. In that case, the link and onsite variables are independent, making the phenomenology richer than the exact conservation case. In low dimensions, the network coherence grows exponentially, for any value of the variances σV[2] [and][ σ]ξ[2][. The system is then] highly sensitive to multiplicative noise and there is no stability margin. Obviously in that case, no stationary solution is reachable and the network coherence becomes infinite at large time. In higher d > 2, a stability margin appears, the system is then capable to perform the average for small noise. The network coherence is then finite below the noise threshold and for σ[2] ⩾ _σc[2]_ [the network coherence grows exponentially again, see] 17 ----- Fig .(7). Another quantity that might be extremely insightful in the context of network consensus algorithms is the variance with respect to random initial conditions ui(t0) at time t0 _N_ � � _uk(t0)_ _._ (70) _k=1_ _CN_ (t, t0) := [1] _N_ _N_ � � var _ui(t) −_ [1] _N_ _i=1_ This quantity tells us how the system forgets (or not) about the initial state values on the network and how time correlations grow in the system. In our formalism, the calculation boils down to compute G(r, t, t0) = ⟨u(r, t)u(0, t0)⟩. Those correlations have not been studied in details in the control literature but are well known in the ageing literature [40]. A problem not addressed in this work is the case of a static (quench) disorder, which may have also a lot of interesting applications in averaging systems and consensus algorithms. A similar threshold, known in the context of Anderson localization [41], appears in those systems as well, and it will be interesting to see how localization emerges in distributed systems with quench disorder. The SHE equation with this type of disorder is often called the parabolic Anderson problem, see for example [42]. One of the most ambitious perspective would be to show that the noise threshold that we have observed in those classical systems, would also appear in quantum systems. Indeed it is well known in the field of quantum computation [19] that noise is the main obstacle to efficient computation, and that above a certain value, computation is no longer possible. This well-known result is called the ”quantum threshold theorem” [43] and it has later been shown that this phenomenon is actually a phase transition [44]. It would of great interest if our results could be extended to quantum systems, and prove that a similar threshold theorem would hold for quantum distributed systems and quantum consensus [18]. This work is partially supported by NSF Awards PHY-1344069 and EECS-1408442. NA is grateful to Thimoth´ee Thiery for discussions during the KPZ program at KITP. CS is grateful to the Labex NEXT, the CSHL, and the MUSE IDEX Toulouse contract for funding his visit to NY. 18 ----- [1] M. Mezard and A. Montanari, Information, physics, and computation. Oxford University Press, 2009. [2] R. Olfati-Saber, “Flocking for multi-agent dynamic systems: Algorithms and theory,” IEEE Transactions on automatic _control, vol. 51, no. 3, pp. 401–420, 2006._ [3] D. Chowdhury, L. Santen, and A. Schadschneider, “Statistical physics of vehicular traffic and some related systems,” _Physics Reports, vol. 329, no. 4, pp. 199–329, 2000._ [4] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation in networked multi-agent systems,” Proceedings _of the IEEE, vol. 95, no. 1, pp. 215–233, 2007._ [5] Y.-Y. Liu, J.-J. Slotine, and A.-L. Barab´asi, “Controllability of complex networks,” Nature, vol. 473, no. 7346, pp. 167–173, 2011. [6] T. Vicsek and A. Zafeiris, “Collective motion,” Physics Reports, vol. 517, no. 3, pp. 71–140, 2012. [7] C. Castellano, S. Fortunato, and V. Loreto, “Statistical physics of social dynamics,” Reviews of modern physics, vol. 81, no. 2, p. 591, 2009. [8] J.-P. Bouchaud and M. Potters, Theory of financial risk and derivative pricing: from statistical physics to risk management. Cambridge university press, 2003. [9] R. N. Mantegna and H. E. Stanley, Introduction to econophysics: correlations and complexity in finance. Cambridge university press, 1999. [10] F. Bullo, Lectures on Network Systems. Version 0.85, 2016. With contributions by J. Cortes, F. Dorfler, and S. Martinez. [11] Y. Kuramoto, Chemical oscillations, waves, and turbulence, vol. 19. Springer Science & Business Media, 2012. [12] J. A. Acebr´on, L. L. Bonilla, C. J. P. Vicente, F. Ritort, and R. Spigler, “The kuramoto model: A simple paradigm for synchronization phenomena,” Reviews of modern physics, vol. 77, no. 1, p. 137, 2005. [13] J. Cortes, S. Martinez, T. Karatas, and F. Bullo, “Coverage control for mobile sensing networks,” in Robotics and _Automation, 2002. Proceedings. ICRA’02. IEEE International Conference on, vol. 2, pp. 1327–1332, IEEE, 2002._ [14] L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,” Systems & Control Letters, vol. 53, no. 1, pp. 65–78, 2004. [15] J. Lin, A. S. Morse, and B. D. Anderson, “The multi-agent rendezvous problem-the asynchronous case,” in Decision and _Control, 2004. CDC. 43rd IEEE Conference on, vol. 2, pp. 1926–1931, IEEE, 2004._ [16] J. Lin, A. S. Morse, and B. D. Anderson, “The multi-agent rendezvous problem. part 2: The asynchronous case,” SIAM _Journal on Control and Optimization, vol. 46, no. 6, pp. 2120–2147, 2007._ [17] W. Ren and R. W. Beard, Distributed consensus in multi-vehicle cooperative control. Springer, 2008. [18] L. Mazzarella, A. Sarlette, and F. Ticozzi, “Consensus for quantum networks: Symmetry from gossip interactions,” IEEE _Transactions on Automatic Control, vol. 60, no. 1, pp. 158–172, 2015._ [19] M. A. Nielsen and I. L. Chuang, Quantum computation and quantum information. Cambridge university press, 2010. [20] L. Xiao, S. Boyd, and S.-J. Kim, “Distributed average consensus with least-mean-square deviation,” Journal of Parallel _and Distributed Computing, vol. 67, no. 1, pp. 33–46, 2007._ [21] M. Huang and J. H. Manton, “Coordination and consensus of networked agents with noisy measurements: stochastic algorithms and asymptotic behavior,” SIAM Journal on Control and Optimization, vol. 48, no. 1, pp. 134–161, 2009. [22] S. Patterson and B. Bamieh, “Consensus and coherence in fractal networks,” IEEE Transactions on Control of Network _Systems, vol. 1, no. 4, pp. 338–348, 2014._ [23] B. Bamieh, M. R. Jovanovic, P. Mitra, and S. Patterson, “Coherence in large-scale networks: Dimension-dependent limitations of local feedback,” IEEE Transactions on Automatic Control, vol. 57, no. 9, pp. 2235–2249, 2012. [24] F. Fagnani and S. Zampieri, “Randomized consensus algorithms over large scale networks,” IEEE Journal on Selected _Areas in Communications, vol. 26, no. 4, pp. 634–649, 2008._ [25] R. Carli, F. Fagnani, P. Frasca, and S. Zampieri, “Gossip consensus algorithms via quantized communication,” Automatica, vol. 46, no. 1, pp. 70–80, 2010. [26] S. Patterson, B. Bamieh, and A. El Abbadi, “Convergence rates of distributed average consensus with stochastic link failures,” IEEE Transactions on Automatic Control, vol. 55, no. 4, pp. 880–892, 2010. [27] J. Wang and N. Elia, “Distributed averaging under constraints on information exchange: emergence of l´evy flights,” IEEE _Transactions on Automatic Control, vol. 57, no. 10, pp. 2435–2449, 2012._ [28] J. Wang and N. Elia, “Consensus over networks with dynamic channels,” International Journal of Systems, Control and _Communications, vol. 2, no. 1-3, pp. 275–297, 2010._ [29] G. Cybenko, “Dynamic load balancing for distributed memory multiprocessors,” Journal of parallel and distributed _computing, vol. 7, no. 2, pp. 279–301, 1989._ [30] J. E. Boillat, “Load balancing and poisson equation in a graph,” Concurrency and Computation: Practice and Experience, vol. 2, no. 4, pp. 289–313, 1990. [31] J. C. Doyle, B. A. Francis, and A. R. Tannenbaum, Feedback control theory. Courier Corporation, 2013. [32] A. Crisanti, G. Paladin, and A. Vulpiani, Products of Random Matrices: in Statistical Physics, vol. 104. Springer Science & Business Media, 2012. [33] B. Touri, Product of random stochastic matrices and distributed averaging. Springer Science & Business Media, 2012. [34] R. R. Muller, “On the asymptotic eigenvalue distribution of concatenated vector-valued fading channels,” IEEE _Transactions on Information Theory, vol. 48, no. 7, pp. 2086–2091, 2002._ [35] M. L. Mehta, Random matrices, vol. 142. Academic press, 2004. [36] F. Family and T. Vicsek, “Scaling of the active zone in the eden process on percolation networks and the ballistic deposition model,” Journal of Physics A: Mathematical and General, vol. 18, no. 2, p. L75, 1985. [37] A.-L. Barab´asi and H. E. Stanley, Fractal concepts in surface growth. Cambridge university press, 1995. [38] J. Krug and H. Spohn, “Kinetic roughening of growing surfaces,” Cambridge University Press, Cambridge, vol. 1, no. 99, p. 1, 1991. 19 ----- [39] T. Halpin-Healy and Y.-C. Zhang, “Kinetic roughening phenomena, stochastic growth, directed polymers and all that. aspects of multidisciplinary statistical mechanics,” Physics reports, vol. 254, no. 4, pp. 215–414, 1995. [40] M. Henkel, H. Hinrichsen, S. L¨ubeck, and M. Pleimling, Non-equilibrium phase transitions, vol. 1. Springer, 2008. [41] P. W. Anderson, “Absence of diffusion in certain random lattices,” Physical review, vol. 109, no. 5, p. 1492, 1958. [42] Y. B. Zel’Dovich, S. Molchanov, A. Ruzmaikin, and D. D. Sokolov, “Intermittency in random media,” Physics-Uspekhi, vol. 30, no. 5, pp. 353–369, 1987. [43] D. Aharonov and M. Ben-Or, “Fault-tolerant quantum computation with constant error,” in Proceedings of the twenty_ninth annual ACM symposium on Theory of computing, pp. 176–188, ACM, 1997._ [44] D. Aharonov, “Quantum to classical phase transition in noisy quantum computers,” Physical Review A, vol. 62, no. 6, p. 062311, 2000. 20 -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1610.00653, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://arxiv.org/pdf/1610.00653" }
2,016
[ "JournalArticle" ]
true
2016-10-03T00:00:00
[ { "paperId": "26815359d454f17bfdd1ff46e482182b6ff85929", "title": "Extreme value statistics of correlated random variables" }, { "paperId": "ba3844f0112a761fc9354582f5254e8e9db63d28", "title": "Consensus and Coherence in Fractal Networks" }, { "paperId": "68074c123229f29fa742e354a2c0a8b6af8057aa", "title": "Consensus for Quantum Networks: Symmetry From Gossip Interactions" }, { "paperId": "ddbf9bc7a13e503f9afcaa4aea1a6495afb41dc8", "title": "Quantum Computation and Quantum Information" }, { "paperId": "7c588932ea3c80975f063afefee653732084c717", "title": "Product of Random Stochastic Matrices and Distributed Averaging" }, { "paperId": "46db537b24e707bf6229c3b74afbaf1b867e27e6", "title": "Distributed Averaging Under Constraints on Information Exchange: Emergence of Lévy Flights" }, { "paperId": "2784dcfdc3fc5726bc2426692dbad06a5b3338ce", "title": "Coherence in Large-Scale Networks: Dimension-Dependent Limitations of Local Feedback" }, { "paperId": "3b62c331be63a928278282418de40f63ba9b0e48", "title": "Controllability of complex networks" }, { "paperId": "04a7df9214f1c1165dd7e169953ae4d5de412566", "title": "Diffusion and multiplication in random media" }, { "paperId": "cd67aaa9b8bc9ad53a2f0b6392e3ceb8f8c589f7", "title": "SPECTRAL DISTRIBUTIONS OF ADJACENCY AND LAPLACIAN MATRICES OF RANDOM GRAPHS" }, { "paperId": "d015a49ef3685ba48a765a17a442ba2c6507d67a", "title": "Free-energy distribution of the directed polymer at high temperature" }, { "paperId": "dd409a309ba4768c910688efbb0a76ee71e2885d", "title": "Convergence Rates of Distributed Average Consensus With Stochastic Link Failures" }, { "paperId": "24f02bb525da2091346984890401a312a20703fc", "title": "Gossip consensus algorithms via quantized communication" }, { "paperId": "3533a122e4db6b9c8f97dfe6516c7ccc9661e164", "title": "Coordination and Consensus of Networked Agents with Noisy Measurements: Stochastic Algorithms and Asymptotic Behavior" }, { "paperId": "ac01e2bfe4961bef7af7322437bb39b5f0feaa26", "title": "Consensus over networks with dynamic channels" }, { "paperId": "bbadd4370cb353cd31a260668824f27a41c9af56", "title": "Distributed Consensus in Multi-vehicle Cooperative Control - Theory and Applications" }, { "paperId": "6cf90d5847351130d978c2097bfe21957a024f0a", "title": "Distributed average consensus with stochastic communication failures" }, { "paperId": "6ed59b022b11aaa03ed58ec9fb70e16ea64ab9c4", "title": "Randomized consensus algorithms over large scale networks" }, { "paperId": "e419cfbbdd1de7f9a2ed6bb2d5392840dcb2a4fd", "title": "Statistical physics of social dynamics" }, { "paperId": "aa6be519b394b44ab24c6ad964f8a2c6a9b23571", "title": "Consensus and Cooperation in Networked Multi-Agent Systems" }, { "paperId": "bf4c74068a75b3b01497f29c0bfb090cd8124510", "title": "The Multi-Agent Rendezvous Problem. Part 2: The Asynchronous Case" }, { "paperId": "d12c412eaf6c8e6be3814e546c36a8776d5e8fff", "title": "Flocking for multi-agent dynamic systems: algorithms and theory" }, { "paperId": "6ce6aa385045b19268f0620c9192028df51f65e6", "title": "The Kuramoto model: A simple paradigm for synchronization phenomena" }, { "paperId": "8da11c1b11b0d62c68992d019fd3d2d2ed063bfb", "title": "The Parabolic Anderson Model" }, { "paperId": "48372b9fdbe64ec8d619babaf7f7ee734b00127c", "title": "Fast linear iterations for distributed averaging" }, { "paperId": "028440361b95d0377e850da47175198ae227a6f0", "title": "Coverage control for mobile sensing networks" }, { "paperId": "64bda138bb95e4d3796dfc291461f0496a564e33", "title": "Products of random matrices." }, { "paperId": "f0400ba52bf050d36782fe721d394371eba81f83", "title": "Theory Of Financial Risk And Derivative Pricing" }, { "paperId": "85b04239d38468e276deb89f8e02d3e4b5f2cf01", "title": "Random Incidence Matrices: Moments of the Spectral Density" }, { "paperId": "292897bb00885a38d54f5425512031d905676be5", "title": "Statistical physics of vehicular traffic and some related systems" }, { "paperId": "ed19bd9fd44a1b5e4aaede814e1eeba0925b12f8", "title": "An Introduction to Econophysics: Correlations and Complexity in Finance" }, { "paperId": "35a8766be2b5862e99397a77b37f076c42a18d07", "title": "Quantum to classical phase transition in noisy quantum computers" }, { "paperId": "6b811738b111869d50b643d978f63adb72092223", "title": "Collective Motion" }, { "paperId": "24d228df6b7c2e196482a4251c22eadff85eb475", "title": "Fault-tolerant quantum computation with constant error" }, { "paperId": "07a333cf0c75ad7ba4321f24af1a49b98b725792", "title": "Kinetic roughening phenomena, stochastic growth, directed polymers and all that. Aspects of multidisciplinary statistical mechanics" }, { "paperId": "651a41f7078f86c9320bd128ad13718f7721bb9e", "title": "Generic behavior in linear systems with multiplicative noise." }, { "paperId": "b8183abf0a3c9b364caf33b4d75699231a5dc580", "title": "Replica-scaling analysis of diffusion in quenched correlated random media." }, { "paperId": "73d8775c6bb21b2faa882107b3cd59d83e7328cf", "title": "Load Balancing and Poisson Equation in a Graph" }, { "paperId": "1d1efce717535589753503a09eb7ff2c876c0128", "title": "Anomalous diffusion in disordered media: Statistical mechanisms, models and physical applications" }, { "paperId": "706af6fa043dcbed058fdfed8fe939084661f7f7", "title": "Diffusion in a random catalytic environment, polymers in random media, and stochastically growing interfaces." }, { "paperId": "ccc91e6fc8c43f7b3c845ce271acedaef2c515eb", "title": "Dynamic Load Balancing for Distributed Memory Multiprocessors" }, { "paperId": "51eb4d3880943ce8c0c3dc5823f3f5b1257b9542", "title": "Multidimensional diffusion in random potentials." }, { "paperId": "08f5226065e0646e27e27061390570036dadc8bf", "title": "Scaling of directed polymers in random media." }, { "paperId": "919736effeef76a8edaa11a843394680d976707b", "title": "Intermittency in random media" }, { "paperId": "fe3e2ab70e3d106128f85f239d302c00c97d613f", "title": "Scaling of the active zone in the Eden process on percolation networks and the ballistic deposition model" }, { "paperId": "f6f215bf741217b7acf66ffe301dd11b2a4f4d2b", "title": "Chemical Oscillations, Waves, and Turbulence" }, { "paperId": "0c8e7919b31430753d1d3951a21a06a8b2c4f6c2", "title": "Absence of Diffusion in Certain Random Lattices" }, { "paperId": "f657f01c94c62117145c579700cb03beafc2fe9c", "title": "Stochastic differential equations" }, { "paperId": null, "title": "Lectures on Network Systems . Version 0.85" }, { "paperId": "a970d706c7c43f4330f334117041cb316000e87d", "title": "Theory of Financial Risk and Derivative Pricing: From Statistical Physics to Risk Management" }, { "paperId": "0b0cc2c941c10ff180e036c068d51c0d6cc1676c", "title": "Feedback Control Theory" }, { "paperId": null, "title": "Non-Equilibrium Phase Transitions vol 1 (Berlin: Springer" }, { "paperId": "5359fb2362ee22a18a5cc1bf9ff7f69d7ce533bf", "title": "Distributed average consensus with least-mean-square deviation" }, { "paperId": "d70fc00c99028c1f29f53b849903230a9482d5e4", "title": "The multi-agent rendezvous problem - the asynchronous case" }, { "paperId": null, "title": "Random Matrices vol 142 (New York: Academic" }, { "paperId": null, "title": "Random matrices Academic press" }, { "paperId": "69b2c55f2cec2a9144e5bd90b47ba02722f44cc4", "title": "On the asymptotic eigenvalue distribution of concatenated vector-valued fading channels" }, { "paperId": "655c280a2e526b8af5bc03bc6b5fd51940263c5a", "title": "Coverage control for mobile sensing networks" }, { "paperId": "9d6641e3dba49d6b7cecf42d42101af05e295f17", "title": "An Introduction to Econophysics: Contents" }, { "paperId": "310cfa8ebe987b4e0a6c4f361f7e0dbd34dda615", "title": "Information, physics, and computation" }, { "paperId": "4c149382eff152778ba58a8ee8c91732364686e2", "title": "Fractal Concepts in Surface Growth" }, { "paperId": "d3e8569557ad7fd551addab72d32573b22d259d4", "title": "Fractal Concepts in Surface Growth: Frontmatter" }, { "paperId": "beeffbf0b18653ceab7f5bb80a2187b84c18c71c", "title": "Products of random matrices in statistical physics" }, { "paperId": null, "title": "Kinetic roughening of growing surfaces" }, { "paperId": null, "title": "Let us remark that V = 1 − ξK1 is not completely satisfied in that case because our example is not periodic" }, { "paperId": null, "title": "In the case of symmetric links another term σ" }, { "paperId": null, "title": "would hold for quantum distributed systems and quantum consensus [18]. This work is partially supported by" } ]
20,565
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00f77951251798ae562e1ebe066753594f4b37be
[ "Computer Science" ]
0.866705
Private Stream Aggregation with Labels in the Standard Model
00f77951251798ae562e1ebe066753594f4b37be
Proceedings on Privacy Enhancing Technologies
[ { "authorId": "2053351159", "name": "J. Ernst" }, { "authorId": "2059727806", "name": "Alexander Koch" } ]
{ "alternate_issns": null, "alternate_names": [ "Proc Priv Enhancing Technol" ], "alternate_urls": null, "id": "d5dc4224-e4c3-43c9-918a-bd6326650b5b", "issn": "2299-0984", "name": "Proceedings on Privacy Enhancing Technologies", "type": null, "url": "https://www.degruyter.com/view/j/popets" }
Abstract A private stream aggregation (PSA) scheme is a protocol of n clients and one aggregator. At every time step, the clients send an encrypted value to the (untrusted) aggregator, who is able to compute the sum of all client values, but cannot learn the values of individual clients. One possible application of PSA is privacy-preserving smart-metering, where a power supplier can learn the total power consumption, but not the consumption of individual households. We construct a simple PSA scheme that supports labels and which we prove to be secure in the standard model. Labels are useful to restrict the access of the aggregator, because it prevents the aggregator from combining ciphertexts with different labels (or from different time-steps) and thus avoids leaking information about values of individual clients. The scheme is based on key-homomorphic pseudorandom functions (PRFs) as the only primitive, supports a large message space, scales well for a large number of users and has small ciphertexts. We provide an implementation of the scheme with a lattice-based key-homomorphic PRF (secure in the ROM) and measure the performance of the implementation. Furthermore, we discuss practical issues such as how to avoid a trusted party during the setup and how to cope with clients joining or leaving the system.
### Johannes Ernst* and Alexander Koch # Private Stream Aggregation with Labels in the Standard Model[†] **Abstract: A private stream aggregation (PSA) scheme** is a protocol of n clients and one aggregator. At every time step, the clients send an encrypted value to the (untrusted) aggregator, who is able to compute the sum of all client values, but cannot learn the values of individual clients. One possible application of PSA is privacy-preserving smart-metering, where a power supplier can learn the total power consumption, but not the consumption of individual households. We construct a simple PSA scheme that supports labels and which we prove to be secure in the standard model. Labels are useful to restrict the access of the aggregator, because it prevents the aggregator from combining ciphertexts with different labels (or from different timesteps) and thus avoids leaking information about values of individual clients. The scheme is based on key-homomorphic pseudorandom functions (PRFs) as the only primitive, supports a large message space, scales well for a large number of users and has small ciphertexts. We provide an implementation of the scheme with a lattice-based key-homomorphic PRF (secure in the ROM) and measure the performance of the implementation. Furthermore, we discuss practical issues such as how to avoid a trusted party during the setup and how to cope with clients joining or leaving the system. **Keywords: Private Stream Aggregation, Aggregator** Obliviousness, Standard Model, Pseudorandom Function, Lattice-Based Cryptography, Learning With Rounding, Smart-Meters DOI 10.2478/popets-2021-0063 Received 2021-02-28; revised 2021-06-15; accepted 2021-06-16. ***Corresponding Author: Johannes Ernst: University of** St. Gallen (most of the work done while at KIT, Karlsruhe) **Alexander Koch: Competence Center for Applied Security** Technology (KASTEL), Karlsruhe Institute of Technology † An extended abstract of this work appeared in [19] ## 1 Introduction Smart meters are becoming more and more ubiquitous in many countries. This has advantages for the power suppliers, because they get near real-time power consumptions from their clients, which they can use for load-balancing and prediction in their networks. However, this raises the question of privacy. Sensitive information like the work schedule can easily be guessed from the variation in the power consumption of a household. In practice, it is often sufficient for the power supplier to know the sum of the consumptions of all clients within a certain area, and this is exactly what a protocol for private stream aggregation (PSA) [22] can offer. PSA considers the scenario where an aggregator wishes to periodically compute the sum of values that are supplied by different clients. The values are encrypted in such a way that the aggregator can only compute their sum, but not the individual values. This is captured in the game-based security definition of aggregator obliviousness (AO), which is given in Definition 3. It is desirable for PSA schemes to support the use of labels. Labels restrict the aggregator to only be able to compute the sum of values which were encrypted under the same label. This prevents the aggregator from mixing ciphertexts of different time steps and thereby learning more about the individual values than would otherwise be possible. A clear advantage of PSA schemes is that they do not require the clients to exchange messages, nor does the aggregator need to send messages to the clients. After the keys have been distributed, the only messages in the protocol are the ciphertexts which the clients send to the aggregator. PSA stays secure even if the aggregator colludes with an arbitrary subset of the clients. In that case, the aggregator only learns the sum of the non-colluding clients’ values. When we apply a PSA protocol to the smart meter scenario, this means that every smart meter encrypts (e.g. every fifteen minutes) its current power consumption and sends it to the supplier. The supplier then is able to compute the sum and thereby learns the power consumption of all households in the specific area. Because of the way the clients’ values are encrypted, the ----- _only information the power supplier gets is the sum of_ all values. To further protect the privacy of each client, techniques from differential privacy can be used. In this case, it means that every client adds a small amount of noise to their value before encrypting it. This induces a small error in the resulting sum, but in many cases a small error is tolerable. However, in this paper we focus on the encryption part. Differential privacy can then be added by standard techniques, e.g. as described in [22]. Apart from privacy-preserving smart metering, PSA has a lot more of possible applications. For example it can be used in federated learning in a similar way to the protocol of [11], to compute a global model update from the local updates that are supplied by the clients. This can help to prevent an adversary from using the model to infer information on the, possibly sensitive, data which the clients used to train the model. PSA (without differential privacy) can be seen as special case of inner-product multi-client functional encryption (IP-MCFE) first introduced by [16]. In innerproduct multi-client functional encryption there are several clients and one or more aggregators. The aggregators can ask for functional decryption keys associated with an arbitrarily chosen vector y. The functional decryption key then enables them to compute the inner product of the clients’ values with the vector y. When we only allow the vector y = (1, . . ., 1), then this is exactly the case of PSA (without differential privacy). The main challenge in IP-MCFE is that the scheme must be secure, although the vectors y are not known at the beginning and an aggregator can hold keys for many different vectors. ### 1.1 Contribution **Provably secure PSA scheme with labels in the** **standard model: The scheme we propose is the first** PSA scheme that both supports labels and is proven to be aggregator oblivious (AO) with adaptive corruptions in the standard model. Although, strictly speaking, the IP-MCFE scheme of [1] can also be used as PSA scheme with the same properties, compared to their scheme, ours is more efficient. The size of each user secret key and the length of the ciphertexts in their scheme grows linearly with the number of clients, whereas ours are constant, i.e. independent of the number of clients. Becker et al. have also proposed a PSA scheme in the standard model, but without labels [9]. They roughly explain how to extend their scheme to support labels, but they provide no security proof of this extension. Furthermore, their scheme seems to be subject to a patent [8]. Our scheme is very similar to the PSA scheme of [24] who used key-homomorphic weak PRFs but only proved non-adaptive security in the standard model. Also, as opposed to our scheme, the labels have to be precomputed at setup time and distributed to all clients. Thus, the scheme does not support an unbounded number of labels and the key size grows linearly with the number of labels. Our scheme needs as its only building block a keyhomomorphic pseudorandom function (PRF) whose output space is ZR for some integer R. It thus relies only on secret-key primitives and is quite flexible. It makes efficient use of the key-homomorphic PRF, as both encryption and decryption only require one PRF evaluation. The scheme can be instantiated with a latticebased key-homomorphic PRF, which are assumed to be secure against quantum adversaries. Additionally, we implemented the scheme and a simple key-homomorphic PRF based on the learning with rounding (LWR) problem. For simplicity and efficiency, we chose a PRF that relies on the random oracle model (ROM). The performance tests show that both encryption and decryption are very efficient, in this case. One restriction of our scheme is that it only supports the encryption of one message per label. However, this is a mild restriction, because any (correct) PSA scheme leaks information about individual messages, if a user encrypts more than one message per label. Furthermore, we concretely describe how the scheme can be used for privacy-preserving smart-meter aggregation. We discuss issues that arise in that setting, such as clients joining or leaving the system and how to execute the setup without a trusted party. ### 1.2 Related Work In this section, we give an overview of other work that is related to ours. **1.2.1 Privacy Preserving Aggregation** Shi et al. [22] were the first to formalize the notion of PSA together with the security definition of AO. They propose a scheme that is based on the Decisional Diffie– Hellman (DDH) problem and prove it to be aggregator oblivious in the ROM. Despite its simplicity, the decryption procedure is inefficient because it has to compute ----- a discrete logarithm. This limits the size of the message space such that the discrete logarithm can be computed in reasonable time. Subsequently, Benhamouda et al. [10] propose a general way to build PSA schemes from key-homomorphic smooth projective hash functions (SPHF). Their construction yields schemes that are aggregator oblivious in the ROM. They give concrete instantiations from the DDH-, DCR- and several other assumptions. An advantage over previous schemes is the low reduction loss, which does not depend on the number of users, but only on the maximum number of labels used. This allows for smaller keys and thereby makes the schemes more efficient. Valovich constructs a PSA scheme from keyhomomorphic weak PRFs [24]. In contrast to the other PSA schemes, the author only considers semi-honest adversaries. He proves the scheme to be aggregator oblivious for non-adaptive corruptions in the standard model. For proving adaptive security, the author resorts to the ROM. The main differences to our scheme are that [24] has a weaker security model and that the author uses _weak PRFs. Because of the use of weak PRFs, the labels_ need to be uniformly random. Therefore, the set of labels is created at setup time and given to all parties, to ensure that everyone uses the same labels. This means that the scheme does not support an unbounded number of labels and that the key size grows linearly with the number of labels. Becker et al. propose a generic PSA scheme [9] that can be instantiated with an additively homomorphic encryption scheme, where the addition of ciphertexts corresponds to the addition of plaintexts, with the additional property that the ciphertexts are indistinguishable from random strings. The security of their scheme relies on the Learning with Errors (LWE) assumption, or its ring variant, and is proven to be secure in the standard model. Two advantages over the scheme of Shi et al. [22] are the prospective post-quantum security and the more efficient decryption algorithm. In contrast to the other PSA schemes, they provide an implementation and give performance results. Their implementation has a message space of size 2[16]. The scheme does not directly support labels which limits the practical use cases. Although the authors sketch how to extend the scheme to work with labels, they provide no security proof for that. Except for [9] all other PSA schemes, including ours, have the restriction that every client must only encrypt one message per label. However, this restriction is also reasonable from a security perspective, because even a perfectly secure PSA scheme leaks information about individual client values, if a client encrypts more than one message per label. We elaborate on this in Section 2.4.1. The advantage of our scheme over [22] and the DDH version of [10] is that the size of the message space is not restricted. The advantage over [10, 22, 24] is that we prove our scheme to be secure under adaptive corruptions in the standard model. A disadvantage of our scheme is the larger key size. Benhamouda et al. [10] report key sizes of 592 bit for 128 bit security for their DDH based scheme when using elliptic curves. For our choices of parameters our scheme provides a security of 114 bit[1] with key-sizes of 268288 bit. Note however, that this is mainly due to the use of a post-quantum secure PRF. We give more details on this in Section 4.4. Advantages of our scheme over [9] are that our scheme has a security proof for the case that labels are used and that it is much simpler. Our scheme relies on keyhomomorphic PRFs while theirs needs an additively homomorphic public key encryption scheme with ciphertexts indistinguishable from random. The following works are not directly comparable to ours, because their focus is a bit different. Emura considers the verifiability of aggregated sums [18]. This means that, when the aggregator publishes the sum, they must provide a publicly verifiable proof that the sum is correct. The author proposes two schemes, which are based on the DDH version of [10] and require pairings. In this paper we do not consider the verifiability of the result. Ács and Castelluccia [3] propose an aggregation scheme that uses similar pair-wise masking as the MCFE scheme of [1]. The users in their scheme agree on shared keys via the Diffie–Hellman key-exchange. Additionally, the scheme offers a mechanism by which the aggregator can decrypt the sum, even when some clients drop out. As opposed to PSA, this mechanism needs interaction between the clients and the aggregator. The authors do not provide a security proof, but argue that the scheme is secure against certain attacks. Bonawitz et al. construct a protocol for the aggregation of model updates in distributed machine learning [11]. Their protocol is resistant to user failures and is proven secure against malicious adversaries in the ROM. They use pairwise masks that are set up by Diffie–Hellman key exchanges between the users. The protocol has four rounds of communication to aggre **1 when used with 1000000 clients. With 10000 clients the security** is 132 bit) ----- gate one model update. The key difference to PSA is that their protocol is resistant to user failures, but requires several rounds of communication, whereas PSA is non-interactive. **1.2.2 Multi-Client Functional Encryption** Chotard et al. [16] are the first to define (decentralized) multi-client functional encryption (DMCFE) and show two instantiations for the inner product functionality. Their schemes have several practical limitations and rely on pairings and the ROM. Abdalla et al. [2] address these limitations and remove the need for pairings, while still relying on the ROM. Finally, Abdalla et al. [1] construct the first MCFE scheme with labels that is secure in the standard model. Their scheme works with any PRF and (single-input) functional encryption scheme for inner products. Due to that, their scheme can be based upon many different mathematical problems including LWE, DDH and DCR. Our security proof strongly relies on techniques from the security proof of their MCFE scheme. The main differences between PSA and MCFE are that in MCFE the aggregator(s) can compute inner products with many different vectors and not just the vector consisting of one-entries. Furthermore, in MCFE these vectors can be chosen by the aggregator(s) adaptively during the protocol execution. This make schemes for MCFE harder to construct and usually less efficient. Nevertheless, both areas have a lot of techniques in common. ### 1.3 Concurrent Work Independent of and concurrent to our work, two more papers on PSA have been published in online pre-print archives recently. The authors of [23] propose two PSA schemes based on variants of two fully homomorphic encryption schemes. The security of both schemes relies on the ringLWE assumption. They also implement the schemes and provide a performance analysis. According to this analysis our scheme seems to be faster, however. Waldner et al. propose a PSA scheme based on PRFs [25]. As opposed to our scheme, the PRF does not need to be key-homomorphic. This enables the use of very efficient PRFs. In [25], the authors use AES and SHA3. However, this comes at the cost of requiring n evaluations of the PRF for encrypting one mes sage, where n is the number of users. The authors also provide an implementation and performance results. In Section 4, we compare the running time of the aforementioned schemes with our implementation. Table 1 shows a comparison of the different properties of the PSA schemes that we described in this section. **Scheme** **Proof** **in** **Number** **Adaptive** **Encryption** **standard** **supported** **corrup-** **cost** **per** **model** **labels** **tions** **client** Shi et al. [22] ✗ unbounded ✓ _O(1)_ Benhamouda ✗ unbounded ✓ _O(1)_ et al. [10] Valovich [24] ✓ bounded ✗ _O(1)_ (1[st] scheme) Valovich [24] ✗ bounded ✓ _O(1)_ (2[nd] scheme) LaPS [9] ✓ none ✓ _O(1)_ SLAP [23] ✗ unbounded ✓ _O(1)_ LaSS [25] ✓ unbounded ✓ _O(n)_ Our scheme ✓ unbounded ✓ _O(1)_ **Table 1. Comparison of PSA schemes. Note that the DDH based** schemes in [22] and [10] have the limitation that the message space needs to be small in order to allow taking a discrete logarithm in reasonable time. ### 1.4 Outline We give the necessary definitions and background in Section 2. In Section 3 we explain our PSA scheme and prove its security according to the game-based security definition of AO. In Section 4 we describe the implementation and choices of parameters and give performance results. In Section 5 we discuss several issues related to the deployment of the scheme in practice, with a focus on smart-meters. The last section summarizes our paper. ## 2 Preliminaries In this section, we explain our basic notation and define the cryptographic problems and primitives we use in this paper. |Scheme|Proof in standard model|Number supported labels|Adaptive corrup- tions|Encryption cost per client| |---|---|---|---|---| |Shi et al. [22]|✗|unbounded|✓|O(1)| |Benhamouda et al. [10]|✗|unbounded|✓|O(1)| |Valovich [24] (1st scheme)|✓|bounded|✗|O(1)| |Valovich [24] (2nd scheme)|✗|bounded|✓|O(1)| |LaPS [9]|✓|none|✓|O(1)| |SLAP [23]|✗|unbounded|✓|O(1)| |LaSS [25]|✓|unbounded|✓|O(n)| |Our scheme|✓|unbounded|✓|O(1)| ----- ### 2.1 Notation Here, we quickly explain some of the notation that we use. By [n] we denote the set {1, . . ., n} and with [n]0 we mean {0, . . ., n}. With log(x) we mean the logarithm to base 2 and with ln(x) we mean the logarithm to base e. As security parameter we use λ. Lower-case boldface letters such as v denote vectors. We use the terms _client and user synonymously. By a PPT Turing ma-_ _chine, we mean a probabilistic Turing machine that runs_ in polynomial time. By x ←$ X we mean that x is chosen uniformly random from the set X . With ⟨x, y⟩ we denote the inner-product of two vectors x and y. Let _q, p ∈_ N with q > p. Then, for a value x ∈ Zq we define _⌊x⌋p := ⌊x · p/q⌋._ ### 2.2 Learning With Rounding (LWR) We will define the learning with rounding (LWR) problem, which can be seen as a deterministic version of the learning with errors (LWE) problem. Learning with rounding was introduced in [7] and has turned out to be very useful to construct secret-key primitives such as pseudorandom functions. **Definition 1. Let λ, q, p ∈** N, with q > p and s ←$ Z[λ]q [.] Let Ls be the following distribution over Z[λ]q _[×][Z][p][: Choose]_ **a ←$ Z[λ]q** [and output][ (][a][,][ ⌊⟨][a][,][ s][⟩⌋][p][)][. The (decision) LWR] problem then is to distinguish between the distribution _Ls and the uniform distribution over Z[λ]q_ _[×][ Z][p][.]_ We will use a key-homomorphic pseudorandom function that is based on the LWR problem to instantiate our scheme. Next we define pseudorandom functions and key-homomorphic pseudorandom functions. ### 2.3 Pseudorandom Functions Intuitively, a pseudorandom function (PRF) is a function that is indistinguishable from a random function (RF). A random function is a function that returns truly random values on all distinct inputs. We use pseudorandom functions PRFk : X →Y that are indexed by a key _k ∈K. For a PPT-adversary A, we define A’s advantage_ in distinguishing a pseudorandom function PRF from a random function as Adv[prf]A,PRF[(][λ][)] :=|Pr[Exp[PRF](λ, A) = 1] − Pr[Exp[RF](λ, A) = 1]|, where the experiments Exp[PRF](λ, A), and Exp[RF](λ, A) are defined as follows: Exp[PRF](λ, A) Exp[RF](λ, A) 1 : _k ←$ K_ 1 : _k ←$ K_ 2 : _b ←A[PRF][k][(][·][)]_ 2 : _b ←A[RF][(][·][)]_ 3 : **return b** 3 : **return b** In the first case, has oracle access to PRF indexed _A_ by a random key k, whereas in the second case A has oracle access to a random function. Intuitively, ’s goal _A_ can be seen as finding out whether they are in Exp[PRF] or Exp[RF]. A pseudorandom-function is computationally _indistinguishable from a random function, if for all PPT-_ adversaries, there exists a negligible function negl such _A_ that for all sufficiently large λ it holds that Adv[prf]A,PRF[(][λ][)][ ≤] [negl][(][λ][)][.] **2.3.1 Key-Homomorphic Pseudorandom Functions** A useful special case of pseudorandom functions are keyhomomorphic pseudorandom functions. A pseudorandom function PRFk : X →Y is key-homomorphic, if for all x ∈X, PRF(·)(x) is a group homomorphism between the key space K and Y. To define this formally, let (K, ∗) and (Y, •) be groups. Then, for all x ∈X _, k1, k2 ∈K_ PRFk1 (x) • PRFk2 (x) = PRFk1∗k2 (x) must hold. A key-homomorphic PRF must fulfill the same security definition as a PRF. A PRF is _almost_ key-homomorphic if PRFk1+k2 (x) = PRFk1 (x) + PRFk2 (x) + e for a small _e ∈_ N. The PRF we use as main building block in our PSA scheme is almost key-homomorphic with e ∈{0, 1}. ### 2.4 Private Stream Aggregation In private stream aggregation (PSA) we have an aggregator and several clients. In each time step, the clients send an encrypted value to the aggregator. The aggregator is then able to compute the sum of these values but no individual client value. It is important that the aggregator can only compute the sum of values that were encrypted under the same time-stamp or label. There is no further interaction beyond the messages that the clients send to the aggregator. Often in PSA differential privacy is additionally used. For this, every client adds a small amount of noise ----- to their value before encrypting it. The aggregator can then compute the resulting noisy sum of the plaintexts. When the noise is chosen appropriately, and enough clients honestly add noise, the noisy sum maintains differential privacy. However, in this paper we are only concerned with the encryption and will leave out the noise in the definition of PSA. Nevertheless, it is no problem to add differential privacy via standard techniques (e.g. as in [22]). Our definition roughly follows the definition of [10], as it is also without noise. **Definition 2 (Private Stream Aggregation). A private** _stream aggregation scheme PSA over ZR (for R ∈_ N) and label space, consists of the following three PPT algo_L_ rithms for the setup, the encryption and the decryption of the aggregate sum: – Setup(1[λ], 1[n]): Given the security parameter λ and the number of users n in unary, it outputs public parameters pp and n + 1 keys (ki)i∈[n]0 . The key k0 is the (secret) key of the aggregator, and each ki is a (secret) key of a user i ∈ [n]. – Enc(pp, ki, l, xi): Given the public parameters pp, a key ki of user i ∈ [n], a label l ∈L and a value _xi_ ZR, it outputs an encryption ci of xi under _∈_ key ki with label l. This algorithm is supposed to be executed by each user at every time step, where the time step is used as label. The user then sends _ci to the aggregator._ – AggrDec(pp, k0, l, {ci}i∈[n]): Given the public parameters pp, the aggregator’s key k0, a label l ∈L, and a set of n ciphertexts {ci}i∈[n] that were encrypted under the same label l, it outputs [∑]i∈[n] _[x][i][ (mod][ R][)][.]_ We additionally require PSA = (Setup, Enc, AggrDec) to satisfy _correctness,_ i.e. that for any _n,_ _λ_ _∈_ N, x1, . . ., xn _∈_ ZR and any label l _∈_ _L, that_ for (pp, {ki}i∈[n]0 ) _←_ Setup(1[λ], 1[n]), and _ci_ _←_ Enc(pp, ki, l, xi), we have #### ∑ AggrDec(pp, k0, l, {ci}i∈[n]) = _xi mod R._ _i∈[n]_ In most PSA schemes (including ours) the sum is computed modulo a public integer R. When the goal is to compute the sum over Z instead of ZR then the clients must be restricted to only encrypt values smaller than a certain value ω and R must be chosen to be greater than n · ω. This difference can be important, because some proofs only go through, when the message space is a group. However, in our proofs it makes no difference whether the clients are allowed to encrypt values from ZR or {0, . . ., ω}. Usually in a PSA scheme, a trusted third party executes the setup algorithm and gives the secret keys to the clients and the aggregator. The clients then regularly encrypt some value and send the ciphertext to the aggregator. By calling AggrDec the aggregator is then able to decrypt the sum of the values. In Section 5.1 we will describe approaches how the trusted setup can be avoided. Next we define the security notion of aggregator obliviousness. We only define encrypt-once security, which is security in the case that every client encrypts only one message per label. This is a reasonable restriction, because it can be easily enforced in practice. Furthermore, encrypting two messages per label leaks the difference of the messages as explained in Section 2.4.1. The PSA schemes of [22] and [10] both have this restriction as well. **Definition 3 (Aggregator obliviousness).** The gamebased security notion of aggregator obliviousness (AO) is defined via the following security experiment AOb(λ, n, A), b ∈{0, 1} given in Figure 1. First, the challenger runs Setup and passes the public parameters pp to the adversary . Then, can adaptively ask queries _A_ _A_ to the following oracles: QEnc(i, xi, l): Given a user index i ∈ [n], a value xi ∈ ZR, and a label l, it answers with c = Enc(pp, ki, l, xi). QCorrupt(i): Given a user index i ∈ [n]0 (including the aggregator’s index 0), it returns the secret key ki. QChallenge(U, {x[0]i _[}][i][∈U]_ _[,][ {][x]i[1][}][i][∈U]_ _[, l][∗][)][: The adversary]_ specifies a set of users U ⊆ [n], a label l[∗] and two challenge messages for each user from . The ora_U_ cle answers with encryptions of x[b]i [, that is][ {][c][i] _←_ Enc(pp, ki, l[∗], x[b]i [)][}][i][∈U] [. This oracle can only be queried] once during the game. (If it is not queried, we set = _U_ _∅_ in the discussion below.) At the end, A outputs a guess α, of whether b = 0 or b = 1. AOb(λ, n, A) (pp, {ki}i∈[n]0 ) ← Setup(1[λ], 1[n]) _α ←A[QCor][(][·][)][,][QEnc][(][·][,][·][,][·][)][,][QChallenge][(][·][,][·][,][·][,][·][)](pp)_ ``` if condition (∗) is satisfied (see p. 123) output α else output 0 ``` **Fig. 1. Aggregator obliviousness experiment for PSA schemes.** Depending on the bit b, the oracle QChallenge answers with encryptions of x[0]i [or][ x]i[1][.] ----- To formally define the condition ( ), we introduce the _∗_ following sets: – Let El ⊆ [n] be the set of all users for which A has asked an encryption query on label l. – Let CS ⊆ [n] be the set of users for which A has asked a corruption query. Even if the aggregator is corrupted, we define this set to only contain the corrupted users and not the aggregator. – Let Ql∗ := U ∪El∗ be the set of users for which A asked a challenge or encryption query on label l[∗]. We say that condition ( ) is satisfied (as used in Fig_∗_ ure 1), if all of the following conditions are satisfied: – = . This means that all users for which re_U ∩CS_ _∅_ _A_ ceives a challenge ciphertext must stay uncorrupted during the entire game. – A has not queried QEnc(i, xi, l) twice for the same #### (i, l). Doing so would violate the encrypt-once restriction. – U ∩El∗ = ∅. This means that A is not allowed to get a challenge ciphertext from users for which they ask an encryption query on the challenge label l[∗]. Doing this would violate the encrypt-once restriction. – If A has corrupted the aggregator and Ql∗ _∪CS = [n]_ then we require that #### ∑ ∑ _x[0]i_ [=] _x[1]i_ _[.]_ _i∈U_ _i∈U_ We will call this condition the balance-condition. The balance condition captures the fact that if has _A_ corrupted the aggregator and received a ciphertext from every uncorrupted user, then they can compute the sum of the plaintexts. If the plaintexts submitted in the challenge query would sum to different values, then could _A_ trivially win the game by using their aggregation capability. Note that the balance-condition does not apply if there is a single honest user for which did not get a _A_ ciphertext on label l[∗]. We say that corruptions are adaptive, because can _A_ ask corruption queries depending on previously asked queries. If has to decide at the beginning of the game _A_ which users they want to corrupt, the term static corruptions is used in the literature. In this paper we only consider adaptive corruptions, because it is a more realistic assumption and because security under adaptive corruptions implies security under static corruptions. We define ’s advantage as _A_ Adv[AO]A,PSA[(][λ, n][) =][|][Pr[][AO][0][(][λ, n,][ A][) = 1]] _−_ Pr[AO1(λ, n, A) = 1]|. A PSA scheme is aggregator oblivious, if for every PPT adversary there is a negligible function negl such that _A_ for all sufficiently large λ Adv[AO]A,PSA[(][λ, n][)][ ≤] [negl][(][λ][)][.] **2.4.1 Inherent Leakage of Sum Queries** Here we will briefly explain why it is dangerous in a PSA scheme, when a client encrypts more than one message per label, even though, the scheme may be formally secure. Imagine that user i ∈ [n] encrypts both xi and x[′]i as ciphertexts ci and c[′]i[, respectively, with the same la-] bel l. When the aggregator got ciphertexts for the same label l from the other users as well, they can use AggrDec to compute AggrDec(pp, k0, l, (c1, . . ., ci, . . ., cn)) _−AggrDec(pp, k0, l, (c1, . . ., c[′]i[, . . ., c][n][))]_ #### ∑ ∑ = ( _xj) + xi −_ ( _xj) −_ _x[′]i_ [=][ x][i] _[−]_ _[x]i[′]_ _[.]_ _j∈[n]\{i}_ _j∈[n]\{i}_ With this, the aggregator learns the difference of the two messages of user i. It also means that if the aggregator knows one of the two messages, they can compute the other one. If the aggregator has two ciphertexts from more than one client, then they can combine them in arbitrary ways to get even more information. This leakage cannot be avoided, because it is leaked by the sum functionality itself. This is also a reason why, in this paper, we restrict the clients to only encrypt one message per label (encrypt-once). ## 3 Adaptively Secure PSA In this section, we construct a scheme for private stream aggregation and prove that it is aggregator oblivious under adaptive corruptions in the standard model. We will define the scheme without noise. The noise can be added via standard techniques (e.g. as in [22]), to ensure differential privacy. In the security proofs, we use diagrams to illustrate the game hops. Figure 2 shows how to read these diagrams. In this example there are the four games G0 to G3. In game G0, only the unmodified lines are executed, that is the lines which are neither framed nor gray. Thus, in G0 only line 1 is executed. Game G1 additionally executes the lines that are framed by a rectangular box, but that are not gray. In our example, G1 executes ----- **Fig. 2. Figure showing how to read the game hop diagrams.** lines 1 and 2. Game G2 executes all unmodified lines, all framed lines and all gray lines, which in this case are all four lines. In game G3, only the unmodified lines are executed and the lines that are gray, but not framed. Therefore, in G3 the lines 1 and 4 are executed. ### 3.1 The Construction Our scheme makes use of a key-homomorphic PRF to create pseudorandom pads which are added to the messages as encryption. Let PRFk : X _→_ _Y_ be the key-homomorphic PRF, where the key spaces (K, +) and (Y, +) are abelian groups. Thus, we have that for all x _∈_ _X_ _,_ [∑]i [PRF][k][i] [(][x][) =][ PRF][∑]i _[k][i]_ [(][x][)][ holds. For our use we] require that (Y, +) is the group (ZR, +), for some integer R. In Section 4, we describe how to instantiate the scheme with a lattice-based key-homomorphic PRF. Throughout this section we will often write [∑]i∈[n] _[x][i]_ and omit the mod R, when it is clear from the context. That is, [∑]i∈[n] _[x][i][ denotes the sum in][ Z][R][ and not the]_ sum in Z. Also we will write [∑]i∈[n] _[k][i][, with which we]_ mean the sum in the group . _K_ We define the PSA scheme PSA = #### (Setup, Enc, AggrDec) in Figure 3. The setup algorithm chooses n random keys for the key-homomorphic PRF and defines the aggregation key as the sum of the client keys. The encryption algorithm uses the keyhomomorphic PRF to create a pseudorandom pad and adds it to the message modulo the public modulus R. The decryption algorithm sums together all ciphertexts which yields the sum of the client values plus the sum of the pseudorandom pads. Because the key of the aggregator is the sum of the client keys, the aggregator can compute the sum of the pseudorandom pads and subtract it from the ciphertexts’ sum to obtain the sum of the plaintexts. In Section 3.2, we show that this construction is aggregator oblivious under adaptive corruptions if the key homomorphic PRF is indistinguishable from a random function. If the PRF is secure in the standard model, then our construction is also secure in the standard model. Setup(1[λ], 1[n]): `for i ∈` [n]: ki ←$ K _k0 :=_ [∑]i∈[n] _[k][i]_ pp := R (the modulus) `return (pp, {ki}i∈[n]0` ) Enc(pp, ki, l, xi): _ti,l := PRFki_ (l) _ci,l := xi + ti,l mod R_ ``` return ci,l ``` AggrDec(pp, l, {ci,l}i∈[n]): _t0,l = PRFk0_ (l) `return` [∑]i∈[n] _[c][i,l][ −]_ _[t][0][,l][ mod][ R]_ **Fig. 3. The PSA scheme that uses key-homomorphic PRFs.** Next, we show that PSA is correct. Note that because PRF is key-homomorphic, we have #### ∑ _i∈[n]_ [PRF][k][i] [(][l][) =][ PRF][∑]i∈[n] _[k][i]_ [(][l][) =][ PRF][k][0] [(][l][)][.] **Correctness: Let n, λ ∈** N, (pp, {ki}i∈[n]0 ) ← Setup(1[λ], 1[n]), l ∈L, xi ∈ ZR, ci,l ← Enc(pp, ki, l, xi). Then we have AggrDec(pp, k0, {ci,l}i∈[n], l) = [∑]i∈[n] _[c][i,l][ −]_ [PRF][k][0] [(][l][) mod][ R] = [∑]i∈[n][(][x][i][ +][ PRF][k][i] [(][l][))][ −] [PRF][k][0] [(][l][) mod][ R] = [∑]i∈[n] _[x][i][ +][ ∑]i∈[n]_ [PRF][k][i] [(][l][)][ −] [PRF][k][0] [(][l][) mod][ R] = [∑]i∈[n] _[x][i][ +][ PRF][∑]i∈[n]_ _[k][i]_ [(][l][)][ −] [PRF][k][0] [(][l][) mod][ R] = [∑]i∈[n] _[x][i][ mod][ R]_ If [∑]i∈[n] _[x][i][ < R][, then we get the sum over the integers]_ #### ∑ _i∈[n]_ _[x][i][ as result.]_ ### 3.2 Security Proof In this section, we prove the aggregator obliviousness of the above scheme. For this, we follow the proof strategy ----- of [1], who used this strategy to show the security of an inner-product MCFE scheme.[2] In the proof of this theorem we show that PSA is aggregator oblivious, if the key-homomorphic PRF is indistinguishable from a random function. **Theorem 1. For any PPT adversary** on the aggre_A_ gator obliviousness game, there is a PPT adversary _B_ on the PRF such that Adv[AO]A,PSA[(][λ][)] _≤_ 2(4n[2](n − 1) + 2n(n − 1) + 2n[2]) · Adv[prf]B,PRF[(][λ][)] _≤_ (8n[3] + 8n[2]) · Adv[prf]B,PRF[(][λ][)][,] where n is the number of users. The adversary B has roughly the same running time as . _A_ _Proof. We use four intermediate games to go from AO0_ to AO1. A description of the games is depicted in Figure 4. We provide the lemmas for the transition between the games in Appendix A. **Game G0: This is the AO0 game, in which the chal-** lenge query is answered with encryptions of x[0]i [.] **Game G1: This game still answers the challenge** query with encryptions of x[0]i [, but changes the pseudo-] random pads that are used for the encryption. For correct decryption, these changes must not affect the sum of the pseudorandom pads. Therefore, to each pseudorandom pad, we add a share of a perfect η-out-of-η secret sharing of zero, where η is the number of users in the challenge query. This makes the ciphertexts useless, unless they are all summed together. This fact enables us to make the change of the next game. **Game G2: This game answers the challenge query** with encryptions of x[1]i [instead of][ x]i[0][. This is possible] because the secret shares of the previous game hide all information on the individual ciphertexts. **Game G3: Here we remove the secret shares from** the pseudorandom pads again. Therefore, this game is **2 The definition of the games G0 to G3 is very similar to theirs.** However, there are differences in the proof. Because in the game of aggregator obliviousness, the adversary is only allowed to ask one challenge query, the games do not have to guess the number of honest users. Also we need one game less than [1], because we do not have a layer of functional encryption. Because we use keyhomomorphic PRFs instead of general PRFs, the transition from _G0 to G1 is also different. Lastly, we directly prove aggregator_ obliviousness without relying on lemmas to upgrade the security. Doing so adds an extra case distinction to the proof, but reduces the reduction loss. This is possible, because PSA is a simpler primitive than MCFE. **Fig. 4. Game hops of the proof of Theorem 1.** identical to AO1, in which the challenge query is answered with encryptions of x[1]i [.] We distinguish two cases. In the first case the adversary corrupts the aggregator. This is the more chal_A_ lenging case, because it allows the adversary to decrypt the sum of ciphertexts. Care must be taken to introduce the changes for the games in a way that cannot be recognized by the adversary. In Lemma 1, we use a hybrid argument over all users and several intermediate games to show that _| Pr[G0(λ, A) = 1] −_ Pr[G1(λ, A) = 1]| _≤_ 2n[2](n − 1) · Adv[prf]B,PRF[(][λ][)][.] To get from G1 to G2 we show in Lemma 2 that _| Pr[G1(λ, A) = 1] −_ Pr[G2(λ, A) = 1]| _≤_ 2n(n − 1) · Adv[prf]B,PRF[(][λ][)][.] ----- Finally, to get from G2 to G3, we apply Lemma 3 in which we show that _| Pr[G2(λ, A) = 1] −_ Pr[G3(λ, A) = 1]| _≤_ 2n[2](n − 1) · Adv[prf]B,PRF[(][λ][)][.] In the second case, does not corrupt the aggregator. _A_ This enables us to directly go from G0 to G3 by a hybrid argument over all users. Thus, in Lemma 4 we show that _| Pr[G0(λ, A) = 1] −_ Pr[G3(λ, A) = 1]| _≤_ 2n[2] _· Adv[prf]B,PRF[(][λ][)][.]_ The reduction uses an unbiased coin to decide whether _B_ to simulate case 1 or case 2, so in conclusion we get Adv[AO]A,PSA[(][λ][)] _≤_ 2(4n[2](n − 1) + 2n(n − 1) + 2n[2]) · Adv[prf]B,PRF[(][λ][)] _≤(8n[3]_ + 8n[2]) · Adv[prf]B,PRF[(][λ][)][.] In this section, we proposed a PSA scheme that is based on key-homomorphic PRFs. We proved that the scheme is aggregator oblivious in the standard model. In the next section, we describe how to instantiate the scheme with a lattice-based PRF and explain our implementation. ## 4 Implementation In this section, we describe the implementation of our scheme, the choice of parameters and performance results. The implementation in Go can be found here [https://github.com/johanernst/khPRF-PSA. Both the en-](https://github.com/johanernst/khPRF-PSA) cryption and the decryption algorithm are fast, so that they can also be executed on computationally limited devices such as smart-meters. The setup algorithm is slower, because for our parameters it needs to draw _λ = 2096 random numbers per client. However it is only_ executed very rarely. ### 4.1 Choice of the Pseudorandom Function For the implementation, we chose to use an almost keyhomomorphic PRF mentioned in [12]. It relies on the LWR assumption and is secure in the random oracle model. Therefore, with this concrete instantiation our scheme is also only secure in the ROM. We chose a ROM-based PRF for its simplicity and efficiency. The standard model PRF of Boneh et al. [12] requires quite large parameters to be secure. The public parameters are two λ[′] _× λ[′]_ matrices. Because the matrices are sampled from {0, 1}[λ][′][×][λ][′] instead of Z[λ]q _[′][×][λ][′]_, the dimension of the matrices must be increased by a factor of log2(q). Since we use λ = 2096 and q = 2[128], this means that _λ[′]_ = 2096 · 128, thus the square matrix would have _λ[′][2]_ _≈_ 7 · 10[10] entries. Even if each entry is stored as single bit, these are 70 gigabits. Also, for evaluating the PRF, the matrices need to be multiplied together multiple times, whereby the intermediate entries which need to be kept in memory get much larger. Thus, this PRF does not seem practical because of both running time and memory constraints. While the key-homomorphic PRF of Banerjee and Peikert [6] is more efficient, it is also more complex and thus, more prone to implementation errors, which endangers the security of the implementation. Kim proposes an approach that allows a smaller modulus q at the cost of larger keys [21]. Since we need q > p = 2[85] for a message space of 2[64], we would not gain very much from a smaller modulus. When our scheme is instantiated with any of the above mentioned standard model key-homomorphic PRFs, then we obtain a PSA scheme that is secure in the standard model. Next, we describe the ROM-based keyhomomorphic PRF of [12], which we use in the following. For λ, q, p ∈ N, with q > p, k ∈ Z[λ]q [and a hash function] _H : X ↦→_ Z[λ]q [, the PRF is defined as:] _Fk(x) := ⌊⟨H(x), k⟩⌋p._ Because in [12] there is no security proof for this function, we provide a short proof in Appendix B. Because the output of F is from Zp, we set the public modulus _R in our scheme equal to p._ The hash function H is required to map to Z[λ]q [. We] construct such a hash function using a standard hash function H _[′], such as SHA3, as follows:_ where byte converts its argument to a byte array. The space between x and i ∈{1, . . ., λ} is necessary to ensure that all inputs to H _[′]_ are different. Note that if q does not divide the size of the output space of H _[′], extra analysis_ is needed to make sure that the mod q operation does not induce any bias. However in our case we choose q as power of 2, whereby it divides the size of the output space of H _[′]. In our implementation we used SHA3-512_ as H _[′]._ _H_ _[′](byte(x ‖ " " ‖ "1")) mod q_ ... _H_ _[′](byte(x ‖ " " ‖ "λ")) mod q_ ####   (1) [,] _H(x) :=_ ####    ----- The rounding function ⌊a⌋ is not exactly linear, but almost linear, which means that: _⌊a + b⌋_ = ⌊a⌋ + ⌊b⌋ + e, for e ∈{0, 1}. This entails that the PRF is only almost key-homomorphic: _Fk1+k2_ (x) = ⌊⟨H(x), k1 + k2⟩⌋p = ⌊⟨H(x), k1⟩ + ⟨H(x), k2⟩⌋p = ⌊⟨H(x), k1⟩⌋p + ⌊⟨H(x), k2⟩⌋p + e = Fk1 (x) + Fk2 (x) + e for e ∈{0, 1}. For our use-case this is not a problem. Because in the decryption algorithm the PRF values of _n clients are summed, the error from the non-linearity_ is at most n − 1. The idea is to use a larger message space where all legitimate messages have a difference larger than n. The decrypted message that potentially contains an error of up to n − 1 is then rounded to the next legitimate message. In Section 4.3, we describe this in more detail. ### 4.2 Choice of Parameters In this section, we describe how we chose the parameters for the PRF and approximately which security level we get from these parameters. The PRF is parameterized by λ, q, p ∈ N and reduces tightly to LWRλ,q,p. We used the LWE-estimator from [5] to estimate the security level for certain choices of parameters. The value of 1/p corresponds to the error rate α in LWE, so we need to choose λ, q, p ∈ N such that LWEλ,q,1/p is hard. For λ = 2096, q = 2[128] and p = 2[85], the program estimates a hardness of over 2[178]. Note that these parameters also satisfy the recommendation of [7] _√_ that q/p > _λ._ According to Theorem 1, the reduction loss is less than 8n[3] +8n[2]. When we suppose we have n = 2[20] users, then the reduction loss is less than 2[64]. This yields a security of 178 64 = 114 bit.[3] _−_ ### 4.3 Concrete Instantiation As described in the previous section, we set λ := 2096, _q := 2[128]_ and p := 2[85]. In our implementation, we set **3 For the 10000 users in our implementation the security is at** least 132 bit the public modulus R = p = 2[85] and PRF := Fk(x) = _⌊⟨H(x), k⟩⌋p with key space Z[λ]q_ [and][ H][ as defined in (1).] As labels we use strings that are converted to byte arrays before given to the hash function. Next, we describe how to mitigate the error introduced by the non-linearity of the rounding function: Because [∑]i∈[n] **[k][i][ =][ k][0][, we have]** #### ∑ _Fki_ (l) + e mod R = Fk0 (l), _i∈[n]_ for e ∈{0, . . ., n − 1}. This means that #### ∑ ∑ (xi + Fki (l)) − _Fk0_ (l) mod R = _xi −_ _e mod R._ _i∈[n]_ _i∈[n]_ To ensure correctness, each client, before calling Enc(pp, ki, l, x[′]i[)][, computes][ x][′]i [:][=][ n][ ·][ x][i][ + 1][. The multi-] plication with n ensures that all legitimate messages are apart by n − 1 and the addition of 1 ensures that the non-linearity error does not cause an underflow mod R. The aggregator, after executing ¯s = AggrDec(pp, k0, l, {ci}i∈[n]), rounds ¯s up to the next multiple s[′] of n and computes s := (s[′] _−_ _n)/n._ **Correctness: After encryption we have** _ci = n · xi + 1 + Fki_ (l) mod R. AggrDec(pp, k0, l, {ci}i∈[n]) yields _s¯ =_ [∑]i∈[n][(][n][ ·][ x][i][ + 1 +][ F][k]i [(][l][))][ −] _[F][k]0_ [(][l][) mod][ R] = [∑]i∈[n][(][n][ ·][ x][i][ + 1)][ −] _[e][ mod][ R]_ = n · [∑]i∈[n] _[x][i][ +][ n][ −]_ _[e][ mod][ R]_ = n · [∑]i∈[n] _[x][i][ +][ n][ −]_ _[e.]_ For the last equality to hold, we need that 0 ≤ _n ·_ [∑]i∈[n] _[x][i][ +][ n][ −]_ _[e < R,]_ which means that e neither creates an underflow nor an overflow modR. Because e is at most n − 1, we have #### 0 < n · [∑]i∈[n] [x][i][ +][ n][ −] [e][ and if][ ∑]i∈[n] [x][i][ <][ (][R][ −] [n][)][/n][,] then we also have #### ∑ _n ·_ _xi + n −_ _e < R._ _i∈[n]_ When we suppose we have at most n = 2[20] users, this means that [∑]i∈[n] _[x][i][ must be smaller than][ (][R][ −]_ 2[20])/2[20] = 2[85]/2[20] _−1 = 2[65]_ _−1. In the next step, the ag-_ gregator rounds ¯s up to the next multiple of n, which is _s[′]_ = n·[∑]i∈[n] _[x][i][+][n][. After computing][ s][ = (][s][′][−][n][)][/n][, they]_ get s = [∑]i∈[n] _[x][i][, which is the desired result. Therefore,]_ ----- if the total sum is at most 2[64], the scheme works correctly. **Security: Because the clients input n · xi + 1 to the** encryption algorithm of the PSA scheme, Theorem 1 guarantees that only [∑]i∈[n][(][n][ ·][ x][i][ + 1)][ mod][ R][ can be] computed by the aggregator. Because n is known publicly, this value does not contain more information than #### ∑ _i∈[n]_ _[x][i][. Thus, our instantiation inherits the security]_ of the general scheme. ### 4.4 Performance In this section, we analyze the performance of our scheme both in theory and by performing running time measurements. Every secret key consists of λ = 2096 elements from Zq = Z2128, which means that every secret key needs 2096 128 = 268288 bits of memory. These are _·_ roughly 33.5 kilobyte. For a security level of 128 bit, Benhamouda et al. [10] report a key size of 592 bit for their scheme and of 416 bit for the scheme of Shi et al. [22]. The size of a ciphertext in our scheme is log(p) = 85 bit. Again for a security level of 128 bit, Benhamouda et al. [10] report a ciphertext size of 296 bit for their PSA scheme and of 416 bit for the scheme of [22]. Both values were computed for 2[20] users and 2[20] labels. Neither Takeshita et al. [23] nor Waldner et al. [25] report their key sizes, however the key size of [25] increases linearly with the number of clients, because every client needs a shared secret with every other client. In many cases, smaller ciphertexts are preferable over small keys, because the ciphertexts have to be sent over the network at every time-step. In our scheme the cost for encryption mainly is the cost for evaluating the PRF. The PRF needs 2096 evaluations of the underlying hash function, #### 2096 modulo operations and 2096 additions and multiplications for evaluating the inner product of the hash and the key. We executed the performance tests on a laptop on a single thread of an Intel Core i5-10210U CPU. We measured the running time of both our scheme and [25]. For a better comparison we executed their scheme with a message space of 2[64] and without noise. We executed the tests 40 times and took the average. In every test we run the encryption algorithm once for every client and executed the decryption algorithm 1000 times. Figure 5 shows the average running time of a single execution of the encryption and decryption algorithm of both our scheme and [25]. As expected the running time for the encryption algorithm of our scheme does not depend on the number of users, whereas the running time of [25] grows linearly. Somewhat surprisingly the running time of our decryption algorithm increases with the number of users. This means that the running time is not completely dominated by the cost of evaluating the PRF, but summing together all ciphertexts also takes significant time. As the figure clearly shows, our scheme outperforms [25] starting from about 3500 clients. Table 2 shows the exact numbers of the average running time of our scheme and [25] for different numbers of users. Table 3 shows the running time of [23] and [9] taken from the respective papers. In both, our scheme and [25], the evaluation of the PRF does not depend on the plaintext. Therefore, encryption could be sped up by computing the PRF beforehand. Then, when the plaintext is available, encryption only consists of adding the PRF output to the plaintext and one modulo operation. The same can be done for decryption as well. **Our scheme** **LaSS (AES variant)** Users Enc Dec Enc Dec 1000 0.913(0.010) 0.875(0.002) 0.295(0.011) 0.277(0.007) 5000 0.929(0.007) 1.209(0.006) 1.590(0.022) 1.508(0.042) 10000 0.901(0.004) 1.805(0.007) 3.941(0.046) 3.643(0.201) **Table 2. Running time in milliseconds of one execution of the** encryption/decryption algorithm of our scheme and the AES version of LaSS [25]. The value in parentheses is the standard deviation. For both schemes we executed 40 measurements and took the average. **SLAPBGV** **LaPS** Users Enc Dec Enc Dec 1000 1.17 3.26 3.724 1.964 **Table 3. Running time in milliseconds of the optimized version** of SLAPBGV [23] and of LaPS [9]. This version of LaPS only provides a security of 80 bit. For a security of 128 bit Becker et al. [9] report a running time of 77.33ms for encryption and 67.62 for decryption. Both implementations were measured with a message space of size 2[16]. Note that the numbers are taken from the respective papers. Thus, a comparison is not entirely reliable. The setup algorithm of our scheme has to draw 2096 random elements from Zq for each user and then compute the sum of the user keys to get the key for the aggregator. As shown in Figure 6, the running time of |Col1|Our scheme|LaSS (AES variant)| |---|---|---| |Users|Enc Dec|Enc Dec| |1000|0.913(0.010) 0.875(0.002)|0.295(0.011) 0.277(0.007)| |5000|0.929(0.007) 1.209(0.006)|1.590(0.022) 1.508(0.042)| |10000|0.901(0.004) 1.805(0.007)|3.941(0.046) 3.643(0.201)| |Col1|SLAPBGV|LaPS| |---|---|---| |Users|Enc Dec|Enc Dec| |1000|1.17 3.26|3.724 1.964| ----- our setup algorithm grows linearly in the number of users. The running time of [25] grows quadratically with the number of users, because every pair of users needs a shared key. The generation of the random numbers can trivially be parallelized and thereby be made much faster in practice. Also, since the setup algorithm is executed very rarely, its running time is not as critical as the running time of the encryption or decryption algorithm. **Fig. 5. This figure shows the running time of one execution of** the encryption and decryption algorithm of both our scheme and [25]. The vertical bars show the standard deviation, which is large enough to be seen only for the decryption algorithm of LaSS. **Fig. 6. This figure shows the running time of the setup algorithm** of both our scheme and LaSS. As in Figure 5 the error bars are too small to be seen. ## 5 Deployment Considerations In this section, we discuss practical issues when deploying our scheme, with a special focus on the smart-meter application. ### 5.1 Setup and Key Management In the PSA literature the setup procedure is usually considered to be executed by a trusted party who distributes secret keys to the clients and the aggregator. This often means that the trusted party is able to decrypt all messages sent by the clients. In the following, we discuss techniques to overcome this limitation. One approach to achieve a decentralized setup is to use techniques very similar to the ones of Chotard et al. [17]. The idea is that we let each client choose their PRF key ki at random. The aggregator’s key is supposed to be the sum of the keys chosen by the clients. So we only need to let the aggregator know the sum of the client keys in a secure way. The solution for this is to combine non-interactive key exchange (NIKE) ([13], [20]) with a technique from [15] as done in [17]. First the clients execute the NIKE, i.e. each client generates a public key and a secret key and uploads the public key to a key server. Each client downloads the public keys of all other clients and uses each other client’s public key together with their own secret key to derive a shared secret with that client. Then the clients use the shared pairwise keys to generate random pads with the property that the pads sum to zero. The clients then add these pads to their PSA secret keys and send the resulting ciphertext to the aggregator. The aggregator will obtain the sum of the client keys by adding all ciphertexts, but learns nothing else about the keys. The process is essentially the same as the DSum functionality of [17], but without the layer of All-or-Nothing Encapsulation and is shown in Figure 7. In principle, one can use the same key-homomorphic PRF as in our PSA scheme. However, the property of key-homomorphism is not needed here. Hence, we recommend using a block cipher such as AES as PRF. One may now argue, that we do not need the rest of the PSA scheme anymore, because we already have a way of letting the aggregator know the sum of client values. Indeed this would basically give us the PSA scheme of [25]. However then the encryption of each value requires n invocations of the PRF, where n is the number of users. Doing this for every encryption becomes ----- **Fig. 7. The decentralized setup algorithm as executed by every** client. The algorithm takes as input the client’s PSA secret key _ki, the NIKE secret key ski and the NIKE public keys of the_ other clients pkj . inefficient as the number of clients becomes large. So it is preferable to only execute this step once for the decentralized setup and then continue with our PSA scheme that only requires one invocation of the keyhomomorphic PRF in the encryption algorithm. For illustration, let us present a simple example of how a NIKE scheme can be built from the Diffie– Hellman key exchange. Every client i publishes their public key pki := g[x][i] and downloads the public keys of all other clients. To compute a shared key with client j, client i computes pk[x]j _[i]_ = g[x][j] _[x][i]_ and hashes this together with their identities H(i, j, g[x][j] _[x][i]_ ). For more details on constructions and security models see [13] and [20]. The most efficient way to distribute the public keys is to use a key server which all clients use to upload and download their public keys. The key server needs to be semi-honest, i.e. it can share all its knowledge with the adversary without compromising security, but must follow the protocol. A malicious key server could perform a person-in-the-middle attack by replacing all client public keys with its own public keys. The key server would then compute a shared secret with every client and would be able to decrypt all messages. However such an attack can be detected when clients manually compare their keys with each other. The communication costs for the decentralized setup of one client is uploading their NIKE public-key to the key-server, downloading n − 1 public-keys and sending one aggregation key share (the encrypted PSA secret key) to the aggregator. The computational costs are n − 1 calls to NIKE.SharedKey to compute the pairwise shared keys and n − 1 PRF or AES evaluations to compute the pseudorandom pad that encrypts the PSA secret key. These operations only have to be executed for the setup and have no influence on the cost for encryption, which is still independent of the number of clients. An alternative to relying on a key server would be to let the clients broadcast their public key to all other clients. However, here a person-in-the-middle attack is also possible. Furthermore this approach would cause roughly n[2] messages in the setup phase, which can be too much in the smart-meter scenario, where the number of clients can be large. Having a decentralized setup as described above has the additional advantage that the setup can be repeated at regular intervals to achieve some sort of forward secrecy. Repeating a centralized setup would mean that the trusted party would have to be available at each time the setup is repeated. Only relying on a semi-honest key server which needs to be online once every interval is much easier to assure. Another approach from literature to decentralize the setup is adhoc multi-input functional encryption (adhoc-MIFE) [4]. In adhoc-MIFE the clients also generate shares of the aggregation key in a decentralized way, by getting as input the public keys of the other parties. The authors consider the computation of innerproducts, which can be seen as a generalization of the computation of sums in our setting. They use 2-round MPC protocols with specific properties in their construction. However their construction does not support labels and is less efficient due to the use of MPC. ### 5.2 Client Failures If one client fails to submit a valid ciphertext, then the aggregator cannot compute the overall sum, because that client’s PRF term is missing. The result will then look like a random value. In the context of smartmetering this is a problem, because there are many devices and it is quite possible that one device fails due to technical problems, loss of network connection or active manipulation by the user. We have several options to cope with such failures. A straightforward approach is to partition the users in several groups and run one instance of the PSA scheme for each group. For example, if we have 1000 users, we can divide them in ten groups of 100 users each. When one user fails, we only lose the values of 100 users, instead of 1000. Two disadvantages of this approach are that a few failing users can cause a lot of lost values and that it reduces the privacy of the users, because their values are only aggregated with a smaller set of other users. To solve this problem, Chan et al. [14] have proposed a generic solution that incorporates differential privacy in an essential way. They let the aggregator learn partial sums of the users’ inputs in such a way ----- that the sum of the values of the non-failing users can always be reconstructed. They use a binary tree, where the clients are the leaf nodes. Each inner node corresponds to the partial sum of the values of the clients beneath that node. So each client produces log(n) ciphertexts of which each one corresponds to an inner node on the path from that client to the root of the tree. When all clients send ciphertexts, then the aggregator uses the ciphertexts corresponding to the root node to compute the total sum. If some clients fail, the aggregator has to use ciphertexts corresponding to the inner nodes to be able to reconstruct the sum of the remaining users as shown in Figure 8. The noise for differential privacy is essential here, because otherwise the aggregator would get all the users’ values in the clear. [0,3] [0,1] [2,3] 0 1 2 3 **Fig. 8. If client 2 fails, the aggregator uses the ciphertexts corre-** sponding to the black nodes to compute the sum of the remaining clients’ values. The notation [i,j] means the (noisy) sum of the values of clients {i,. . .,j}. This figure is a smaller version of Figure 1b from [14]. With this approach each client has log(n) secret keys, one corresponding to each node on the path to the root. Also, instead of sending one message, each client sends log(n) messages. The aggregator holds one aggregation key for each inner node in the tree. Another way to view this is that we are running one instance of the PSA scheme for each inner node. This approach does not add additional rounds of communication. The aggregator will always be able to compute the overall sum, no matter how many clients fail, however, when more clients fail, the resulting sum is noisier. The scheme is generic and does not pose any special requirements on the underlying PSA scheme, so it can be used to make our scheme failure tolerant. Only one small adaption has to be made. The scheme of Chan et al. only considers user values in {0, 1}, whereas in our case the values are from {0, . . ., ∆}, where ∆ is the largest reasonable power consumption. To accommodate for this we need to multiply the ϵ and δ parameters for differential pri vacy with 1/∆. In Appendix C we describe in a bit more detail how the adapted encryption algorithm works and also provide figures with pseudocode. ### 5.3 Client Joins and Leaves If clients leave the system, for example if they changed their power supplier, we can treat them as permanently failed (cf. Section 5.2), as suggested by [14]. This will slightly increase the noise relative to the number of remaining clients. Therefore, when many clients have left, we can repeat the setup for the remaining clients and start over with a new tree. Repeating the setup is more practical when using the decentralized setup described in Section 5.1, because then we do not need to entrust a third party with the key generation. To accommodate for joining clients, [14] propose to create a tree that has more leaves than there are clients, where the trusted party creates secret keys for every leave node. The not-yet-present clients are treated as failed, until they join. When a new client joins, they get the secret keys corresponding to their leave node from the trusted party. This has the advantage that neither the other clients nor the aggregator need to be notified when a client joins. However it has the disadvantage that the trusted party needs to be available whenever a new client joins. In the following we describe how client-joins can work together with the decentralized setup from Section 5.1. Since we are using a binary tree now, we are essentially running one instance of the PSA scheme for every inner node, where the clients of each instance are the clients below the respective inner node. This means that in the beginning, each client creates log(n) aggregation key shares and sends them to the aggregator. Whenever a new client joins they broadcast their public key to the other clients by uploading it to the key server and downloads the public keys of the other clients and computes a shared key with each of them. The other clients download the public key of the new client and use it to compute a shared secret with the new client. Then a new aggregation key is created for each node on the path from the new client to the root. This means that each client, which shares an inner node with the new client, chooses a new secret key for the respective node and sends the aggregation key share to the aggregator. The aggregator receives all the aggregation key shares for each node on the path from the new client to the root and combines them to obtain a completely new aggregation key for all these log(n) nodes. ----- Another way to view this is that whenever a new client joins, (a part of) the decentralized setup is repeated for the inner nodes on the path from the new client to the root. The cost for the new (n + 1)th client is computing a shared secret with the n other clients and computing aggregation key shares for each node on the path to the root. The number of PRF/AES evaluations depends on the position of the new node in the tree, but is at least n (for the aggregation key associated with the root node) and at most 2n. The other clients have to compute one new shared secret. The number of PRF/AES evaluations depends on their position in relation to the new node, but is again at least n and at most 2n. ### 5.4 Concrete Smart-Meter Example In this section we summarize how our PSA scheme together with the above extensions can be used for privacy preserving smart-metering. Every smart-meter is preconfigured with the IP addresses of the key server and the power-supplier, the public key of the powersupplier and with an upper bound on the number of other smart-meter devices that are expected to join the system[4]. Every smart-meter creates a random key for the PSA scheme and a public and secret key for the NIKE scheme. The smart-meters then upload their public keys to the key-server and compute pairwise shared keys as described in Section 5.1. They then compute their key share as described in Figure 7, encrypt it with the power-suppliers public key and send the resulting ciphertext to the power-supplier. The power-supplier is then able to compute the aggregation key for each inner node. This concludes the setup. At every time step (e.g. every fifteen minutes) the smart-meters encrypt their current power consumption and send it to the power-supplier. There are two straightforward ways of choosing the label which is needed as input for the PRF. One possibility is to use a counter that starts with 0 and is incremented at every time step. This has the disadvantage that if a client malfunctions and misses a time step, its counter becomes asynchronous to the counters of the other smart-meters, whereby its ciphertexts cannot be decrypted anymore. The other possibility provides a solution for this. The label can be chosen as the current date and hour and a **4 When more users than expected join the system, then the** (non-interactive) setup can simply be repeated value between 0 and 3 which indicates the quarter of the hour we are currently in. Then every smart-meter can deduce the next label from the current time. The clocks of the smart-meters only need to have the same time up to fifteen minutes precision and the current time can be easily obtained via the network. Note that this ensures that every label is only used once, whereby the encryptonce condition is satisfied and thus, our security proof applies. To accommodate for users joining the system, every smart-meter checks the key server for new public keys at regular intervals such as once every day. If there are new public keys, the clients recompute some of their key-shares as described in Section 5.3 and send them to the aggregator. Failing or leaving smart-meters are also treated as described in Section 5.3. Since we use the generic approach of [14] to provide fault-tolerance, the resulting power-consumption, which is decrypted by the power-supplier, contains a small amount of noise. However as stated by [14] the additive error is only slightly larger than logarithmic in the number of users. The smart-meters can also be configured to repeat the setup at regular intervals to provide better forward secrecy or to reduce the noise, when too many smartmeters have left the system. ## 6 Conclusion PSA is a useful protocol for privately letting an aggregator know the sum of client-supplied values. It has applications from privacy-preserving smart metering to distributed machine learning. In this paper, we proposed a PSA scheme that is based on key-homomorphic PRFs as the only building block. It supports a large message space and scales well for a large number of users. Also it has very small ciphertexts (85 bit in our implementation with a message space of 2[64]). Both encryption and decryption mostly consist of only one PRF evaluation. We proved the security of our scheme in the standard model. Furthermore we implemented our scheme using a lattice-based key-homomorphic PRF in the ROM. We analyzed the performance both in theory and by measuring the running time in practice and thereby showed that the scheme is quite efficient. Moreover we discussed possible solutions for practical issues as how to decentralize the setup and how to accommodate for clients joining or leaving the system. ----- ### Acknowledgments We would like to thank the anonymous reviewers. Alexander Koch was supported by the Competence Center for Applied Security Technology (KASTEL). ## References [1] M. Abdalla, F. Benhamouda, and R. Gay. From singleinput to multi-client inner-product functional encryption. In S. D. Galbraith and S. Moriai, editors, ASIACRYPT 2019, _Proceedings, Part III, volume 11923 of LNCS, pages 552–_ [582. Springer, 2019. 10.1007/978-3-030-34618-8_19.](https://doi.org/10.1007/978-3-030-34618-8_19) [2] M. Abdalla, F. Benhamouda, M. Kohlweiss, and H. Waldner. Decentralizing inner-product functional encryption. In D. Lin and K. Sako, editors, Public-Key Cryptography, PKC 2019, _Proceedings, Part II, volume 11443 of LNCS, pages 128–157._ [Springer, 2019. 10.1007/978-3-030-17259-6_5.](https://doi.org/10.1007/978-3-030-17259-6_5) [3] G. Ács and C. Castelluccia. I have a DREAM! (DiffeRentially privatE smArt Metering). In T. Filler, T. Pevný, S. Craver, and A. D. Ker, editors, Information Hiding, IH 2011, volume [6958 of LNCS, pages 118–132. Springer, 2011. 10.1007/978-](https://doi.org/10.1007/978-3-642-24178-9_9) [3-642-24178-9_9.](https://doi.org/10.1007/978-3-642-24178-9_9) [4] S. Agrawal, M. Clear, O. Frieder, S. Garg, A. O’Neill, and J. Thaler. Ad hoc multi-input functional encryption. In T. Vidick, editor, Innovations in Theoretical Computer Sci_ence Conference, ITCS 2020, volume 151 of LIPIcs, pages_ 40:1–40:41. Schloss Dagstuhl – Leibniz-Zentrum für Infor[matik, 2020. 10.4230/LIPIcs.ITCS.2020.40.](https://doi.org/10.4230/LIPIcs.ITCS.2020.40) [5] M. R. Albrecht, R. Player, and S. Scott. On the concrete hardness of Learning with Errors. Journal of Mathematical _[Cryptology, 9(3):169–203, 2015. 10.1515/jmc-2015-0016.](https://doi.org/10.1515/jmc-2015-0016)_ [6] A. Banerjee and C. Peikert. New and improved keyhomomorphic pseudorandom functions. In J. A. Garay and R. Gennaro, editors, CRYPTO 2014, Proceedings, Part _I, volume 8616 of LNCS, pages 353–370. Springer, 2014._ [10.1007/978-3-662-44371-2_20.](https://doi.org/10.1007/978-3-662-44371-2_20) [7] A. Banerjee, C. Peikert, and A. Rosen. Pseudorandom functions and lattices. In D. Pointcheval and T. Johansson, editors, EUROCRYPT 2012, Proceedings, volume 7237 of _[LNCS, pages 719–737. Springer, 2012. 10.1007/978-3-642-](https://doi.org/10.1007/978-3-642-29011-4_42)_ [29011-4_42.](https://doi.org/10.1007/978-3-642-29011-4_42) [8] D. Becker and J. G. Merchan. Post-quantum secure private stream aggregation, Apr. 21 2020. [URL https://](https://patents.google.com/patent/US10630655B2/en) [patents.google.com/patent/US10630655B2/en. US Patent](https://patents.google.com/patent/US10630655B2/en) 10,630,655. [9] D. Becker, J. Guajardo, and K.-H. Zimmermann. Revisiting private stream aggregation: Lattice-based PSA. In NDSS _[2018. The Internet Society, 2018. URL https://www.ndss-](https://www.ndss-symposium.org/wp-content/uploads/2018/02/ndss2018_02B-3_Becker_paper.pdf)_ [symposium.org/wp-content/uploads/2018/02/ndss2018 02B-](https://www.ndss-symposium.org/wp-content/uploads/2018/02/ndss2018_02B-3_Becker_paper.pdf) [3 Becker paper.pdf.](https://www.ndss-symposium.org/wp-content/uploads/2018/02/ndss2018_02B-3_Becker_paper.pdf) [10] F. Benhamouda, M. Joye, and B. Libert. A new framework for privacy-preserving aggregation of time-series data. ACM _Transactions on Information and System Security (TISSEC),_ [18(3):10:1–10:21, 2016. 10.1145/2873069.](https://doi.org/10.1145/2873069) [11] K. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth. Practical secure aggregation for privacy-preserving machine learning. In B. M. Thuraisingham, D. Evans, T. Malkin, and D. Xu, editors, ACM SIGSAC Conference on Computer _and Communications Security, CCS 2017, pages 1175–1191._ [ACM, 2017. 10.1145/3133956.3133982.](https://doi.org/10.1145/3133956.3133982) [12] D. Boneh, K. Lewi, H. Montgomery, and A. Raghunathan. Key homomorphic prfs and their applications. In R. Canetti and J. A. Garay, editors, CRYPTO 2013, Proceedings, Part _I, volume 8042 of LNCS, pages 410–428. Springer, 2013._ [10.1007/978-3-642-40041-4_23.](https://doi.org/10.1007/978-3-642-40041-4_23) [13] D. Cash, E. Kiltz, and V. Shoup. The Twin Diffie-Hellman problem and applications. In N. P. Smart, editor, EURO_CRYPT 2008, Proceedings, volume 4965 of LNCS, pages_ [127–145. Springer, 2008. 10.1007/978-3-540-78967-3_8.](https://doi.org/10.1007/978-3-540-78967-3_8) [14] T.-H. H. Chan, E. Shi, and D. Song. Privacy-preserving stream aggregation with fault tolerance. In A. D. Keromytis, editor, Financial Cryptography and Data Security, FC 2012, volume 7397 of LNCS, pages 200–214. Springer, 2012. [10.1007/978-3-642-32946-3_15.](https://doi.org/10.1007/978-3-642-32946-3_15) [15] M. Chase and S. S. Chow. Improving privacy and security in multi-authority attribute-based encryption. In E. Al-Shaer, S. Jha, and A. D. Keromytis, editors, ACM Conference on _Computer and Communications Security, CCS 2009, pages_ [121–130. ACM, 2009. 10.1145/1653662.1653678.](https://doi.org/10.1145/1653662.1653678) [16] J. Chotard, E. D. Sans, R. Gay, D. H. Phan, and D. Pointcheval. Decentralized multi-client functional encryption for inner product. In T. Peyrin and S. D. Galbraith, editors, ASIACRYPT 2018, Proceedings, Part II, volume [11273 of LNCS, pages 703–732. Springer, 2018. 10.1007/978-](https://doi.org/10.1007/978-3-030-03329-3_24) [3-030-03329-3_24.](https://doi.org/10.1007/978-3-030-03329-3_24) [17] J. Chotard, E. Dufour-Sans, R. Gay, D. H. Phan, and D. Pointcheval. Dynamic decentralized functional encryption. In D. Micciancio and T. Ristenpart, editors, CRYPTO 2020, _Proceedings, Part I, volume 12170 of LNCS, pages 747–775._ [Springer, 2020. 10.1007/978-3-030-56784-2_25.](https://doi.org/10.1007/978-3-030-56784-2_25) [18] K. Emura. Privacy-preserving aggregation of time-series data with public verifiability from simple assumptions. In J. Pieprzyk and S. Suriadi, editors, Australasian Conference _on Information Security and Privacy, ACISP 2017, Pro-_ _ceedings, Part II, volume 10343 of LNCS, pages 193–213._ [Springer, 2017. 10.1007/978-3-319-59870-3_11.](https://doi.org/10.1007/978-3-319-59870-3_11) [19] J. Ernst and A. Koch. Efficient private stream aggregation with labels in the standard model. In S.-L. Gazdag, D. Loebenberger, and M. Nüsken, editors, crypto day mat_ters 32, Bonn, 2021. Gesellschaft für Informatik e.V. / FG_ [KRYPTO. 10.18420/cdm-2021-32-16.](https://doi.org/10.18420/cdm-2021-32-16) [20] E. S. Freire, D. Hofheinz, E. Kiltz, and K. G. Paterson. Noninteractive key exchange. In K. Kurosawa and G. Hanaoka, editors, Public-Key Cryptography, PKC 2013, Proceed_ings, volume 7778 of LNCS, pages 254–271. Springer, 2013._ [10.1007/978-3-642-36362-7_17.](https://doi.org/10.1007/978-3-642-36362-7_17) [21] S. Kim. Key-homomorphic pseudorandom functions from LWE with small modulus. In A. Canteaut and Y. Ishai, editors, EUROCRYPT 2020, Proceedings, Part II, volume [12106 of LNCS, pages 576–607. Springer, 2020. 10.1007/978-](https://doi.org/10.1007/978-3-030-45724-2_20) [3-030-45724-2_20.](https://doi.org/10.1007/978-3-030-45724-2_20) [22] E. Shi, T. H. Chan, E. G. Rieffel, R. Chow, and D. Song. Privacy-preserving aggregation of time-series data. In NDSS _[2011. The Internet Society, 2011. URL https://www.ndss-](https://www.ndss-symposium.org/ndss2011/privacy-preserving-aggregation-of-time-series-data)_ [symposium.org/ndss2011/privacy-preserving-aggregation-of-](https://www.ndss-symposium.org/ndss2011/privacy-preserving-aggregation-of-time-series-data) ----- [time-series-data.](https://www.ndss-symposium.org/ndss2011/privacy-preserving-aggregation-of-time-series-data) [23] J. Takeshita, R. Karl, T. Gong, and T. Jung. SLAP: Simple lattice-based private stream aggregation protocol. Cryptology _[ePrint Archive, Report 2020/1611, 2020. URL https://](https://eprint.iacr.org/2020/1611/20210513:151621)_ [eprint.iacr.org/2020/1611/20210513:151621.](https://eprint.iacr.org/2020/1611/20210513:151621) [24] F. Valovich. Aggregation of time-series data under differential privacy. In T. Lange and O. Dunkelman, editors, LATIN_CRYPT 2017, Revised Selected Papers, volume 11368 of_ _[LNCS, pages 249–270. Springer, 2017. 10.1007/978-3-030-](https://doi.org/10.1007/978-3-030-25283-0_14)_ [25283-0_14.](https://doi.org/10.1007/978-3-030-25283-0_14) [25] H. Waldner, T. Marc, M. Stopar, and M. Abdalla. Private stream aggregation from labeled secret sharing schemes. _Cryptology ePrint Archive, Report 2020/81, 2021. URL_ [https://eprint.iacr.org/2020/081.](https://eprint.iacr.org/2020/081) ## A Lemmas for the Security Proof of the PSA Scheme **Lemma 1. (Transition from G0 to G1): For every PPT** adversary, which corrupts the aggregator, there is a _A_ PPT adversary, on the PRF with _B_ _| Pr[G0(λ, A) = 1] −_ Pr[G1(λ, A) = 1]| _≤_ 2n[2](n − 1) · Adv[prf]B,PRF[(][λ][)][.] _Proof. We prove the transition by a hybrid argument._ As in the proof of [1], the goal in each hybrid step is to add a random pad to the PRF value of one more user. Each hybrid step consists of three game hops as depicted in Figure 9. We define hybrid games G0,µ for µ ∈ [n]. Let η = |U| and U = {i1, . . ., iη} be the set of users for which A asks the challenge query. Let Θ = min(η, µ). If Θ ≥ 2, in hybrid step µ, the game adds random pads to the PRF values of the first Θ users in . The condition that _U_ #### Θ 2 is necessary, because we need two users in to _≥_ _U_ change the PRF to a RF, because we must not change the overall sum of the ciphertexts. The pads are set up such that they are a perfect Θ-out-of-Θ secret sharing of 0. This makes sure that the pads have no effect on the sum of the ciphertexts. If µ > η, then there are already random pads on the PRF values of all users from, _U_ therefore, the subsequent games are the same. We have that G0 = G0,1 and G0,n = G1. In the following we will shortly describe the different games of the transition from G0,µ−1 to G0,µ: **Game G0,µ−1: This is step µ** _−_ 1 of the hybrid argument between G0 and G1. Let Θ = min(η, µ). If Θ ≥ 2, then in this game there are random pads added to the PRF-values of the first Θ 1 honest users. _−_ **Fig. 9. Game hops for one step of the hybrid argument** ----- **Game G[′]0,µ−1[: In this game, the PRF of user][ i][1][ is]** replaced by a RF. The pseudorandom pad tiΘ,l of user _iΘ is set such that_ [∑]i∈[n] _[t][i,l][ =][ t][0][,l][ still holds. The game]_ has to guess the first and Θ’th user of before seeing _U_ the challenge query, because it must be able to answer encryption queries before seeing the challenge query. **Game G[′′]0,µ−1[: Here we still have the RF of the]** previous game. When answering the challenge query, the game adds a random pad uΘ to the RF. Now, the first Θ honest users have a random pad added to their PRF/RF values. **Game G0,µ: This game again uses a PRF instead of** a RF for honest users i1 and iΘ. Because of the changes in the previous games, there are random pads added to the PRF value of the first Θ users of, so this is the _U_ _µ’th hybrid game._ Next we give the reductions for the transitions between the games. **Transition from G0,µ−1 to G[′]0,µ−1[: The difficulty]** in this step is that [∑]i∈[n] [PRF][k][i] [(][l][∗][) =][ PRF][k][0] [(][l][∗][)][. Re-] placing one PRF with a RF while holding the other keys fixed would violate this property. This is the reason why in each hybrid step of the transition from G0 to _G1 the game guesses two honest users. The pad of one_ user is determined by the answer of the PRF challenger and the pad of the other user is set, such that the pads still sum to PRFk0 (l[∗]). Now we show the indistinguishability of G0,µ−1 and _G[′]0,µ−1_ [by a reduction to the security of the PRF. We] assume that there is an adversary which can distin_A_ guish G0,µ−1 and G[′]0,µ−1[. The reduction][ B][ guesses the] honest users i[∗]1 [and][ i]Θ[∗] [. They generate the secret keys] _{ki}i ∈_ [n]0 \{i[∗]1[, i][∗]Θ[}][ for all users and the aggregator ex-] cept of users i[∗]1 [and][ i]Θ[∗] [. They send the public parameters] pp to and answer the queries as follows: _A_ QCorrupt(i): If the guess of i[∗]1 [and][ i]Θ[∗] [was correct,] these two users stay uncorrupted. Since generated the _B_ keys of all the other clients, they can simply answer with the corresponding secret key. If i = 0, then B returns the aggregation key k0. Note that here k0 is not the sum of the client keys, but chosen randomly as well. The pads of i[∗]1 [and][ i]Θ[∗] [will be chosen accordingly such that] #### ∑ _i∈[n]_ _[t][i,l][ =][ t][0][,l][ still holds.]_ QEnc(i, xi, l): If i = i[∗]1 [or][ i][ =][ i]Θ[∗] [,][ B][ queries][ l][ to] their own PRF challenger and receives al, which is either PRFk′ (l), for some k[′], or RF(l). They set ti[∗]1 _[,l][ :][=][ a][l][ and]_ _ti[∗]Θ[,l][ :][=][ PRF][k][0][ −]_ [∑]j∈[n]\{i[∗]1 _[,i][∗]Θ[}][ PRF][k][j]_ [(][l][)][ −] _[a][l][ =][ t][0][,l][ −]_ #### ∑ _j∈[n]\{i[∗]Θ[}][ t][j,l][. This ensures that all][ t][i,l][ still sum to]_ _t0,l. Note that ti[∗]Θ[,l][ =][ t][0][,l][ −]_ [∑]j∈[n]\{i[∗]Θ[}][ t][j,l][ also holds] in the unmodified scheme (Figure 3) due to the key homomorphism of the PRF. If i = i[∗]1[,][ B][ sends][ x][i][ +][ t][i][∗]1 _[,l]_ to A. If i = i[∗]Θ[,][ B][ sends][ x][i][ +][ t][i][∗]Θ[,l][ to][ A][ and stores][ t][i]1[∗][,l] until A asks an encryption query for i[∗]1 [and label][ l][. For] the other clients knows the corresponding secret keys _B_ and can, therefore, answer the queries without asking their PRF challenger. QChallenge(U, {x[0]i _[}][i][∈U]_ _[,][ {][x]i[1][}][i][∈U]_ _[, l][∗][)][:]_ Here _B_ encrypts x[0]i [the same way as in the][ QEnc][ queries.] If the PRF challenger uses PRFk′ instead of a RF then _ki[∗]1_ [is implicitly set to][ k][′][ and][ k][i]Θ[∗] [is set such that] #### ∑ _i∈[n]_ _[k][i][ =][ k][0][. In that case the][ k][i][ are a perfect secret]_ sharing of k0, which is exactly as in G0,µ−1. So in that case, B perfectly simulates G0,µ−1. If the PRF challenger uses a RF, then ti[∗]1 _[,l][ =][ RF][(][l][)]_ and ti[∗]µ [is set such that all][ t][i,l][ sum to][ t][0][,l][. So in this case] _B perfectly simulates game G[′]0,µ−1[.]_ **Transition from G[′]0,µ−1** **[to][ G]0[′′],µ−1** [In this step] the goal is to add random pads uΘ to the RF of users i1 and iΘ in the answers of the challenge query. Because we consider encrypt-once security, cannot ask an en_A_ cryption query for user i1 or iΘ on label l[∗]. Therefore, the only information that A has about RF(l[∗]), comes from the answer to the challenge query. Since RF(l[∗]) is identically distributed as RF(l[∗]) + uΘ, A cannot realize that they received RF(l[∗]) + uΘ instead of RF(l[∗]). Thus, _G[′]0,µ−1_ [and][ G]0[′′],µ−1 [are perfectly indistinguishable.] **Transition from G[′′]0,µ−1** **[to][ G][0][,µ][ In this step we]** need to change back the RF of users i1 and iΘ to a PRF. This works exactly as the transition from G0,µ−1 to G[′]0,µ−1[.] After η _−_ 1 of these steps we reached Game G1. Now in G1 the challenge query (U, {x[0]i _[}][i][∈][U]_ _[,][ {][x]i[1][}][i][∈U]_ _[, l][∗][)][ is an-]_ swered with x[0]i [+][ t][i,l][∗] [+ ¯][u][i][, where] _u¯i =_ #### ⎧ ⎪⎪⎨∑j∈U\{i1} [u][j] if i = i1, _−ui_ if i ∈U \ {i1}, #### ⎪⎪⎩0 else. Therefore, the {u¯i}i∈U are a perfect η out of η secret sharing of 0. The guessing of the two honest users in each hybrid step incurs a reduction loss of n(n − 1) and using n hybrid games leads to a loss of n[2](n − 1) for the hybrid argument. Since the hybrid argument is applied twice, once to add and once to remove the random pads, we get a total reduction loss of 2n[2](n − 1). **Lemma 2. (Transition from G1 to G2): For every PPT** adversary, which corrupts the aggregator, there is a _A_ ----- PPT adversary on the PRF with _B_ _| Pr[G1(λ, A) = 1] −_ Pr[G2(λ, A) = 1]| _≤_ 2n(n − 1) · Adv[prf]B,PRF[(][λ][)][.] _Proof. The goal in this step is to change the answer of_ the challenge query from encryptions of x[0]i [to encryp-] tions of x[1]i [. We distinguish two cases here. Remember] that Ql∗ is the set of clients for which A has received a ciphertext on label l[∗], either by an encryption or a challenge query. In the first case Ql∗ = HS, which means that gets a ciphertext of every honest user for the chal_A_ lenge label l[∗]. Then we have Ql∗ _∪CS = [n], whereby A’s_ challenge query is restricted by the balance-condition. We then argue that, since [∑] _x[0]i_ [=][ ∑] _[x]i[1][, the change is]_ covered by the ¯ui. In the second case there is an honest user of whom _A does not get a ciphertext on label l[∗]. Therefore, the_ challenge messages are not restricted by the balance condition. Here we use the fact that is lacking a cipher_A_ text of an honest user iq and thereby has no information about PRFkiq (l[∗]). Case 1 (Ql∗ = HS): In this case A knows k0, because they corrupted the aggregator. Furthermore _Ql∗_ = HS, whereby A’s challenge messages are restricted by the balance-condition. We argue as the authors in [1]. For t0,l := PRFk0 (l), we have _ti,l∗_ = t0,l∗ #### [∑] and that {u¯i}i∈U is a perfect η out of η secret sharing of 0. Therefore, {x[0]i [+] _[t][i,l][∗]_ [+][ ¯][u][i][}][i][∈U][ and][ {][x]i[1] [+] _[t][i,l][∗]_ [+][ ¯][u][i][}][i][∈U] are perfect secret sharings of [∑]i∈U [(][x]i[0] [+][ t][i,l][∗] [)][ and] #### ∑ _i∈U_ [(][x]i[1] [+] _[t][i,l][∗]_ [)][, respectively. The balance-condition re-] quires that [∑]i∈U _[x]i[0]_ [=][ ∑]i∈U _[x]i[1][. So][ {][x]i[0]_ [+][ t][i,l][∗] [+][ ¯][u][i][}][i][∈U] and {x[1]i [+][ t][i,l][∗] [+][ ¯][u][i][}][i][∈U][ are both perfect secret sharings] of the same value and are, therefore, perfectly indistinguishable. Case 2 (Ql∗ _̸= HS): Other than in case 1, A’s mes-_ sages in the challenge query are not restricted by the balance-condition, because there is at least one honest user for which A has no ciphertext on label l[∗]. Therefore, in this case we need another reduction. If asks no chal_A_ lenge query or a challenge query with =, has no _U_ _{}_ _A_ information about the challenge bit and, therefore, ’s _A_ advantage is 0. Thus, we can assume that contains _U_ at least one user. Additionally, in this case Ql∗ ≠ _HS,_ therefore, HS \ Ql∗ contains an honest user that is different from the user in . The reduction guesses the _U_ _B_ users i[∗]q _[∈HS \][ Q][l][∗]_ [and][ i]u[∗] _[∈U][. Then][ B][ changes the]_ PRF of these two users to a RF by setting ti∗u[,l][ :][=][ RF][(][l][)] and ti∗q _[,l][ :][=][ t][0][,l][ −]_ [∑]i∈[n]\iq∗ _[t][i,l][. This is done the same]_ way as in the transition from G0,µ−1 to G[′]0,µ−1[.] In the next step we change all ciphertexts in the challenge query from x[0]i [+][ t][i,l][∗] [to][ x]i[1] [+][ t][i,l][∗] [. Because of] the ¯ui, {x[0]i [+][ t][i,l][∗] [+][ ¯][u][i][}][i][∈U][ and][ {][x]i[1] [+][ t][i,l][∗] [+][ ¯][u][i][}][i][∈U][ are] perfect η out of η secret sharings of [∑]i∈U [(][x]i[0] [+][t][i,l][∗] [)][ and] #### ∑ _i∈U_ [(][x]i[1] [+][ t][i,l][∗] [)][ respectively. For both][ b][ = 0][ and][ b][ = 1] we have [∑]i∈U [(][x]i[b] [+][ t][i,l][∗] [) =][ ∑]i∈U\{i[∗]u[}][(][x]i[b] [+][ t][i,l][) +][ x][i]u[∗] [+] _ti∗u[,l][∗]_ [. Because][ t][i][∗]u[,l][∗] [=][ RF][(][l][∗][)][, both][ {][x]i[0] [+][ t][i,l][∗] [+][ ¯][u][i][}][i][∈U] and {x[1]i [+][ t][i,l][∗] [+][ ¯][u][i][}][i][∈U][ are secret sharings of a truly] random value and are, therefore, perfectly indistinguishable. In the last step we change back the RF to a PRF again. The reduction loss of n(n − 1) comes from guessing the two users i[∗]u [and][ i]q[∗][. The factor of two is there,] because the RF needs to be changed back into a PRF. Therefore, the total reduction loss is 2n(n − 1). **Lemma 3. (Transition from G2 to G3): For every PPT** adversary, that corrupts the aggregator, there is a _A_ PPT adversary on the PRF with _B_ _| Pr[G2(λ, A) = 1] −_ Pr[G3(λ, A) = 1]| _≤_ 2n[2](n − 1) · Adv[prf]B,PRF[(][λ][)] _Proof. This transition is just applying the G0_ _G1 tran-_ _−_ sition backwards. **Lemma 4. For every PPT adversary**, which does not _A_ corrupt the aggregator, there is a PPT adversary on _B_ the PRF with _| Pr[G0(λ, A) = 1] −_ Pr[G3(λ, A) = 1]| _≤_ 2n[2] _· Adv[prf]B,PRF[(][λ][)][.]_ _Proof. In this case we can directly go from G0 to G3 via_ a hybrid argument over all users. Let {i1, . . ., iη} = U be the set of users specified in the challenge query. In hybrid game Hµ the challenge query for (i1, . . ., iµ) is answered with an encryption of x[1]i [, whereas for the other] users it is answered with an encryption of x[0]i [.] Formally, in Hµ we have: QChallenge(U = {i1, . . ., iη}, {x[0]i _[}][i][∈U]_ _[,][ {][x]i[1][}][i][∈U]_ _[, l][∗][):]_ #### �Enc(pp, ki, x[0]i [)] if i = iτ for τ > µ _ci,l∗_ := Enc(pp, ki, x[1]i [)] if i = iτ for τ ≤ _µ_ `return {ci,l∗` _}i∈U_ We have that G0 = H0 and G3 = Hn. To get from _Hµ−1 to Hµ we use the two intermediate games Hµ[′]_ _−1_ and Hµ[′′]−1[. In][ H]µ[′] _−1[, the PRF of user][ i][µ][ is replaced by]_ a RF and in Hµ[′′]−1 [the challenge query of user][ i][µ][+1][ is] ----- answered with an encryption of x[1]q+1 [instead of][ x]q[0]+1[. In] Lemma 1 we needed two honest users in order to change the PRF to a RF. This was necessary, because we had to make sure that the sum of the ciphertexts remained unchanged. Here we are in the case that does not corrupt _A_ the aggregator and, therefore, is unable to recognize _A_ when the sum of the ciphertexts changes. Thus, we only need one honest user to exchange the PRF for a RF. We now describe the games in a bit more detail: **Game Hµ−1: In this game the challenge query for** users iτ with τ ≤ _µ −_ 1 is answered with encryptions of _x[1]iτ_ [, whereas the challenge query for the users][ i][τ][ with] _τ > µ −_ 1 is answered with encryptions of x[0]iτ [.] **Game Hµ[′]** _−1[: This game guesses user][ i][µ][ and uses]_ a RF instead of a PRF to answer the encryption and challenge queries for this user. **Game Hµ[′′]−1[: In this game the challenge query for]** user iµ is answered with an encryption of x[1]iµ [instead of] _x[0]iµ_ [.] **Game Hµ: This game uses a PRF instead of a RF** for user iµ again. The answer to the challenge query is an encryption of x[1]iτ [for all users with][ τ][ ≤] _[µ][.]_ Next, we prove the transitions between the games by reductions to the security of the PRF. **Transition from Hµ−1 to Hµ[′]** _−1[: The reduction][ B]_ guesses the user i[∗]µ [and generates keys][ k][i] [for all users,] including the aggregator, except of user i[∗]µ[. Then][ B][ an-] swers the queries as follows: QCorrupt(i): If i ̸= i[∗]µ [then][ B][ returns the self gener-] ated key ki. If i = i[∗]µ [then the guess that][ i]µ[∗] [would be] honest was wrong and B aborts. If i = 0, i.e. A wishes to corrupt the aggregator, aborts, because we are in _B_ the case where does not corrupt the aggregator. _A_ QEnc(i, xi, l): If i ̸= i[∗]µ [then][ B][ simply answers with] _xi + PRFki_ (l) mod R. If i = i[∗]µ [then][ A][ asks][ l][ to their] PRF challenger CPRF, gets the answer al and sends xi[∗]µ [+] _al mod R to A._ QChallenge(U = {i1, . . ., iη}, {x[0]i _[}][i][∈U]_ _[,][ {][x]i[1][}][i][∈U]_ _[, l][∗][):]_ The reduction asks l[∗] to CPRF and receives the answer _al∗_ . Then B answers with x[1]iτ [+][ PRF][k]iτ [(][l][∗][)][ for][ τ < µ][,] with x[0]iτ [+][ PRF][k]iτ [(][l][∗][)][ for][ τ > µ][ and with][ x]i[0]τ [+][ a][l][∗] [for] _τ = µ. If iµ ̸= i[∗]µ[, then][ B][’s guess of][ i][∗]µ_ [was wrong and] they abort the game. The reduction can directly use ’s output, as _B_ _A_ guess for CPRF. If CPRF used a PRF to generate its answers, then B perfectly simulates Hµ−1 and if CPRF used a RF, then B perfectly simulates Hµ[′] _−1[.]_ **Transition from Hµ[′]** _−1_ **[to][ H]µ[′′]−1[: The games][ H]µ[′]** _−1_ and Hµ[′′]−1 [are perfectly indistinguishable, because][ RF][ +] _x[0]µ_ [is identically distributed as][ RF][ +][ x]µ[1] [.] **Transition from Hµ[′′]−1** **[to][ H][µ][: Here we only need]** to change back the RF of user iµ to a PRF. So this transition is the transition from Hµ−1 to Hµ[′] _−1_ [applied] backwards. Guessing user i[∗]µ [entails a reduction loss of][ n][, for] both the transition from Hµ−1 to Hµ[′] _−1_ [and from][ H]µ[′′]−1 to Hµ. Together with the n hybrid steps leads to a reduction loss of 2n[2]. ## B Security Proofs of the PRFs In this section we prove that the PRF used in Section 4.1 is secure. **Theorem 2. Let λ, q, p ∈** N with q > p, k _←$_ Zλq [and] _H : X ↦→_ Z[λ]q [. When we model][ H][ as a random oracle,] then for any PPT adversary A on the PRF Fk(x) := _⌊⟨H(x), k⟩⌋p, there is an adversary B on LWRλ,q,p with_ Adv[prf]A,F [(][λ][)][ ≤] [Adv][LWR]B (λ). _Proof. The idea is for an LWR sample (a, b) ∈_ Z[λ]q _[×]_ Zp, to interpret a as H(x) for some x and the LWR secret s as the PRF-key k. Then b = ⌊⟨H(x), k⟩⌋p. This is possible, because can program the random oracle _B_ accordingly. The reduction works as follows. maintains a _B_ _B_ table of triples (·, ·, ·). If A asks a PRF query for value _x, B looks for an entry (x, a, b) in the table and returns b_ if such an entry is present. If there is no such entry, then requests an LWR sample from their LWR challenger, _B_ receives (a, b), stores (x, a, b) in the table and returns b to . _A_ If A asks a RO query for value x, B again looks for an entry (x, a, b) in the table, but this time returns a if an entry is found. If there is no such entry, again _B_ queries their LWR challenger, stores the answer (a, b) in the table as (x, a, b) and returns a to B. In LWR the values of a are uniformly random, therefore, ’s answers to the RO queries of are uniformly _B_ _A_ random as well. Furthermore, by maintaining the table of triples (x, a, b), B ensures that the answer to the RO queries are consistent with the answers of the PRF queries. If the LWR challenger returns random values _b, then B perfectly simulates a random function. If the_ LWR challenger returns actual LWR samples then per_B_ fectly simulates the PRF Fk(x) := ⌊⟨H(x), k⟩⌋p, where **k is the LWR secret s. Therefore,** can directly forward _B_ ’s guess to their LWR challenger and wins the game, _A_ if wins. _A_ ----- **Fig. 11. The diluted geometric mechanism from [14]. The func-** tion Geom(α) returns a value k with probability _[α]α[−]+1[1]_ _[α][−|][k][|][.]_ **Fig. 10. The encrypt procedure for the fault-tolerant scheme,** where B(i) is the set of all log2(n) nodes of the tree from client i to the root. For each of these nodes it calls PSA.Encrypt which is the encrypt algorithm of our proposed scheme in Figure 3. ## C Noise and Fault Tolerance In this section we quickly describe how the noise is added in the fault tolerant version of our scheme. This is only a slight adaption of the method of Chan et al. especially of their Figure 2. They only considered user values in {0, 1}, whereas we consider values in {0, . . ., ∆}, where ∆ is the largest possible power consumption for which privacy shall still hold. Therefore we have an additional factor of 1/∆ in the computation of ϵ0, the rest remains unchanged. Note that the factor of 1/∆ is also present in [22], which introduced PSA. In the literature of differential privacy, ∆ is also called the sensitivity of the function. Figure 10 shows the encryption algorithm of the fault tolerant scheme. For each node on the path from client i to the root, i.e. for each instance of the PSA scheme in which client i is part of, the algorithm creates one ciphertext by calling the encrypt function from our PSA scheme in Figure 3. The parameters ϵ and δ are the same for all clients. One execution of the encrypt algorithm in Figure 10 provides (ϵ, δ) computational differential privacy. The values ϵ0 and δ0 are then set to accommodate for the fact, that the aggregator gets K ciphertexts per client. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.2478/popets-2021-0063?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.2478/popets-2021-0063, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBYNCND", "status": "HYBRID", "url": "https://petsymposium.org/popets/2021/popets-2021-0063.pdf" }
2,021
[ "JournalArticle" ]
true
2021-07-23T00:00:00
[ { "paperId": "cf93b2ac8cfac8ce0bf42f685ca7a5e207756d70", "title": "Dynamic Decentralized Functional Encryption" }, { "paperId": "f7da5a6484d7a5e8caa1e16d2741045ed600054a", "title": "Key-Homomorphic Pseudorandom Functions from LWE with Small Modulus" }, { "paperId": "43d41a9ee94259440e114d203ceb2c514e49db7f", "title": "From Single-Input to Multi-Client Inner-Product Functional Encryption" }, { "paperId": "95235de776ce51eb2340c9836cafa5a87d051e35", "title": "Decentralizing Inner-Product Functional Encryption" }, { "paperId": "b32788a3bd83dbf5f3eba411d57c9fca8ced4c45", "title": "Decentralized Multi-Client Functional Encryption for Inner Product" }, { "paperId": "db0cc2f21b20cbc0ab8946090967399c25709614", "title": "Practical Secure Aggregation for Privacy-Preserving Machine Learning" }, { "paperId": "7ee8a1f2739f6879be0a9a1b5cce4a700c168826", "title": "Aggregation of Time-Series Data Under Differential Privacy" }, { "paperId": "d705cc4656b997f0b9cde0ac5fb5e476625f0e03", "title": "Privacy-Preserving Aggregation of Time-Series Data with Public Verifiability from Simple Assumptions" }, { "paperId": "c5bc008a9da6b493db9375ee05e21aaffbf6e238", "title": "A New Framework for Privacy-Preserving Aggregation of Time-Series Data" }, { "paperId": "3365f31550db350ce54b7ccfd9dff3cb7715185c", "title": "On the concrete hardness of Learning with Errors" }, { "paperId": "0fb2489b324c29df8ab19fe26a878bbb3cfaeb6e", "title": "New and Improved Key-Homomorphic Pseudorandom Functions" }, { "paperId": "b32cea998f4842abc06f13b8190b8bf1cce46aac", "title": "Key Homomorphic PRFs and Their Applications" }, { "paperId": "05b56733becfe17946766b0069799180d1260df8", "title": "Non-Interactive Key Exchange" }, { "paperId": "683432f4c531a23dcf342584219c52da97c11476", "title": "Pseudorandom Functions and Lattices" }, { "paperId": "555a239220db7e3fdcbf43b4de7f2bb7c6d101fe", "title": "Privacy-Preserving Stream Aggregation with Fault Tolerance" }, { "paperId": "902ddd904effe50d228ea25ab23c16ae664d3ce5", "title": "I Have a DREAM! (DiffeRentially privatE smArt Metering)" }, { "paperId": "c509296ec59da5c2a711aeeceb09c4b171e1dcca", "title": "Improving privacy and security in multi-authority attribute-based encryption" }, { "paperId": "573630fdb9ce67c267caa043429f5e78b41e471f", "title": "The Twin Diffie-Hellman Problem and Applications" }, { "paperId": "c84d340c6a738446170ce06344f38a42ab51e883", "title": "Private Stream Aggregation from Labeled Secret Sharing Schemes" }, { "paperId": "8950f1e0beaebc3968ff536052243751294a3fcc", "title": "SLAP: Simple Lattice-Based Private Stream Aggregation Protocol" }, { "paperId": "c4be6ae449b80ac2666532c5b01c389ca516b421", "title": "Ad Hoc Multi-Input Functional Encryption" }, { "paperId": "810eb064e3f914554dfc46d0ab0ded71a62bff73", "title": "Revisiting Private Stream Aggregation: Lattice-Based PSA" }, { "paperId": "7cc53ef35f8398181bd09755ecc2fa8f52d0da1d", "title": "Privacy-Preserving Aggregation of Time-Series Data" }, { "paperId": null, "title": "[ 0 ,µ − 1 G ′′ 0 ,µ − 1 G 0 ,µ the game can guess i ∗ Θ , without yet" }, { "paperId": null, "title": "the value of Θ. i ∗ 1 ← $ [ n ] , i ∗ Θ ← $ [ n ] \\ { i ∗ 1 } ( pp , { k i } i ∈ [ n ] 0 ) ← Setup (1 λ , 1 n ) α ← A QEnc ( · , · , · ) , QCorrupt ( · ) , QChallenge ( · , · , · , · )" } ]
30,310
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00f84366664c896ee112940dc12eaacef6189f89
[ "Computer Science" ]
0.886323
Distributed volumetric scene geometry reconstruction with a network of distributed smart cameras
00f84366664c896ee112940dc12eaacef6189f89
2009 IEEE Conference on Computer Vision and Pattern Recognition
[ { "authorId": "2108215905", "name": "Shubao Liu" }, { "authorId": "1712504", "name": "Kongbin Kang" }, { "authorId": "1682784", "name": "Jean-Philippe Tarel" }, { "authorId": "145700259", "name": "D. Cooper" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
# Distributed Volumetric Scene Geometry Reconstruction With a Network of Distributed Smart Cameras ## Shubao Liu † Kongbin Kang † Jean-Philippe Tarel ‡ David B. Cooper † †Division of Engineering, Brown University, Providence, RI 02912 {sbliu, kk, cooper}@lems.brown.edu ## ‡Laboratoire central des Ponts et Chauss´ees (LCPC), Paris, France jean-philippe.tarel@lcpc.fr ## Abstract _Central to many problems in scene understanding based_ _on using a network of tens, hundreds or even thousands_ _of randomly distributed cameras with on-board processing_ _and wireless communication capability is the “efficient”_ _reconstruction of the 3D geometry structure in the scene._ _What is meant by “efficient” reconstruction? In this pa-_ _per we investigate this from different aspects in the con-_ _text of visual sensor networks and offer a distributed recon-_ _struction algorithm roughly meeting the following goals: 1._ _Close to achievable 3D reconstruction accuracy and robust-_ _ness; 2. Minimization of the processing time by adaptive_ _computing-job distribution among all the cameras in the_ _network and asynchronous parallel processing; 3. Com-_ _munication Optimization and minimization of the (battery-_ _stored) energy, by reducing and localizing the communica-_ _tions between cameras. A volumetric representation of the_ _scene is reconstructed with a shape from apparent contour_ _algorithm, which is suitable for distributed processing be-_ _cause it is essentially a local operation in terms of the in-_ _volved cameras, and apparent contours are robust to our-_ _door illumination conditions. Each camera processes its_ _own image and performs the computation for a small sub-_ _set of voxels, and updates the voxels through collaborat-_ _ing with its neighbor cameras. By exploring the structure_ _of the reconstruction algorithm, we design the minimum-_ _spanning-tree (MST) message passing protocol in order to_ _minimize the communication. Of interest is that the result-_ _ing system is an example of “swarm behavior”. 3D recon-_ _struction is illustrated using two real image sets, running_ _on a single computer. The iterative computations used in_ **_the single processor experiment are exactly the same as_** **_are those used in the network computations. Distributed_** _concepts and algorithms for network control and communi-_ _cation performance are theoretical designs and estimates._ ## 1. An Overview of the System ### 1.1. Motivation With the recent development of cheap and powerful visual sensors, wireless chips and embedded systems, cameras have enough computing power to do some on-board “smart” processing. These “smart cameras” can form a network to collaboratively monitor, track and analyze the scenes of interest. This area have drawn a lot of attention in both academia and industry over the past years (see [1] [11] and [13] for an overview). However compared with the maturity and availability of the camera network hardware, the software capable of fully utilizing the huge amount of visual information is greatly under-developed. This has become the bottleneck for the wide deployment of the smart camera network (also called visual sensor network, VSN). There is an obvious demand to synchronize the recent development of vision algorithms with the development of the visual sensor network hardware. Our paper presents a completely new and natural approach to 3D reconstruction within a smart camera network. ### 1.2. The Goal Our goal is minimum-error 3D scene reconstruction based on edge information with N calibrated smart cameras through their collaborated distributed processing, as illustrated in Fig. 1. A Bayesian approach is taken to 3D reconstruction, where the surface is treated as a stochastic process modelling the smoothness of the surface. Thanks to the apparent contours’ robustness to environmental factors, shape-from-apparent-contours is more suitable for outdoor distributed camera applications than the intensitybased multi-view reconstruction. The representation for the estimated surface is a discretized level set function defined on a grid of voxels. The cost function to be minimized is the sum of the area of the 3D surface and the integral of “consistency” between the apparent contour of the current surface estimate and the image edges. The object surface is to be reconstructed distributedly with N smart cameras 1 ----- co-operating to minimize both the processing time and the consumed on-board battery energy. Computation and communication load-balancing are investigated to make battery usage roughly at the same level over all the cameras. ### 1.3. 3D Surface Reconstruction The proposed surface estimation procedure is iterative through numerical solution of the first order variation of the energy functional (i.e., the cost function). It turns out that each iteration is a linear incremental change of the current estimated surface. All of the computation takes place within a thin band around the estimated surface. For each camera c, the data and information available for the (t + 1)th iteration is: its projection matrix; the edges in its image; and a subset of voxels that this camera maintains. The incremental update is the sum of two increments. The first increment comes from the contribution of the image edge data. This increment attempts to align the contour generators of the estimated 3D surface with the edge-data in the observed image. The second increment is the contribution of the a priori stochastic model of the 3D surface. Hence, for each voxel on the estimated 3D surface at the start of an incremental surface update-iteration, the cameras contributing to the voxel update are the ones whose contour generators are close to that voxel. A voxel is in the primary responsibility set (PRS) of each camera whose image information contributes to the voxel’s updating. Each contributes to the first update-increment. One of these cameras takes responsibility for computing the second update-increment. This group of cameras each has a record of the changes made, and therefore of the total update change made. For a 3D surface voxel not contributed by any camera at the start of an update-iteration, there is no first incremental-update, and one of the cameras takes responsibility for computing and communicating the second incremental-update. This voxel is in the second set of the responsible cameras (SRS). Fig. 2 illustrates these concepts on a sphere shape. ### 1.4. Distributed Processing It is know that that the battery power for two wireless cameras to communicate is approximately proportional to the square of their distance. Hence, rather than two cameras communicating directly, the signals from the transmitting camera travels to the receiving camera through a sequence of relays where it travels from one camera to a camera that is close, then to another close camera, etc. In this way, communication power increases linearly with distance between cameras. This is a camera communication network (CCN) optimization problem that finds the best routing for each communication, i.e. how to send messages. For our purpose, only the end-to-end communication in the application layer is considered. Inspired by the observation that the communication in the proposed algorithm works more like broadcast (although not exactly, which will be discussed later) than point-to-point ad-hoc communication, we can optimize the communication further by decid Figure 1. Volumetric world, smart cameras and their observations. ing what to send and who to send to, instead of only optimizing how to send. This results in an efficient message passing protocol based on the minimum-spanning-tree of the cam_era reconstruction network (CRN, the exact meaning will_ be discussed later.) Distributing the voxel updating job among all the smart cameras to enable parallel processing is achieved by each camera processing those voxels in its primary and second responsibility sets. These sets for the various cameras are close enough in size such that the partition results in balanced parallel processing. A camera determines its secondary responsibility set through negotiating the boundaries with its neighbor cameras in the CRN. Also incurred is battery energy for the communications in determining the secondary responsibility set. Rough minimization of communications battery energy is achieved by routing communications over paths contained in an MST (Minimum Spanning Tree). Also some communication is required among cameras having primary sets that are close in order for the cameras to figure out their secondary responsibility sets. ## 2. Shape From Apparent Contours A shape-from-apparent-contours algorithm is first developed to reconstruct the 3D shape from apparent edges in different views. The algorithm also incorporates the prior knowledge about the surface (e.g., surface smoothness) to produce a complete shape. The proposed algorithm combines the ideas in 2D active contoursand variational surface reconstruction [7, 6, 15, 9] based on implicit surface deformation. In active contour fitting, the best curve C [⋆] is found by deforming a curve C(s) to make it fit the object boundaries: C [⋆](s) = arg min C(s) [E][(][C][(][s][))][.] The functional E(C(s)) is usually defined as � 1 � 1 E(C(s)) = µ C(s)ds − ||∇GI(C(s)||ds (1) 0 0 ----- Figure 2. Illustration of the key concepts (including contour generators, band, PRS, SRS) in the distributed reconstruction algorithm, with a simple setting (a sphere shape and evenly distributed cameras around the equator of the sphere. where ds is the infinitesimal curve length, ∇GI(C(s)) = ∇(G ∗ I(C(s))) is the data term measuring the influence of the image intensity gradient along the fitted curve, G ∗ I is the convolution of the intensity image I with a Gaussian filter G, µ is a scalar value controlling the influence of the length of the curve. Apparent contours are curves coming from contour generators on the surface through perspective projection. So instead of assuming that the contours can be deformed freely, we constrain them with a 3D surface: Ci(s) = Πi(Gi(s)), (2) where Πi is the ith camera’s perspective projection, which maps a 3D point X to a 2D image point xi; Ci(s) is the apparent contour in image i, Gi(s) is the corresponding contour generator on the surface, as shown in Fig. 2. Notice that � 1 � ||∇GI(Ci(s))||ds = 1Gi(X)||∇GIi(Πi(X))||dA 0 S (3) where S is the surface. Equation (3) turns the line integral to a surface integral with the introduction of the contour generator indicator function 1Gi(X), which is a impulse function. (In experiment, it is approximated with a Gaussian function.) Through the occluding geometry relationship between the surface normal N and the tangent plane N[¯] i (got from back-projecting of the tangent line of the apparent contours), we extend (3) to � 1Gi(X)||∇GI(xi)|| · |N[¯] [T]i [N][|][dA] (4) S to further enforce the tangency constraint. The higher order term |N[¯] [T]i [N][|][ make sthe shape evolution converge faster and] more accurate. The surface to be reconstructed S [⋆] is the optimal surface that minimizes an energy functional in the form of a weighted area, with the weights depending on the M observed images as in (4) and a prior term: E(S) = �S [Φ(][X][,][ N][)][dA] M � = �S ��i [1][G]i[(][X][)][||∇][G][I][(][x][i][)][|| · |][N][ ¯] [T]i [N][|][ +][ µ] dA, (5) where dA is an infinitesimal area element, µ is a parameter controlling the smoothness of the surface. Interpreted in Bayesian language, the prior energy term �S [µdA][ corre-] sponds to a prior probability Z[1] [e][−][µArea][((][S][))][, which is the] energy representation of a 1st-order Continuous Markov Random Field, encouraging smooth surfaces instead of bumpy ones. The functional (5) is minimized through gradient descent methods by computing the first order variation. The gradient descent flow for (5) can be written as [7]: St = F N, (6) F = 2κΦ −⟨ΦX, N⟩− 2κ⟨ΦN, N⟩. (7) where κ is the mean curvature of the surface S. With the level set representation, S = {X : φ(X) = 0}, the above evolution equation can be rewritten as: φt = F ||∇φ||. (8) Through some calculus derivation, we get the speed function for (8) as F = 2µκ − M �⟨ΦiX, N⟩. (9) i=1 ## 3. Distributed Algorithm for Scene Recon- struction In the above, we have briefly described a centralized algorithm for shape from apparent contours, where one central processor collects data from all cameras and processes them in batch. In the visual sensor network applications, distributed algorithms are preferred, where each smart camera runs identical programs but with different states and different image inputs. In this section we show that the proposed algorithm can be run distributedly on the smart camera network by augmenting the algorithm with a job division module and a communication module. Principally the algorithm can be extended distributedly because: (1) the algorithm reconstructs the contour generators, and the other part of the surface is interpolated through the prior energy, equivalently, a priori stochastic model for the 3D surface. (2) It has been shown that the contour generators can be reconstructed locally by studying the differential geometry of the apparent contour change [4, 3, 10]. ----- Fv[c] a scalar representing camera c’s contribution to voxel v’s updating PRSc the primary responsible set of camera c SRSc the secondary responsible set of camera c, PRSc ∪ SRSc = Vc Table 1. Main notation summary voxel ID voxel value neighbor cameras in MST 1001 1.302 {2, 30} 2187 -2.630 ...{10} ... ... ... PRS: {1001, ...} SRS: {2187, ...} watching voxel list: {...} boundary voxel list: {...} Table 2. An example of the data structures that each camera maintains. To highlight the structure of the reconstruction procedure, we summarize each voxel’s updating with this formula: |V c|the set of voxels that the camera c maintains| |---|---| |C v|the set of cameras that maintains voxel v| |F vc|a scalar representing camera c’s contribution to voxel v’s upd| |PRS c|the primary responsible set of camera c| |SRS c|the secondary responsible set of camera c, PRS ∪SRS = c c| |voxel ID|voxel value|neighbor cameras in MST| |---|---|---| |1001 1.302 {2, 30} 2187 -2.630 ...{10} ... ... ... PRS: {1001, ...} SRS: {2187, ...} watching voxel list: {...}||| φ[t]v[+∆][t] = φ[t]v [+ (2][µκ][ +] � Fv[c][)][||∇][φ][||][∆][t,] (10) c∈Cv where Cv is the set of cameras c that has Fv[c] [̸][= 0][ for voxel] v, Fv[c] [is the speed contribution from camera][ c][ to voxel][ v][:] Fv[c] [=][ −⟨][Φ]iX(v)[,][ N][(][v][)][⟩][.] (11) Formula (10) describes the updating operation for each voxel. A na¨ıve parallel implementation of the algorithm is to divide the entire set of voxels into M (the number of smart cameras) subsets, and each camera takes care of one subset of the voxels. The problem with this na¨ıve approach is that (1) each camera needs to maintain a copy of all the other cameras’ observed images. This implies a huge amount of data communication, which prevents the algorithm scaling up to a large camera network; (2) the contour generators dynamically change as the surface shape evolves. So fixing the set of voxels that each camera maintains requires distant cameras to exchange information about voxels’ states and image observations. This prevents the communication between cameras from being localized. Instead we build a camera-centric distributed algorithm, in which each camera c maintains a gradually changing dynamic subset Vc of voxels around the current estimated contour generators seen by this camera. Algorithm 1 describes the over-all procedure in a high level, with each subroutine being discussed in detail later in Algorithms 2 and 5. Through each camera maintaining a subset of voxels Vc and localizing the computation and communication, Algorithm 1 has good scalability with respect to the number of cameras and the resolution of the volumetric representation. In the following, we elaborate on different aspects of the distributed algorithm, including complete surface coverage, Figure 3. Illustration of job distribution scheme in 2D case. The light-green strip indicates the narrow band. The dark blue indicates the PRS of camera 1; the shallow blue indicates the SRS voxels of camera 1. The dark red indicates the PRS of camera 2; the shallow red indicates the SRS of camera 2. computation load balancing among cameras, communication optimization, etc. For the sake of clarity, Table 1 summarizes the main notation used in the following discussion; And Table 2 shows the main data structures that each camera maintains to support the distributed algorithm. The usages of these data structures is discussed below. **Algorithm 1 Camera-centric distributed algorithm for** scene geometry reconstruction 1: for each smart camera c, do 2: Compute the incremental updates Fv[c][,][ ∀][v][ ∈V][c][,] according to formula (11). If maxv∈Vc |Fv[c][|][ < ε] (where ε is a stop criterion threshold), then terminate. 3: Send Fv[c] [to all the cameras in][ C][v][,][ ∀][v][ ∈V][c] [through] a minimum-spanning-tree (MST) message passing protocol as described in Algorithm 5. 4: Update each voxel’s level set value according to formula (10), after receiving messages from the other tree branches of this node in the MST, as described in Algorithm 5. 5: Update the voxel set Vc as described in Algorithm 2. 6: end for ----- ### 3.1. Job Distribution Scheme In the level set method ([12, 14]), a narrow band implementation is commonly used to save memory and computation. It is based on the fact that only the voxels around the surface (zero-level set) contribute to the shape evolution. So in the implementation, a band Ω around the surface S is defined with an interval [DL, DH] on each voxel’ level set function value and only the voxels inside the band are updated (see Fig. 2). The price for this is that after each iteration the band should be updated to keep the new surface always inside the band through keeping a watching list of voxels, which keep track of the boundary of the narrow band. As illustrated in Fig. 2 (3D version) and Fig. 3 (2D version), we need to further divide the band into patches so that each camera takes care of one patch. Each patch should contain at least all the “core voxels” — those voxels around its contour generator defined by the contour generator indicator function. The set of “core voxels” are called the Pri_mary Responsible Set (PRS); (2) Each patch should include_ some “free voxels” — those voxels around the core voxels that are not taken care of by any other cameras. These “free voxels” hosted by camera c belong to the Secondary Re_sponsible Set (SRS) of camera c. To effectively distribute_ the reconstruction job among the cameras, there are three criteria that the job division scheme should address: PRSc ∈Vc (correctness) (12) ∪cVc = Ω (complete coverage)(13) |Vc| is approximately equal (load balance) (14) Eqn. (12) guarantees the correctness of the speed computation Fv[c][; Eqn. (13) ensures that all voxels inside the narrow] band Ω are updated. With the satisfaction of (12) and (13), the “free” voxels are distributed with the consideration of load balance among cameras with the Algorithm (4). **Algorithm 2 Update the voxel set Vc for each camera c** 1: Update the PRS of camera c as described in Algorithm 3. 2: Update the SRS of camera c as described in Algorithm 4. As described in Algorithm 1, after each iteration, for each camera c, its voxel set Vc (composed of PRS and SRS) should be updated. First each camera’s new PRS can be computed easily, given the new detected contour generator, through narrow band updating, as described in Algorithm 3. Besides PRS, there are other portions of the surface that are not covered by any camera. To ensure that these “free” voxels are updated correctly, we need to assign them to some host cameras. These “free” voxels are put in the SRS of their corresponding host cameras. There are two considerations in these voxels’ distribution: These voxels may belong to neighbor cameras’ PRS in the next iterations, so if we could put these voxels to these potential cameras then we can save the communications later; Another concern is the **Algorithm 3 Update the PRS** 1: % update the narrow band 2: for each voxel v in the watching list, do 3: **if its level set function value φ(v) ∈** [DL, DH] then 4: expand the boundary voxels by adding the neighbor voxels whose level set function’s absolute values are greater than |φ(v)|. 5: **else** 6: delete this voxel. 7: **end if** 8: end for 9: Update the contour generator indicator values 1Gc (v), ∀v ∈Vc for camera c. Put voxels whose indicator value is above a threshold TG into the new PRS. **Algorithm 4 Update the SRS for camera c** 1: % update the boundary list 2: for each boundary voxel v, do 3: **for each c[′]** ∈Cv, do 4: **if v ∈** PRSc′ then 5: delete v from Vc; Add its neighbors to the boundary voxel list. 6: **end if** 7: **end for** 8: end for 9: % At this stage, each boundary voxel has only two hosts. 10: % Now start pairwise load balance. 11: for each voxel v in the boundary list, do 12: c[′] = Cv\c, 13: **if |Vc′** | < |Vc| then 14: delete v from Vc, and add its neighbors to the boundary voxel list. 15: **end if** 16: end for load balance. Due to the non-uniformity of the surface and the distribution of the cameras, the size of the PRS for each camera is different. The existence of these “free” voxels provides us a leverage to balance the workload among cameras. The PRSs are fixed for the given surface and the cameras’ locations; The SRSs are flexible as long as together with PRS they cover the whole surface. We can take advantage of this to assign these “free” voxels to the cameras that have relatively small PRS’s. The workload balances are negotiated pairwisely by neighbor cameras that share boundaries, as described in Algorithm 4. The communications in Algorithm 4 happens in two steps: 1) communication between c and c[′] when checking v ∈ PRSc[′]; 2) communication between c and c[′] when checking |Vc′ | < |Vc|. Since this operation is performed for each boundary voxel, the communication cost is proportional to the number of boundary voxels. ----- (a) (b) (c) Figure 4. Illustration of a simple communication case. (a) the virtual communication path in the na¨ıve approach; (b) the physical communication path in the na¨ıve approach; The communication cost is 8 units; (c) the physical communication path in the MST case; The communication cost is 4 units. ### 3.2. Communication Optimization As discussed above, cameras need to communicate with each other locally to share information about their common voxels and dynamically assign work loads among cameras. Here we examine the problem of optimizing the communications between these cameras. From the above description (especially in (10)), we know that each voxel’s incremental update is composed of the summation of the participating cameras’ contributions. So the basic communication job is: sending each camera c’s incremental updating contribution Fc[v] [to all the other cameras in][ C][v][. Now let us analyze the] communication cost of the na¨ıve approach – each camera sends its own value Fv[c] [to all other cameras in the set][ C][v] directly. Suppose the communication cost between neighbor cameras in the graph is 1 unit. For a random graph, the average communication complexity for one message passing is O(D) = O(log(N )), where D is the diameter of the communication graph of the network. Then, the total average communication complexity is O(N [2]log(N )). The worst case for one message passing is N, with the worst total communication complexity being N [3]. Instead of sending Fv[c] [directly to all the other cameras] in Cv, there exists a more efficient way. Look at what each camera needs — the summation of Fc[v] [from all the partic-] ipating cameras c ∈Cv. Based on this observation, our solution is the tree message passing protocol, as described in Algorithm 5 and illustrated in Fig. 5. We store this tree representation of the CRN distributedly, through each camera maintaining a list of directly connected camera nodes for each voxel, as shown in Table 2. Why does the message passing work correctly for the tree structure? This is because there is “no loop” in the tree, which guarantees that cutting each edge will separate the tree into two separate subtree. And the message sent through the edge is all the summed information from the subtree. In this way, each node’s value is contributed to other nodes exactly once. Take the tree in Fig. 5 for example. For node j, it will receive message from k, l and i. And each message from Figure 5. Illustration of the minimum spanning tree message passing. Each node sends a message to one of its edges given the message from the other edges have arrived. k, l, i is the summation of the values in their subtrees {k}, {l}, and {i, m, n, o, p, q}. **Algorithm 5 Tree message passing protocol** 1: for each node in the tree, do 2: Compute and send message to one edge if the messages from the other edges have been received; 3: Otherwise, wait. 4: end for Next the communication cost of the tree message passing scheme is analyzed. For a tree with N nodes, there are (N − 1) edges and since we send information bidirectionally, the communication cost is 2(N − 1) units. Given the set of cameras Cv for a fixed voxel v, there are many trees that can be constructed; which one is the best? Given a weighted undirected graph G, we define a minimum span_ning tree (MST) as a connected subgraph of G for which_ the combined weight of all the included edges is minimized. In our case, the minimum spanning tree is the one that has the minimum communication cost. Since voxel updating is a key operation in the algorithm, the improvement on this operation will greatly speed up the algorithm. The tree message passing is very useful for distributed smart camera systems in general, since it is a common operation to summarize message in one subgraph and send it to the other branch. The tree message passing ensures that the protocol described above works correctly for tree topology structure. The MST can also be constructed and updated distributedly (See [2, 8, 5] for more details.) With this MST protocol described in Algorithm 5, we can see that each camera updates its own copy of voxels only after receiving messages from all its neighbor cameras. By this, there is no need to synchronize among the cameras after each iteration. Each camera runs its own algorithm and updates its own state only after it receives all the information needed asynchronously. And the synchronization is implicitly controlled by the message passing. ----- Figure 10. Shape evolution path of the David bust This dataset is challenging in two aspects. First the object is textureless, non-Lambertian, and the illumination changes (due to flash light), which challenges most multiview stereo algorithms based on intensity matching. Secondly the object is embedded in an natural indoor background. In the experiment on this dataset, the level set function is defined on a 64 × 64 × 64 grid, and the parameter µ is set as 0.05. The projection error for the David dataset is 0.37 pixels. The projected 3D reconstructed apparent contours are shown in Fig. 9. Fig. 10 shows the whole evolution path, starting from a cubic bounding box. It can be seen that the shape evolution process does converge to the object even though the background in the image is complex. A rough estimate of communication load and hence battery energy expenditure. As discussed previously, neighbor cameras communicate with each other 1) to figure out the ownership of the “free” voxels; and 2) to exchange information about the voxels’ update values. In these two experiments, the total number of iterations is set as 200. The narrow band-width is 6, so the number of voxels in the narrow band is the surface area times the narrow band width (approximately 3 × 10[4] for the Dinosuar dataset). Each voxel is maintained by 3 cameras on average. So the total number of voxels that all the cameras take care of is about three times that number: 6 × 10[4]. Since 20-23 cameras can cover the surface tightly, the number of “free” voxels is small compared to the size of the union of the PRS. So the message exchanged is dominated by the voxel updating messages. As shown in section 3.2, the total number of update values exchanged is 2(N − 1) ≈ 1.2 × 10[5]. Each update value is 4 bytes (stored in single precision format), the total number of communication bytes for each iteration is 4.8 × 10[5] = 48KB. With 200 iteration, the total data exchanged is about 13.6 MB. This order of communication cost is affordable in VSNs. ## 5. Conclusions and Discussion In this paper we define the problem and present a solution to the 3D reconstruction of an object, indoors or outdoors, from silhouettes in images taken by a network of _many randomly distributed battery powered cameras hav-_ Figure 6. Two sample images of the Toy Dinosaur dataset Figure 7. The reconstructed dinosaur shape Figure 8. Shape evolution path of the Toy Dinosaur ## 4. Experimental Results We first test the proposed algorithm on a public dataset, Toy Dinosaur [1]. In Fig. 6, two sample images out of a total of 23 images are shown. In this dataset, the background is relatively simple. The level set function is defined on a 56 × 120 × 96 grid, µ is set as 0.01 (a small value to prevent smoothing out the dinosaur’s high curvature parts). Fig. 7 shows the reconstructed shape after 200 iterations. From the results, we see that the overall shape is successfully reconstructed. Fig. 8 shows the whole shape evolution process, starting from a bounding rectangular box. It successfully converges to the concave parts, e.g. recovering the two hands, and separating two legs, etc. The reconstruction accuracy is measured by the projection error, which is defined as the distance between the projected apparent contour and the image apparant contour. The average projection error for this dataset is 0.21 pixels. The next experiment is on the David bust dataset which consists of 20 calibrated images taken by one moving real camera. Fig. 9 shows samples of the image sequence. 1This dataset is available at http://www-cvr.ai.uiuc.edu/ ponce_grp/data/mview/. ----- Figure 9. (top) Image sequence with indoor background; (bottom) Projected silhouette contours (in red) estimated by the algorithm, during the 3D reconstruction process, overlapped with the image edges ing onboard processing and wireless communication. The goal is reconstruction to close to the achievable accuracy while roughly minimizing processing time and battery usage. (More generally, other constraints may be present, e.g., communication bandwidth limitation.) The challenge is to use few image pixels in an image, to communicate as little data as possible, and for each camera to communicate to as few other cameras as possible. Our solution involves maximum a posteriori probability estimation for achieving close to optimal accuracy, introducing and using a dynamically changing vision graph for assigning computation tasks to the various cameras for achieving minimum computation time, and routing camera communications over a minimum spanning tree (MST) for achieving minimum communications battery usage. The main contribution of this work is the distributed processing of shape-from-contours, including the regionspecific vision graph, job division schemes, and the MST message passing protocol. The region-specific vision graph and MST message passing developed in the paper can be applied to other distributed vision tasks generally. The job division scheme is linked to the shape-from-contours approach more tightly, but the principals developed here, including the three constraints (correctness, complete cov_erage and load balance), can be extended to other vi-_ sion problems. For example, for shape from texture and contours, similar job division schemes can be developed by selecting most valuable image observations. The distributed algorithm proposed in the paper is not only applicable to smart camera network but also applicable to multiprocessor systems such as many-core CPUs and GPUs nowadays. We compute rough estimates of the amount of required computation and the required communication cost. The approach is appropriate for networks of very large numbers of cameras. ## References [1] I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks: a survey. Computer Networks, 38:393–422, 2002. [2] B. Awerbuch. Optimal distributed algorithms for minimum weight spanning tree, counting, leader election and related problems. In Proc. 19th Symp on Theory of Computing, pages 230–240, May 1987. [3] M. Brand, K. Kang, and D. Cooper. Algebraic solution to visual hull. In CVPR, 2004. [4] R. Cipolla and P. Giblin. Visual Motion of Curves and Sur_faces. Cambridge University Press, 2000._ [5] B. Das and V. Loui. Reconstructing a minimum spanning tree after deletion of any node. Algorithmica, 31:530–547, 2001. [6] O. Faugeras, J. Gomes, and R. Keriven. Geometric Level Set _Methods in Imaging, Vision and Graphics. Osher and Para-_ _gios Eds., chapter Variational Principles in Computational_ Stereo. 2003. [7] O. Faugeras and R. Keriven. Variational principles, surface evolution, PDE’s, level set methods and the stereo problem. _IEEE Trans. Image Processing, 7(3):336–344, 1998._ [8] R. Gallager, P. Humblet, and P. Spira. A distributed algorithm for minimum weight spanning tree. ACM Trans. on _Programming Languages and Systems, 5(1):66–77, January_ 1983. [9] P. Gargallo, E. Prados, and P. Sturm. Minimizing the reprojection error in surface reconstruction from images. In ICCV, pages 1–8, 2007. [10] S. Liu, K. Kang, J.-P. Tarel, and D. Cooper. Free-form object reconstruction from occluding edges and texture edges: A unified and robust operator based on duality. _PAMI,_ 30(1):131–146, January 2008. [11] K. Obraczka, R. Manduchi, and J. Garcia-Luna-Aveces. Managing the information flow in visual sensor networks. In _5th Symp. Wireless Personal Multimedia Communications,_ volume 3, pages 1177–1181, 2002. [12] S. Osher and R. Fedkiw. Level Set Methods and Dynamic _Implicit Surfaces. Springer-Verlag, New York, 2002._ [13] B. Rinner and W. Wolf. An introduction to distributed smart cameras. Proceedings of the IEEE, 96:1565–1575, 2008. [14] J. A. Sethian. Level Set Methods and Fast Marching Meth_ods. Cambridge University Press, 1999._ [15] A. J. Yezzi and S. Soatto. Stereoscopic segmentation. In _ICCV, 2001._ -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/CVPR.2009.5206589?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/CVPR.2009.5206589, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "http://perso.lcpc.fr/tarel.jean-philippe/publis/jpt-cvpr09.pdf" }
2,009
[ "JournalArticle", "Conference" ]
true
2009-06-20T00:00:00
[]
8,753
en
[ { "category": "Business", "source": "external" }, { "category": "Environmental Science", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00fe625102de79328a9ee3bf867afa05f8be82d7
[ "Business" ]
0.907663
Blockchain-based multi-organization taxonomy for smart cities
00fe625102de79328a9ee3bf867afa05f8be82d7
SN Applied Sciences
[ { "authorId": "1581528660", "name": "Ekleen Kaur" }, { "authorId": "72563368", "name": "Anshul Oza" } ]
{ "alternate_issns": null, "alternate_names": [ "SN Appl Sci" ], "alternate_urls": null, "id": "4766beb3-bec6-4b4a-a5fc-cf12ba2339b6", "issn": "2523-3963", "name": "SN Applied Sciences", "type": null, "url": "https://www.springer.com/engineering/journal/42452" }
null
**Research Article** # Blockchain‑based multi‑organization taxonomy for smart cities **Ekleen Kaur[1] · Anshul Oza[1]** Received: 14 October 2019 / Accepted: 5 February 2020 / Published online: 18 February 2020 © Springer Nature Switzerland AG 2020 **Abstract** With the tremendous development in distributed ledger technology, assimilation of tokenization in sustainable assets has been a proof of concept. This paper is a documentation of ERC20 standards where cryptocontracts are an evident implementation of WRC tokens. Issuing and earning of these token credits rely on aggregate hazardous waste released into water as a by-product and careful monitoring of water quality standards (after treatment) using IOT sensors. The minting of this token currency is put forth by exchanging ether. This research is twofold attempt to eliminate the differences between Small and Medium Enterprise and large-scale Enterprises and to establish a business ground with equal opportunity of earning credits based on recycled wastewater. **Keywords Blockchain · Water analysis · IOT sensing · Permissioned blockchain network** ## 1 Introduction Distributed ledger technology has progressed after Satoshi Nakamoto’s offered solution to double spending problem by implementing peer-to-peer transaction after mining the first block [1]. Blockchain technology has continuously evolved since then; blockchain is nothing but cryptographically linked blocks storing information. Each block contains a cryptographic hash of the previous block, a timestamp, i.e., the time at which the transaction was confirmed after successful mining of blocks and transaction data. With the advent of blockchain came the concerns about the anonymity of miners and the organizational theory of decentralization [2]. Decentralized governance revolves around transparency and trust among the members. Blockchain has 4 major features: _Immutability Once the data of a transaction is confirmed_ and recorded, it can never be changed or amended. The same asset can undergo various other transactions adding to the list of confirmed recorded transactions, but the state of those confirmed transactions remains immutable. _Provenance Immutability of recorded transaction gives_ provenance of assets, providing an entire history of the asset, including where it has been, who or how many members have owned the asset previously and so on. _Consensus This weeds out potentially fraudulent trans-_ actions out of the database. A transaction cannot be confirmed on a blockchain without consensus. The consensus is very important for the validation of a transaction. _Distributed It urges various organizations to share and_ exchange data. _Smart contracts are the building pillars of blockchain_ in a business organization. Inside distributed ledger technology, we audit the smart contracts built for blockchain, at which it is designed to be implemented [3]. Ethereum is a well-known public as well as private blockchain network. Ethereum demonstrates a successful implementation of a complex Merkle tree, i.e., Merkle Patricia tree. Ethereum works on PoW (proof of work), PoS (proof of stake) and PoA (proof of author - Ekleen Kaur, ekleenkaur17@gmail.com; Anshul Oza, ansuhuloza@ipsacademy.org | [1]IPS IES Academy, Indore, India. SN Applied Sciences (2020) 2:440 | https://doi.org/10.1007/s42452-020-2187-4 V l (0123456789) ----- ity) consensus algorithms. In terms of efficiency, PoA is considered more robust than PoS [4]. PoA allows non-consecutive block approval from any one authority. Ethereum is usually taken to a permissionless blockchain network just like bitcoin. A permissionless blockchain is a network in which anyone can become a member and take part in consensus, add a new block and confirm new transactions. A permissioned blockchain allows ‘any node’ to take part if the identity and role are defined, whereas a private blockchain only allows ‘known nodes’ to participate in the network. ERC20 is the standard used for smart contracts on the ethereum blockchain for implementing tokens. Crypto tokens are a special kind of virtual currency that represent an asset or utility. Tokenization of assets [5] on blockchain is creating tokens using smart contracts based on some standards, i.e., ERC20 in this case and regulate those tokens by writing functions in the contract that allow the token to be traded and exchanged in the form of currency or in the form of an asset. Each token item is a hash value representing a crypto asset, which is under the consensus of all consortium members. Typically, a token contains both direct access and user authenticated access. Hash value in the token is made available only on the regulator’s or owner’s authentication. ERC20 token standard is complemented by ERC223 [6] and ERC777 [7]. Just like ERC20, there are many other smart contract-based token standards such as ERC721, ERC1155, NEP5, NEP11, QRC20 although ERC223 and ERC777 are very analogous to ERC20. Tokens dealing with smart contracts involve two function calls under ERC20, and on the other hand tokens working on ERC223 and ERC777 fashion involve only a single function call. Despite a few advantages, ERC20 is still prominent over ERC223 and ERC777. Now the implementation of code in a smart contract is done using solidity. Solidity is a high-level contractoriented programming language. It is used to build smart contracts. Solidity works on the principles of object-oriented programming just like any other programming language like C++, Java, Python, etc. Smart contracts are necessary for the regulation of a business network for the purpose it was meant to fulfill. Tokenizing sustainable assets are being developed under great examination, such sustainable assets contribute to waste management and establish sanitation. Improperly treated wastewater is one of the largest sources of water pollution in India [8]. Water quality and pollution are generally measured in terms of concentration [9]. So, waste management mainly revolves around the non-biodegradable constituents which are required to be treated. Overview of some major types of wastewater chemical contaminants include: (i.) _Total Dissolved Solids (TDS) Comprises of inorganic_ salts and small amounts of organic matter dissolved in water. (ii.) _Biochemical Oxygen demand (BOD) Amount of oxy-_ gen required by aerobic microorganisms. (iii.) _Chemical Oxygen demand (COD) Oxygen equivalent_ of organic matter content susceptible to oxidation by a strong chemical oxidant. Identification of water quality is done, based on some standards: - _PH Scale used to specify the acidity of a solution. Nor-_ mal water has a PH nearing 7. - _Turbidity It is the measure of the degree to which_ water loses its transparency due to the presence of suspended particulates (size greater than > 1000 nm). Turbidity is measured in NTU’s which stands for Nephelometric Turbidity Units. - _Biochemical Oxygen demand (BOD) 3–5 particulates per_ million for normal water. - _Total Dissolved Solids (TDS) For normal water, the value_ of total dissolved solids is 300–500 mg/liter. - _Temperature (°C) This property has a varying range_ based on the type of microorganisms perpetuating. Industrial water has psychrophiles in 0–20 °C, mesophiles in 10–45 °C. Normal water body has a temperature of 13 °C. - _Hardness and oil/grease Normal water ranges less_ between 45–46 mg/l and 5–6 mg/l correspondingly. Now that certain quality standards have been established, the question arises about what to monitor the water quality standards with? As sensor technology advances the whole system of everyday challenges is easily implemented with advanced solutions. Internet of Things (IoT) possess ability to transfer data over a network at high speed with great accuracy. IoT sensors are devices that record data and provide new insights by monitoring data. IoT sensors play a big role in monitoring the water quality, the reason why it is significant to convert the data recorded by the sensors into useful information and to deploy this information for interaction between stakeholders [10]. ### 1.1 Challenges with wastewater management - _Inadequate treatment of water Since there are organi-_ zations that are already recycling wastewater, these ----- organizations do not realize the need for amending the current systems. Due to the inappropriate treatment of wastewater, the water quality standards are never met that can be reused by the company to profitable amounts. - _Lack of careful monitoring The major reasons for inad-_ equate treatment of water is due to the lack of an efficiently established monitoring system that can regulate these organizations and regularly keep the water quality standards under check. - _No incentive earned As of now, India lacks a well-estab-_ lished wastewater monitoring system that can provide an incentive to the organizations that are a part of the business network. - _Freshwater Dependency still intact There is no major_ monitoring of the treatment plants, so the resultant recycled water is of no major use to the organization itself, due to which the freshwater dependency of such organizations is still intact. Correspondingly, the probability rises that the major pollutants are still present in water. ## 2 Literature review Swan [11]: Melanie Swan writes the manifestation of blockchain and how it is soon going to take over the economy, the crucial necessity of decentralization in business, also token exchange that has been evolving since the arrival of blockchain and how vast the magnitude of the functionality of a smart contract in business is. However, blockchain technology is only limited to theoretical terms, and its potential is defined without a scope of the actual implementation. The book is a beginning ground for the new economy that majorly lacks the actual complexities of guidelines to implement blockchain tokenization on a protuberant scale; the real-time challenges are just briefly mentioned. Mougayar [12]: William Mougayar surrounds blockchain around exploring the new ways in which it is capable of extending its service in business and economies, solving the contortions of Melanie Swan’s [11] work. In our research, our real-time asset, i.e., wastewater is getting monitored by IOT applications and traded as tokens; we cite this book to further blockchain’s immense capability in business to resolve larger problems that current centralized authorities are facing in the present time such as maintaining water contamination and establishing real standards to monitor the treatment of wastewater. William Mougayar’s paper doesn’t discuss the shortcomings of extending blockchain as a service or provide real-time solutions to those shortcomings. Duca et al. [13]: The paper states the self-purification of water, hydrochemicals that are essential for determining water quality and how the concept of mineralization works, a very detailed illustration of ­H2O2 as an oxidizing and reducing agent and radical formation across chemical bonds, modeling of these chemical links in the interim of redox reactions along with the role of transition metals in quasi-reducing state. The author pens down the fact that the dependency of self-purification on OH radical is directly proportional and biotic components in water play a key role in self-purification. Gheorge Duca’s Paper in the year 2008 points to the efficiency of the detailed water purification processes; blockchain’s architecture for smart cities as discussed in our research empathizes with this fact of monitoring the water purification standards while growing the business economy. We engross the acceptance sampling to target the set of organizations with previously set up water purification management systems, and our research aims at monitoring this efficiency in a business network as we discuss in Sect. 3 of this paper. Echard [14]: The data in a business network is prone to attack; the author talks about the importance of encryption and the chances of vulnerable data, the need for security keys and the decryption of data. The conclusion speaks of hazy ideas regarding the security of data over a network. Due to the presence of malicious crackers, data are never too safe or completely secure on a network. Schaad [15]: This paper is an additional focus on devices that form a major part of Internet of Things today. Data modeling using binary data, the signing and encryption standards of objects (COSE), after amending from JOSE, well-defined basic COSE structure, multi-signature on objects have been a part of the modern updates. Signing and verifying altogether with encryption algorithms, the author also explains signature algorithms while defining security aspects and authentication. COSE’s registries and Media Type explain the usage of keys in IOT-based applications and are very necessary while maintaining security standards. Echard [14] mentions data security on a network connecting IOT devices; Schaad [15] advances his work by COSE protocol encryption for cryptographic signing of data, enhancing data security; COSE protocol follows layer encryption, i.e., content and recipient layer. Our work uses cryptographic encryption for ethereum blockchain which uses Elliptic Curve Digital Signature algorithm as a key pair encryption standard for unique public and private key pairs owned by each organization. Hong et al. [16] exemplifies policy-making complication given a local government that stands as a regulator, emphasizing schemes and target-based production results. The game model theory puts wealth and welfare with equal importance. The paper promotes the reduction ----- of hazardous emissions at the stake of adopting green technology inside big firms. The complexity of the problem case is analyzed depending upon the initial allowances which are expected to be clearly defined. Hybrid algorithm composes of polynomial dynamic programming (PDP), genetic algorithm (GA), and it also consists of binary search in matters relating to the efficiency of the business model on the ground of policy making. In the business trade, the paper states about firms exchanging allowances which are used either to sell or to buy; this theory is functioning in our research. Consensus on blockchain ensures decentralization, so in a decentralized business network exchanging of these allowances is aiding to the circulation of the token economy that resolves the big question of the token trade between organizations. Zecchini [17]: This paper extends the above-proposed business model governed by a regulator in its research. The author clearly defines the problem statement of water impurity standards and how IOT can help in revolutionizing data elucidation. The merits and demerits of a regulator organization like the government are well elaborated and authentically structured; the entire thesis of the storage of data and transmission is through low power wide area network and efficient comparisons with other data transmission techniques and its architecture. Cryptographically encrypting data storage while laying out the merits of blockchains, the analysis of a permissioned blockchain over a permissionless blockchain, the rules for establishing quality credits to offer incentive in exchange of water quality has been just theoretically defined by the author. His future works include the implementation of permissioned blockchain for quality credit. Our research on multi-organization blockchain taxonomy establishes proof for practical implementation of smart contracts over permissioned blockchain network and a profit formula for anticipating quality credit along with the token price. Silvestre [18]: supply chain ambiguity and stimulating factors are certain management challenges faced by the economies which eventually hinder the sustainable growth of such economies. The research has a description on the geographical factors that affect the social demands and their roles in supply chain. Even though the rise of globalization in complex business networks is facing high uncertainty, all are promoting stages in aggrandizing green technology in sustainable business networks. Klassen and Vereecke [19]: Silvestre’s [18] work in sustainable development of economies in supply chain results in expanding business which is associated to certain complexities. Compared to Silvestre’s work, this paper proposes that supply chain risks and social responsibility are inter-related terms. The social intervention leads to increasing risks alongside opportunities. Social responsibility must include close monitoring of customers’ demands and supplies. Collaborative decentralization offers flexibility of workflows and processes; organizations with too many regulations face human health challenges. Innovation strategies involve management policies and product innovations that legitimize performance. Responsibility extends to auditing social standards by stakeholders that include alignment of risks and profits combined. We differ in this context because auditing social standards in order to increase sustainable business needs to be maintained; smart contract in our business model maintains those standards. The most important factor is which organization owns what quantified amount of the sustainable asset chosen in the model, i.e., the waste water token. Li et al. [20]: Tokens are classified depending upon the work it was supposed to fulfill. Utility tokens are the currency base of applications, determining access control of the application. Security tokens derive its value on the top of blockchain, externally, but it can be subject to mandatory surveillance and supervision. Asset-backed are tokens converted from real and virtual assets under open asset protocol, a policy backed token offering privacy conservancies on data. Recorded data on the ledger can be only traded. Our research uses a real-world asset as an assetbased token implemented on ERC20 standard with the token contract laying the rules of ownership. The contract provides proof of ownership which can be transferred if the ownership of the recorded data is specified with its quantified amount given that the amount is less than the total available token balance of that particular time stamp. Zhou [21]: Data ownership in a permissionless blockchain is proposed using a Dlattice architecture of a Double DAG, as the presence of an Account DAG data security is immune to influence from other accounts. DPOS-BA-DAG protocol establishes decentralization by consensus in the presence of forks. Tokenizing data is given because of Dlattice in the paper with consensus in the time period of just 10 s. The author compares the economic incentives of Algorand protocol, a large amount of signature data, and establishes comparisons between various consensus protocols like fault Tolerance, Ouroboros, etc. Dlattice aims to extend healthcare with the help of IOT, exemplifying our research on sustainable assets; however, IOT supervision in our business model is a part of permissioned network that uses proof of authority consensus. According to the paper, the Account DAG structure allows same user to own multiple accounts where each account has a separate public key as an identity, but this is not the current possibility in our blockchain architecture because multiple accounts owned by the same organization enterprise is irrelevant and it may lead to loss of track on records on monthly quality credits and penalties issued to the organization. Sánchez-Corcuera [22]: The smart city application faces many challenges; sustainable assets cater risks that ----- can overweigh its benefits as discussed by Klassen’s [19] research in 2012; these risks are only going to augment in the future. Urban planning is comprehended by sustainable factors revolving around the environment. Waste management, as stated by the paper, discusses different approaches like smart containers and new infrastructures but does not include authority access and data transparency, which is a major feature evident for smart cities. Data tokenization, consensus on blockchain, is a big solution to these problems. Kundu [23]: Kundu Debasish supports and proves Sánchez-Corcuera Ruben’s [22] work; transparency is a key factor for an efficient smart city despite the available set of technologies carrying the society in the imminent time. To establish transparency and trust which is the building of any business contract, smart contracts or cryptocontracts direct the workflows and access control of transactions, which we implemented for tokenizing WRC—wastewater recycling certificates. Access control has a lot of contortions in token trade, defining the rules of data ownership [20], such as - Who has the right to mint ether, - Who has the right to withdraw back from the token amount back to ether, - Who can send and regulate transactions, - What is the minimum token price each organization must own in order to be a part of the network. The author elucidates the importance of transactions without third-party intervention liquid economy that tokenizes land base assets allowing the transferring of such assets at a much faster and safer state. We focus our findings on the basis of the four layer topology as discussed by Kundu Debasish. Kouhizadeh [24]: Circular economy in a supply chain using blockchain technology can help monitor and keep a record of previously deleted products, by storing the record of each transaction on the ledger and helps keep track of all updates. Circular economy subsidizes exchanging of products. Blockchain’s shift over renewable energy has helped reduce emissions on environment. The author states about waste exchanges supplementing blockchain. Accessing products, transaction records, final stage data, by-products are all potential aspects of blockchain based waste management. Data tokenization managed by local governments or semi-government bodies namely regulator is the prime proposal of our research. Detailed elucidation of complex inter-relationships advancing toward development in circular economy which are defined in terms of participant owned assets and regulator regulated transactions. One of the major differences is that the author claims on using blockchain as an alternative for product deletion, whereas our major closure is on using blockchain as a restate for resource elimination; wastewater cycle extends the capacity of these potential resources for a more optimized utilization by organizations. Savelyev [25]: Kouhizadeh Mahtab proves discarded products in a supply chain can be regulated as a circular economy on blockchain [24]; Savelyev discusses these blockchain economies in association with the legal aspects ascending within such businesses. The paper clearly discusses tokenization of assets that can be registered on a network working on blockchain; tokens function according to the rules essential for the business network in a blockchain architecture. What is evident is that the author discusses about the possibilities of providing tokenized assets some cognizable rights that are recognized for every token economy to inherit and transfer. It is lucid from the paper about the existence of certain anomalies related to the ownership of blockchain tokens that are still being analyzed to legalize tokenized assets in the future. Our research manifests blockchain mining; sustainable asset solves the problem of ownership [20] of the tokenized asset at the time of transactions which eliminates a major research gap, since our research is not tokenizing private property or objects owned by organizations. Blemus et al. [26]: Increased transparency with the intervention of distributed ledger technology has improved governance. Token economy is raising ICO (Initial Coin Offering) investment; ever since the distributed ledger technology has stepped into business the amount raised has been progressing on the graph. The author mentions token rights specific to the issuing of tokens; our research is apropos for providing rights to identity of ownership of tokens that are issued by the regulator symbolizing a successful quality check and utilization of wastewater. Corporate governance using tokenization on blockchain is just a theoretical survey in the paper, without signifying major protocols involved. Similarly the paper briefly mentions about the establishment of a relationship between tokens and economic/noneconomic ecosystem by proposing IOT as a solution medium to develop that link but there is no further progress in the research in this regard. Roth et al. [27]: Blemus et al. [26] and Li et al. [20] classify tokens as security, utility and crypto tokens similarly. The author makes an interactive comparison between UTXObased, layer-based and smart contract-based tokens. The paper contemplates the impact of tokenizing assets on blockchain by eliminating the need of a platform provider or a third-party authorization for transactions to enforce. The paper briefly mentions regulatory issues of legal policies but majorly focuses on crowdfunding equity using blockchain. Wastewater is not only quenchable to the need of crowdfunding in our research but also retorts the problem of asset ownership even though it is not a ----- catechizing factor in our research. Our research highlights toward the innovation of using wastewater as a circulating business economy. Crowdfunding is a major problem in the business economy; the author proposes solutions to the problem of crowdfunding using blockchain. Blemus et al. [26] points to the monopoly of token stakeholders that form a major part in any corporate governance, and Roth et al. [27] contravenes this fact by providing an analysis of market captalization of various blockchains with ethereum token market capitalization at the maximum; his findings justify the safety of assets from arbitrary manipulation and monopoly of token holders in a network using blockchain; the end of this section sums up our differences with this research. Sanghavi et al. [28]: The author proposes to employ the asset tokenization in quotidian phases and decision making in an organization, with just a glimpses on the fact of an actual possibility. Our research to bolster the above facts focuses on implementing the amalgamation property of cryptocontracts and proposes a touchstone to enhance governance in the regulation of enterprises or organizations that lobotomize over time. Taking the above facts into consideration, the author acknowledges possibilities without results. The author lacks significant results regarding decision making process within the business network and makes predictions without any expert review which has no value. Prince Michael von und zu Liechtenstein [29]: According to the paper policies to be put forth under the blockchain law, as planned and argued by the Liechtenstein government, majorly involves the embryonic idea of tokenbased economy. The vast ground of asset tokenization has extended to a large scale in the context of business; cognizable blockchain rights ameliorate degree of trust amidst organizations. Apart from immutable data and security due to data enciphering, blockchain law contradicts all the belied theories of losing copious amounts of money due to bugs or lack of proper auditing of smart contracts that in the past has resulted in the loss of millions of dollars. The author extrapolates tokenization as a method of establishing a trust to bolster entrepreneurial participation in blockchain environments. There is a major research gap; it doesn’t provide an actual plan on how the token economy will be regulated once the law is imposed. Davydov and Khalilova [30]: Unlike Liechtenstein’s [29] proposal of cognizable rights for blockchain, the author mentions the incrementing possibilities of money acquisition for the banks and also P2P market services by proposing a business model creating entrepreneurial opportunities. The research gap is lucid from the paper for not discussing any of the legal aspects indulged with the bank loan amount upon tokenization. The author’s finding attenuates the research because of a lack of sufficient result that could actually supplement the theory proposed or some amount of implemented result that could dwell the readers’ trust on the less probability of money lost due to tokenization, which the author fails to accomplish. There is a significant probability of money getting lost in the proposed business model; it is not reliable under load conditions. Levin [31]: The author mentions about the low liquidity property of smart contracts. The four steps involved in asset offering include the process of digitization, tokenization, asset trading and dealing. With the instance of APT tokens in the paper, the author mentions the contingency of escrow accounts that hold in actual assets compatible with the ERC20 standard. Some key emphasis is on information transparency that enhances trust and promotes the sales market; tokenization is cogent in the present-day economy because it deals with the problem of high capital investment. Escrowing assets save the property from rental scams. The author discusses the potential of cryptocurrencies to be traded as real-time assets but the research gap is that it lacks many possibilities and the instances fail to sufficiently extend the theory. The author has provided no relevance of multi-signature wallet in escrow account holdings to uphold its credibility. To sum up, the above papers extend various aspects of blockchain; they deal with many of the incoming challenges rising with the diversification of blockchain in the modern day ecosystem. This research is another aspect to extend blockchain’s capability for business; however, none of the aforementioned research builds token economy with IOT using a sustainable asset. We tokenize wastewater on ethereum due to its largest market capitalization; to the best of our knowledge, this is the first research that builds such a business model on private ethereum network using POA consensus. We deal with the challenges on introspection; IOT monitoring of water quality in a POA network has a regulator; Proof of Authority Consensus demands the presence of authorities to approve transactions for the issuing of quality credits. For this regard, we concern those organization having favorable outcome in the quotidian scrutiny. Section 3 of this paper provides our work for smart city modeling; finally, we provide our discussions in the conclusion section that highlight the crux and summarize the research. ## 3 Methodology Sustainable energy credits are composing into business. Tokenizing these energy credits for business on ERC20 fashion has some verdicts that need scrutiny. India has many organizations utilizing freshwater for commercialism ----- with detailed and well-structured water treatment platforms. Problem case - The dependency on freshwater is still intact. - There has been no considerate curtailing of waste emission in water bodies. - There is a need for regulator [16] that can monitor and examine the water quality standards. - The organizations purifying wastewater need to reuse that treated water for merchandising. - We need a profit formula for the reuse of corresponding recycled water. - Large-Small organizations combined have a huge difference in water intake, so there is a need to standardize profits. Based on these conglomerate quandaries, this research tokenizes WRC using cryptocontracts from ether. WRC is the token minted from ether. Any token created from ether using ERC20 mandates a token price. Token price is the price of token exchange for minting coins, while minting ether into WRC tokens. So, while trying to convert 1 ether into WRC coins, the token price is taken to be 14th power of 10. Minting 10000 WRC coins for one Ether. _Tp = 100000000000000 Wei_ _Ts = WRC​_ _Tn = wastewater Reprocessing Coins_ where Tp is the token price, Ts is the token symbol, Tn is the token name. Organizations exchanging coins by minting and withdrawing WRC standardizes trade in the business network. The problem case mentions a profit formula that issues this WRC as quality credits for reusing wastewater. Regulator or the local government benchmark for issuing these tokens is the regulator set ground zero, minimum percentage reuse of recycled wastewater. The regulator set ground zero is the arbitrator of quality credits. Such qualified organizations contribute to the maximum supply of coins and trade tokens amidst the organizations that fail to meet the quality benchmark. For the quality benchmark to be set, the benchmark must be decided by auditing data of water quality acquired from the treatment plants. The data readings are received every hour with lead acid battery backups to restrain it from losing contact with the server because of major challenges while storing data is the problem of data tampering. It is very important for the IOT sensors to stay in contact with the server for reading data; this data are gathered in the database, and the percentage reuse is calculated based on this hourly received data. The percent reuse determines the quality credits earned. The break in data readings is the fact behind data tampering resulting in heavy penalties or no issued coins for the month. In this way, efficient monitoring of water quality along with data tampering is accompanied by earning tokens. The user interface provides the percentage to the cryptocontract where the transactions are defined. The profit formula varies according to the proportion of varying organizations above and below the ground zero. Percentage reuse standardizes the phenomenon of equal competence and acquisition of quality credits, which helps big large firms with the annual dependency of several thousand gallons of water trade with SME’s (Small Medium Enterprises) requiring only a few gallons of water intake. For instance: Given a set of sample data to elucidate the profit formula. Organization 1: This is a chemical industry that requires 25,000 gallons of freshwater, emits 10,000 gallons of wastewater and reuses 6000 gallons after recycling the water that has met the quality standards. The percentage reuse of this Chemical Industry sums up to an approximate of 60%. Organization 2: This is a leather industry that requires 8000 gallons of freshwater, emits 2000 gallons of wastewater and reuses 800 gallons of water after recycling the water that has met the quality standards. The percentage reuse of this Leather Industry sums up to an approximate of 40%. Organization 3: This is pharmaceutical industry that requires 2000 gallons of freshwater, emits 800 gallons of wastewater and reuses 300 gallons of water after recycling the water that has met the quality standards. ----- The percentage reuse of this Pharmaceutical Industry sums up to an approximate of 37.5%. The above statistics bolster the fact that, - any type of corporate sector depending upon freshwater and - any firm of large/small/medium scale contribute equal opportunity of fair trade in the business network. The water quality monitoring is examined in the database and for the organization fitting into those quality checks the regulator either issues monthly exchange of WRC or receives a token amount of penalty. Internet of Things administers water quality using sensors. Sensors record the data to check it in the realm reach of normal water quality; the UI connects to the blockchain node and the frontend to the database. Some common sensors used in this perspective are: - Turbidity sensor - BOD—biochemical oxygen demand sensor - PH sensor - Volume sensor - Temperature sensor - TDS—total dissolved solids sensor There is a detailed reason behind using individual sensors over multi-sensor water quality sensing appliances like YSI ProDSS/Pro 20, H198193, Proteus, etc. 1. The cost and maintenance of a multi-sensor are comparatively higher than installing individual sensors. 2. The range of individual water quality-testing sensors is higher than a multi-sensor. 3. Every individual sensor is dependent on one another. If one stops working others cannot function; to repair the faulty sensor even the other sensors can’t be used for sensing to record hourly data. The mechanism of IOT sensing is that on any physical change received during the working of the entire model, the sensor is calibrated to detect and send the digital/ analog signal to the central unit, i.e., IC ESP32, which in turn further connects to the webserver using wifi chip and updates the data result at the unit interval of time. What concerns the regulating of participant organizations is the fluctuation around regulator set at ground zero, the arrival of an organization B below ground zero and the arrival of organization C above the ground zero. This fluctuation is the bottom layer architecture of the cryptostocks, this architecture determines the present value of the token price. Understanding profits comes with the count of organizations. The profit gain formula can be considered as a ratio as follows: let Pr be the Profit Ratio given as under, Pr = Oa∕(Ua + Oa) given Pr > 50% where _Oa is the grouping count of over-achieving_ organizations, Ua is the grouping count of under-achieving organizations. _Loss ratio: This damage is equivalent to the above profit_ ratio extending in one criterion. P L r = r where Pr < 50% So, ----- Lr = Oa∕ [(]Ua + Oa Lr < 50 % ), The deviations in the above ratio are linked to the cryptostock. WRC is a token on ERC20; therefore, this token corresponds to a token price. The regulator defines the token price as a firm value standardized by the value corresponding to the ground zero percentage ratio. Let Sp be the stock price, _Sp has a constant value at ground zero equivalent to Tp,_ when the Pr ratio is equivalent to 50% S T P p = p ± r _where Pr is > or < 50%, when Pr < 50% then Pr = Lr._ For instance: Organization B in the bar graph is counted as an underachiever with a percentage reuse of 20%; similarly, organization C is counted as overachiever with the percentage reuse of 60%. assuming this data the stock price won’t be affected by Pr as Oa = 1 Ua = 1 Pr = Oa∕ (Ua + Oa) = 1∕2 i.e. 50 % Note: The above formula for calculating Sp is valid only when Pr is not equal to 50%, i.e., Pr ≠ 50% The issuing of quality credits in tokens will be provided to the respective groups (Oa or Ua) depending upon the percentage reuse calculated. Organization IOT Meters 1 1 Server Organization IOT Meters 2 2 Organization Database IOT Meters N N The application of IOT sensor in the wastewater monitoring integration involves an analysis of the water quality standard for normal wastewater after treatment. The volumetric analysis is done before and after the treatment of wastewater. The conformity of the percentage reuse of wastewater in the issuing of tokens is in regard to this set of gathered readings from each of the participant enterprises. Enterprises in the business network earn quality token credits, based on this endorsement ratio. The set of gathered readings from each of the sensors is required to pass a quality probe in every unit time. The concluding volumetric analysis is the average of those volumetric readings that pass the quality check during the entire month in every standard unit of time intervals. Percentage reuse is given by The ratio of volumetric analysis before and after the treatment of wastewater. Let Re be the required ratio. Re = [(]V ∕ VA) ∗ 100 where V volume of wastewater, VA volume after treatment, _VA is the average of those readings that pass the quality_ probe given the sample table to elucidate the sensing process. VA = 횺(volume analyzed with acceptable quality)∕n where n is the No. of times the quality was maintained out of the total no. of readings (N) taken, such that n > N/2, i.e., 50%. Vol- Result ume Oil sensor TDS sensor COD sensor BOD sensor Temperature sensor S. Turno. bidity sensor PH sensor 1 ✓ ✗ ✓ ✗ ✗ ✓ ✗ _Va_ Fail 2 ✓ ✓ ✗ ✓ ✓ ✓ ✓ _Vb_ Pass 3 ✓ ✓ ✓ ✓ ✓ ✓ ✓ _Vc_ Pass An overview glimpse of the deployment of IOT sensors in the business model to test the water quality probe and the application of IOT sensing can be taken into consideration by the following the diagram; the feature of esp32 microcontroller is that it can be programmed with wifi and bluetooth. For obtaining digital signal output to the analog sensor, we use an ADC converter, IOT meter port to connect our sensors: turbidity, ph, temperature, etc. Triac T1 BT138-V for output pulse modulation. IC2 7805T for voltage regulation. |Organization 1|IOT Meters 1| |---|---| |Organization 2|IOT Meters 2| |---|---| ----- ## 4 Limitations - The token is built on ERC20 which means its dependency on ether prohibiting full user access. - The cost of ethereum nodes mining the blocks vary according to demand and supply; this indirectly affects the minting of the tokens at varying prices according to Ether. - The token price fluctuation is explicitly elucidated, but _Pr is not the only deciding factor affecting Sp. So token_ price is not completely determined by factors inside the business network. - The cost associated with the setting up of IOT sensors for a large number of organizations might be huge. This involves a large amount of capital invested for IOT sensors; these sensors are then fitted in the pipe holes and openings to record and monitor the task it is calibrated to perform. - After huge capital investment, the IOT sensors also include the cost of maintenance; setup cost is not the prerequisite alone. To attain the desired set of efficient results and accurately measure data, the IOT sensors need a different budget for its maintenance. The architecture after set up needs to be periodically inspected and checked, any requirement regarding waste accumulation around the sensors has to be dealt with the immediate action, so that the recording of data doesn’t get affected and business network prolongates to serve the fair trade. - PH sensors used in sensing might not efficiently record data as it was expected to after 2–3 months because of the accumulation of layers of embedded salt around it. This problem can only be either replacing the IOT sensor or ensuring regular inspection and cleaning around the sensor body. - The organization where the sensors are deployed needs to have a continual internet network to stay connected to the server while reading the data. - The research does not implement or provide a solution to the cross border challenges for business in countries with varying legal regulations and degree of acceptance for tokenizing blockchain economy [32]. ## 5 Conclusion With any observed physical change using the IoT sensors, the record of our data maintaining the water quality gets amended with the server. This eventually is directly concerned with the marking of standards in the growth statistics of each of the business organizations, composing a part of the network. The contribution of quality credits is merchandising blockchain for business on a protuberant scale. Encrypting the identities of organizations by the cryptographic hash of the account address in ethereum manifests security. This research on ethereum blockchain coalesce trade is rooting on sustainable assets like wastewater. A private network on ethereum provides limited record of transactions happening only within the network. Our readings put in proportion the percent endorsement of reuse from the regulator set ground zero. The tokens are issued based upon this proportion, i.e., the percentage reuse of wastewater, featuring the highlighted part of this research alongside the factors affecting Tp. What needs a careful thought is the prominence of the profits for the use cases that involve organizational trade. After the successful completion of this research, we pen down the holistic reported results. We validate our business logic deployed on a permissioned ethereum network on the POA consensus. ----- Using the quality benchmark of 35%, we establish the relationship and dependency of two important token parameters, i.e., Sp and Tp in the business network. According to the current business logic, all the organizations can mint and withdraw tokens; however, the purpose of research was to implement this innovation to build smart cities, so as per real-time deployment challenges and for better administration over decentralization, the business logic is expected to concede with the regulator’s consent for minting tokens from ethers at the organizational level of deployment. Future works are on 2D/3D data visualization on the IOT meter data clusters to minimize the probability of tampering; however, these real times challenges are not a hinderance to the research at this level. A proportion of each organization’s profit with the volumetric dependency on fresh water, after successful reuse of treated water is another fact into consideration. Further research might consider the use of Zigbee and 6LoWPAN protocol for connectivity and information transfer between the IOT sensors for featuring automation in this corporate taxonomy. ECP32 is just demonstrating an overview of the sensing application, a strong determining factor of the incorporate future changes will occur according to the location of deployment. ### Compliance with ethical standards **Conflict of interest On behalf of all authors, the corresponding au-** thor states that there is no conflict of interest. ## References 1. Nakamoto S (2009) Bitcoin: a peer-to-peer electronic cash sys[tem. Cryptography Mailing list at https​://metzd​owd.com](https://metzdowd.com) 2. Atzori M (2017) Blockchain technology and decentralized gov[ernance: is the state still necessary? J Gov Regul. https​://doi.](https://doi.org/10.22495/jgr_v6_i1_p5) [org/10.22495​/jgr_v6_i1_p5](https://doi.org/10.22495/jgr_v6_i1_p5) 3. Udokwu C, Kormiltsyn A, Thangalimodzi K, Norta A (2018) An exploration of blockchain enabled smart-contracts application [in the enterprise. https​://doi.org/10.13140​/rg.2.2.36464​.97287​](https://doi.org/10.13140/rg.2.2.36464.97287) 4. De Angelis S et al, PBFT vs proof-of-authority: applying CAP theorum to permissioned Blockchain 5. Li X, Wu X, Pei X, Yao Z (2019) Tokenization: open asset protocol on blockchain. In: 2019 IEEE 2nd international conference on information and computer technologies (ICICT), Kahului, HI, [USA, pp 204-209. https​://doi.org/10.1109/infoc​t.2019.87110​21](https://doi.org/10.1109/infoct.2019.8711021) [6. See https​://githu​b.com/ether​eum/EIPs/issue​s/223](https://github.com/ethereum/EIPs/issues/223) [7. See http://eips.ether​eum.org/EIPS/eip-777](http://eips.ethereum.org/EIPS/eip-777) 8. Amoatey P, Bani R (2011) Wastewater management 9. Chakraborty D, Water pollution in India: an input-output analysis 10. Frank B et al (2019) How urban storm- and wastewater management prepares for emerging opportunities and threats: digital transformation, ubiquitous sensing, new data sources, and beyond: a horizon scan. Environ Sci Technol 53(15):8488–8498. [https​://doi.org/10.1021/acs.est.8b064​81](https://doi.org/10.1021/acs.est.8b06481) 11. Swan M (2015) Blockchain: blueprint for a new economy. O’Reilly. P. Vii 12. Mougayar W (2016) The Business blockchain: promise, practice, and application of the next internet technology. Wiley, London 13. Duca G, Bunduchi E, Gladchi V, Romanciuc L, Goreaceva N (2008) Estimation of the natural water self-purification capacity from the kinetic standpoint 14. Echard C (2017) Ensuring software integrity in IoT devices 15. Schaad J (2019) CBOR, Object Signing and Encryption (COSE), Accessed 2019 16. Hong Z, Chu C, Zhang LL, Yu Y (2017) Optimizing an emission trading scheme for local governments: a Stackelberg game mode and hybrid algorithm 17. Zecchini M (2019) Data collection, storage and processing for water monitoring based on IoT, and blockchain technologies. zecchini.1596071@studenti.uniroma1.it 18. Silvestre BS (2015) Sustainable supply chain management in emerging economies: environmental turbulence, institutional voids and sustainability trajectories. Int J Prod Econ 167:156–169 19. Klassen RD, Vereecke A (2012) Social issues in supply chains: capabilities link responsibility, risk (opportunity), and performance. Int J Prod Econ 140(1):103–115 20. Li X, Wu X, Pei X, Yao Z (2019) Tokenization: open asset protocol on blockchain. In: 2019 IEEE 2nd international conference on information and computer technologies, pp 204–209 21. Zhou T et al (2019) DLattice: permission-less blockchain based on DPoS-BA-DAG consensus for data tokenization. IEEE Access. [https​://doi.org/10.1109/ACCES​S.2019.29066​37](https://doi.org/10.1109/ACCESS.2019.2906637) 22. Sánchez-Corcuera R et al (2019) Smart cities survey: technologies, application domains and challenges for the cities of the future. Int J Distrib Sens Netw 15(6) 23. Kundu D (2019) Blockchain and trust in a Smart City. Environ Urban Asia 10(1):31–43 24. Kouhizadeh M et al (2019) At the Nexus of Blockchain technology, the circular economy, and product deletion. Appl Sci 9:1712 25. Savelyev A (2018) Some risks of tokenization and blockchainizai[tion of private law. Comput Law Secur Rev 34(4):863–869. https​](https://doi.org/10.1016/j.clsr.2018.05.010) [://doi.org/10.1016/j.clsr.2018.05.010](https://doi.org/10.1016/j.clsr.2018.05.010) 26. Blemus S et al (2019) Initial crypto-asset offerings (ICOs), tokenization and corporate governance, 1905.03340 27. Roth J, Schär F, Schöpfer A (2019) The Tokenization of assets: using blockchains for equity crowdfunding. Available at SSRN 3443382 28. Sanghavi V et al (2018) International Journal of Research in Engineering, IT and Social Sciences, 08(11), 60–64. ISSN 2250-0588 29. Prince Michael von und zu Liechtenstein HSH (2019) The tokenization of assets and property rights. Trusts Trustees 25(6):630–632 30. Davydov V, Khalilova M (2019) Business model of creating digital platform for tokenization of assets on financial markets. In: Conference series: materials science and engineering 31. Levin B (2018) Potential for cryptocurrency to fund investment in sustainable real assets 32. Barsan IM (2017) Legal challenges of initial coin offerings (ICO). Revue Trimestrielle de Droit Financier (RTDF) 3:54–65 **Publisher’s Note Springer Nature remains neutral with regard to** jurisdictional claims in published maps and institutional affiliations. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/s42452-020-2187-4?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/s42452-020-2187-4, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://link.springer.com/content/pdf/10.1007/s42452-020-2187-4.pdf" }
2,020
[]
true
2020-02-18T00:00:00
[]
11,641
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/00ff024e0800db79a8f7c661b47357fc98e6c0af
[ "Computer Science" ]
0.874122
Adversarial-Playground: A visualization suite showing how adversarial examples fool deep learning
00ff024e0800db79a8f7c661b47357fc98e6c0af
Visualization for Computer Security
[ { "authorId": "2064028216", "name": "Andrew P. Norton" }, { "authorId": "121817403", "name": "Yanjun Qi" } ]
{ "alternate_issns": null, "alternate_names": [ "Vis Comput Secur", "VizSEC" ], "alternate_urls": null, "id": "55cbbee6-4000-432f-af6a-c0ffd25ba0c0", "issn": null, "name": "Visualization for Computer Security", "type": "conference", "url": null }
Recent studies have shown that attackers can force deep learning models to misclassify so-called “adversarial examples:” maliciously generated images formed by making imperceptible modifications to pixel values. With growing interest in deep learning for security applications, it is important for security experts and users of machine learning to recognize how learning systems may be attacked. Due to the complex nature of deep learning, it is challenging to understand how deep models can be fooled by adversarial examples. Thus, we present a web-based visualization tool, Adversarial-Playground, to demonstrate the efficacy of common adversarial methods against a convolutional neural network (CNN) system. Adversarial-Playground is educational, modular and interactive. (1) It enables non-experts to compare examples visually and to understand why an adversarial example can fool a CNN-based image classifier. (2) It can help security experts explore more vulnerability of deep learning as a software module. (3) Building an interactive visualization is challenging in this domain due to the large feature space of image classification (generating adversarial examples is slow in general and visualizing images are costly). Through multiple novel design choices, our tool can provide fast and accurate responses to user requests. Empirically, we find that our client-server division strategy reduced the response time by an average of 1.5 seconds per sample. Our other innovation, a faster variant of JSMA evasion algorithm, empirically performed twice as fast as JSMA and yet maintains a comparable evasion rate1.
## ADVERSARIAL-PLAYGROUND: A Visualization Suite Showing How Adversarial Examples Fool Deep Learning #### Andrew P. Norton[*] Yanjun Qi[†] Department of Computer Science, University of Virginia **ABSTRACT** Recent studies have shown that attackers can force deep learning models to misclassify so-called “adversarial examples:” maliciously generated images formed by making imperceptible modifications to pixel values. With growing interest in deep learning for security applications, it is important for security experts and users of machine learning to recognize how learning systems may be attacked. Due to the complex nature of deep learning, it is challenging to understand how deep models can be fooled by adversarial examples. Thus, we present a web-based visualization tool, ADVERSARIALPLAYGROUND, to demonstrate the efficacy of common adversarial methods against a convolutional neural network (CNN) system. ADVERSARIAL-PLAYGROUND is educational, modular and interactive. (1) It enables non-experts to compare examples visually and to understand why an adversarial example can fool a CNN-based image classifier. (2) It can help security experts explore more vulnerability of deep learning as a software module. (3) Building an interactive visualization is challenging in this domain due to the large feature space of image classification (generating adversarial examples is slow in general and visualizing images are costly). Through multiple novel design choices, our tool can provide fast and accurate responses to user requests. Empirically, we find that our client-server division strategy reduced the response time by an average of 1.5 seconds per sample. Our other innovation, a faster variant of JSMA evasion algorithm, empirically performed twice as fast as JSMA and yet maintains a comparable evasion rate[1]. **Index Terms:** I.2.6 [Artificial Intelligence]: Learning— Connectionism and neural nets; K.6.5 [Management of Computing and Information Systems]: Security and Protection—Unauthorized access **1** **INTRODUCTION** Adversarial examples for Deep Neural Network (DNN) models are usually crafted through an optimization procedure that searches for small, yet effective, perturbations of the original image (details in Sect. 2). Understanding why a DNN model performs as it does is quite challenging, and even moreso to understand how such a model can be fooled by adversarial examples. With growing interest in adversarial deep learning, it is important for security experts and users of DNN systems to understand how DNN models may be attacked in face of an adversary. This paper introduces a visualization tool, ADVERSARIAL-PLAYGROUND, to enable better understanding of how different types of adversarial examples fool DNN systems. ADVERSARIAL-PLAYGROUND provides a simple and intuitive interface to let users visually explore the impact of three attack algorithms that generate adversarial examples. Users may specify parameters for a variety of attack types and generate new samples on-demand. The interface displays the resulting adversarial example compared to the original alongside classification likelihoods on both images from the DNN. ADVERSARIAL-PLAYGROUND provides the following benefits: - Educational: Fig. 1 shows a screen-shot of the visualization. This intuitive and simple visualization helps practitioners of deep learning to understand how their models misclassify and how adversarial examples by various algorithm differ. - Interactive: We add two novel strategies in ADVERSARIALPLAYGROUND to make it respond to users’ requests in a sufficiently quick manner. The interactive visualization allow users to gain deeper intuition about the behavior of DNN classification through maliciously generated inputs. Investigating the behavior of machine learning systems in adversarial environments is an emerging topic at the junction of computer security and machine learning [1]. While machine learning models may appear to be effective for many security tasks like malware classification [7] and facial recognition [9], these classification techniques were not designed to withstand manipulations made by intelligent and adaptive adversaries. In contrast with applications of machine learning to other fields, security tasks involve adversaries that may respond maliciously to the classifier [1]. Recent studies show that intelligent attackers can force machine learning systems to misclassify samples by performing nearly imperceptible modifications to the sample before attempting classification [4,10]. These samples, named as “adversarial examples,” have effectively fooled many state-of-the-art deep learning models. *e-mail: apn4za@virginia.edu †e-mail: yanjun@virginia.edu - Modular: Security experts can easily plug ADVERSARIALPLAYGROUND into their benchmarking frameworks as a module. The design also allows experts to easily add other DNN models or more algorithms of generating adversarial examples in the visualization. To the authors’ best knowledge, this is the first visualization platform showing how adversarial examples are generated and how they fool a DNN system. The rest of this paper takes the following structure: Sect. 2 discusses the relevant backgrounds, Sect. 3 introduces the system organization and software design of ADVERSARIAL-PLAYGROUND, Sect. 4 presents an empirical evaluation with respect to different design choices, and Sect. 5 concludes the paper by discussing possible extensions. **2** **BACKGROUND** 1Project source code and data from our experiments available at: https: ``` //github.com/QData/AdversarialDNN-Playground. ``` **2.1** **Adversarial Examples and More** Studies regarding the behavior of machine learning models in adversarial environments generally fall into one of three categories: (1) _poisoning attacks, in which specially crafted samples are injected_ into the training of a learning model, (2) privacy-aware learning, which aim to preserve the privacy of information in data samples, or (3) evasion attacks, in which the adversary aims to create inputs that are misclassified by a target classifier. Generating adversarial examples is part of this last category. ----- Figure 1: ADVERSARIAL-PLAYGROUND User Interface The goal of adversarial example generation is to craft an input for a particular classifier that, while improperly classified, reveals only slight alteration on the input. To formalize the extent of allowed alterations, evasion algorithms minimize the difference between the “seed” input and the resulting adversarial example based on a predefined norm (a function measuring the distance between two inputs). In some cases, the adversary specifies the “target” class of an adversarial sample — for example, the adversary may desire an image that looks like a “6” to be classified as a “5” (as in Fig. 1). This is referred to as a targeted attack. Conversely, if the adversary does not specify the desired class, the algorithm is considered to be _untargeted._ Formally, let us denote f : X → _C to be a classifier that maps_ the set of all possible inputs, X, to a finite set of classes, C. Then, given a target class yt ∈ _C, a seed sample x ∈_ _X, and a norm function_ _∥·∥, the goal of generating a targeted adversarial example is to find_ _x[′]_ _∈_ _X such that:_ _x[′]_ = argmin _{∥x_ _−_ _s∥_ : f (s) = yt _}_ (1) _s∈X_ Similarly, in the untargeted case, the goal is to find x[′] such that: _x[′]_ = argmin _{∥x_ _−_ _s∥_ : f (s) ̸= f (x)} (2) _s∈X_ In this formalization, we see there are two key degrees of freedom in creating a new evasion algorithm: targeted vs. untargeted attacks and the choice of norm functions. The latter category provides a useful grouping scheme for algorithms generating adversarial inputs, suggested by Carlini and Wagner [2]. ADVERSARIAL-PLAYGROUND uses two evasion algorithms provided by the cleverhans library [3]: the Fast Gradient Sign Method (FGSM) based on the L[∞] norm, and the Jacobian Saliency Map Approach (JSMA) based on the L[0] norm [8]. **2.2** **DNNs and the MNIST Dataset** DNNs can efficiently learn highly-accurate models in many domains [5,7]. Convolutional Neural Networks (CNNs), first popularized by LeCun et al. [6], perform exceptionally well on image classification. ADVERSARIAL-PLAYGROUND uses a state-of-the-art CNN model on the popular MNIST “handwritten digits” dataset for visualizing evasion attacks. This dataset contains 70,000 images of hand-written digits (0 through 9). Of these, 60,000 images are used as training data and the remaining 10,000 images are used for testing. Each sample is a 28 _×_ 28 pixel, 8-bit grayscale image. Users of our system are presented with a collection of seed images, selected from each of the 10 classes in the testing set (see the right side of Fig. 1). **2.3** **TensorFlow Playground** Our proposed package follows the spirit of TensorFlow Playground — a web-based educational tool that helps users understand how neural networks work [11]. TensorFlow Playground has been used in many classes as a pedagogical aid and helps the self-guided student learn more. Its impact inspires us to visualize adversarial examples through ADVERSARIAL-PLAYGROUND. Our web-based visualization tool assists users in understanding and comparing the impact of standard evasion techniques on deep learning models. **3** **ADVERSARIAL-PLAYGROUND: A MODULAR AND INTER-** **ACTIVE VISUALIZATION SUITE** In creating our system, we made several design decisions to make ADVERSARIAL-PLAYGROUND educational, modular and interactive. Here, we present the four major system-level decisions we made: (1) building ADVERSARIAL-PLAYGROUND as a web-based application, (2) utilizing both client- and server-side code, (3) rendering images with the client rather than the server and (4) implementing a faster variation of JSMA attack. We released all project code on GitHub in the interest of providing a high-quality, easy-to-use software package to demonstrate how adversarial examples fool deep learning. **3.1** **A Web-based Visualization Interface** ADVERSARIAL-PLAYGROUND provides quick and effective visualizations of adversarial examples through an interactive webapp as shown by Fig. 1. The user selects one attacking algorithm from the navigation bar at the top of the webapp. On the right-hand pane, the user sets the attacking strength the algorithm using the slider, selects a seed image, and (if applicable) a target class. (Fig. 1 at right.) Selecting a seed image immediately loads the image to the left-hand display and displays the output of the CNN classifier in a bar chart below. After setting the parameters, the user clicks “Generate Adversarial Sample.” This runs the chosen adversarial algorithm in real-time to attempt generating an adversarial sample. The sample is displayed in the primary pane to the left of the controls (Fig. 1 at center). The generated sample is fed through the CNN classifier, and then the likelihoods are displayed in a bar chart below the sample. Finally, the classification of generated sample is displayed below the controls at right. This web-based visualization generates adversarial examples “ondemand” from user-specified parameters. Therefore users can see the impact of different adversarial algorithms with varying configurations. Developing ADVERSARIAL-PLAYGROUND as a web-based (as opposed to a local) application enables a large number of users to utilize the application without requiring an installation process on each computer. By eliminating an installation step, we encourage potential users who may be only casually interested in adversarial machine learning to explore what it is. This supports the pedagogical goals of the software package. **3.2** **A Modular Design with Client-Server Division** Two key features of ADVERSARIAL-PLAYGROUND are its modular design and the division of the functionality between the client and server; the client handles user interaction and visualization, while the server handles more computationally intensive tasks. Fig. 2 diagrams the interaction between each component of our system. At the upper right, we have the user who may specify hyperparameters for the evasion algorithm. Moving counter-clockwise, these parameters are transferred to the server, where the appropriate adversarial algorithm module is selected and run against the pre-trained CNN module. TensorFlow is used to reduce computation time and improve compatibility. Finally, the resulting sample is sent to the client and plotted using the JavaScript library Plotly.JS. ----- Figure 2: ADVERSARIAL-PLAYGROUND System Sketch Users running a local copy of the webapp may easily customize the tool to their needs; by separating the deep learning model and the evasion methods from the main visualization and interface codebase, changing or adding DNN models or adding new adversarial algorithms is straightforward. **TensorFlow based Server-side:** Our inspiration, the TensorFlow Playground, was written entirely in JavaScript and other clientside technologies, allowing a lightweight server to host the service for many users. Unfortunately, adversarial examples are usually generated on larger, deeper networks than those created by users of TensorFlow Playground, and this makes a JavaScript-only approach prohibitively slow. Instead, we chose to use a GPU-enabled server running Python with TensorFlow to generate the adversarial examples on the backend (server-side), then send the image data to the client. This provides increased speed (aiding interactivity), adds compatibility with other TensorFlow-based deep learning models and allows the flexibility of evasion algorithms (promoting modularity). **Server-side Configuration: The server-side of ADVERSARIAL-** PLAYGROUND requires a computer with Python 3.5, TensorFlow 1.0 (or higher), the standard SciPy stack, and the Python package ``` Flask. We have tested the code on Windows, Linux, and Mac ``` operating systems. To install, clone the GitHub repository and install the prerequisites via pip3 -r install requirements.txt. A pre-trained MNIST model is already stored in the GitHub repository; all that is needed to start the webapp is to run python3 run.py. Once the app is started, it will run on localhost:9000. **3.3** **Visualizing Sample through Client-side Rendering** As shown in Fig. 2, through the client, the user adjusts hyperparameters and submits a request to generate an adversarial sample to the server. Once the TensorFlow back-end generates the adversarial image and classification likelihoods, the server returns this data to the client. Finally, this information is displayed graphically to the user through use of the Plotly JavaScript library. As we generate adversarial samples on the server-side, it was tempting to produce the output images on the server as well. In our prototype, we used server-side rendering of these images with the Python library matplotlib, then downloaded the image for display on the client. However, we ultimately decided to assign all visualization tasks to the client, using JavaScript and the Plotly.JS library, after realizing this approach was faster. This is because generating images on the server with the default ``` matplotlib utilities required creating a full PNG image, writing ``` it to disk, then transferring the image to the client; this took time and increased latency. Fortunately, client-side rendering of images required transmission of far less data; only pixel values for the 28 _×_ 28 MNIST images and the 10 values for classification likelihoods needed to be sent. Additionally, the Plotly.JS library provided interactive plots that enable users to view the underlying values for each pixel. Empirically, switching to a client-side rendering of images reduced response time by approximately 1.5 seconds. (Sect. 4.2.) ADVERSARIAL-PLAYGROUND’s modularity extends into the visualization code, too. Although it may be hosted on any machine that supports TensorFlow, the web-based client/server division of the webapp allows the computationally intensive “back-end” to be hosted on a powerful server while the visualizations may be accessed from any device. **3.4** **Faster Variant of JSMA Attack** While dividing the computation and visualization steps between the client and server saved some time, actually generating the adversarial example is where the most time is consumed (Table 1). In particular, the Jacobian Saliency Map Approach (JSMA) algorithm by Papernot et al. [8] can take more than two-thirds of a second to generate a single adversarial output. In order to provide an interactive experience, our web app must generate adversarial samples quickly. We therefore introduce a new, faster variant of the JSMA that maintains a comparable evasion rate to the original, but can take half as much time. **JSMA Background: Most state-of-the-art evasion algorithms** are slow due to the expensive optimization and the large search space involved in image classification [2,3]. The original JSMA algorithm is a targeted attack that uses the L[0] norm in Equation 1. To generate x[′] from x, JSMA iteratively selects the “most influential” combination of two features to alter. To rank features by their influence, JSMA uses a saliency map of the forward derivative of the classifier. The ranking and alteration process is repeated until the altered sample is successfully classified as yt or the L[0] distance between the altered and seed samples exceeds a provided threshold, ϒ. The largest consumption of time in JSAM is the combinatorial search over all feature pairs to determine the “best” pair to alter; if there are M features in a given sample, JSMA must evaluate Θ(M[2]) candidates at each iteration. When working on high-dimensional data, this can become prohibitively expensive. We introduce a new, faster variant of JSMA that maintains a comparable evasion rate to the original, which we call Fast Jacobian Saliency Map Apriori (FJSMA). **FJSMA Improvement: Our FJSMA approach is an approxima-** tion of JSMA that uses an a priori heuristic to significantly reduce the search space. Instead of considering all pairs of features (p, _q),_ our improvement only considers such pairs where p is in the top k features when ranked by the derivative in the p-coordinate, where _k is a small constant. (See red-bolded modifications to JSMA in_ Algorithm 1.) If we denote the set consisting of the top k elements in A as ranked by f by argtopx∈A ( f (x); k), then the loop in our Fast Jacobian Saliency Map Apriori (FJSMA) selection routine is Θ(k _·|Γ|), where_ _k ≪|Γ| and |Γ| = M is the size of the feature set. Since determining_ the top k features can be done in linear time, this is considerable improvement in asymptotic terms. This modification improves the runtime from Θ(M[2]) to Θ(M · _k),_ where M is the feature size and k is some small constant. Our experiments show k may be as little as 15% of M and still maintain the same efficacy in terms of evasion rate as JSMA. **4** **PERFORMANCE TESTING** We conducted a series of timing tests to quantify how our design choices have influenced the speed of interactive responses by ADVERSARIAL-PLAYGROUND. First, we consider the impact of relegating the visualization code to the client (from Sect. 3.2); then, ----- **Algorithm 1 Fast Jacobian Saliency Map Apriori Selection** ∇F(X) is the forward derivative, Γ the features still in the search space, t the target class, and k is a small constant **Input: ∇F(X), Γ, t, k** � � 1: K = argtopp∈Γ _−_ _[∂]_ _∂[F][t]X[(][X]p_ [)] [;][ k] _▷_ Changed for FJSMA 2: for each pair (p, _q) ∈_ _K ×_ Γ, p ̸= q do ▷ Changed for FJSMA 3: _α = ∑i=p,q_ _∂_ **F∂tX(Xi** ) 4: _β = ∑i=p,q ∑_ _j≠_ _t_ _∂_ **F∂ jX(Xi** ) 5: **if α < 0 and β > 0 and −α ×** _β > max then_ 6: _p1, p2 ←_ _p,_ _q_ 7: _max ←−α ×_ _β_ 8: **end if** 9: end for 10: return p1, p2 we show that FJSMA is faster and just as accurate as the JSMA implementation by cleverhans package. **4.1** **Client-side Rendering Improves Response Speed** Rendering done on... Server Image Download Total Server-side 4472 350 4821 Client-side 3335 — 3335 **Difference** 1137 350 1486 Table 1: Latency with and without client-side visualization. A time profiling of the latency experienced by the user when 1) the server handled all computation and visualization and 2) the visualization was offloaded to the client. The “Server” column denotes time taken for the server to respond, while the “Image Download” column shows the additional time taken to transfer each image (only applicable for server-side rendering). We first conducted timing tests to evaluate how the choice of client-side rendering (Sect. 3.2) has influenced the speed of responding to users’ requests. We loaded the webapp and measured the response time of the server for a variety of seed images and target classes. We repeated these for a total of between 10 and 16 times (depending on the algorithm), averaged the response time, and reported the result in Table 1. The majority of the time for both with and without client-side visualization is in the server computation. However, offloading the visualization to the client resulted in a nearly 1.5-second speedup (an approximately 30% difference). Interestingly, not only did the image download time get eliminated, but the server computation time was reduced as well. This is, in part, due to the reduction of I/O operations and image generation required by the server when visualization is done by the client. **4.2** **FJSMA Improvement** We propose a faster approximation of the JSMA attacking algorithm: FJSMA in Sect. 3.4. Using the same CNN model we used in ADVERSARIAL-PLAYGROUND for the MNIST dataset, we compared FJSMA with JSMA through two metrics: (1) the “wall clock” time needed for successfully generating an adversarial example, and (2) the evasion rate — a standard metric that reports the percentage of seed images that were successfully converted into adversarial samples. This comparison was conducted in a batch manner. We ran both evasion attacks on the 10000-sample MNIST testing set for a range of values of the ϒ parameter for both algorithms. For FJSMA, we also varied the value of input parameter k (the percentage of the feature-set size). Intuitively, this k value is a control on how tight of an approximation FJSMA is to JSMA; as k grows larger, we should |ϒ|10%|15%|20%|25%| |---|---|---|---|---| |JSMA Evasion Rate|0.658|0.824|0.867|0.879| |FJSMA Evasion Rate [k = 10%]|0.583|0.777|0.823|0.826| |FJSMA Evasion Rate [k = 15%]|0.613|0.816|0.867|0.871| |FJSMA Evasion Rate [k = 20%]|0.633|0.833|0.878|0.887| |FJSMA Evasion Rate [k = 30%]|0.638|0.844|0.896|0.901| |JSMA Time (s)|0.606|0.745|0.807|0.803| |FJSMA Time [k = 10%] (s)|0.411|0.468|0.490|0.485| |FJSMA Time [k = 15%] (s)|0.414|0.473|0.483|0.484| |FJSMA Time [k = 20%] (s)|0.415|0.466|0.482|0.483| |FJSMA Time [k = 30%] (s)|0.415|0.464|0.490|0.485| Table 2: JSMA and FJSMA Comparison. Each column represents a test run with a particular value of ϒ (the maximum allowed perturbation for a sample). The top half of the table provides the average evasion rate for each algorithm, while the bottom half provides the average time (in seconds) that it took to generate an adversarial example. The FJSMA algorithm was run with multiple values of k, where k was 10%, 15%, 20%, and 30% of the feature space size. expect the performance of the two approaches to converge to each other. Results of this experiment are summarized in Table 2. The ``` cleverhans JSMA and the proposed FJSMA attack achieve similar ``` evasion rates for all tested values of ϒ and k, with larger values of _k increasing the evasion rate. Curiously, for k ≥_ 20%, our implementation of FJSMA even outperforms that of cleverhans JSMA; this is likely due to implementation details. The average time to form an evasive sample from a seed benign sample is given in the second half of the table. Our FJSMA approach greatly improves upon the speed of JSMA. However, varying the value of k does not produce a significant variation in runtime per sample; we conjecture this is because of the small feature space of MNIST and that searching 30% of the feature space likely does not dominate the runtime. In summary, FJSMA achieves a significant improvement in speed, while maintaining essentially the same evasion rate — an important advantage for interactive visualization. **5** **DISCUSSION AND FUTURE WORK** The study of evasion attacks on machine learning models is a rapidly growing field. In this paper, we present a web-based tool ADVERSARIAL-PLAYGROUND for visualizing the performance of adversarial examples against deep neural networks. ADVERSARIALPLAYGROUND enables non-experts to compare adversarial examples visually and can help security experts explore more vulnerability of deep learning. It is modular and interactive. To our knowledge, our platform is the first visualization-focused package for adversarial machine learning. A straightforward extension of this work is to increase the variety of supported evasion methods. For example, including the new attacks based on L[0], L[2], and L[∞] norms from Carlini and Wagner’s recent paper [2] would be a good next step in comparing the performance of multiple evasion strategies. However, expansion in this manner presents an additional issue of latency. To generate evading samples “on-demand,” the adversarial algorithm must run quickly; these other algorithms take much longer to execute than those we selected, so some time-saving techniques must be explored. Another direction for development is to provide more choices of classifiers and datasets. Allowing the user to select from CIFAR, ImageNet, and MNIST data would highlight the similarities and differences between how a single attack method deals with different data. Similarly, providing the user with a choice of multiple pre-trained models — possibly hardened against attack through adversarial training — would help distinguish artifacts of model choice from the behavior of the attack. These two extensions would help users more fully understand the behavior of an adversarial algorithm. |Rendering done on...|Server|Image Download|Total| |---|---|---|---| |Server-side|4472|350|4821| |Client-side|3335|—|3335| |Difference|1137|350|1486| ----- **REFERENCES** [1] M. Barreno, B. Nelson, A. D. Joseph, and J. Tygar. The Security of Machine Learning. Machine Learning, 81(2):121–148, 2010. [2] N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. CoRR, abs/1608.04644, 2016. [3] I. J. Goodfellow, N. Papernot, and P. D. McDaniel. cleverhans v0.1: an adversarial machine learning library. CoRR, abs/1610.00768, 2016. [4] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. [5] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Advances in Neural _Information Processing Systems, pp. 1097–1105, 2012._ [6] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, November 1998. [7] Microsoft Corporation. Microsoft Malware Competition Challenge. https://www.kaggle.com/c/malware-classification, 2015. [8] N. Papernot, P. D. McDaniel, S. Jha, M. Fredrikson, Z. B. Celik, and A. Swami. The limitations of deep learning in adversarial settings. _CoRR, abs/1511.07528, 2015._ [9] O. M. Parkhi, A. Vedaldi, and A. Zisserman. Deep Face Recognition. In British Machine Vision Conference, 2015. [10] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. J. Goodfellow, and R. Fergus. Intriguing properties of neural networks. CoRR, abs/1312.6199, 2013. [11] J. Yosinski, J. Clune, A. M. Nguyen, T. J. Fuchs, and H. Lipson. Understanding neural networks through deep visualization. CoRR, abs/1506.06579, 2015. -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1708.00807, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://arxiv.org/pdf/1708.00807" }
2,017
[ "JournalArticle" ]
true
2017-08-01T00:00:00
[ { "paperId": "ed46777287b287969571f4e9bffbbf3741ebb769", "title": "Cleverhans V0.1: an Adversarial Machine Learning Library" }, { "paperId": "df40ce107a71b770c9d0354b78fdd8989da80d2f", "title": "Towards Evaluating the Robustness of Neural Networks" }, { "paperId": "819167ace2f0caae7745d2f25a803979be5fbfae", "title": "The Limitations of Deep Learning in Adversarial Settings" }, { "paperId": "1b5a24639fa80056d1a17b15f6997d10e76cc731", "title": "Understanding Neural Networks Through Deep Visualization" }, { "paperId": "bee044c8e8903fb67523c1f8c105ab4718600cdb", "title": "Explaining and Harnessing Adversarial Examples" }, { "paperId": "d891dc72cbd40ffaeefdc79f2e7afe1e530a23ad", "title": "Intriguing properties of neural networks" }, { "paperId": "abd1c342495432171beb7ca8fd9551ef13cbd0ff", "title": "ImageNet classification with deep convolutional neural networks" }, { "paperId": "afd0859a858481d2f36109f68090aebd77456b7f", "title": "The security of machine learning" }, { "paperId": "162ea969d1929ed180cc6de9f0bf116993ff6e06", "title": "Deep Face Recognition" }, { "paperId": null, "title": "Microsoft Malware Competition Challenge" }, { "paperId": "162d958ff885f1462aeda91cd72582323fd6a1f4", "title": "Gradient-based learning applied to document recognition" } ]
7,086
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0100d00584f7b6a2e3969fd57a9d164103f39a69
[ "Computer Science" ]
0.910426
ENHANCED DYNAMIC RESOURCE ALLOCATION SCHEME BASED ON PACKAGE LEVEL ACCESS IN CLOUD COMPUTING : A REVIEW
0100d00584f7b6a2e3969fd57a9d164103f39a69
Bioinformatics
[ { "authorId": "144912228", "name": "Manpreet Kaur" }, { "authorId": "145713817", "name": "Rajinder Singh" } ]
{ "alternate_issns": [ "1367-4811" ], "alternate_names": [ "Int Conf Bioinform", "International Conference on Bioinformatics", "BIOINFORMATICS" ], "alternate_urls": null, "id": "15d4205f-903b-403c-9a2e-906f02ce04d8", "issn": "1367-4803", "name": "Bioinformatics", "type": "journal", "url": "http://bioinformatics.oxfordjournals.org/" }
Cloud computing is distributed computing, storing, sharing and accessing data over the Internet. It provides a pool of shared resources to the users available on the basis of pay as you go service that means users pay only for those services which are used by him according to their access times. This research work deals with the balancing of work load in cloud environment. Load balancing is one of the essential factors to enhance the working performance of the cloud service provider. It would consume a lot of cost to maintain load information, since the system is too huge to timely disperse load. Load balancing is one of the main challenges in cloud computing which is required to distribute the dynamic workload across multiple nodes to ensure that no single node is overwhelmed. It helps in optimal utilization of resources and hence in enhancing the performance of the system. We propose an improved load balancing algorithm for job scheduling in the cloud environment using load distribution table in which the current status, current package, VM Capacity and the number of cloudlets submitted to each and every virtual machine will be stored. Submit the job of the user to the datacenter broker. Data center broker will first find the suitable Vm according to the requirements of the cloudlet and will match and find the most suitable Vm according to its availability or the machine with least load in the distribution table. Multiple number of experiments have been conducted by taking different configurations of cloudlets and virtual machine. Various parameters like waiting time, execution time, turnaround time and the usage cost have been computed inside the cloudsim environment to demonstrate the results. The main contributions of the research work is to balance the entire system load while trying to minimize the make span of a given set of jobs. Compared with the other job scheduling algorithms, the improved load balancing algorithm can outperform them according to the experimental results.
# ENHANCED DYNAMIC RESOURCE ALLOCATION SCHEME BASED ON PACKAGE LEVEL ACCESS IN CLOUD COMPUTING : A REVIEW ## Manpreet Kaur [(1)], Dr. Rajinder Singh[(2)] (1) Research Scholar, Department of Computer Science & Engineering, GGSCET, Bathinda, Punjab. ### manpreet12395@gmail.com (2) Assistant Professor, Department of Computer Science & Engineering, GGSCET, Bathinda, Punjab. ### rajneel2807@gmail.com ## ABSTRACT Cloud computing is Internet based development and use of computer technology. It is a style of computing in which dynamically scalable and often virtualized resources are provided as a service over the Internet. Users need not have knowledge of, expertise in, or control over the technology infrastructure "in the cloud" that supports them. Scheduling is one of the core steps to efficiently exploit the capabilities of heterogeneous computing systems. On cloud computing platform, load balancing of the entire system can be dynamically handled by using virtualization technology through which it becomes possible to remap virtual machine and physical resources according to the change in load. However, in order to improve performance, the virtual machines have to fully utilize its resources and services by adapting to computing environment dynamically. The load balancing with proper allocation of resources must be guaranteed in order to improve resource utility. ## Keywords Cloud Computing, Load Balancing, Virtual Machine, Packages, Leases ## INTRODUCTION Cloud Computing (CC)[1] is an emerging technology that has abstruse connection to Grid Computing (GC) paradigm and other relevant technologies such as utility computing, distributed computing and cluster computing. The aim of both GC and CC is to achieve resource virtualization. In spite of the aim being similar, GC and CC have significant differences. The main emphasis of GC is to achieve maximum computing, while that of CC is to optimize the overall computing capacity. CC also provides a way to handle wide range of organizational needs by providing dynamically scalable servers and application to work with. Leading CC service providers such as Amazon, IBM, `Dropbox', Apple's `iCloud',Google's applications, Microsoft's `Azure', etc., are able to attract normal users through out the world. CC have introduced a new paradigm, which helps its users to store or develop applications dynamically and access them from anywhere and anytime just by connecting to an application using Internet. Depending on customer's requirement CC provides easy and customizable services to access or work with cloud applications. Based on the user requirement CC can be used to provide platform for designing applications, infrastructure to store and work on company's data and also provide applications to do user's routine tasks. When a customer chooses to use cloud services, data stored in the local repositories will be sent to a remote data center. This data in remote locations can be accessed or managed with the help of services provided by cloud service providers. This makes clear that for a user to store or process a piece of data in cloud, he/she needs to transmit the data to a remote server over a channel (internet). This data processing and storage needs to be done with utmost care to avoid data breaches. It is the model for convenient on-demand network access, with minimum management efforts for easy and fast network access to resources that are ready to use. It is an upcoming paradigm that offers tremendous advantages in economic aspects, such as reduced time to market, flexible computing capabilities, and limitless computing power. Popularity of cloud computing is increasing day by day in distributed computing environment. There is a growing trend of using cloud environments for storage and data processing needs. To use the full potential of cloud computing, data is transferred, processed, retrieved and stored by external cloud providers. However, data owners are very skeptical to place their data outside their own control sphere. **Figure 1. Cloud Computing** ### 6207 | P ----- V o l u m e 1 6 N u m b e r 2 I N T E R N A T I O N A L J O U R N A L O F C O M P U T E R S & T E C H N O L O G Y ### BENFITS OF CLOUD COMPUTING Some common benefits of cloud computing are: - **Reduced Cost: Since cloud technology is implemented incrementally (step-by-step), it saves organizations total** expenditure. **• Increased Storage: When compared to private computer systems, huge amounts of data can be stored than usual.** **• Flexibility: Compared to traditional computing methods, cloud computing allows an entire organizational segment or** portion of it to be outsourced. - **Greater mobility: Accessing information, whenever and wherever needed unlike traditional systems (storing data in** personal computers and accessing only when near it). - **Shift of IT focus:** Organizations can focus on innovation (i.e., implementing new products strategies in organization) rather than worrying about maintenance issues such as software updates or computing issues. These benefits of cloud computing draw lot of attention from Information and Technology Community (ITC). A survey by ITC in the year 2008, 2009 shows that many companies and individuals are noticing that CC is proving to be helpful when compared to traditional computing methods. **Figure 2. Benefits of Cloud Computing** ### CLOUD COMPUTING: SERVICE MODELS Cloud computing can be accessed through a set of services models. These services are designed to exhibit certain characteristics and to satisfy the organizational requirements. From this, a best suited service can be selected and customized for an organization's use. Some of the common distinctions in cloud computing services are Software-as-aService (SaaS), Platform-as-a-Service (PaaS), Infrastructureas-a-Service (IaaS), Hardware-as-a-Service (HaaS) and Data storage-as-a-Service (DaaS). Service model details are as follows: **• Software as a Service (SaaS)[4]: The service provider in this context provides capability to use one or more** applications running on a cloud infrastructure. These applications can be accessed from various thin client interfaces such as web browsers. A user for this service need not maintain, manage or control the underlying cloud infrastructure (i.e. network, operating systems, storage etc.). Examples for SaaS cloud's are Salesforce, NetSuite. **• Platform as a Service (PaaS)[5]: The service provider in this context provides user resources to deploy onto cloud** infrastructure, supported applications that are designed or acquired by user. A user using this service has control over deployed applications and application hosting environment, but has no control over infrastructure such as network, storage, servers, operating systems etc. Examples for PaaS cloud's are Google App Engine, Microsoft Azure, Heroku. **• Infrastructure as a Service (IaaS): The consumer is provided with power to control process, manage storage, network** and other fundamental computing resources which are helpful to manage arbitrary software and this can include operating system and applications. By using this kind of service, user has control over operating system, storage, deployed applications and possible limited control over selected networking components. Examples for IaaS cloud's are Eucalyptus (The Eucalyptus Opensource Cloud-computing System), Amazon EC2, Rackspace, Nimbus. ### 6208 | P a g e ----- V o l u m e 1 6 N u m b e r 2 I N T E R N A T I O N A L J O U R N A L O F C O M P U T E R S & T E C H N O L O G Y **Figure 3. Cloud Computing Service Model** ### CLOUD COMPUTING: DEPLOYMENT MODELS Among the service models explained above, SaaS, PaaS and IaaS are popular among providers and users. These services can be deployed on one or more deployment models such as, public cloud, private cloud, community cloud and hybrid cloud to use features of cloud computing. Each of these deployment models are explained as follows: - Public cloud: This type of infrastructure is made available to large industrial groups or public. These are maintained and owned by organization selling cloud services. - Private cloud: This type of cloud deployment is just kept accessible to the organization that designs it. Private clouds can be managed by third party or the organization itself. In this scenario, cloud servers may or may not exist in the same place where the organization is located. - Hybrid cloud: With in this deployment model there can be two or more clouds like private, public or a community. These constituting clouds (combinations of clouds used, such as `private and public', `public and community', etc.) remain different but yet bound together by standardized or preparatory technology that enables application and data portability. - **Community cloud: This type of cloud infrastructure is shared by several organizations and supports a specific** community with shared concerns. This can be managed by an organization or third party and can be deployed off or in the organizational premise. Usage of deployments models and services modeled provided by CC changes how systems are connected and work is done in an organization. It adds up dynamically expandable nature to the applications, platforms, infrastructure or any other resource that is ordered and used in CC. **Figure 4. Types of Cloud** ## LOAD BALANCING One of the foremost usually used applications of load balancing is to produce quality of service from multiple servers, typically called a server data center. Usually load-balanced systems are properly working inside popular internet sites, big chat networks, high-bandwidth file transfer protocol sites, and domain name System (DNS) servers. It additionally ### 6209 | P a g e ----- V o l u m e 1 6 N u m b e r 2 I N T E R N A T I O N A L J O U R N A L O F C O M P U T E R S & T E C H N O L O G Y prevents the clients from contacting back-end servers directly, which can have security advantages by hiding the structure of the inner network. Some load balancers give a mechanism for improving the one parameter specially within back end server Load balancing offers the IT team an opportunity to attain a considerably higher fault tolerance. It will mechanically give the capability required to handle any increase or decrease of application traffic. It is additionally necessary that the load balancer itself doesn't become the cause of failure. Sometimes load balancers enforced in high-availability servers can additionally replicate the user’s session needed by the application. Load balancing is dividing work load between a set of computers in order to receive the good response time and all the nodes are equally loaded and, in general, all users get served quicker. Load balancing may be enforced with hardware, software, or a mix of each. Typically, load balancing is that the main reason for server’s unbalanced response time. Load balancing plans to optimize the usage of resources, maximize overall success ratio, minimize waiting time interval, and evade overloading of the resources. By the utilization of multiple algorithms and mechanisms with load balancing rather than one algorithm might increase reliability and efficiency. Load balancing within the cloud differs from classical thinking on load balancing design and implementation by misusage of data center servers to perform the requests on the basis of first come first serve basis. The older load balancing algorithm allocates the requests according to the incoming requests of the client. ## RELATED WORK Nguyen Khac Chien et al. (2016) has proposed a load balancing algorithm which is used to enhance the performance of the cloud environment based on the method of estimating the end of service time. They have succeeded in enhancing the service time and response time of the user. Ankit Kumar et al (2016) focuses on the load balancing algorithm which distributes the incoming jobs among VMs optimally in cloud data centers. The proposed algorithm in this research work has been implemented using Cloud Analyst simulator and the performance of the proposed algorithm is compared with the three algorithms which are preexists on the basis of response time. In the cloud computing milieu, the cloud data centers and the users of the cloud-computing are globally situated, therefore it is a big challenge for cloud data centers to efficiently handle the requests which are coming from millions of users and service them in an efficient manner. S.Yakhchi et al. (2015) discusses that the energy consumption has become a major challenge in cloud computing infrastructures. They proposed a novel power aware load balancing method, named ICAMMT to manage power consumption in cloud computing data centers. We have exploited the Imperialism Competitive Algorithm (ICA) for detecting over utilized hosts and then we migrate one or several virtual machines of these hosts to the other hosts to decrease their utilization. Finally, we consider other hosts as underutilized host and if it is possible, migrate all of their VMs to the other hosts and switch them to the sleep mode. Surbhi Kapoor et al. (2015) aims at achieving high user satisfaction by minimizing response time of the tasks and improving resource utilization through even and fair allocation of cloud resources. The traditional Throttled load balancing algorithm is a good approach for load balancing in cloud computing as it distributes the incoming jobs evenly among the VMs. But the major drawback is that this algorithm works well for environments with homogeneous VMS, does not considers the resource specific demands of the tasks and has additional overhead of scanning the entire list of VMs every time a task comes. The issues have been addressed by proposing an algorithm Cluster based load balancing which works well in heterogeneous nodes environment, considers resource specific demands of the tasks and reduces scanning overhead by dividing the machines into clusters. Shikha Garg et al. (2015) aims to distribute workload among multiple cloud systems or nodes to get better resource utilization. It is the prominent means to achieve efficient resource sharing and utilization. Load balancing has become a challenge issue now in cloud computing systems. To meets the user’s huge number of demands, there is a need of distributed solution because practically it is not always possible or cost efficient to handle one or more idle services. Servers cannot be assigned to particular clients individually. Cloud Computing comprises of a large network and components that are present throughout a wide area. Hence, there is a need of load balancing on its different servers or virtual machines. They have proposed an algorithm that focuses on load balancing to reduce the situation of overload or under load on virtual machines that leads to improve the performance of cloud substantially. Reena Panwar et al. (2015) describes that the cloud computing has become essential buzzword in the Information Technology and is a next stage the evolution of Internet, The Load balancing problem of cloud computing is an important problem and critical component adequate operations in cloud computing system and it can also prevent the rapid development of cloud computing. Many clients from all around the world are demanding the various services rapid rate in the recent time. Although various load balancing algorithms have been designed that are efficient in request allocation by the selection of correct virtual machines. A dynamic load management algorithm has been proposed for distribution of the entire incoming request among the virtual machines effectively. Mohamed Belkhouraf et al. (2015) aims to deliver different services for users, such as infrastructure, platform or software with a reasonable and more and more decreasing cost for the clients. To achieve those goals, some matters have to be addressed, mainly using the available resources in an effective way in order to improve the overall performance, while taking into consideration the security and the availability sides of the cloud. Hence, one of the most studied aspects by researchers is load balancing in cloud computing especially for the big distributed cloud systems that deal with many clients and big amounts of data and requests. The proposed approach mainly ensures a better overall performance with efficient load balancing, the continuous availability and a security aspect. ### 6210 | P a g e ----- V o l u m e 1 6 N u m b e r 2 I N T E R N A T I O N A L J O U R N A L O F C O M P U T E R S & T E C H N O L O G Y Lu Kang et al. (2015) improves the weighted least connections scheduling algorithm, and designs the Adaptive Scheduling Algorithm Based on Minimum Traffic (ASAMT). ASAMT conducts the real-time minimum load scheduling to the node service requests and configures the available idle resources in advance to ensure the service QoS requirements. Being adopted for simulation of the traffic scheduling algorithm, OPNET is applied to the cloud computing architecture. Hiren H. Bhatt et al. (2015) presents a Flexible load sharing algorithm (FLS) which introduce the third function. The third function makes partition the system in to domain. This function is helpful for the selection of other nodes which are present in the same domain. By applying the flexible load sharing to the particular domains in to the distribute system, the performance can be improved when any node is in overloaded situation. ## RESEARCH GAP Cloud computing thus involving distributed technologies to satisfy a variety of applications and user needs. Sharing resources, software, information via internet are the main functions of cloud computing with an objective to reduced capital and operational cost, better performance in terms of response time and data processing time, maintain the system stability and to accommodate future modification in the system .So there are various technical challenges that needs to be addressed like Virtual machine migration, server consolidation, fault tolerance, high availability and scalability but central issue is the load balancing, it is the mechanism of distributing the load among various nodes of a distributed system to improve both resource utilization and job response time while also avoiding a situation where some of the nodes are heavily loaded while other nodes are idle or doing very little work. It also ensures that all the processor in the system or every node in the network does approximately the equal amount of work at any instant of time. To make the final determination, the load balancer retrieves information about the candidate server’s health and current workload in order to verify its ability to respond to that request. Load balancing solutions can be divided into software-based load balancers and hardware-based load balancers. Hardware-based load balancers are specialized boxes that include Application Specific Integrated Circuits (ASICs) customized for a specific use. They have the ability to handle the high-speed network traffic whereas Software-based load balancers run on standard operating systems and standard hardware components. ## PROBLEM FORMULATION Clients request for the virtual machine and cloud broker handles the client request according to available virtual machine. If the VM is idle, then broker allocate that VM to user for the processing and if VM is not free then incoming requests of the client are send into the waiting state until the resources are free. 1. The current algorithm will work only in the homogeneous cloud system where all the resources are of same configuration. 2. Cloudlets are not assigned to the virtual machine according to their capacities. There may be a scenario where a cloudlet with highest priority is assigned to the machine with lowest capacity in the host. 3. The processing capacity (No of processors / MIPS) is not considered for assigning the VM to a job. 4. Every time before allocation, extra overhead is involved in parsing the table of virtual machines from top to bottom. ### CLOUD SIM Cloud service providers charge users depending upon the space or service provided. In R&D [16], it is not always possible to have the actual cloud infrastructure for performing experiments. For any research scholar, academician or scientist, it is not feasible to hire cloud services every time and then execute their algorithms or implementations. For the purpose of research, development and testing, open source libraries are available, which give the feel of cloud services. Nowadays, in the research market, cloud simulators are widely used by research scholars and practitioners, without the need to pay any amount to a cloud service provider. **Tasks performed by cloud simulators :** The following tasks can be performed with the help of cloud simulators: - Modelling and simulation of large scale cloud computing data centres. - Modelling and simulation of virtualised server hosts, with customisable policies for provisioning host resources to VMs. - Modelling and simulation of energy-aware computational resources. - Modelling and simulation of data centre [18] network topologies and message-passing applications. - Modelling and simulation of federated clouds. - Dynamic insertion of simulation elements, stopping and resuming simulation. - User-defned policies for allocation of hosts to VMs, and policies for allotting host resources to VMs. **The scope and features of cloud simulations include:** - Data centres - Load balancing ### 6211 | P a g e ----- V o l u m e 1 6 N u m b e r 2 I N T E R N A T I O N A L J O U R N A L O F C O M P U T E R S & T E C H N O L O G Y - Creation and execution of cloudlets - Resource provisioning - Scheduling of tasks - Storage and cost factors ## CONCLUSION This paper is based on cloud computing technology which has a very vast potential and is still unexplored. The capabilities of cloud computing are endless. Cloud computing provides everything to the user as a service which includes platform as a service, application as a service, infrastructure as a service. One of the major issues of cloud computing is load balancing because overloading of a system may lead to poor performance which can make the technology unsuccessful. So there is always a requirement of efficient load balancing algorithm for efficient utilization of resources. Our paper focuses on the various load balancing algorithms and their applicability in cloud computing environment. ## REFERENCES [1] S. Yakhchi, S. Ghafari, M. Yakhchi, M. Fazeli and A. Patooghy, "ICA-MMT: A Load Balancing Method in Cloud Computing Environment," IEEE, 2015. [2] S. Kapoor and D. C. Dabas, "Cluster Based Load Balancing in Cloud Computing," IEEE, 2015. [3] S. Garg, R. Kumar and H. Chauhan, "Ef?cient Utilization of Virtual Machines in Cloud Computing using Synchronized Throttled Load Balancing," 1st International Conference on Next Generation Computing Technologies (NGCT-2015), pp. 77-80, 2015. [4] R. Panwar and D. B. Mallick, "Load Balancing in Cloud Computing Using Dynamic Load Management Algorithm," IEEE, pp. 773-778, 2015. [5] M. Belkhouraf, A. Kartit, H. Ouahmane, H. K. Idrissi,, Z. Kartit and M. . E. Marraki, "A secured load balancing architecture for cloud computing based on multiple clusters," IEEE, 2015. [6] L. Kang and X. Ting, "Application of Adaptive Load Balancing Algorithm Based on Minimum Traffic in Cloud Computing Architecture," IEEE, 2015. [7] N. K. Chien, N. H. Son and H. D. Loc, "Load Balancing Algorithm Based on Estimating Finish Time of Services in Cloud Computing," ICACT, pp. 228-233, 2016. [8] H. H. Bhatt and H. A. Bheda, "Enhance Load Balancing using Flexible Load Sharing in Cloud Computing," IEEE, pp. 72-76, 2015. [9] S. S. MOHARANA, R. D. RAMESH and D. POWAR, "ANALYSIS OF LOAD BALANCERS IN CLOUD COMPUTING," International Journal of Computer Sciencand Engineering (IJCSE), pp. 102-107, 2013. [10] M. P. V. Patel, H. D. Patel and . P. J. Patel, "A Survey On Load Balancing In Cloud Computing," International Journal of Engineering Research & Technology (IJERT), pp. 1-5, 2012. [11] R. Kaur and P. Luthra, "LOAD BALANCING IN CLOUD COMPUTING," Int. J. of Network Security,, pp. 1-11, 2013. [12] Kumar Nishant,, P. Sharma, V. Krishna, Nitin and R. Rastogi, "Load Balancing of Nodes in Cloud Using Ant Colony Optimization," IEEE, pp. 3-9, 2012. [13] Y. Xu, L. Wu, L. Guo,, Z. Chen, L. Yang and Z. Shi, "An Intelligent Load Balancing Algorithm Towards Ef?cient Cloud Computing," AI for Data Center Management and Cloud Computing: Papers from the 2011 AAAI Workshop (WS-11-08), pp. 27-32, 2011. [14] A. K. Sidhu and S. Kinger, "Analysis of Load Balancing Techniques in Cloud Computing," International Journal of Computers & Technology Volume 4 No. 2, March-April, 2013, ISSN 2277-3061, pp. 737-741, 2013. [15] O. M. Elzeki, M. Z. Reshad and M. A. Elsoud, "Improved Max-Min Algorithm in Cloud Computing," International Journal of Computer Applications (0975 – 8887), pp. 22-27, 2012. ### 6212 | P a g e -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.24297/ijct.v16i2.6046?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.24297/ijct.v16i2.6046, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "HYBRID", "url": "https://cirworld.com/index.php/ijct/article/download/6046/5981" }
2,017
[ "Review" ]
true
2017-04-06T00:00:00
[ { "paperId": "7d49c7f1dc2187077873b8561c0c92277ef17288", "title": "Load balancing in cloud computing using dynamic load management algorithm" }, { "paperId": "88866c141a5cbcf802bf1e64861885928dd2a8db", "title": "Efficient utilization of virtual machines in cloud computing using Synchronized Throttled Load Balancing" }, { "paperId": "9fdd53cb97a70ccb19589475a2b831948f08441c", "title": "Enhance load balancing using Flexible load sharing in cloud computing" }, { "paperId": "68780ce5ce5dbcd23bc084e06d397767d0288d19", "title": "Cluster based load balancing in cloud computing" }, { "paperId": "fbf1a24d685097a963e66d9ab52e6f32b22d3d2b", "title": "Application of adaptive load balancing algorithm based on minimum traffic in cloud computing architecture" }, { "paperId": "e0756b7a204ee52b82ca5a2824c74bd546467534", "title": "A secured load balancing architecture for cloud computing based on multiple clusters" }, { "paperId": "94b1f5eef5f37799a005efd24bafefaef97b7a17", "title": "ICA-MMT: A load balancing method in cloud computing environment" }, { "paperId": "b20ece3866c93c2e8d4fa6ce8cf4db91cf13e4e5", "title": "Improved Max-Min Algorithm in Cloud Computing" }, { "paperId": "8a64a37f426f030266d22eaac0b9ca18bde798e1", "title": "Load Balancing of Nodes in Cloud Using Ant Colony Optimization" }, { "paperId": "d632593703a604d9ab27e9310ecd9a849d405346", "title": "Analysis of Load Balancing Techniques in Cloud Computing" }, { "paperId": "3aa6121d76b8661d7911ebce4fa8f88f20e49def", "title": "Load balancing algorithm based on estimating finish time of services in cloud computing" }, { "paperId": "8583671da2e3a598f8c1842fdf6da053a7f9ec29", "title": "Load Balancing in Cloud Computing" }, { "paperId": "d0af10255083ce3fd567ca67192542c0b36eab53", "title": "ANALYSIS OF LOAD BALANCERS IN CLOUD COMPUTING" }, { "paperId": "49f26118a6728ea07412328873e726560c0b658a", "title": "An Intelligent Load Balancing Algorithm Towards Efficient Cloud Computing" } ]
5,719
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0101445aec81d2dec8562a83e656ac6ccd633ee2
[ "Computer Science" ]
0.87755
Founding Cryptography on Tamper-Proof Hardware Tokens
0101445aec81d2dec8562a83e656ac6ccd633ee2
IACR Cryptology ePrint Archive
[ { "authorId": "1707396", "name": "Vipul Goyal" }, { "authorId": "1688856", "name": "Yuval Ishai" }, { "authorId": "1695851", "name": "A. Sahai" }, { "authorId": "144002578", "name": "R. Venkatesan" }, { "authorId": "144506573", "name": "A. Wadia" } ]
{ "alternate_issns": null, "alternate_names": [ "IACR Cryptol eprint Arch" ], "alternate_urls": null, "id": "166fd2b5-a928-4a98-a449-3b90935cc101", "issn": null, "name": "IACR Cryptology ePrint Archive", "type": "journal", "url": "http://eprint.iacr.org/" }
null
# Founding Cryptography on Tamper-Proof Hardware Tokens Vipul Goyal[1][,⋆], Yuval Ishai[2][,⋆⋆], Amit Sahai[3][,⋆⋆⋆], Ramarathnam Venkatesan[4], and Akshay Wadia[3] 1 UCLA and MSR India vipul.goyal@gmail.com 2 Technion and UCLA yuvali@cs.technion.ac.il 3 UCLA _{sahai,awadia}@cs.ucla.edu_ 4 Microsoft Research, India and Redmond venkie@microsoft.com **Abstract. A number of works have investigated using tamper-proof** hardware tokens as tools to achieve a variety of cryptographic tasks. In particular, Goldreich and Ostrovsky considered the problem of software protection via oblivious RAM. Goldwasser, Kalai, and Rothblum introduced the concept of one-time programs: in a one-time program, an honest sender sends a set of simple hardware tokens to a (potentially malicious) receiver. The hardware tokens allow the receiver to execute a secret program specified by the sender’s tokens exactly once (or, more generally, up to a fixed t times). A recent line of work initiated by Katz examined the problem of achieving UC-secure computation using hardware tokens. Motivated by the goal of unifying and strengthening these previous notions, we consider the general question of basing secure computation on hardware tokens. We show that the following tasks, which cannot be realized in the “plain” model, become feasible if the parties are allowed to generate and exchange tamper-proof hardware tokens. **– Unconditional and non-interactive secure computation. We** show that by exchanging simple stateful hardware tokens, any functionality can be realized with unconditional security against malicious parties. In the case of two-party functionalities f (x, y) which take their inputs from a sender and a receiver and deliver their output to the receiver, our protocol is non-interactive and only requires a unidirectional communication of simple stateful tokens from the The original version of this chapter was revised: The copyright line was incorrect. This has been [corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-642-11799-2_36](http://dx.doi.org/10.1007/978-3-642-11799-2_36) _⋆_ Research supported in part by a Microsoft Research Graduate Fellowship and the grants of Amit Sahai mentioned below. _⋆⋆_ Supported in part by ISF grant 1310/06, BSF grants 2008411, and NSF grants 0830803, 0716835, 0627781. _⋆⋆⋆_ Research supported in part from NSF grants 0916574, 0830803, 0627781, and 0716389, BSF grant 2008411, an equipment grant from Intel, and an Okawa Foundation Research Grant. ----- sender to the receiver. This strengthens previous feasibility results for one-time programs both by providing unconditional security and by offering general protection against malicious senders. As is typically the case for unconditionally secure protocols, our protocol is in fact UC-secure. This improves over previous works on UC-secure computation based on hardware tokens, which provided computational security under cryptographic assumptions. **– Interactive secure computation from** stateless tokens based on one-way functions. We show that stateless hardware tokens are sufficient to base general secure (in fact, UC-secure) computation on the existence of one-way functions. **– Obfuscation from stateless tokens. We consider the problem** of realizing non-interactive secure computation from stateless tokens for functionalities which allow the receiver to provide an arbitrary number of inputs (these are the only functionalities one can hope to realize non-interactively with stateless tokens). By building on recent techniques for resettably secure computation, we obtain a general positive result under standard cryptographic assumptions. This gives the first general feasibility result for program obfuscation using stateless tokens, while strengthening the standard notion of obfuscation by providing security against a malicious sender. ## 1 Introduction A number of works (e.g. [1,2,3,4,5,6,7,8,9,10,11,12,13]) have investigated using tamper-proof hardware tokens[1] as tools to achieve a variety of cryptographic goals. There has been a surge of research activity on this front of late. In particular, the recent work of Katz [9] examined the problem of achieving UC-secure [14] two party computation using tamper-proof hardware tokens. A number of followup papers [10,11,12] have further investigated this problem. In another separate (but related) work, Goldwasser et al. [13] introduced the concept of one-time _programs: in a one-time program, a (semi-honest) sender sends a set of very_ simple hardware tokens to a (potentially malicious) receiver. The hardware tokens allow the receiver to execute a program specified by the sender’s tokens exactly once (or, more generally, up to a fixed t times). This question is related to the more general goal of software protection using hardware tokens, which was first addressed by Goldreich and Ostrovsky [1] using the framework of oblivious RAM. The present work is motivated by the observation that several of these pre vious goals and concepts can be presented in a unified way as instances of one general goal: realizing secure computation using tamper-proof hardware tokens. The lines of work mentioned above differ in the types of functionalities being 1 Informally, a tamper-proof hardware token provides the holder of the token with black-box access to the functionality of the token. We will often omit the words “tamper-proof” when referring to hardware tokens, but all of the hardware tokens referred to in this paper are assumed to be tamper-proof. ----- considered (e.g., non-reactive vs. reactive), the type of interaction between the parties (interactive vs. non-interactive protocols), the type of hardware tokens (stateful vs. stateless, simple vs. complex), and the precise security model (standalone vs. UC, semi-honest vs. malicious parties). This unified point of view also gives rise to strictly stronger notions than those previously considered, which in turn give rise to new feasibility questions in this area. The introduction of tamper-proof hardware tokens to the model of secure com putation, as formalized in [9], invalidates many of the fundamental impossibility results in cryptography. Taking a step back to look at this general model from a foundational perspective, we find that a number of natural feasibility questions regarding secure computation with hardware tokens remain open. In this work we address several of these questions, focusing on goals that are impossible to realize in the plain model without tamper-proof hardware tokens: **– Is it possible to achieve unconditional security for secure computa-** **tion with hardware tokens? We note that this problem is open even for** stand-alone security, let alone UC security, and impossible in the plain model [15]. While in the semi-honest model this question is easy to settle by relying on unconditional protocols based on oblivious transfer (OT) [16,17,18,19], this question is more challenging when both parties as well as the tokens they generate can be malicious. (See Sections 1.2 and 3.1 for relevant discussion.) In the case of stateless tokens, which may be much easier to implement, security against unbounded adversaries cannot be generally achieved, since an unbounded adversary can “learn” the entire description of the token. A natural question in this case is whether stateless tokens can be used **to realize (UC) secure computation based on the assumption that** **one-way functions exist.** Previous positive results for secure two-party computation with hardware tokens relied either on specific number theoretic assumptions [9] or the existence of oblivious transfer protocols in the plain model [10,11], or alternatively offered weaker notions of security [20]. A related question is: is it possible to obtain unconditionally secure **one-time programs for all polynomial-time computable functions?** The previous work of [13] required the existence one-way functions in order to construct one-time programs. **– Is it possible to realize non-interactive secure two-party computa-** **tion with simple hardware tokens? Again, this problem is open[2]** even for stand-alone security, and impossible in the plain model. Constructions of oblivious RAM [1] and one-time programs [13] provide partial solutions to 2 All the previous questions were open even without any restriction on the size of the tokens. In the current and the following questions we restrict the tokens to be simple in the sense that the size of each token can only depend on the security parameter. This rules out a trivial solution of building a token which realizes a party in a secure two-party computation protocol. ----- this problem; however, in these models the sender is semi-honest.[3] Thus, in the context of one-time programs we ask: is it possible to achieve one**time programs tolerating a malicious sender? We note that [13] make** partial progress towards this question by constructing one-time zero knowledge proofs, where the prover can be malicious. However, in the setting of hardware tokens, the GMW paradigm [21] of using zero knowledge proofs to compile semi-honest protocols into protocols tolerating malicious behavior does not apply, since one would potentially need to prove statements about hardware tokens (as opposed to ordinary NP statements). **– Which notions of program obfuscation can be realized using simple** **hardware tokens? Again, this problem can be captured in an elegant way** within the framework of secure two-party computation, except that here we need to consider reactive functionalities which may take a single input from the “sender” and a sequence of (possibly adaptively chosen) inputs from the “receiver”. Obfuscation can be viewed as a non-interactive secure realization of such functionalities. While this general goal is in some sense realized by the construction of oblivious RAM [1] (which employs stateful tokens), several natural questions remain: Is it possible to achieve obfuscation using **only stateless tokens? Is it possible to offer a general protection** **against a malicious sender using stateless or even stateful tokens?** To illustrate the motivation for the latter question, consider the goal of obfuscating a poker-playing program. The receiver of the obfuscated program would like to be assured that the sender did not violate the rules of the game (and in particular cannot bias the choice of the cards). **– What are the simplest kinds of tamper-proof hardware tokens** **needed to realize the above goals? For example, Goldwasser et al. [13] in-** troduce a very simple kind of stateful token that they call an OTM (one-time memory) token.[4] An OTM token stores two strings s0 and s1, takes a single bit b as input, and then outputs sb and stops working (or self-destructs). Note that an OTM token essentially implements the one-out-of-two string OT functionality; a subtle distinction between OTM and traditional OT is discussed in Section 3.1. An even simpler type of token is a bit-OTM token, where the strings s0 and s1 are restricted to be single bits. Is it possible **to realize unconditional, non-interactive, or UC-secure two-party** **computation using only bit-OTM tokens? We note that previous works** on secure two-party computation with hardware tokens [9,10,11,20] all make use of more complicated hardware tokens. 3 In these models, the sender is allowed to arbitrarily specify the functionality of the oblivious RAM or the one-time program, and the receiver knows nothing about this functionality except an upper bound on its circuit size or running time. (Thus, the issue of dishonest senders does not arise in these models.) In the present work, by a one-time program tolerating a malicious sender, we mean that the receiver knows some partial specification of the functionality – modeled in the usual paradigm of secure two-party computation. 4 The use of OTM tokens in [13] is motivated in part by the goal of achieving leakage _resilience, a feature that our constructions using such tokens inherit as well._ ----- **1.1** **Our Results** We show that the following tasks, which cannot be realized in the “plain” model, become feasible if the parties are allowed to generate and exchange simple tamper-proof hardware tokens. **– Unconditional non-interactive secure computation. We show that by** exchanging stateful hardware tokens, any functionality can be realized with _unconditional security against malicious parties. In the case of two-party_ functionalities f (x, y) which take their inputs from a sender and a receiver and deliver their output to the receiver, our protocol is non-interactive and only requires a unidirectional communication of tokens from the sender to the receiver (in case an output has to be given to both parties, adding a reply from the receiver to the sender is sufficient). This result strengthens previous feasibility results for one-time programs by providing unconditional security, by offering general protection against malicious senders, and by using only bit-OTM tokens. As is typically the case for unconditionally secure protocols, our protocol is in fact UC-secure. This improves over previous works on UC-secure computation based on hardware tokens, which provided computational security under cryptographic assumptions. See Sections 3.1 and 3.2 for details of this result and a high level overview of techniques. **– Interactive secure computation from stateless tokens based on one-** **way functions. We show that stateless hardware tokens are sufficient to** base general secure (in fact, UC-secure) computation on the existence of one_way functions. One cannot hope for security against unbounded adversaries_ with stateless tokens since an unbounded adversary could query the token multiple times to “learn” the functionality it contains. See Section 4 for details. **– Obfuscation from stateless tokens. We consider the problem of real-** izing non-interactive secure computation from stateless tokens for reactive functionalities which take a single input from the sender and an arbitrary sequence of inputs from the receiver (these are the only functionalities one can hope to realize non-interactively with stateless tokens). By building on recent techniques for resettably secure computation [22], we obtain a general positive result under standard cryptographic assumptions. This gives the first general feasibility result for program obfuscation using stateless tokens, while strengthening the standard notion of obfuscation by providing security against a malicious sender. We also propose constructions of non-interactive secure computation for general reactive functionalities with stateful tokens. See the full version for details. In all of the above results, the size of each hardware token is either constant or polynomial in the security parameter, and its code is independent of the inputs of the parties. Thus, the tokens could theoretically be “mass-produced” before being used in any particular protocol with any particular inputs. ----- We stress that in contrast to some previous results along this line (most no tably, [1,13,20]), our focus is almost entirely on feasibility questions, while only briefly discussing more refined efficiency considerations. However, in most cases our stronger feasibility results can be realized while also meeting the main efficiency goals pursued in previous works. The first two results above are obtained by utilizing previous protocols for secure computation based on OT [18,19], and thus a main ingredient in our constructions is showing how to securely implement OT using hardware tokens. Note that in the case of non-interactive secure computation, additional tools are needed since the protocols of [18,19] are (necessarily) interactive. **1.2** **Related Work** The use of tamper-proof hardware tokens for cryptographic purposes was first explored by Goldreich and Ostrovsky [1] in the context of software protection (one-time programs [13] is a relaxation of this goal, generally called program obfuscation [23]), and by Chaum, Pederson, Brands, and Cramer [2,3,4] in the context of e-cash. Ishai, Sahai, and Wagner [5] and Ishai, Prabhakaran, Sahai and Wagner [24] consider the question of how to construct tamper-proof hardware tokens when the hardware itself does not guarantee complete protection against tampering. Gennaro, Lysyanskaya, Malkin, Micali, and Rabin [6] consider a similar question, when the underlying hardware guarantees that part of the hardware is tamper-proof but readable, while the other part of the hardware is unreadable but susceptible to tampering. Moran and Naor [8] considered a relaxation of tamper-proof hardware called “tamper-evident seals,” and given number of constructions of graphic tasks based on this relaxed notion. Hofheinz, M¨uller-Quade, and Unruh [25] consider a model similar to [9] in the context of UC-secure protocols where tamper-proof hardware tokens (signature cards) are issued by a trusted central authority. The model that we primarily build on here is due to Katz [9], who considers a setting in which users can create and exchange tamper-proof hardware tokens where malicious users have full control over the functionality realized by each token they create. The main result of [9] is a general protocol for UC-secure twoparty computation using stateful tokens, under the DDH assumption. Chandran, Goyal, Sahai [10] implement UC-secure two-party computation using stateless tokens, under the assumption that oblivious transfer protocols exist in the plain model. Aside from just considering stateless tokens, [10] also introduce a variant of the model of [9] that allows for the adversary to pass along tokens, and in general allows the adversary not to know the code of the tokens he produces. We do not consider this model here. Moran and Segev [11] also implement UC-secure two-party computation under the same assumption as [10], but using stateful tokens, and only requiring tokens to be passed in one direction. Damg˚ard, Nielsen, and Wichs [12] show how to relax the “isolation” requirement of tamper-proof hardware tokens, and consider a model in which tokens can communicate a fixed number of bits back to its creator. Hazay and Lindell [20] propose constructions of practical protocols for various problems of interest using trusted stateful ----- tokens. Very recently and independently of our work, practical oblivious transfer protocols using stateless tokens and relying only on one-way functions were suggested by Kolesnikov [26]. In contrast to the corresponding feasibility result from our work, these protocols either provide a weaker security guarantee or assume that tokens are well-formed, but on the other hand they offer better practical efficiency. Goldwasser, Kalai, and Rothblum [13] introduced the notion of one-time pro grams, and showed how to realize it under the assumption that one-way functions exist, as we have already discussed. They also construct one-time zero-knowledge proofs under the same assumption. Their results focus mainly on achieving efficiency in terms of the number of tokens needed, and a non-adaptive use of the tokens by the receiver. Finally, in a seemingly unrelated work which is motivated by quantum physics, Buhrman, Christandl, Unger, Wehner and Winter [27] consider the application of non-local boxes to cryptography. Using non-local boxes, Buhrman et al. show an unconditional construction for oblivious transfer in the interactive setting. A non-local box implements a trusted functionality taking input and giving output to both the parties (as opposed to OTM tokens which could be prepared maliciously). However, the key problem faced by Buhrman et al. is similar to a problem we face as well: delayed invocation of the non-local box by a malicious party. Indeed, one can give a simple interactive protocol (omitted here) for building a trusted non-local-box using OTM tokens. This provides an alternative to the interactive variant of our construction of unconditional secure computation from hardware tokens described in Section 3.1. ## 2 Preliminaries In this section we briefly discuss some of the underlying definitions and concepts. The reader is referred to the full version for the details. We use the UC-framework of Canetti [28] to capture the general notion of se cure computation of (possibly reactive) functionalities. Our main focus is on the two-party case. We will usually refer to one party as a “sender” and to another as a “receiver”. A non-reactive functionality may receive an input from each party and deliver output to each party (or only to the receiver). A reactive functionality may have several rounds of inputs and outputs, possibly maintaining state information between rounds. Our model for tamper-proof hardware is similar to that of Katz [9]. As we consider both stateful and stateless tokens, we define different ideal functionalities for the two. By Fwrap[single] we denote an ideal functionality that allows a sender to generate a “one-time token” which can be invoked by its designated receiver. A one-time token is a stateful token which takes an input from the receiver and returns a function which is specified in advance by the sender. (Note that if the sender is malicious, this function can be arbitrary.) After being invoked by the receiver, such a token “self-destructs”. Thus, the only state these tokens keep is a flag which indicates whether the token has been run or not. Simple tokens of this type were used in [13]. ----- We also define an ideal functionality Fwrap[stateless] for stateless tokens. Here the token computes some (deterministic) function specified by the sender, and the receiver can query the token an unbounded number of times. Note that this makes stateless tokens useless if the receiver has enough resources to “learn” the token’s description (either because the token is too small or the receiver is too powerful). [5] By a non-interactive protocol we refer to a protocol in which the communi cation only involves a single batch of tokens, possibly along with an additional message, communicated from a sender to a receiver. ## 3 Unconditional Non-interactive Secure Computation Using Stateful Tokens In this section we establish the feasibility of unconditionally non-interactive secure computation based on stateful hardware tokens. As is typically the case for unconditionally secure protocols, our protocols are in fact UC secure. This section is organized as follows. In Subsection 3.1 we present an interactive protocol for arbitrary functionalities, which requires the parties to engage in multiple rounds of interaction. This gives an unconditional version of previous protocols for UC-secure computation based on hardware tokens [9,10,11], which all relied on computational assumptions.[6] This subsection also introduces some useful building blocks that are used for the non-interactive solution in the next subsection. In Subsection 3.2 we consider the case of secure evaluation of two-party func tionalities which deliver output to only one of the parties (the “receiver”). We strengthen the previous result in two ways. First, we show that in this case interaction can be completely eliminated: it suffices for the sender to non-interactively send tokens to the receiver, without any additional communication. Second, we show that even very simple, constant-size stateful tokens are sufficient for this purpose. This strengthens previous feasibility results for one-time programs [13] by providing unconditional security (in fact, UC-security), by offering general protection against malicious senders, and by using constant-size tokens. **3.1** **The Interactive Setting** Unconditionally secure two-party computation is impossible to realize for most nontrivial functionalities, even with semi-honest parties [29,30]. However, if the parties are given oracle access to a simple ideal functionality such as Oblivious 5 While the formal definition of this functionality forces a malicious sender to also use only stateless tokens, this requirement can be relaxed without affecting the security of our protocols. See Section 4 for details. 6 The work of [11] realizes an unconditionally UC-secure commitment from stateful to kens. This does not directly yield protocols for secure computation without additional computational assumptions. ----- Transfer (OT) [16,17], then it becomes possible not only to obtain unconditionally secure computation with semi-honest parties [31,32,33], but also unconditional UC-security against malicious parties [18,19]. This serves as a natural starting point for our construction. In the OT-hybrid model, the two parties are given access to the following ideal OT functionality: the input of P1 (the “sender”) consists of a pair of k-bit strings (s0, s1), the input of P2 (the “receiver”) is a choice bit c, and the receiver’s output is the chosen string sc. The natural way to implement a single OT call using stateful hardware tokens is by having the sender send to the receiver a token which, on input c, outputs sc and erases s1−c from its internal state. The use of such hardware tokens was first suggested in the context of one-time programs [13]. Following the terminology of [13], we refer to such tokens as OTM (one-time-memory) tokens. An appealing feature of OTM tokens is their simplicity, which can also lead to better resistance against side-channel attacks (see [13] for discussion). This simplicity feature served as the main motivation for using OTM tokens as a basis for one-time programs. Another appealing feature, which is particularly important in our context, is that the OTM functionality does not leave room for bad sender strategies: whatever badly formed token a malicious sender may send is equivalent from the point of view of an honest receiver to having the sender send a well-formed OTM token picked from some probability distribution. (This is not the case for tokens implementing more complex functionalities, such as 2-out-of-3 OT or the extended OTM functionality discussed below, for which badly formed tokens may not correspond to any distribution over well-formed tokens.) Given the above, it is tempting to hope that our goal can be achieved by simply taking any unconditionally secure protocol in the OT-hybrid model, and using OTM tokens to implement OT calls. However, as observed in [13], there is a subtle but important distinction between the OT-hybrid model and the OTM-hybrid model: while in the former model the sender knows the point in the protocol in which the receiver has already made its choice and received its output, in the latter model invoking the token is entirely at the discretion of the receiver. This may give rise to attacks in which the receiver adaptively invokes the OTM tokens “out of order,” and such attacks may have a devastating effect on the security of protocols even in the case of unconditional security. A more detailed discussion of such attacks and simple solution ideas (that do not work) is included in the full version. **Extending the OTM functionality. To solve the above problem, we will** realize an extended OTM functionality which takes from the sender a pair of strings (s0, s1) along with an auxiliary string r, takes from the receiver a choice bit c, and delivers to the receiver both sc and r. We denote this functionality by ExtOTM. What makes the ExtOTM functionality nontrivial to realize using hardware tokens is the need to protect the receiver from a malicious sender who may try to make the received r depend on the choice bit c while at the same ----- _time protecting the sender from a malicious receiver who may try to postpone_ its choice c until after it learns r. Using the ExtOTM functionality, it is easy to realize a UC-style version of the OT functionality which not only delivers the chosen string to the receiver (as in the OTM functionality) but also delivers an acknowledgement to the sender. This flavor of the OT functionality, which we denote by, can be _F_ [OT] realized by having the sender invoke ExtOTM with (s0, s1) and a randomly chosen r, and having the receiver send r to the sender. In contrast to OTM, the functionality allows the sender to force any subset of the OT calls to _F_ [OT] be completed before proceeding with the protocol. This suffices for instantiating the OT calls in the unconditionally secure protocols from [18,19]. We refer the reader to the full version of this paper for a UC-style definition of the OTM, ExtOTM, and functionalities. _F_ [OT] **Realizing ExtOTM using general[7]** **stateful tokens. As discussed above,** we cannot directly use a stateful token for realizing the ExtOTM functionality, because this allows the sender to correlate the delivered r with the choice bit _c. On the other hand, we cannot allow the sender to directly reveal r to the_ receiver, because this will allow the receiver to postpone its choice until after it learns r. In the following we sketch our protocol for realizing ExtOTM using stateful tokens. This protocol is non-interactive (i.e., it only involves tokens sent from the sender to the receiver) and will also be used as a building block towards the stronger results in the next subsection. We refer the reader to the full version of this paper for a formal description of the protocol and its proof of security. Below we include a detailed overview. As mentioned above, at a high level, the challenge we face is to prevent un wanted correlations in an information-theoretic way for both malicious senders and malicious receivers. This is a more complex situation than a typical similar situation where only one side needs to be protected against (c.f. [34,35]). To accomplish this goal, we make use of secret-sharing techniques combined with additional token-based “verification” techniques to enforce honest behavior. Our ExtOTM protocol ΠExtOTM starts by having the sender break its aux iliary string r into 2k additive shares r[i], and pick 2k pairs of random strings (q0[i] _[, q]1[i]_ [). (][Each][ o][f the str][in][gs][ q]b[i] [a][n][d][ r][i][ i][s][ k][-][b][i][t][ lon][g][, w][here][ k][ i][s a stat][i][st][i][ca][l] security parameter.) It then generates 2k OTM tokens, where the i-th token contains the pair (q0[i] _[◦]_ _[r][i][, q]1[i]_ _[◦]_ _[r][i][) (w][here][ ‘][◦][’ i][s the c][on][cate][n][at][ion o][perat][o][r][). No][te]_ that a malicious sender may generate badly formed OTM tokens which correlate _r[i]_ with the i-th choice of the receiver; we will later implement a token-based verification strategy that convinces an honest receiver that the sender did not cheat (too much) in this step. Now the receiver breaks its choice bit c into 2k additive shares c[i], and invokes the 2k OTM tokens with these choice bits. Let (ˆq[i], ˆr[i]) be the pair of k-bit strings obtained by the receiver from the i-th token. Note that if the sender is honest, the 7 Here, we make use of general tokens. Later in this section, we will show how to achieve the ExtOTM functionality (and in fact every poly-time functionality) using only very simple tokens – just bit OTM tokens. ----- receiver can already learn r. We would like to allow the receiver to learn its chosen string sc while convincing it that the sender did not correlate all of the auxiliary strings ˆr[i] with the corresponding choice bits ci. (The latter guarantee is required to assure an honest receiver that ˆr = [�] _rˆ[i]_ is independent of c as required.) This is done as follows. The sender prepares an additional single-use hardware token which takes from the receiver its 2k received strings ˆq[i], checks that for each ˆq[i] there is a valid selection ˆci such that ˆq[i] = qcˆ[i]i [(o][ther][wi][se the t][o][ke][n][ retur][n][s] _⊥), and finally outputs the chosen string scˆ1⊕...⊕cˆ2k_ . (All tokens in the protocol can be sent to the receiver at one shot.) Note that the additive sharing of r in the first 2k tokens protects an honest sender from a malicious receiver who tries to learn scˆ where ˆc is significantly correlated with r, as it guarantees that the receiver effectively commits to c before obtaining any information about _r. The receiver is protected against a malicious sender because even a badly_ formed token corresponds to some (possibly randomized) ideal-model strategy of choosing (s0, s1). Finally, we need to provide to the receiver the above-mentioned guarantee that a malicious sender cannot correlate the receiver’s auxiliary output ˆr = [�] _rˆ[i]_ with the choice bit c. To explain this part, it is convenient to assume that both the sender and the badly formed tokens are deterministic. (The general case is handled by a standard averaging argument.) In such a case, we call each of the first 2k tokens well-formed if the honest receiver obtains the same r[i] regardless of its choice c[i], and we call it badly formed otherwise. By the additive sharing of c, the only way for a malicious sender to correlate the receiver’s auxiliary output with c is to make all of the first 2k tokens badly formed. To prevent this from happening, we require the sender to send a final token which proves that it knows all of the 2k auxiliary strings ˆr[i] obtained by the receiver. This suffices to convince the receiver that not all of the first 2k tokens are badly formed. Note, however, that we cannot ask the sender to send these 2k strings r[i] in the clear, since this would (again) allow a malicious receiver to postpone its choice c until after it learns r. Instead, the sender generates and sends a token which first verifies that the receiver knows r (by comparing the receiver’s input to the k-bit string r) and only then outputs all 2k shares r[i]. The verification step prevents correlation attacks by a malicious receiver. The final issue to worry about is that the string _r received by the token (which may be correlated with the receiver’s choices ci)_ does not reveal to the sender enough information to pass the test even if all of its first 2k tokens are badly formed. This follows by a simple information-theoretic argument: in order to pass the test, the token must correctly guess all 2k bits _ci, but this cannot be done (except with 2[−][Ω][(][k][)]_ probability) even when given arbitrary k bits of information about the ci. The above protocol shows the following (see full version for proof): _Claim. Protocol ΠExtOTM realizes ExtOTM with statistical UC-security in the_ _Fwrap[single][-][h][y][br][i][d m][o][de][l.]_ We are now ready to prove the main feasibility result of this subsection. ----- **Theorem 1 (Interactive unconditionally secure computation using** **stateful tokens). Let f be a (possibly reactive) polynomial-time computable** _functionality. Then there exists an efficient, statistically UC-secure interactive_ _protocol which realizes f in the Fwrap[single][-hybrid model.]_ _Proof. We compose three reductions. The protocols of [18,19] realize uncondi-_ tionally secure two-party (and multi-party) computation of general functionalities using . A trivial reduction described above reduces to ExtOTM. _F_ [OT] _F_ [OT] Finally, the above Claim reduces ExtOTM to Fwrap[single][.] **3.2** **The Non-interactive Setting** In this subsection we restrict the attention to the case of securely evaluating two-party functionalities f (x, y) which take an input x from the sender and an input y from the receiver, and deliver f (x, y) to the receiver. We refer to such functionalities as being sender-oblivious. Note that here we consider only non_reactive sender-oblivious functionalities, which interact with the sender and the_ receiver in a single round. The reactive case will be discussed in the full version. Unlike the case of general functionalities, here one can hope to obtain non _interactive protocols in which the sender unidirectionally send tokens (possibly_ along with additional messages[8]) to the receiver. For sender-oblivious functionalities, the main result of this subsection strengthens the results of Section 3.1 in two ways. First, it shows that a noninteractive protocol can indeed realize such functionalities using stateful tokens. Second, it pushes the simplicity of the tokens to an extreme, relying only on OTM tokens which contain pairs of bits. Below we provide only a high-level description of the construction and the underlying ideas. We refer the reader to the full version for the full description of the protocols and their analysis. **One-time programs. Our starting point is the concept of a one-time pro-** _gram (OTP) [13]. A one-time program can be viewed in our framework as a_ non-interactive protocol for f (x, y) which uses only OTM tokens, and whose security only needs to hold for the case of a semi-honest sender (and a malicious receiver).[9] The main result of [13] establishes the feasibility of computationallysecure OTPs for any polynomial-time computable f, based on the existence of one-way functions. The construction is based on Yao’s garbled circuit technique [37]. Our initial observation is that if f is restricted to the complexity class NC[1], one can replace Yao’s construction by an efficient perfectly secure variant (cf. [38]). This yields perfectly secure OTPs for NC[1]. Alternatively, we 8 Since our main focus is on establishing feasibility results, the distinction between the “hardware” part and the “software” part is not important for our purposes. 9 The original notion of OTP from [13] is syntactically different in that it views f as a function of the receiver’s input, where a description of f is given to the sender. This can be captured in our framework by letting f (x, y) be a universal functionality. ----- also present a general construction of a OTP from any “decomposable randomized encoding” of f . This can be used to derive perfectly secure OTPs for larger classes of functions (including NL) based on randomized encoding techniques from [39,38]. See the full version for further details. A next natural step is to construct unconditionally secure OTPs for any polynomial-time computable function f . In the full version of this paper, we describe a direct and self-contained construction which uses the perfect OTPs for NC[1] described above to build a statistically secure construction for any f . However, this result will be subsumed by our main result, which can be proved (in a less self-contained way) without relying on the latter construction. **Handling malicious senders. As in Section 3.1, the main ingredient in our** solution is an interactive secure protocol Π for f . The high level idea of our construction is obtain a non-interactive protocol for f which emulates Π by having the sender generate and send a one-time token which computes the sender’s next message function for each round of Π (a similar idea was used in [13] to construct one time proofs). Using the above procedure, we transform Π into a non-interactive protocol Π _[′]_ which uses very complex one-time tokens (for implementing the next message functions of Π). The next idea is that we can break each such complex token into simple OTM tokens by using a one-time program realization of each complex token. More details are provided in the full version. **From the plain model to the OT-hybrid model. So far we assumed the** protocol Π to be secure in the plain model. This rules out unconditional security as well as UC-security, which are our main goals in this section. A natural approach for obtaining unconditional UC-security is to extend the above compiler to protocols in the OT-hybrid model. This introduces a subtle difficulty which was already encountered in Section 3.1: the sender cannot directly implement the OT calls by using OTM tokens. To solve this problem, we build on the (non-interactive) ExtOTM protocol from Section 3.1. See full version for details. **From string-OTM to bit-OTMs. As a final optimization, in the full version** we show how to use an unconditionally UC-secure non-interactive implementation of a string-OTM token using bit-OTM tokens. This yields the following main result of this section: **Theorem 2 (Non-interactive unconditionally secure computation us-** **ing** **bit-OTM** **tokens).** _Let f_ (x, y) be a non-reactive, sender-oblivious, _polynomial-time computable two-party functionality. Then there exists an efficient,_ _statistically UC-secure non-interactive protocol which realizes f in the Fwrap[single][-]_ _hybrid model in which the sender only sends bit-OTM tokens to the receiver._ ## 4 Two-Party Computation with Stateless Tokens In this section, we again address the question of achieving interactive two-party computation protocols, but asking the following questions: (1) Can we rely on ----- _stateless tokens while only assuming that one-way functions exist? (2) Can the_ above be achieved without requiring that the complexity or number of the tokens grows with the complexity of the function being computed, as was the case in the previous section? We show how to positively answer both questions: We use stateless tokens, whose complexity is polynomial in the security parameter, to implement the OT functionality. Since (as discussed earlier) secure protocols for any two-party task exist given OT, this suffices to achieve the claimed result. Before turning to our protocols, we make a few observations about stateless tokens to set the stage. First, we observe that with stateless tokens, it is always possible to have protocols where tokens are exchanged only at the start of the _protocol. This is simply because each party can create a “universal” token that_ takes as input a pair (c, x), where c is a (symmetric authenticated/CCA-secure) encryption[10] of a machine M, and outputs M (x). Then, later in the protocol, instead of sending a new token T, a party only has to send the encryption of the code of the token, and the other party can make use of that encrypted code and the universal token to emulate having the token T . The proof of security and correctness of this construction is straightforward. **Dealing with dishonestly created stateful tokens. The above discussion,** however, assumes that dishonest players also only create stateless tokens. If that is not the case, then re-using a dishonestly created token may cause problems with security. If we allow dishonest players to create stateful tokens, then a simple solution is to repeat the above construction and send separate universal tokens for each future use of any token by the other player, where honest players are instructed to only use each token once. Since this forces all tokens to be used in a stateless manner, this simple fix is easily shown to be correct and secure; however, it may lead to a large number of tokens being exchanged. To deal with this, as was discussed in the previous section, we observe that by Beaver’s OT extension result [36] (which requires only one-way functions), it suffices to implement O(k) OTs, where k is the security parameter, in order to implement any polynomial number of OTs. Thus, it suffices to exchange only a polynomial number of tokens even in the setting where dishonest players may create stateful tokens. **Convention for intuitive protocol descriptions. In light of the previous** discussions, in our protocol descriptions, in order to be as intuitive as possible, we describe tokens as being created at various points during the protocol. However, as noted above, our protocols can be immediately transformed into ones where a bounded number of tokens (or in the model where statelessness is guaranteed, only one token each) are exchanged in an initial setup phase. **4.1** **Protocol Intuition** We now discuss the intuition behind our protocol for realizing OT using stateless tokens; due to the complexity of the protocol, we do not present the intuition 10 An “encrypt-then-MAC” scheme would suffice here. ----- for the entire protocol all at once, but rather build up intuition for the different components of the protocol and why they are needed, one component at a time. For this intuition, we will assume that the sender holds two random strings s0 and s1, and the receiver holds a choice bit b. Note that OT of random strings is equivalent to OT for chosen strings [41]. **The Basic Idea. Note that, since stateless tokens can be re-used by malicious** players, if we naively tried to create a token that output sb on input the receiver’s choice bit b, the receiver could re-use it to discover both s0 and s1. A simple idea to prevent this reuse would be the following protocol, which is our starting point: 1. Receiver sends a commitment c = com(b; r) to its choice bit b. 2. Sender sends a token, that on input (b, r), checks if this is a valid decommit ment of c, and if so, outputs sb. 3. Receiver feeds (b, r) to the token it received, and obtains w = sb **Handling a Malicious Receiver. Similar to the problem discussed in the** previous section, there is a problem that the receiver may choose not to use the token sent by the sender until the end of the protocol (or even later!). In our context, this can be dealt with easily. We can have the sender commit to a random string π at the start of the protocol, and require that the sender’s token must, in addition to outputting sb, also output a valid decommitment to π. We then add a last step where the receiver must report π to the sender. Only upon receipt of the correct π value does the sender consider the protocol complete. **Proving Knowledge. While this protocol seems intuitive, we note that it is** actually insecure for a fairly subtle reason. A dishonest sender could send a token that on input (b, r), simply outputs (b, r) (as a string). This means that at the end of the protocol, the dishonest sender can output a specific commitment c, such that the receiver’s output is a decommitment of c showing that it was a commitment to the receiver’s choice bit b. It is easy to see that this is impossible in the ideal world, where the sender can only call an ideal OT functionality. To address the issue above, we need a way to prevent the sender from creating a token that can adaptively decide what string it will output. Thinking about it in a different way, we want the sender to “prove knowledge” of two strings before he sends his token. We can accomplish this by adding the following preamble to the protocol above: 1. Receiver chooses a pseudo-random function (PRF) fγ : {0, 1}[5][k] _→{0, 1}[k],_ and then sends a token that on input x ∈{0, 1}[5][k], outputs fγ(x). 2. Sender picks two strings x0, x1 ∈{0, 1}[5][k] at random, and feeds them (one at-a-time) to the token it received, and obtains y0 and y1. The sender sends (y0, y1) to the receiver. 3. Sender and receiver execute the original protocol above with x0 and x1 in place of s0 and s1. The receiver checks to see if the string w that it obtains from the sender’s token satisfies fγ(w) = yb, and aborts if not. ----- The crucial feature of the protocol above is that a dishonest sender is effectively committed to two values x0 and x1 after the second step (and in fact the simulator can use the PRF token to extract these values), such that later on it must output xb on input b, or abort. Note that a dishonest receiver may learn k bits of useful information about _x0 and x1 each from its token, but this can be easily eliminated later using the_ Leftover Hash Lemma (or any strong extractor). **Preventing correlated aborts. A final significant subtle obstacle remains,** however. A dishonest sender can still send a token that causes an abort to be correlated with the receiver’s input, e.g. it could choose whether or not to abort based on the inputs chosen by the receiver (see full version for a discussion of why this is a problem). To prevent a dishonest sender from correlating the probability of abort with the receiver’s choice, the input b of the receiver is additively shared into bits _b1, . . ., bk such that b1 +_ _b2 +_ _· · ·_ + _bk = b. The sender, on the other hand, chooses_ strings z1, . . ., zk and r uniformly at random from {0, 1}[5][k]. Then the sender and receiver invoke k parallel copies of the above protocol (which we call the Quasi_OT protocol), where for the ith execution, the sender’s inputs are (zi, zi + r),_ and the receiver’s input is bi. Note that at the end of the protocol, the receiver either holds [�] _zi if b = 0, or r +_ [�] _zi if b = 1._ Intuitively speaking, this reduction (variants of which were previously used by, e.g. [34,35]) forces the dishonest sender to make one of two bad choices: If each token that it sends aborts too often, then with overwhelming probability at least one token will abort and therefore the entire protocol will abort. On the other hand, if few of the sender’s tokens abort, then the simulator will be able to perfectly simulate the probability of abort, since the bits bi are (k − 1)-wise independent (and therefore all but one of the Quasi-OT protocols can be perfectly simulated from the receiver’s perspective). We make the receiver commit to its bits bi using a statistically hiding commitment scheme (which can be constructed from one-way functions [42]) to make this probabilistic argument go through. This completes the intuition behind our protocol. The result of this section is summarized by the following theorem, whose proof appears in full version. **Theorem 3 (Interactive UC-secure computation using stateless to-** **kens). Let f be a (possibly reactive) polynomial-time computable functionality.** _Then, assuming one-way functions exist, there exists a computationally UC-_ _secure interactive protocol which realizes f in the Fwrap[stateless]-hybrid model. Fur-_ _thermore, the protocol only makes a black-box use of the one-way function._ _Oblivious Reactive Functionalities in the Non-Interactive Setting._ In the full version, we generalize our study of non-interactive secure computation to the case of reactive functionalities. Roughly speaking, reactive functionalities are the ones for which in the ideal world, the parties might invoke the ideal trusted party multiple times and this trusted party might possibly keep state between ----- different invocations. For the interactive setting (i.e. when the parties are allowed multiple rounds of interaction in the Fwrap-hybrid models) there are standard techniques using which, given protocol for non-reactive functionality, protocol for securely realizing reactive functionality can be constructed. However, these techniques fail in the non-interactive setting. In the full version, we study what class of reactive functionalities can be securely realized in the non-interactive setting for the case of stateless as well as stateful hardware token. _Acknowledgements. We thank J¨urg Wullschleger for pointing out the relevance_ of [27] and for other helpful comments. We thank Guy Rothblum for useful discussions. ## References 1. Goldreich, O., Ostrovsky, R.: Software protection and simulation on oblivious rams. J. ACM 43(3), 431–473 (1996) 2. Chaum, D., Pedersen, T.P.: Wallet databases with observers. In: Brickell, E.F. (ed.) CRYPTO 1992. LNCS, vol. 740, pp. 89–105. Springer, Heidelberg (1993) 3. Brands, S.: Untraceable off-line cash in wallets with observers (extended abstract). In: Stinson, D.R. (ed.) CRYPTO 1993. LNCS, vol. 773, pp. 302–318. Springer, Heidelberg (1994) 4. Cramer, R., Pedersen, T.P.: Improved privacy in wallets with observers (extended abstract). In: Helleseth, T. (ed.) EUROCRYPT 1993. LNCS, vol. 765, pp. 329–343. Springer, Heidelberg (1994) 5. Ishai, Y., Sahai, A., Wagner, D.: Private circuits: Securing hardware against prob ing attacks. In: Boneh, D. (ed.) CRYPTO 2003. LNCS, vol. 2729, pp. 463–481. Springer, Heidelberg (2003) 6. Gennaro, R., Lysyanskaya, A., Malkin, T., Micali, S., Rabin, T.: Algorithmic tamper-proof (ATP) security: Theoretical foundations for security against hardware tampering. In: Naor, M. (ed.) TCC 2004. LNCS, vol. 2951, pp. 258–277. Springer, Heidelberg (2004) 7. Hofheinz, D., M¨uller-quade, J., Unruh, D.: Universally composable zero-knowledge arguments and commitments from signature cards. In: Proc. of the 5th Central European Conference on Cryptology MoraviaCrypt 2005, Mathematical Publications (2005) 8. Moran, T., Naor, M.: Basing cryptographic protocols on tamper-evident seals. In: Caires, L., Italiano, G.F., Monteiro, L., Palamidessi, C., Yung, M. (eds.) ICALP 2005. LNCS, vol. 3580, pp. 285–297. Springer, Heidelberg (2005) 9. Katz, J.: Universally composable multi-party computation using tamper-proof hardware. In: Naor, M. (ed.) EUROCRYPT 2007. LNCS, vol. 4515, pp. 115–128. Springer, Heidelberg (2007) 10. Chandran, N., Goyal, V., Sahai, A.: New constructions for UC secure computation using tamper-proof hardware. In: Smart, N.P. (ed.) EUROCRYPT 2008. LNCS, vol. 4965, pp. 545–562. Springer, Heidelberg (2008) 11. Moran, T., Segev, G.: David and Goliath commitments: UC computation for asymmetric parties using tamper-proof hardware. In: Smart, N.P. (ed.) EUROCRYPT 2008. LNCS, vol. 4965, pp. 527–544. Springer, Heidelberg (2008) ----- 12. Damg˚ard, I., Nielsen, J.B., Wichs, D.: Isolated proofs of knowledge and isolated zero knowledge. In: Smart, N.P. (ed.) EUROCRYPT 2008. LNCS, vol. 4965, pp. 509–526. Springer, Heidelberg (2008) 13. Goldwasser, S., Kalai, Y.T., Rothblum, G.: One-time programs. In: Wagner, D. (ed.) CRYPTO 2008. LNCS, vol. 5157, pp. 39–56. Springer, Heidelberg (2008) 14. Canetti, R.: Universally composable security: A new paradigm for cryptographic protocols. In: FOCS, pp. 136–145 (2001) 15. Chor, B., Kushilevitz, E.: A zero-one law for boolean privacy. SIAM J. Discrete Math. 4(1), 36–47 (1991) 16. Rabin, M.O.: How to exchange secrets with oblivious transfer (1981) 17. Even, S., Goldreich, O., Lempel, A.: A randomized protocol for signing contracts. Commun. ACM 28(6), 637–647 (1985) 18. Kilian, J.: Founding cryptography on oblivious transfer. In: STOC, pp. 20–31 (1988) 19. Ishai, Y., Prabhakaran, M., Sahai, A.: Founding cryptography on oblivious transfer - efficiently. In: Wagner, D. (ed.) CRYPTO 2008. LNCS, vol. 5157, pp. 572–591. Springer, Heidelberg (2008) 20. Hazay, C., Lindell, Y.: Constructions of truly practical secure protocols using stan dardsmartcards. In: Ning, P., Syverson, P.F., Jha, S. (eds.) ACM Conference on Computer and Communications Security, pp. 491–500. ACM, New York (2008) 21. Goldreich, O., Micali, S., Wigderson, A.: How to play any mental game or a com pleteness theorem for protocols with honest majority. In: STOC, pp. 218–229 (1987) 22. Goyal, V., Sahai, A.: Resettably secure computation. In: Joux, A. (ed.) EUROCRYPT 2009. LNCS, vol. 5479, pp. 54–71. Springer, Heidelberg (2009) 23. Barak, B., Goldreich, O., Impagliazzo, R., Rudich, S., Sahai, A., Vadhan, S.P., Yang, K.: On the (im)possibility of obfuscating programs. In: Kilian, J. (ed.) CRYPTO 2001. LNCS, vol. 2139, pp. 1–18. Springer, Heidelberg (2001) 24. Ishai, Y., Prabhakaran, M., Sahai, A., Wagner, D.: Private circuits ii: Keeping secrets in tamperable circuits. In: Vaudenay, S. (ed.) EUROCRYPT 2006. LNCS, vol. 4004, pp. 308–327. Springer, Heidelberg (2006) 25. Hofheinz, D., M¨uller-Quade, J., Unruh, D.: Universally composable zero-knowledge arguments and commitments from signature cards. In: 5th Central European Con[ference on Cryptology (2005), http://homepages.cwi.nl/~hofheinz/card.pdf](http://homepages.cwi.nl/~hofheinz/card.pdf) 26. Kolesnikov, V.: Truly efficient string oblivious transfer using resettable tamper proof tokens. In: Micciancio, D. (ed.) TCC 2010. LNCS, vol. 5978. Springer, Heidelberg (2010) 27. Buhrman, H., Christandl, M., Unger, F., Wehner, S., Winter, A.: Implications of superstrong nonlocality for cryptography. Proceedings of the Royal Society A 462(2071), 1919–1932 28. Canetti, R.: Universally composable security: A new paradigm for cryptographic protocols. In: FOCS, pp. 136–145 (2001) 29. Ben-Or, M., Goldwasser, S., Wigderson, A.: Completeness theorems for non cryptographic fault-tolerant distributed computation (extended abstract). In: STOC, pp. 1–10 (1988) 30. Kushilevitz, E.: Privacy and communication complexity. SIAM J. Discrete Math. 5(2), 273–284 (1992) 31. Goldreich, O., Vainish, R.: How to solve any protocol problem - an efficiency im provement. In: Pomerance, C. (ed.) CRYPTO 1987. LNCS, vol. 293, pp. 73–86. Springer, Heidelberg (1988) ----- 32. Galil, Z., Haber, S., Yung, M.: Cryptographic computation: Secure faut-tolerant protocols and the public-key model. In: Pomerance, C. (ed.) CRYPTO 1987. LNCS, vol. 293, pp. 135–155. Springer, Heidelberg (1988) 33. Goldreich, O.: Foundations of Cryptography: Basic Applications. Cambridge Uni versity Press, Cambridge (2004) 34. Kilian, J.: Uses of Randomness in Algorithms and Protocols. MIT Press, Cambridge (1990) 35. Lindell, Y., Pinkas, B.: An efficient protocol for secure two-party computation in the presence of malicious adversaries. In: Naor, M. (ed.) EUROCRYPT 2007. LNCS, vol. 4515, pp. 52–78. Springer, Heidelberg (2007) 36. Beaver, D.: Correlated pseudorandomness and the complexity of private computa tions. In: STOC, pp. 479–488 (1996) 37. Yao, A.: How to generate and share secrets. In: FOCS, pp. 162–167 (1986) 38. Ishai, Y., Kushilevitz, E.: Perfect constant-round secure computation via perfect randomizing polynomials. In: Widmayer, P., Triguero, F., Morales, R., Hennessy, M., Eidenbenz, S., Conejo, R. (eds.) ICALP 2002. LNCS, vol. 2380, pp. 244–256. Springer, Heidelberg (2002) 39. Feige, U., Kilian, J., Naor, M.: A minimal model for secure computation (extended abstract). In: STOC, pp. 554–563 (1994) 40. Brassard, G., Cr´epeau, C., Santha, M.: Oblivious transfers and intersecting codes. IEEE Transactions on Information Theory 42(6), 1769–1780 (1996) 41. Beaver, D., Goldwasser, S.: Multiparty computation with faulty majority (extended announcement). In: FOCS, pp. 468–473. IEEE, Los Alamitos (1989) 42. Haitner, I., Reingold, O.: Statistically-hiding commitment from any one-way func tion. In: STOC, pp. 1–10 (2007) 43. Anderson, W.E.: On the secure obfuscation of deterministic nite automata. Cryp tology ePrint Archive, Report 2008/184 (2008) 44. Canetti, R., Goldreich, O., Goldwasser, S., Micali, S.: Resettable zero-knowledge (extended abstract). In: STOC, pp. 235–244 (2000) -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-642-11799-2_19?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-642-11799-2_19, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://link.springer.com/content/pdf/10.1007%2F978-3-642-11799-2_19.pdf" }
2,010
[ "JournalArticle" ]
true
2010-02-09T00:00:00
[]
14,854
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01044c3e265ad414aec8cf608c24e1d1cf406077
[ "Computer Science" ]
0.916699
Hardened Bloom Filters, with an Application to Unobservability
01044c3e265ad414aec8cf608c24e1d1cf406077
Ann. UMCS Informatica
[ { "authorId": "50198404", "name": "Nicolas Bernard" }, { "authorId": "1953911", "name": "Franck Leprévost" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
Classical Bloom filters may be used to elegantly check if an element e belongs to a set S, and, if not, to add e to S. They do not store any data and only provide boolean answers regarding the membership of a given element in the set, with some probability of false positive answers. Bloom filters are often used in caching system to check that some requested data actually exist before doing a costly lookup to retrieve them. However, security issues may arise for some other applications where an active attacker is able to inject data crafted to degrade the filters’ algorithmic properties, resulting for instance in a Denial of Service (DoS) situation. This leads us to the concept of hardened Bloom filters, combining classical Bloom filters with cryptographic hash functions and secret nonces. We show how this approach is successfully used in the TrueNyms unobservability system and protects it against replay attacks.
Annales UMCS Informatica AI XII, 4 (2012) 11–22 DOI: 10.2478/v10065-012-0018-y ### Hardened Bloom Filters, with an Application to Unobservability Nicolas Bernard[1][∗], Franck Leprévost[1][†] ``` 1LACS, University of Luxembourg ``` 162 a, Avenue de la Faïencerie, L-1511 Luxembourg Abstract – Classical Bloom filters may be used to elegantly check if an element e belongs to a set S, and, if not, to add e to S. They do not store any data and only provide boolean answers regarding the membership of a given element in the set, with some probability of false positive answers. Bloom filters are often used in caching system to check that some requested data actually exist before doing a costly lookup to retrieve them. However, security issues may arise for some other applications where an active attacker is able to inject data crafted to degrade the filters’ algorithmic properties, resulting for instance in a Denial of Service (DoS) situation. This leads us to the concept of hardened Bloom filters, combining classical Bloom filters with cryptographic hash functions and secret nonces. We show how this approach is successfully used in the TrueNyms unobservability system and protects it against replay attacks. ### 1 Introduction Many applications in computer science depend on the result of the following problem: check if an element e belongs to a set S, and, if it does not, add e to S. Depending on the application we have in mind, the "match" or "no match" answer will usually lead to additional processing, like for instance in the following two examples: (1) Filtering duplicated packets on a network connection: On a network connection, it can happen that a packet is duplicated. The destination host then receives it twice, so does the application. This is for instance the case on a UDP connection. ∗Nicolas.Bernard@uni.lu †Franck.Leprevost@uni.lu ----- (2) Counting the number of different elements in a collection: If they are not in this set, a counter is increased and the element is added to the set. Bloom filters [1] address these problems in an elegant manner. A Bloom filter is a probabilistic data structure that allows to represent a finite set S without storing the actual elements of the set S. Among their main properties, Bloom filters have small footprints, a fast lookup time, allow to add elements quickly to the represented set S, and the addition of an element cannot fail due to the data structure being “full”. Bloom filters do not store any data and can only provide boolean answers on the membership of a given element in the set, with some probability of false positive answers. They are often used in caching system to check that some requested data actually exist before doing a costly lookup to retrieve them. In the situation of the example (1) above, a Bloom filter at the receiving end could be used to drop the duplicated packets: packets that do not match are processed (i.e., used by the application) and added to the set, while packets that do match are considered as duplicated and discarded. In the situation of example (2), each element of the collection is matched against a Bloom filter representing an “already accounted” set. While the result is only of probabilistic nature, its complexity is O(m) whereas the complexity of a classical algorithm remains O(m log m), where m is the number of elements of the collection. This being said, security issues may be raised for many applications, leading e.g. to Denial of Service (DoS) attacks. The purpose of this article is to provide a solution to these issues by introducing hardened Bloom filters. Moreover, we show their use in the seminal example of the TrueNyms protocol [2], which raised our interest in Bloom filters and motivated the present contribution. This article is organized as follows: in section 2, we briefly explain the underlying concept of a classical Bloom filter. In section 3, we describe the security issues that an external malicious party may exploit, leading to the construction of hardened Bloom filters. In section 4, we briefly describe the TrueNyms unobservability system, and describe how to efficiently use hardened Bloom filters to prevent replay attacks on this system. We conclude this article with some further ideas for the enhancements of our approach, which we plan to develop in due time. ### 2 Classical Bloom filters A Bloom filter (in the classical understanding as defined in [1]) is a probabilistic data structure representing a finite set S. It consists of a bit array A of size 2[n] (in practice n is small, say n < 25), and k distinct hash functions (Hj)1≤j≤k such that Hj(data) = ij ∈ [0, 2[n] − 1]. (1) ----- In other words, ij is an index of A, depending on the data considered. Moreover, k is also small: its chosen value — in a first approach — depends on the allowed probabilistic "false-positive" occurrences according to formula 2 below. The discussion about the (lack of) requirements on hash functions in the context of classical Bloom filters is addressed in part 2.2. 2.1 Construction of S and A Initially, S = ∅ and all the bit values of A are equal to 0. An element e is added to S by setting to 1 all the positions of the array A indexed by the hash values i1 = H1(e), i2 = H2(e), . . ., ik = Hk(e): ∀j ∈ [1, k], A[Hj(e)] ← 1. The test to determine if an element e is already in S is performed by generating the indices for this element. An element e is then probably in S if, and only if: ∀j ∈ [1, k], A [Hj(e)] = 1. The probability in the previous sentence applies only to the "if" part. Indeed, there can be values i, j, e, e[′] s.t. Hi(e) = Hj(e[′]). In other words, an index in the array A could be “part of” multiple elements of S. As a consequence, there is no way to remove elements from S and, once set to 1, a value A[i] is never reset to 0. It implies in particular that, once added, an element belonging to S is always found if matched against the filter. Now, with some probability, the filter can represent an element e as belonging to S although it is not the case: it may indeed happen that all the indices corresponding to e are equal to 1, while e ̸∈ S. Such a “false-positive” occurs with a probability: km[�][k] �1 − �1 − [1] � ≈ �1 − e −2km[n][ �][k], (2) 2[n] where m is the number of elements in S. 2.2 Non-cryptographic hash functions A hash function as used in the context of classical Bloom filters a priori differs strongly from a hash function used in the context of cryptology. It is a function[1]: H : N −→ [0, 2[n] − 1] with good statistical distribution properties for given “normal data”, as described for instance in section 6.4 of [3]. In particular, these hash functions usually lack the 1We consider here any finite word on any finite alphabet as mappable to an element of N, and that distinct words lead to distinct elements of N. ----- compression property (see [4, section 9.2.1]) that is a mandatory and important part of a cryptographic hash function. Such a hash function can be very simple, and usually it is in order to be fast. For instance, it may consist in the modular division by some prime, chosen according to the needed size of the image. In fact, since the hash function does not need to consider all the data given but only a suitable part to obtain a correct distribution, we can even construct hash functions with complexity in O(1). Consequences are multiple, but here we will only note the three following : (1) Recall that we need k distinct hash functions for the classical Bloom filters. We can create many different functions with similar properties by changing a parameter in one fixed scheme. For instance, in a scheme based on modular division, the choice of k distinct appropriate primes leads to k distinct hash functions. (2) It is possible to find preimages : it means that given an i, it is possible to find Dx, Dy, · · · such that H(Dx) = H(Dy) = · · · = i. Indeed, many simple hash functions can be easily inverted. Anyway, given the usual size of the image set, it would be easy to find such values by brute-force. (3) It is usually even possible, given a few such hash functions H1, · · ·, Hj and corresponding indices i1, · · ·, ij, to find a common preimage D such that H1(D) = i1, · · ·, Hj(D) = ij. (3) ### 3 Security issues and Hardened Bloom filters As mentioned in the introduction (section 1), security issues may be raised in some applications. For instance, assume the elements to be matched can be tampered by an external malicious party, say Mallory. Recall then that the probability given in equation 2 applies to “ordinary” elements. Since the hash functions Hk are a priori non-cryptographic ones, Mallory can craft special elements that will fill A with bits set to 1 much faster than random data would[2]. Of course, once all the bits of the array A are equal to 1, each element tried against the filter will match, which results in a denial of service (DoS) attack in the cases given beforehand: all the elements are considered as already in the set, even when they are not. So, Bloom filters must be hardened to prevent such attacks if Mallory controls the incoming data. If an attacker can inject as many elements he wants to, the battle is lost because even if he is restricted to the probability given by equation 2, with m growing, the probability will converge to 1. However, such a case is rare, and most of the time the attacker will find himself unable to add more than a fixed number of elements per time unit. Here, it is possible to fight back, and design appropriate countermeasures. 2The irony being that, while collisions are usually a sign of weakness in cryptographic hash functions, here Mallory has to find non-colliding elements in order to set to 1 all the bits of the array A as fast as possible. ----- 3.1 Protection against index selection attacks To prevent Mallory from just deciding upon a set of indices and creating suitable data to send, the first idea is to use Bloom filters where the k hash functions have some cryptographic properties. Notably, it must be hard — no faster way than brute force — to find preimages, to insure that the attacker will not be able in practice to find a common preimage as defined in equation 3. With such hash functions, it would be far harder for Mallory to find non-colliding packets than simply deciding which bits in the array A he wants to set and generating the corresponding data. The natural choice for a hash function with such cryptographic properties, is to take a cryptographic hash function H [c] [4, page 323]. Note however that the properties of a cryptographic hash function are a superset of what is actually needed: we comment on these aspects in section 5. 3.2 k cryptographic hash functions ? The first difficulty is to find k such functions. As we have seen in section 2.2, it is easy to have many non-cryptographic hash functions. Unfortunately, even for a small relevant k, we cannot find k different standard cryptographic hash functions. The list of such hash functions mainly consist of md5, sha-1, the sha-2 and ripemd families [4, 5], and this list can hardly be extended much further. Nonetheless, there are multiple ways to solve this issue : (1) Conceptually, the easiest way is probably to add the index of the function before the data. In other words, given one cryptographic hash function H [C], and using the | symbol for concatenation, we define the k hash functions as Hi(data) := H [c] (i|data), 1 ≤ i ≤ k. Some variants of this method can be imagined. For instance, the index could be used in the initialization vector of the compression function of the hash function. However this proposal only makes the implementation harder as specifying this vector is usually not possible through the API of the cryptographic libraries providing such functions. (2) One can also think of using the iterated application of the cryptographic hash function H [c] to produce the (Hi)1≤i≤k. More precisely, the k hash functions are defined as Hi(data) := (H [c])[i] (data), 1 ≤ i ≤ k, with (H [c])[i] (data) = � H [c](data) if i = 1, � H [c][ �](H [c])[i][−][1] (data) if 2 ≤ i ≤ k. (3) Another way, is to notice that the fingerprint returned by a cryptographic hash function is a lot longer than an index for the bit array of the Bloom ----- filter. Indeed, the shortest fingerprints are at least 128 bits long, while it is unusual for an index to be more than 25 bits long, as noted in section 2. The idea then would be to see the fingerprint provided by a cryptographic hash function as the concatenation of l indices : H [c](data) = i1|i2|i3| · · · |il|r, where r is an unused “remainder” if the size of the fingerprint is not a multiple of the size of an index, and ij are the indices of equation 1. Of course, it may happen that l < k, then this scheme would need to be combined with one of the previous two to generate the k required indices. However, as there are standard hash functions with fingerprints size up to 512 bits at least, it should be possible to use it alone in most cases. (4) Another possibility that we will not detail here would be to construct custom hash functions using block ciphers [4, section 9.4.1]. Security-wise, there is no evidence that one of the previous schemes has some obvious advantage over the others. Let us then compare them on their speed. The algorithmic complexity of a cryptographic hash function is at least in O(s), where s is the size of the data to be hashed. To simplify, assume that the algorithm complexity of the cryptographic hash function is indeed s, the complexity of the different schemes would then be in : (1) ks for the first one, as the H [c] function is called k times on data of size s + ϵ (ϵ being the size of the index added before the actual data). (2) s + (k − 1)f for the second one, where f is the size of a fingerprint: H [c] is called once on data of size s, then k − 1 times on the fingerprint of size f generated at the previous step. The second scheme is hence faster than the first one if the data size is large. (3) The third one needs only one call to the cryptographic hash function if l ≥ k. If l < k the exact complexity depends on the combination with one of the other schemes, but will be reduced compared to it anyway. The third scheme then seems to be the best choice, since it is the fastest one. It must be noted however that a cryptographic hash function is anyway much slower than a non-cryptographic one. To take an example, the number of operations to hash data of size s can be as low as 1 for a non-cryptographic hash function as described in 2.2, while it would be of the order of 160s for a typical cryptographic hash function like Ripemd-160 [6]. 3.3 Protection against offline attacks Let us recall that the hash functions considered here give a value that is an index for the array A, i.e. a value belonging to [0, 2[n]], with n < 25, and hence preimages can be found by brute force. Moreover, because Bloom filters are deterministic (and the different schemes presented in 3.2 do not change this), the same input will fill two filters in the same way. Mallory can then perform the following offline DoS attack: ----- Brute force the hash functions to create a set of elements that would fill the Bloom filter faster than “normal” data would. Even if he is not anymore able to select indices and craft data to set them specifically, he can still generate a lot of data packets and send the group of them that sets the greatest number of bits in the array A. While such elements would have some collisions on indices, they would still fill the filter a lot faster than the statistical probability predicts. Let us summarize the situation: to insure the protection against index selection attacks (seen in part 3.1), we rely on Bloom filters using cryptographic hash functions. Now, to furthermore insure the protection against an offline attack as described above, we add the utilization of secret nonces. A nonce is a random value, which in our context is generated at the instantiation of a Bloom filter and is then used as a key so that the cryptographic hash functions are in fact replaced by MACs (or keyed hash functions, see [4, page 325]). Instead of giving all details, we provide here the conceptual idea, which amounts to specializing H [c] for each Bloom filter F in something like H [c,][F](data) := H [c](nF|data), (4) where nF is the nonce used for filter F. With such a scheme, Mallory is blinded: he is not able to know the effect of an element and hence cannot craft special elements anymore. As a consequence, an active DoS attack by Mallory against Bloom filters hardened this way does not work, provided that Mallory is only able to add a limited number of elements per second. The main drawback is that it is not possible anymore to take the union of two sets by using a bit-wise OR operation on the arrays of the corresponding bloom filters unless they are using the same nonce. For most applications, this however should not be a significant issue. We define here a hardened Bloom filters as a classical Bloom filter using cryptographically-enhanced hash functions together with a secret nonce, addressing index-selection attacks as well as offline attacks. ### 4 Hardened Bloom filters and TrueNyms We now describe how such hardened Bloom filters are used in the TrueNyms unobservability system [7, 8, 2] as a protection against some forms of active traffic analysis. Let us first recall what TrueNyms[3] is. 3We partially rely on [8] for the wording of some paragraphs of subsections 4.1 and 4.2, as well as for the figures 1 and 2. ----- 4.1 The TrueNyms unobservability system The TrueNyms system allows Alice and Bob to communicate over an IP network without any observer knowing it. More precisely, when parties are using TrueNyms for their communications, an observer, as powerful as he may be, is unable to know who they are communicating with. He is unable to know when a communication occurs. He is even unable to know if a communication occurs at all. This TrueNyms system is a peer-to-peer overlay network based on Onion-Routing [9, 10], to which it adds protection against all forms of traffic analysis, including replay attacks. Its performance is experimentally validated and is appropriate for most uses (e.g. Web browsing and other HTTP-based protocols like RSS, Instant Messaging, file transfers, audio and video streaming, remote shell, . . . ) but the usability of applications requiring a very low end-to-end latency (like for instance telephony over IP) may be degraded. Briefly, Onion-Routing transmits data through nested encrypted tunnels established through multiple relays R1, R2, etc. (see Figure 1 — in the following, a node denotes either a relay or Alice or Bob). These relays accept to take part in an anonymity system, but are not supposed trusted. Indeed, some of them can cooperate with a passive observer Eve or with an active observer Mallory. Relays see only enciphered traffic and know only the previous and next nodes on the route. They do not know if those nodes are other relays or end-points. Fig. 1. In Onion-Routing, to communicate with Bob, Alice creates a set of nested encrypted tunnels. For every packet, each relay removes the outermost encryption layer (hence the name of this scheme). To clarify some terminology used throughout this section, an encrypted tunnel between Alice and one of the nodes is called a connection. Then, a set of nested connections between Alice and Bob is called a route. Despite being created by Alice, those routes are ----- not related to IP source routing or other IP-level routing. Standard IP routing is still used between successive nodes if these nodes are on an IP network as we consider here. At last, in TrueNyms, a communication is a superset of one or more routes between Alice and Bob that are used to transmit data between them. A communication can use multiple routes simultaneously and / or sequentially. 4.2 Replay attacks An issue with standard cryptography modes when used in Onion-Routing is that they allow an active replay attack[4]. Let us examine the situation at a relay at a given time: for instance, let us assume that this specific relay is a part of three routes, as depicted in Figure 2. ## 1 7TEXIF WXOVGR OATGBX FWULFO TTPAXO CFBAQL AUTFYF NAELF2 ETEOPG QXBGFA DM3XRE TUZLFB ## 1 7TEXIF WXOVGR OATGBX FWULFO TTPAXO CFBAQL AUTFYF NAELF2 ETEOPG QXBGFA QXBGFA TUZLFB ABCDEF XNSXAX XNSXAX LAMPFB ORSUAT ZAFPFL ECZAFV ORWCMX CLOCRW VOYUAV 4NBXVE XLDTFH ABCDEF XNSXAX 3NTUBM LAMPFB ORSUAT ZAFPFL ECZAFV ORWCMX CLOCRW VOYUAV 4NBXVE XLDTFH ## A ## A Fig. 2. Cryptography hides connection bindings to a passive observer (left), but not to an active observer able to inject duplicate packets (right). On the left of Figure 2, the observer sees three distinct incoming connections (A, B, C), while there is also three outgoing connections (1, 2, 3). To make the relaying useless, the observer must discover the relationship between the incoming and the outgoing connections, or at least he must discover the outgoing connection corresponding to an incoming one he is interested in. As an encryption layer is removed on each connection, he cannot discover this by a casual glance at the content of the packets. Moreover, in TrueNyms, the packet size and rate are normalized, and care is taken to prevent information leaks when a route is established or closed (as described in [7, 8, 2]). Those standard traffic analysis methods are hence closed to an attacker. However, as cryptography is deterministic, if nothing is done, a given packet entered twice through a same incoming connection would be output twice — in its form with an encryption layer removed — on the corresponding outgoing connection. So Mallory takes a packet and duplicates it, say on connection A, which leads to the right side of Figure 2. He then looks for two identical packets on the output, and finds them ``` 4This is different of the replay attacks well known in cryptography, where an attacker can play ``` part of a protocol back from a recording, and that are usually prevented by the use of nonces or timestamps. ----- on the connection 3, so he learns that connection A and connection 3 are part of the same route. Obviously, depending on the interest of Mallory, he can perform a similar attack on the next relay having the connection 3 as an incoming connection, and then see where it leads ultimately. Or he can perform the same attack on the other incoming connections B and C, and figure out exactly which outgoing connection 1 or 2 corresponds to them. The obvious way to prevent an external attacker to inject packets would be to use node-to-node authentication on a route, but in this case it would not be sufficient since, even if we assume that the replay of an authenticated packet is not possible, the possibility for Mallory to operate a node must also be accounted for. This means there is no way to actually prevent packet injection by an active observer, and so the system has to be designed in a way that makes such injection useless. 4.3 Using hardened Bloom filters to prevent replay attacks Recall that packets between two successive nodes on a route can be replayed by Mallory, and hence will be output on the corresponding outgoing connection to the downstream relay. In the TrueNyms implementation, to prevent such replay attacks, a relay “remembers” all the packets of a transmission and compares each incoming packet on the same connection to them. If it does not match, the packet is forwarded; if it does match, it is dropped (and a dummy packet is forwarded). Of course this approach requires a very fast way to compare a new packet to the previous ones, hence the need for Bloom filters. The situation is then similar to the context described in the example (1) of section 1: an accepted packet is added to the filter if it “was not” already in it. In TrueNyms, as the traffic is shaped, Mallory cannot simply flood the filter as the addition to the filter is only done for transmitted packets, and packets outside the shaping envelope are simply dropped. In order to protect our unobservability system against the security issues raised in section 3, TrueNyms relies on hardened Bloom filters. Notice that, as false positives can occur, legitimate packets may be dropped. This may slightly alter the performance of the system, but is not otherwise an issue as TrueNyms provides end-to-end reliability if needed: the packet will then be resent with another aspect. To ensure this different aspect, unacknowledged packets are buffered unencrypted. If it is necessary to retransmit a packet, a nonce (unrelated to the nonces used in the hardened Bloom filters in part 3.3) it includes is changed before the packet is re-encrypted. As the cipher is used in bi-IGE mode (see below), the new encrypted packet will have no similarities with the old one. ----- Nonetheless, a long term connection would start to swamp the hardened Bloom Filter after some time, and packets would start to be lost more and more. In TrueNyms, this is not an issue due to two distinct features : (1) Even if the communication is long-term, this is not the case of the routes it uses. The lifetime of a route is chosen at random and is fixed before it is used ; (2) Routes are re-keyed from time to time. It means the encryption keys used for the connections are changed. As the same packet entering twice but going through the encryption layer with different keys would give different (and a priori unmatchable without knowing the keys) outputs, the hardened Bloom filters can be replaced by new ones during the key changes. Of course, it only prevents Mallory from replaying identical packets. If let unhindered, he will replay slightly different packets and his attack would be successful because after adding or removing an encryption layer with a standard block cipher mode, the original and replayed packets will have similarities. For the use of hardened Bloom filters to be effective, this attack must be prevented too, for instance by employing a special mode like bi-IGE (which is a bi-directional application of the Infinite Garble Extension mode — Campbell, 1977, [11]) as it is done in TrueNyms. ### 5 Conclusions and further work In this paper, after recalling the functioning and the main properties of classical Bloom filters, we considered the situation where a malicious party may develop indexselection attacks or offline attacks against some applications, leading e.g. to Denial of Service situations. We then designed hardened Bloom filters able to withstand such attacks, combining classical Bloom filters together with cryptographic hash functions and secret nonces. Although these hardened Bloom filters are slower than classical Bloom filters, mostly due to the use of cryptographic hash functions over non-cryptographic ones, we described how they are concretely successfully used in the TrueNyms unobservability system to defend it against active traffic analysis attacks. Should the need arise, performance can probably be improved by further work on the hash functions. Our proposed hardened Bloom filters relies notably on cryptographic hash functions. However, the requirements are probably weaker: for instance, while compression and preimage resistance appear to be needed, it is not obvious that second-preimage and collision-resistance are necessary as well. It may hence be possible to construct custom hash functions with only the mandatory properties, that would be faster than the usual cryptographic hash functions. We intend to study these possibilities in a future work. Finally, multiple variants of Bloom filters have been proposed (Bloomier filters, etc.) over the years, some faster, some using less space, some allowing to remove elements, ----- etc. In a future work, we also intend to study the possibility to similarly harden some of these numerous existing variants of Bloom filters. ### Acknowledgements The FNR/04/01/05/TeSeGrAd grant partially supported this research. ### References [1] Bloom B. H., Space/time trade-offs in hash coding with allowable errors, Communications of the ACM 13 (7) (1970): 422. [2] Bernard N., Leprévost F., Unobservability of low-latency communications: the TrueNyms protocol, work in progress. [3] Knuth D. E., Sorting and Searching,The Art of Computer Programming 3 (1998). [4] Menezes A. J., van Oorschot P. C., Vanstone S. A., Handbook of Applied Cryptography, Discrete Mathematics and its Applications, CRC Press (1997). [5] Anderson R., Security Engineering: A Guide to Building Dependable Distributed Systems, Wiley (2001). [6] Preneel B., Dobbertin H., Bosselaers A., The Cryptographic Hash Function RIPEMD-160, CryptoBytes 3 (2) (1997): 9. [7] Bernard N., Non-observabilité des communications à faible latence, Université du Luxembourg, Université de Grenoble 1 – Joseph Fourier (2008). [8] Bernard N., Leprévost F., Beyond TOR: The TrueNyms Protocol, Security and Intelligent Information Systems 7053 (2012): 68. [9] Goldschlag D. M., Reed M. G., Syverson P. F., Hiding Routing Information, Proceedings of Information Hiding: First International Workshop, Springer-Verlag, LNCS 1174 (1996): 137. [10] Reed M. G., Syverson P. F., Goldschlag D. M., Anonymous connections and Onion Routing, IEEE Journal on Selected Areas in Communications 16(4) (1998): 482. [11] Knudsen L., Block Chaining Modes of Operation, Department of Informatics, University of Bergen (2000); http://www.ii.uib.no/publikasjoner/texrap/ps/2000-207.ps -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.2478/v10065-012-0018-y?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.2478/v10065-012-0018-y, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GREEN", "url": "https://journals.umcs.pl/ai/article/download/3361/2555" }
2,012
[ "JournalArticle" ]
true
null
[ { "paperId": "45ad696d9922ae2aba4a7122091e494953e29607", "title": "Beyond TOR: The TrueNyms Protocol" }, { "paperId": "c60b9518c2559caf7a0f3b9b7f939b2126628ff2", "title": "Non-observabilité des communications à faible latence" }, { "paperId": "1161313527071f5825d65402474e2de0a8fb72e8", "title": "Cryptographic Hash Function" }, { "paperId": "a3336e6f4d5cfe621dd616d9d9bac443ba4f8b6b", "title": "The Art of Computer Programming: Volume 3: Sorting and Searching" }, { "paperId": "d3c44bf122dc45c95496ebedfb47247fe07cd4bb", "title": "Anonymous connections and onion routing" }, { "paperId": "961011a97535f89d386f81926335bdd8196ff300", "title": "Hiding Routing Information" }, { "paperId": "f39a2c11983b21fd5054d5393614959bfbc4e50f", "title": "Space/time trade-offs in hash coding with allowable errors" }, { "paperId": null, "title": "The TrueNyms Protocol, Security and Intelligent Information Systems" }, { "paperId": "c1515b351a216d4896cec060e48bcf57f8b62b8b", "title": "Block Chaining Modes of Operation" }, { "paperId": null, "title": "Handbook" }, { "paperId": "9e0c718def0567d068959cd55b67d15d990748a8", "title": "Sorting and Searching" }, { "paperId": "b4c9a5000cfd1a8717b3a20117eeba34d80df600", "title": "Sorting and searching\" the art of computer programming" }, { "paperId": null, "title": "It is possible to find preimages" }, { "paperId": "e86c10dc325d6777ecf7eb02c113157d9e324677", "title": "1 The Cryptographic Hash Function RIPEMD-160" }, { "paperId": null, "title": "Unobservability of low-latency communications: the TrueNyms protocol" } ]
7,442
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01060a62a51c08248abc0b204b4e255a09731ad8
[]
0.852913
A Framework for Data Privacy Preserving in Supply Chain Management Using Hybrid Meta-Heuristic Algorithm with Ethereum Blockchain Technology
01060a62a51c08248abc0b204b4e255a09731ad8
Electronics
[ { "authorId": "2212173325", "name": "Yedida Venkata Rama Subramanya Viswanadham" }, { "authorId": "3382395", "name": "Kayalvizhi Jayavel" } ]
{ "alternate_issns": [ "2079-9292", "0883-4989" ], "alternate_names": null, "alternate_urls": [ "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-247562", "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-247562", "https://www.mdpi.com/journal/electronics" ], "id": "ccd8e532-73c6-414f-bc91-271bbb2933e2", "issn": "1450-5843", "name": "Electronics", "type": "journal", "url": "http://www.electronics.etfbl.net/" }
Blockchain is a recently developed advanced technology. It has been assisted by a lot of interest in a decentralized and distributed public ledger system integrated as a peer-to-peer network. A tamper-proof digital framework is created for sharing and storing data, where the linked block structure is utilized to verify and store the data. A trusted consensus method has been adopted to synchronize the changes in the original data. However, it is challenging for Ethereum to maintain security at all blockchain levels. As such, “public–private key cryptography” can be utilized to provide privacy over Ethereum networks. Several privacy issues make it difficult to use blockchain approaches over various applications. Another issue is that the existing blockchain systems operate poorly over large-scale data. Owing to these issues, a novel blockchain framework in the Ethereum network with soft computing is proposed. The major intent of the proposed technology is to preserve the data for transmission purposes. This new model is enhanced with the help of a new hybrid algorithm: Adaptive Border Collie Rain Optimization Algorithm (ABC-ROA). This hybrid algorithm generates the optimal key for data restoration and sanitization. Optimal key generation is followed by deriving the multi objective constraints. Here, some of the noteworthy objectives, such as information preservation (IP) rate, degree of modification (DM), false rule (FR) generation, and hiding failure (HF) rate are considered. Finally, the proposed method is successfully implemented, and its results are validated through various measures. The recommended module ensures a higher security level for data sharing.
# electronics _Article_ ## A Framework for Data Privacy Preserving in Supply Chain Management Using Hybrid Meta-Heuristic Algorithm with Ethereum Blockchain Technology **Yedida Venkata Rama Subramanya Viswanadham * and Kayalvizhi Jayavel** Department of Computer Science and Engineering, SRM Institute of Science and Technology, Chennai 603203, Tamil Nadu, India *** Correspondence: yvrsvish@gmail.com** **Citation: Viswanadham, Y.V.R.S.;** Jayavel, K. A Framework for Data Privacy Preserving in Supply Chain Management Using Hybrid Meta-Heuristic Algorithm with Ethereum Blockchain Technology. _[Electronics 2023, 12, 1404. https://](https://doi.org/10.3390/electronics12061404)_ [doi.org/10.3390/electronics12061404](https://doi.org/10.3390/electronics12061404) Academic Editor: Akshya Swain Received: 12 January 2023 Revised: 7 March 2023 Accepted: 8 March 2023 Published: 15 March 2023 **Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **Abstract: Blockchain is a recently developed advanced technology. It has been assisted by a lot of** interest in a decentralized and distributed public ledger system integrated as a peer-to-peer network. A tamper-proof digital framework is created for sharing and storing data, where the linked block structure is utilized to verify and store the data. A trusted consensus method has been adopted to synchronize the changes in the original data. However, it is challenging for Ethereum to maintain security at all blockchain levels. As such, “public–private key cryptography” can be utilized to provide privacy over Ethereum networks. Several privacy issues make it difficult to use blockchain approaches over various applications. Another issue is that the existing blockchain systems operate poorly over large-scale data. Owing to these issues, a novel blockchain framework in the Ethereum network with soft computing is proposed. The major intent of the proposed technology is to preserve the data for transmission purposes. This new model is enhanced with the help of a new hybrid algorithm: Adaptive Border Collie Rain Optimization Algorithm (ABC-ROA). This hybrid algorithm generates the optimal key for data restoration and sanitization. Optimal key generation is followed by deriving the multi objective constraints. Here, some of the noteworthy objectives, such as information preservation (IP) rate, degree of modification (DM), false rule (FR) generation, and hiding failure (HF) rate are considered. Finally, the proposed method is successfully implemented, and its results are validated through various measures. The recommended module ensures a higher security level for data sharing. **Keywords: data privacy preservation system; Ethereum blockchain technology; adaptive border** collie rain optimization; supply chain network; data sanitization and restoration **1. Introduction** Developers can create many distributed apps based on smart contracts over the Ethereum programming platform. For example, voting, financial transactions, business administration, and contract signing are the applications used in the Ethereum platform [1]. During data sharing, the protection of privacy is very important. If shared data fall into the wrong hands, the data could be misused and accessed by intruders, the denial of loans and health insurance, and victimization of a person by financial fraud [2]. However, if the data can be shared with the right users, then there is no information to be stolen and misused by unauthorized entities [3]. The newly developed privacy-preserving data publishing (PPDP) and privacy-preserving data mining (PPDM) approaches are utilized to reduce privacy issues [4]. Data analysis contains commodity data, and mining turns it into economic value [5]. The digital world increases the possibility of losing control over all of one’s own intellectual, emotional, and situational knowledge, breaking the informational privacy area and losing one’s autonomy [6]. The primary problem in this situation is that control of privacy leakage by every individual requires high freedom in the flow of information the technology enables, the connections it facilitates, and the advantage supplied by the ----- _Electronics 2023, 12, 1404_ 2 of 29 information source [7]. Government organizations also create legislative guidelines to protect personal data, including what purposes the particular data is used for, how it is gathered, and how it should be preserved. Corporate privacy issues are also solved by these guidelines [8]. Ethereum is a programming environment that makes it possible to provide privacy preservation with the help of blockchain by building over distributed applications [9]. With “smartcontract”, the contract may be carried out without needing a centralized authority after it has been installed. The smart contract executes intent in a perfect world, and it produces reliable results [10]. Blockchain is a new technology that has gained popularity due to the rise of cryptocurrency such as Ethereum and Bitcoin [11]. It may be considered a distributed ledger with cryptography, and accurately records the results of every transaction. A transaction recorded on the blockchain cannot be changed beyond that point, and everyone can see it [12]. Every network member will reach a consensus on the blockchain, ensuring that all valid transactions are recorded [13]. Secondly, every network member will reach a consensus on the blockchain, ensuring that no invalid transactions are recorded. Thirdly, all transactions recorded on the blockchain are auditable by network members and cannot be tampered with by other issues. Smart contracts are being deployed alongside the advancement of blockchain technology such as Ethereum and Hyperledger to increase the potentiality of blockchain [14]. Addressing a transaction for a certain smart contract’s address will cause it to be activated [15]. When a smart contract is activated, it can run the predefined program independently through a centralized authority. The challenges of the existing supply chain are shown below. Rival actors can be found in the same supply chain, which makes data sharing very difficult [16]. Due to these challenges, research on and implementation of traceability are growing slowly. Although there has been work to create supply chain traceability, it is not practical in the actual world. The work demonstrates several procedures in a centralized system’s data structures and transparent framework. Researchers’ interest in blockchain technology and its application to the supply chain has grown over the last few years [17]. Some limitations of the existing research and their techniques are depicted here. The supply chains’ global character makes it challenging to achieve the desired traceability. Several methods are implemented to provide better privacy-preserved data transfer in the supply chain network. However, these methods fail to address issues with certificate verifiability and raise concerns about the existence of privacy-sensitive information. A hash network links the blockchain in chronological order. For each node, it has a copy of a ledger, and few mutually distrusting nodes often maintain this hash network. This system provides better confidentiality and non-repudiation over the Ethereum networks, but the cost requirement is high and the system’s feasibility low. To overcome these challenges, a novel blockchain-based privacy preservation model is implemented in the Ethereum network to provide better privacy preservation. The contributions of the designed Ethereum-based privacy preservation model are listed below. To design a blockchain-based data privacy preservation model with a hybrid meta _•_ heuristic algorithm over the supply chain network to secure information exchange and guarantee the privacy of data access in the Ethereum platform. Here, the performance improvement of the proposed model is applicable to different applications regarding cryptocurrency, food supply chains, and sealed-bid auctions. _•_ To generate the key with the help of developed ABC-ROA for exchanging the secured data in the supply chain framework using data restoration and data sanitization procedures in the Ethereum environment. The developed ABC-ROA algorithm is utilized to restore the data from the receiver side. Consequently, it helps to access the original data that can be generated from the original key. To implement the hybrid meta-heuristic algorithm known as ABC-ROA for choosing _•_ the best optimal key to maximize the performance of the developed blockchain-based ----- _Electronics 2023, 12, 1404_ 3 of 29 privacy preservation model. Here, the designed ABC-ROA algorithm improves the system’s robustness. It is also used to solve complex issues. _•_ To compare the developed ABC-ROA-based privacy preservation system with existing meta-heuristic algorithms using a variety of metrics to verify the performance of the developed model. This paper is split into the following sections. The merits and demerits of recent privacy preservation in the Ethereum network model based on blockchain are described in Section 2. The suggested dataset for the blockchain-based privacy preservation system in Ethereum and the model explanation offered are covered in Section 3. The procedures used to create the privacy preservation model, such as data restoration and sanitization in the supply chain method, are shown in Section 4. The Ethereum privacy preservation model with the ABC-ROA algorithm, optimal key details, and objective function details are described in Section 5. The acquired outcomes of the recommended method are summarized in Section 6. The paper is concluded in Section 7. **2. Literature Survey** _2.1. Related Works_ In 2021, Lin et al. [18] introduced privacy-preserving blockchain architecture (PPChain), and PPChain’s design has been changed in Ethereum. PPChain’s architecture allowed regulation to provide security. They specifically incorporated cryptographic primitives such as broadcast encryption and group signature into a workable byzantine fault tolerance consensus protocol with a separation mechanism to remove the transaction fee and mine for a reward instead of using the existing mining model. They offered in-depth security and privacy analysis and a performance study to demonstrate the usefulness of PPChain. Examples from the food supply chain, sealed-bid auctions, and cryptocurrency described how the PPChain might be used in regulation applications. In 2022, Rahmadika et al. [19] implemented an efficient architecture for secure misbehavior detection in lightweight IoMT tools in the “artificial pancreas” model (APS). The suggested method used deep learning, which protected privacy, and boosted security by integrating blockchain technology built on the Ethereum smart contract ecosystem. The efficacy of the developed system has been empirically tested for commensurate incentive schemes, exhaustiveness with compact findings, and an untraceable characteristic from a different neural deep learning technique. Consequently, the model has a high recall rate, demonstrating that it is almost completely capable of identifying harmful events in the case being studied. In 2022, Xiong et al. [20] designed a secure privacy preservation authentication mechanism for inter-constellation collaboration. They created both permanent and transient identities for each satellite to protect privacy. The permanent identity was used for interconstellation collaboration, whereas the temporary identity was utilized for communication inside the constellation. For information exchange among cooperative satellite constellations, a consortium blockchain was introduced. A replica storage node mechanism has been suggested to enable effective authentication with minimal resources, where well-resourced satellites cache the duplicated data exchanged across the blockchain. A branch-and-bound approach has been used to address the integer programming issue of choosing the replica storage node. According to a security study, the suggested authentication technique was secure against various possible attacks, which included formal analysis using informal verification and BAN Logic. The suggested system provided efficiency in communication overheads, signaling, and processing with greater performance, based on a comparison between it and other privacy preservation schemes. Evaluations showed that the suggested onboard caching approach achieved minimal storage costs and communication delay. In 2022, Singh et al. [21] implemented a privacy preservation model in smart healthcare using a blockchain-based federated learning method for preserving privacy, which used IoT cloud platforms to provide privacy and security. Scalable machine learning applications such as health-care use federated learning technologies. Users could also utilize a well-trained deep learning system without putting their private information in the ----- _Electronics 2023, 12, 1404_ 4 of 29 cloud. Additionally, it covered the uses of federated learning in a smart city’s distributed secure environment. In 2020, Guo et al. [22] implemented a blockchain-based privacy preservation system in the Ethereum network to provide better privacy preservation data. In that model, they used various mechanisms in the blockchain. Privacy protection and analyzing anonymity methods were used in digital currency. The aim of the encryption mechanism was focused on the privacy protection scheme. In 2022, Mohan et al. [23] designed the proof of authority (PoA) consensus process, which required little computational power, and it has been implemented on a Raspberry Pi network. The elliptic integrated encryption process used a double-encryption process to protect the secrecy of data. With a speed of at least 25 transactions per second, it has been reported to perform well in contrast to previous systems. It was also readily expandable to accommodate various health-care workers. A new range of real-time health monitoring tools with excellent security and data privacy might come from further work on this concept, potentially leading to significant innovation in the IoMT sector. In 2018, Elisa et al. [24] developed a model for a decentralized deep learning e-government system utilizing a blockchain framework that guaranteed data confidentiality and privacy and boosted public sector confidence. A prototype of the suggested system was also provided, supported by a theoretical and in-depth examination of the system’s security and privacy consequences. In 2022, Dewangan et al. [25] suggested a system that used tokens to create pupils’ identities and saved them in a file system. The suggested model used SHA-256 for cryptographic hashing, Edwards-curve digital signature algorithm (EdDSA), and IPFS for digital signature and verification. The results of this suggested method show the transaction speed, the time needed for validating and signing a transaction, and the time needed for each transaction. They evaluated the privacy, transaction costs, huge file storage, blockchain registration, and implementation costs of this system to those of the already built solutions. _2.2. Statement of Problem_ Personal information is stolen by many hackers and intruders, so privacy preservation is much needed nowadays. Personal data information theft, virus threats, and spamming are illegal activities. However, limited hardware resources in IoT applications, including network bandwidth, computing power, and storage pose unique challenges to the blockchain. Therefore, various researchers have developed efficient technique to secure the data along with the blockchain technology. Some of the disadvantages and advantages of the existing blockchain-based privacy preservation techniques are listed in Table 1. PPChain [18] provides qualitative security and efficient privacy over personal data. Additionally, it ignores the correlation between variables to enhance training performance. However, it is not sufficiently developed to achieve conflicting properties, such as regulation, transparency, anonymity, and confidentiality, and is computationally expensive. Bi-LSTM [19] significantly increases the storage cost. Additionally, it achieves higher prediction accuracy, precision, and f1 score, yet it requires more time for training and is slow compared to other convolutional techniques and does not contain expensive hardware to perform complex mathematical calculations. Signs [20] indicate a high percentage of false positives. Additionally, the board caching method achieves low storage costs and communication latency. However, it does not access the service anywhere and does not provide any security system for communication and privacy protection, so it suffers slightly from the security perspective. Federated learning [21] achieves efficiency in computation, communication overheads, and low signaling with more functionality. Additionally, it achieves high authentication with replica storage and limited resources, yet the efficiency is low when compressing the massive number of devices in the security system and it is very expensive. Data encryption [22] efficiently achieves strict rights management, mainly used for several functions. Additionally, it increases the integrity of the data, and it is very cheap to implement. However, it provides less computation for compatibility and the scalability is very low. For the Raspberry Pi network [23], the implementation cost is low. Additionally, the implementation cost is low, yet to handle a large amount of data, ----- _Electronics 2023, 12, 1404_ 5 of 29 more storage is needed in this network, and it does not provide a proper balance between connectivity and storage requirements. Peer-to-peer technology [24] improves computation rationality and identity anonymity. Additionally, it is easy to set up the client data and does not need any special knowledge. However, the data confidentiality is very poor, and the file resources do not centrally organize, so it takes more time. EdDSA [25] improves the immutability and transparency to enhance the privacy system. Additionally, it reduces the computational complexity of the decentralization algorithms, yet while using the enormous data, the private key is sometimes leaked and does not support merging complex data. Therefore, these challenges motivated us to develop an efficient privacy-preserving system with blockchain technology. **Table 1. Features and disadvantages of the existing blockchain-based privacy preservation techniques** with blockchain technology in Ethereum. **Study** **Techniques** **Features** **Disadvantages** Lin et al. [18] PPChain Rahmadika et al. [19] BiLSTM Xiong et al. [20] SGINs Federated Singh et al. [21] Learning Guo et al. [22] Data encryption Raspberry Pi Mohan et al. [23] network Elisa et al. [24] peer-to-peer Dewangan et al. [25] EdDSA _•_ It provides qualitative security and efficient privacy over personal data. _•_ It ignores the correlation between variables to enhance the training performance. _•_ It significantly increases the storage cost. _•_ It achieves higher prediction accuracy, precision and f1 score. _•_ It gives a high percentage of false positive rates. _•_ The board caching scheme achieves low storage costs and low communication latency. _•_ It achieves more functionality attributes regarding the efficiency in signaling, communication overheads, and computation. _•_ It achieves efficient authentication with replica storage and limited resources. _•_ It efficiently achieves strict rights management, mainly used for several applications. _•_ It increases the integrity of the data, and it is very cheap to implement. _•_ The implementation cost is low. _•_ The model solution is more feasible and reliable. _•_ It improves computation rationality and identity anonymity. _•_ It is easy to set up the client data and does not need special knowledge. _•_ It improves the immutability and transparency to enhance the privacy system highly. _•_ It reduces the computational complexity of the decentralization algorithms. _•_ The data confidentiality is very poor. _•_ The file resources do not centrally organize, so it takes more time. _•_ While using enormous data, the private key is sometimes leaked. _•_ It does not support merging complex data. _•_ It is not developed for attaining conflicting parameters such as regulation and confidentiality. Anonymity and transparency. _•_ It is computationally expensive. _•_ It requires more training time and is a slower process than other convolutional techniques. _•_ It does not contain expensive hardware to do complex mathematical calculations. _•_ It does not access the service anywhere. _•_ It does not provide any security system for communication and privacy protection, so it suffers slightly from the security issue. _•_ The efficiency is low when compressing the massive number of devices in the security system. _•_ The federated communication system is very expensive. _•_ It provides less computation for compatibility. _•_ The scalability is very low. _•_ To handle a large amount of data, more storage is needed in this network. _•_ It does not provide a proper balance between connectivity and storage requirements. ----- _Electronics 2023, 12, 1404_ 6 of 29 **3. Privacy Preservation of Supply Chain Management Data: New Meta-Heuristic with** **Ethereum Blockchain** _3.1. Data Used for Privacy Preservation_ Client data are protected to establish a privacy preservation system. SCM gathers input data from a dataset called DataCo Smart Supply Chain for Big Data Analysis. It is available [at https://www.kaggle.com/shivkp/customer-behaviour (accessed on 10 January 2023).](https://www.kaggle.com/shivkp/customer-behaviour) The firm DataCo Global uses these data related to supply networks for their analysis. This dataset comprises registered operations that allow using R software areas and machine learning techniques, such as commercial distribution, sales, production, and supply. It also incorporates the relationship between organized and unstructured data. The gathered data are separated into three subsets: dataset 1, which comprises manufacturer data, dataset 2 counts transmitted data to managers who are present in different nations, and dataset 3 contains data transferred to each firm in each country. _3.2. SCM Privacy Preservation Framework_ SCM is one of the best-known commercial organizations due to its capacity for improving the efficiency of the firm. Supply chains and Ethereum blockchains are coupled to improve the security of supply chain networks. Many methods for ensuring the data privacy preservation and security of the Ethereum network with blockchains were introduced, including a VMI mode system, homomorphic encryption, fully observable supply chain management, PBFT algorithm, and consensus-based collaborative management mechanism. These methods help with cost-effective reconciliation, minimizing dispute settlement, increasing security, reducing dwell time, enhancing transparency, decentralizing data distribution, lowering complexity, increasing throughput, and addressing issues with _Electronics 2023, 12, x FOR PEER REVIEW information sharing and data tracking. The architectural representation of the Ethereum7 of 30_ blockchain technology developed for privacy preservation is given in Figure 1. **Figure 1.** Architectural representation of the privacy preservation platform developed in the **Figure 1.** Architectural representation of the privacy preservation platform developed in the Ethereum blockchain. Ethereum blockchain. Owing to these issues, higher credibility and dependability of data sharing and keeping security in the Ethereum blockchain technology, a data privacy preservation d l i bl k h i h b t bli h d I iti ll th d i d d t th d f ----- _Electronics 2023, 12, 1404_ 7 of 29 Owing to these issues, higher credibility and dependability of data sharing and keeping security in the Ethereum blockchain technology, a data privacy preservation model using blockchain has been established. Initially, the desired data are gathered from publicly available online databases. Data restoration and data sanitization are the two stages of secured data transfer in the developed model. The initial manufacturer data are cleaned during the sanitization procedure to guarantee the security of the data during transmission. Here, the originally collected data undergo sanitization. Then, the optimal key is generated with the ABC-ROA algorithm, which follows multi objective functions, such as HF, IP, FR, and DM. Then, the sanitized data progress to the restoration process, where they are stored. The restored data are transferred to the supply chain framework, where the blockchain is adapted to preserve the data during transmission. These overall processes are done in an Ethereum environment using blockchain technology. This technique sends the sanitized data over a single blockchain to prevent unauthorized access and delay in data transmission. The performance of this blockchain-based privacy preservation model is verified through various heuristic algorithms regarding key sensitivity, cost function, Euclidean distance, and harmonic mean. **4. Supply Chain Network Creation and Privacy Preservation Steps Handled** _4.1. Supply Chain Networks_ The supply chain network is divided into key levels: level 1 indicates the raw materials, level 2 denotes the supplier, level 3 indicates the manufacturer, level 4 denotes the manager, level 5 indicates the delivery, and level 6 denotes the customer. The network is then incorporated into the blockchain to raise the level of information-sharing security. Here, the manufacturer’s data undergo data sanitization to conceal them by employing the “chosen optimal key” created by the ABC-ROA algorithm. Then, in the restoration phase, these cleaned data are restored using the same optimum key through the authenticated users on the receiver end. These actions protect the confidentiality of the data shared in supply chain networks. The proposed data privacy preservation model divides the supply chain network into four main levels: level 1 designates the product’s manufacturer, level 2 the management, level 3 the product delivery, and level 4 the vendor. The phenomena that are a part of the supply chain network are further characterized. Manufacturers in various industries create their databases with information on the price, weight, product manager, delivery method, and suppliers of their manufactured goods. The manufacturer’s data are concealed or cleaned up using an ideal key. Then, the managers upload their data into the blockchain, dividing it into various subchains to increase security. Additionally, the data are sent appropriately in their supply chain subchains when transferred from the first management level to the last delivery level. The data restoration then happens at the vendor level. By restoring the actual data, the vendor utilizes the best key to access the private information. Five manufacturers with their required production goods are part of the supply chain network. Pretend these companies produce leather, cosmetics, electronics, paper, and wooden goods. They create a database based on the product information, which contains information such as item price, description, weight (kg), amount, brand, controlled by vendor manager, and shipment mode (delivery). The price, item quality, brand, and item description are understood to be the sensitive fields in this situation, and they must be sanitized using the created EF-HHO algorithm. The manufacturer’s subchains are shown in Equation (1). _JM1[(][1][)][,][ JM]1[(][2][)][,][ JM]1[(][3][)][,][ JM]1[(][4][)][,][ JM]1[(][5][)]_ (1) The terms JM1[(][1][)][,][ JM]1[(][2][)][,][ JM]1[(][3][)][,][ JM]1[(][4][)] and JM1[(][5][)] are the sub-chains and are calculated using Equation (2). _JM1[(][1][(][n][))][;][n][=][1,2], JM1[(][2][(][n][))][;][n][=][1,2], JM1[(][3][(][n][))][;][n][=][1,2],_ (2) _JM1[(][4][(][n][))][;][n][=][1,2,3], JM1[(][5][(][n][))][;][n][=][1,2,3]_ ----- _Electronics 2023, 12, 1404_ 8 of 29 The single blockchain JM1 is created by combining these subchains. These data are then transferred to the manager level. The managers and their corresponding subchains are shown in Equations (3) and (4). _FR1, FR2, FR3, FR4, FR5_ (3) _JM2[(][1][)][,][ JM]2[(][2][)][,][ JM]2[(][3][)][,][ JM]2[(][4][)][,][ JM]2[(][5][)]_ (4) The subchains of the manager are given in Equation (5). _JM2[(][1][(][n][))][;][n][=][1,2], JM2[(][2][(][n][))][;][n][=][1,2], JM2[(][3][(][n][))][;][n][=][1,2],_ (5) _JM2[(][4][(][n][))][;][n][=][1,2,3], JM2[(][5][(][n][))][;][n][=][1,2,3]_ The vendors and their corresponding subchains are measured by Equation (6) and Equation (7), respectively. The single blockchain is indicated by JM2. _WT1, WT2, WT3, WT4, WT5_ (6) _JM2[(][1][)][,][ JM]2[(][2][)][,][ JM]2[(][3][)][,][ JM]2[(][4][)][,][ JM]2[(][5][)]_ (7) Then, the subchains of the manager are denoted in below Equation (8). _JM3[(][1][(][n][))][;][n][=][1,2], JM3[(][2][(][n][))][;][n][=][1,2], JM3[(][3][(][n][))][;][n][=][1,2],_ (8) _JM3[(][4][(][n][))][;][n][=][1,2,3], JM3[(][5][(][n][))][;][n][=][1,2,3]_ _Electronics 2023, 12, x FOR PEER REVIEW_ Finally, the delivery level is represented in Equation (9). 9 of 30 _DE1, DE2, DE3, DE4, DE5_ (9) _DE1,_ _DE2_, _DE3,_ _DE4_, _DE5_ (9) Here, the term DE1 is a single delivery level. The term WT1 is vendor value in the supply chain.Here, the term _DE1 is a single delivery level. The term_ _WT1 is vendor value in the_ supply chain. The representation of the supply chain network with blockchain is given in Figure 2. The representation of the supply chain network with blockchain is given in Figure 2. **Figure 2. Supply chain network with blockchain.** **Figure 2. Supply chain network with blockchain.** **_4.2. Data Sanitization and Data Restoration_** The sanitization and restoration techniques used in this blockchain-based privacy ----- _Electronics 2023, 12, 1404_ 9 of 29 _4.2. Data Sanitization and Data Restoration_ The sanitization and restoration techniques used in this blockchain-based privacy preservation system are described below. Sanitization [26]: The data sanitization phase over the recommended data privacy preservation system in the Ethereum network is described in this section. The collected data from the real databases undergo data sanitization. In most cases, the sanitization process occurs at the manufacturer level, and blockchain sanitization also happens. The sensitive data in the blockchains’ subblocks are cleaned after being separated into subblocks. It is not necessary to sanitize the non-sensitive data in the subblocks. The term JM1[∗] [is a] blockchain-sanitized database calculated by Equation (10). _JM1[∗]_ [= (][C][2] _[⊕]_ _[JM][1][) +][ 1]_ (10) The term JM1 is actual data, and the binarized sanitized data is JM1[∗][. The term][ U]Y1[(][1][)]×X is a sensitive field, and the corresponding subchain is represented by JM1[(][4][(][n][))][;][n][=][1,2,3,4]. The data columns are to be sanitized in the blockchain JM1[(][1][)][. The sensitive field is given in the] blockchain matrix JM1[(][4][(][n][))][;][n][=][1]. The blockchain matrix is given in Equation (11). 1 1 4 3 2 2 3 5 5 4 4 1 2 3 _M1_ _M3_ _F1_ _F3_ _V1_ _V3_ _KF1_ _KF3_ _Q1_ _Q3_ _D1_ _D3_ _U1_ _U3_     _JM1[(][4][(][n][))][;][n][=][1]_ =   =   (11) The term UY1(1)×X [is a sensitive matrix. It is from the blockchain matrix][ JM]1[(][4][(][n][))][;][n][=][1] and is given in Equation (12). _KF1_ _KF8_ _Q1_ _Q8_ _D1_ _D8_ _U1_ _U8_   3 5 5 4 4 1 2 3   =   (12) _U_ (1) _Y1_ _×X_ [=]   Here, the result of the sanitized matrix in the blockchain matrix based on the rule is given. The sensitive data is present in the binarized data C2. The subblock is indicated by _JM1[(][4][(][n][))][;][n][=][1], and the identity matrix is denoted by YPS. In addition, the term JM1[(][4][(][n][))][;][n][=][1]_ and YPS are added to the sanitized data JM1[(][∗][4][(][n][))][;][n][=][1]. Restoration: The restoration process is very efficient for a privacy-preserving system. When employing the ABC-ROA algorithm on the receiver side, the receiver can access the original data using the generated optimal key. The first step is to binarize the blockchain. The key generation methods, the data C2, and the sanitized blockchain JM1[∗] [are binarized.] The sanitized data’s binarized form is further altered through the unit phase. To extract the restored data JM3, the binarized key matrix JM1[∗] [and binarized][ YPS][ are subtracted. The] data restoration process is measured using Equation (13). � � _JM3[(][n][(][n][))]_ = _JM([∗]n[(][1])[(][n][))]_ _−_ 1 _⊕_ _A2_ (13) Then, the newly designed ABC-ROA algorithm is used to recreate the sanitized key _A2. The restored data are denoted by JM3,and hence the lossless function can be performed._ The data sanitization and restoration process in the Ethereum blockchain network for the data privacy preservation system is depicted in Figure 3. ----- _ElectronicsElectronics 20232023, 12, 12, 1404, x FOR PEER REVIEW_ 11 of 30 10 of 29 Proposed ABC-ROA Subtract a unit step Subtract a unit step Sanitization key Sanitized data Actual database **Figure 3. Data sanitization and data restoration process in the blockchain-based privacy preserva-** **Figure 3. Data sanitization and data restoration process in the blockchain-based privacy preserva-** tion system. tion system. **5. Adaptive Border Collie Rain Optimization Algorithm with Ethereum Blockchain for5. Adaptive Border Collie Rain Optimization Algorithm with Ethereum Blockchain** **SCM Data Privacy Preservationfor SCM Data Privacy Preservation** _5.1. Optimal Key Generation_ _5.1. Optimal Key Generation_ In the above section, the selected sensitive fields are chosen based on the sanitization In the above section, the selected sensitive fields are chosen based on the sanitization process. The ABC-ROA optimization algorithm chooses the optimal key in sensitive fields. process. The ABC-ROA optimization algorithm chooses the optimal key in sensitive fields. The termis adapted to transfer the solution. The optimal key is recreated with the matrix dimensionThe term AAn is an optimal key value. In the key generation phase, the Khatri–Rao productn is an optimal key value. In the key generation phase, the Khatri–Rao product is adapted to transfer the solution. The optimal key is recreated with the matrix dimension and calculated using Equation (14). and calculated using Equation (14). 5 1 1 1 5 1 1 1 7 2  4 3  7 2 4 3 7 3  2 2  7 3 2 2 _JM1[(][∗][4][(][n][))]JM[;][n][=]1(∗[1]4( )=n_ );n=168= 126, A11, =A1 5= 5353 545 (14)(14) 1 48 2 45 14     2 6 2 3 1 4 4 1 [[√]M[n]JM[×][JM][max][]]     Here, the term JM1[∗] [is sanitized data. The rule-hiding technique is used in the sanitized]2 6 2 3  _M_ _JMn_ ×JM max  database. The ABC-ROA algorithm is adapted to optimize the key values betweenAn, respectively. The length of the key value is the same as that of the number of sensitiveHere, the term _JM is sanitized data. The rule-hiding technique is used in the sani-1*_ _A1 and_ fields. The key length is determined by usingtized database. The ABC-ROA algorithm is adapted to optimize the key values between�M[n]JM[. Finally, the optimal key is formed by] using the ABC-ROA algorithm.[A]1[ and] _An, respectively. The length of the key value is the same as that of the number of_ sensitive fields. The key length is determined by using _M_ _JMn_ . Finally, the optimal key is formed by using the ABC-ROA algorithm. Proposed ABC-ROA Original database Sanitized data ----- _Electronics 2023, 12, 1404_ 11 of 29 _5.2. Objective Function_ The ABC-ROA-based privacy preservation system chooses the optimal key in the restoration and sanitization process. These methods are used to solve constraints such _H1, H2, H3&H4. The objective function of the model is calculated by Equation (15)._ _GG = argmin(H1 + H2 + H3 + H4)_ (15) _{An}_ The selected optimal key is denoted by An. The term IG is normalized data and measured using Equation (16). _h1_ _H1 =_ (16) max(h) _iters_ _∀_ = _[no][.][o f sensitiveJM][∗]_ (17) _no.o f sensitiveJM_ The number of sensitive rules is given in Equation (18). The ratio of sensitive rules to the number of sensitive rules is presented below in Equation (19). _h1 =_ ���J1��� _∩_ _ST_ (18) _H1 =_ ��J1�� _∩_ _ST_ (19) _ST_ The information preservation ratio is calculated using Equation (20). _h2_ _H2 =_ (20) max(h2)∀iters = _[no][.][o f non][ −]_ _[sensitive wronghiddnJM][∗]_ (21) _no.o f nonsensitiveJM_ The information loss is measured using Equation (22). _h2 = 1 −_ ��J − _J1��_ (22) _J_ _|_ _|_ Here, the term H3 is a false rule generation. The false rules generation is calculated by Equation (23). _h3_ _H3 =_ (23) max(h3)∀iters = _[no][.][o f dataouto f bounceJM][∗]_ (24) _total no.o f recordJM_ _h3 =_ ��J − _J1��_ (25) _J_ _|_ _|_ The modified degree is measured in Equation (26). _h4_ _H4 =_ (26) max(h4)∀iters _h4 = dist(JM, JM[∗])_ (27) The optimal key is generated in the blockchain-based privacy preservation system. The optimal key selection improves the performance of the model. ----- _Electronics 2023, 12, 1404_ 12 of 29 The optimal key is generated in the blockchain-based privacy preservation system. The optimal key selection improves the performance of the model. _5.3. Ethereum Blockchain Technology5.3. Ethereum Blockchain Technology_ A blockchain is a decentralized ledger of data kept by network nodes, and a singleA blockchain is a decentralized ledger of data kept by network nodes, and a single entity does not control it. The blockchain data blocks are connected using cryptographicentity does not control it. The blockchain data blocks are connected using cryptographic concepts. Everyone on a blockchain is responsible for their actions since the transactionconcepts. Everyone on a blockchain is responsible for their actions since the transaction data are immutable and open to the public. A blockchain-based application is transparentdata are immutable and open to the public. A blockchain-based application is transparent and attack-resistant. Ethereum is an open-source blockchain framework for decentralizedand attack-resistant. Ethereum is an open-source blockchain framework for decentralized applications that manage digital wealth. “Smart contracts” refers to the applications thatapplications that manage digital wealth. “Smart contracts” refers to the applications that operate on the Ethereum virtual machine (EVM). Two widely used scripting languages foroperate on the Ethereum virtual machine (EVM). Two widely used scripting languages creating smart contracts on Ethereum are Solidity and Vyper. The two types of accountsfor creating smart contracts on Ethereum are Solidity and Vyper. The two types of in Ethereum are contract accounts and externally owned accounts (EOA). Every account accounts in Ethereum are contract accounts and externally owned accounts (EOA). Every type contains a different 20-byte hexadecimal string-based unique address. The data are account type contains a different 20-byte hexadecimal string-based unique address. The transmitted with the help of the owner’s private key and thus, it controls the EOA, which data are transmitted with the help of the owner’s private key and thus, it controls the has an ether balance (i.e., sending data to prompt the initiation of a smart contract). There is EOA, which has an ether balance (i.e., sending data to prompt the initiation of a smart no code associated with an EOA. On the other hand, the corresponding code for a contract contract). There is no code associated with an EOA. On the other hand, the corresponding account with an ether balance is triggered by a transaction or another smart contract. A few code for a contract account with an ether balance is triggered by a transaction or another benefits of this model’s use of Ethereum blockchain technology include restricted access smart contract. A few benefits of this model’s use of Ethereum blockchain technology to the consumer’s or generator’s private data, practical calculation techniques that may include restricted access to the consumer’s or generator’s private data, practical be implemented on the smart contract, and complete decentralization, achieving trans calculation techniques that may be implemented on the smart contract, and complete parent on-chain market clearing. The architecture representation of Ethereum blockchain decentralization, achieving transparent on-chain market clearing. The architecture technology is given in Figure 4. representation of Ethereum blockchain technology is given in Figure 4. **Figure 4. Privacy preservation framework in Ethereum blockchain technology.** **Figure 4. Privacy preservation framework in Ethereum blockchain technology.** _5.4. Proposed ABC-ROA_ The ABC-ROA algorithm is designed to enhance the effectiveness of the developed data privacy preservation system to select the optimal key in the data sanitization and data restoration process in the SCM. The existing algorithms face a few challenges in privacy preservation in Ethereum blockchain technology, where the existing heuristic algorithms have a limited number of resources to store the data. Hence, they face security and scalability challenges. The existing ROA and BCO algorithms are utilized for this work. Here, the BCO algorithm is selected in the developed method to increase the performance and robustness of the model. Additionally, it reduces the errors to increase the effectiveness concerning precision, f1score, and accuracy. The BCO is solving the multi objective combinational optimization problems. However, it gains low scalability and ----- _Electronics 2023, 12, 1404_ 13 of 29 utility. Due to the issues in the BCO algorithm, the ABC-ROA algorithm is developed by combining the ROA. ROA is chosen in the implemented model since it can perform in parallel computing and is mostly helped for high privacy protection. It is used to improve the user experience. These advantages in both BCO and ROA make the efficient performance in the suggested blockchain-based privacy preservation system by overcoming these conventional issues. The ABC-ROA algorithm is used to improve the performance of the developed blockchain-based data privacy preservation system. The ABC-ROA algorithm is implemented based on the current fitness and average fitness function. The term defines the current fitness Fitcr, and the average fitness is denoted by the term Fitavg. If Fitcr < Fitavg, choose the random parameter s = 2 otherwise, select s = 3 as the random parameter value. However, in the conventional algorithm, the random parameter s is selected randomly in the interval between [0, 1]. If s == 2 then, select the ROA algorithm otherwise, select the BCO algorithm. The developed ABC-ROA algorithm increases the fitness function. BCO [27]: Three dogs and sheep are considered in the Border collie optimization process. In a real-world scenario, one dog can manage the herd by himself. However, three dogs are considered because the search space for certain optimization issues might be large. When the algorithm is started, a group of three dogs and some lambs are exhibited here. The dogs are in charge of returning the sheep to the farm after they have gone out in different directions for grazing. Random variables are used to initialize the positions of sheep and dogs. According to their positions, the dogs are designated lead, left, and right. From the front, the lead dog directs the herd. The person with the fitness Fitg is chosen to be the dog in front of the herd or the lead dog. The major task of these dogs is to observe and stalk the herd. The terms Fitsj and _Fitm f denotes the fitness values. The fitness of the sheep is known as Fitt. The velocity of_ the lead dog is calculated using Equation (28). � _Wg(u + 1) =_ _Wg(u)[2]_ + 2 × NCg(u) × Pg(u) (28) Then, the velocity of the left dog is measured using the given Equation (29). � _Wsj(u + 1) =_ _Wsj(u)[2]_ + 2 × NCsj(u) × Psj(u) (29) The term NC is the acceleration of the dog and the term P is the dog’s position. The term _W indicates the velocity of the dog. The right dog velocity is calculated using Equation (30)._ � _Wm f (u + 1) =_ _Wm f (u)[2]_ + 2 × NCm f (u) × Pm f (u) (30) Here, the variables Wsj(u + 1), Wm f (u + 1) and Wg(u + 1) is denoted by the velocity of time at (u + 1) for the right, left, and lead dogs. Moreover, the terms NCg(u),NCsj(u), and NCm f (u) denotes the acceleration for lead, right, and left dogs. Pg(u),Psj(u), and _Pm f (u) describes the position of lead, right, and left dogs._ Gathering: In the sheep gathering method, the updated sheep velocity is considered. The approaches of stalking, gathering and eyeing are used in this algorithm. The sheep near the lead dog follow the lead dog’s direction. The sheep that is closer to the lead dog is determined using Equation (31). Here, the term Eh determines the positive shows the sheep nearer to the lead dog. _Eh =_ �Fitg − _Fitt�_ _−_ �� _Fitm f −2_ _Fitsj_ � _−_ _Fitt_ � (31) ----- _Electronics 2023, 12, 1404_ 14 of 29 Equation (32) indicates the sheep’s velocity. The term Pg is the current sheep location. � _Wth(u + 1) =_ _Wg(u + 1)[2]_ + 2 × NCg(u) × Pg(u) (32) The variable Wth is defined as the velocity of the sheep that is influenced by the lead dog. Stalking: To keep the dogs to guide, they must stalk the sheep closer to the left and right dogs. The stalked sheep velocity is updated using Equations (33) and (34), respectively. _Wsj =_ ��Wsj(u + 1) tan(s1)�2 + 2 × NCsj(u) × Psj(u) (33) _Wm f =_ 2 �� � _Wm f (u + 1) tan(s2)_ + 2 × NCm f (u) × Pm f (u) (34) _Wm f + Wsj_ _Wtt(u + 1) =_ (35) 2 Consequently, the term Wtt represents the velocities of left and right dogs. Hence, the traversing angles of s1 and s2 is considered randomly. Eyeing: In this scenario, it is anticipated that the least fit dog will follow the sheep and give them a close look. The velocity of the left dog is given in Equation (36). The variable _Wm f and NCm f is described the velocity and acceleration of the left dog. Additionally,_ the term Wsj and NCsj is defined as the velocity and acceleration of the right dog. Hence, the term Psj defines the collection of sheep that are presented in the current location. Additionally, the average time of individual can be represented as e. � _Wt f (u + 1) =_ _Wm f (u + 1)[2]_ _−_ 2 × NCm f (u) × Pm f (u) (36) The velocity of the right dog is given in Equation (37). � _Wt f (u + 1) =_ _Wsj(u + 1)[2]_ _−_ 2 × NCsj(u) × Psj(u) (37) The updated acceleration of the sheep and dog is calculated by Equation (38). � _NCj(u + 1) =_ � _Wj(u + 1) −_ _Wj(u)_ _Timej(u)_ (38) The updated time of the sheep and dog is measured by Equation (39). _e_ _Timej(u + 1) = Avg∑_ _j=1_ _Wj(u + 1) −_ _Wj(u)_ (39) _NCj(u + 1)_ The lead dog’s position is updated using Equation (40). _Pg(u + 1) = Wg(u + 1) × Timeg(u + 1)_ (40) + [1]2 _[NC][g][(][u][ +][ 1][)][ ×][ Time][g][(][u][ +][ 1][)][2]_ The left dog’s position is updated and calculated by Equation (41). _Pm f (u + 1) = Wm f (u + 1) × Timem f (u + 1)_ (41) + [1]2 _[NC][m f][ (][u][ +][ 1][)][ ×][ Time][m f][ (][u][ +][ 1][)][2]_ The position of the right dog is updated using Equation (42). _Psj(u + 1) = Wsj(u + 1) × Timesj(u + 1)_ (42) + [1]2 _[NC][sj][(][u][ +][ 1][)][ ×][ Time][sj][(][u][ +][ 1][)][2]_ ----- _Electronics 2023, 12, 1404_ 15 of 29 The updated locations of the sheep are determined using Equation (43) and Equation (44), respectively. _Pth(u + 1) = Wth(u + 1) × Timeth(u + 1)_ (43) + [1]2 _[NC][th][(][u][ +][ 1][)][ ×][ Time][th][(][u][ +][ 1][)][2]_ _Ptt(u + 1) = Wtt(u + 1) × Timett(u + 1)_ (44) _−_ [1]2 _[NC][tt][(][u][ +][ 1][)][ ×][ Time][tt][(][u][ +][ 1][)][2]_ The eyed sheep are updated, and it is determined using below Equation (45). _Pt f (u + 1) = Wt f (u + 1) × Timet f (u + 1)_ (45) _−_ [1]2 _[NC][t f][ (][u][ +][ 1][)][ ×][ Time][t f][ (][u][ +][ 1][)][2]_ Then, the sheep go to the track with the help of dog guidance, which is given in Equation (46). _Pg(u + 1) = Wg(u + 1) × Timeg(u + 1)_ + [1]2 (Wg(uTime+1)j−(uW)g(u)) _NCt f (u + 1) × Timet f (u + 1)[2]_ (46) The stalking, gathering and eyeing behavior over the sheep by a dog is described. By substituting the value of NC(u + 1) in Equations (46) and (47), the population values are attained based on the gathered sheep, left dog, stalked sheep, eyed sheep, and right dog. ROA [28]: Raindrops fall on the ground randomly. A raindrop can serve as a metaphor for each possible solution. As raindrops fall randomly on the ground, certain places in the solution space can be chosen randomly. Each raindrop’s radius is the most distinguishing characteristic. As time passes and a raindrop is joined to other droplets, its radius can decrease time. The radius of each droplet can decrease the time and also increase the connectivity of other droplets within a suitable range after the first population of replies is generated. Every droplet checks its nearest neighborhood based on its size at each cycle. Check for the end of the area that a single droplet has covered if it is still unconnected to any other droplet. Every droplet has variables while we are addressing a problem in n dimensions. Here, the term S is a large drop in radius. Then, the radius S1 and S2 makes a large form of the raindrop. The term m defines the variables in each droplet and is calculated using Equation (47). 1 _S = (S1[m]_ [+][ S]2[m][)] _m_ (47) Therefore, by increasing the number of iterations, weak droplets disappear, or the droplets create strong droplets. The initial population will decrease continuously, causing a speed of determining the correct answer. The term γ represents the soil characteristic given in Equation (48). 1 _S = (γS1[m][)]_ _m_ (48) Here, the variable s1 is the radius that does not move on the properties of the soil, which is depicted as γ. As a result, the droplet’s radius will be used to establish the lower and upper bounds of the variable in the initial stage. Two endpoints of the variable are examined in the next stage, and so on until the last variable. The term PDp is an ordering cost and is measured using Equation (49). _M_ _PDp =_ ∑ _j=1_ _ESj(U)_ (49) _Rj_ ----- _Electronics 2023, 12, 1404_ 16 of 29 The initial droplet cost would be adjusted at this point. The inventory holding cost is indicated by IDps, and is given in Equation (50). _M_ _IDps =_ ∑ _j=1_ _Ij_ 2Rj �Rj − _Tj�2_ (50) The shortage of the cost is denoted by TDp and is shown in Equation (51). _M_ _TDp =_ ∑ _j=1_ _Mj_ 2Rj _Tj[2]_ (51) The term UsDp is a transportation cost measured using Equation (52). _M_ _UsDp =_ ∑ _ESjTIDpj_ (52) _j=1_ The objective of the solution is calculated by Equation (53). _UpDp = (EDp + CDp + PDp + IDps + TDp + UsDp)_ (53) Here, the term UpDp denotes the total cost. The total investment is indicated by InV, and it is shown in Equation (54). _M_ _Jkj_ _InV =_ ∑ ∑ _GMJjk_ _YMJjk_ (54) _k=1_ _Jkj=1_ Here, the term YMJjk is an inventory capacity level. The fixed cost is represented by GMJjk. The term EU is total time calculated by Equation (55). _m_ ### ∑ _k=1_ _o_ ### ∑ (UMO + UYMO) ESj·ZYMO (55) _j=1_ _EU =_ _y_ ### ∑ _y=1_ The number of raindrops is denoted by y and the number of warehouses is represented by m. For each droplet, this situation would be repeated. Nearby droplets in their route may interact with one another, greatly speeding up the process. A droplet’s radius continuously decreased at the lowest point, greatly improving the accuracy of the re ----- _Electronics 2023, 12, 1404_ 17 of 29 sponse. The pseudocode of the implemented ABC-ROA is presented in Algorithm 1. **Algorithm 1: Developed ABC-ROA** Initialize the population and acceleration value Find the fitness solution Calculate the velocity Using Equation (29). For j = 1 to Maxiter For k = 1 to PoP If (CurrentFit < avgFit) Assign the value of s = 2 Else Assign the value of s = 3 End if If (s = = 2) Select the radius of the raindrop using Equation (50). **Update the solution with the ROA algorithm using Equation (51).** Else **Update the solution with the BCO algorithm using Equation (38).** Determine the best fitness of the sheep Update the velocity of the sheep in the BCO algorithm Evaluate the sheep’s position Update the position of the sheep using Equation (32). End if End End **Obtain the best position** End **6. Results and Discussion** _6.1. Simulation Setting_ The developed ABC-ROA-based privacy preservation model over the Ethereum network by blockchain technology was implemented in the MATLAB environment. In this developed system, the chromosome length was set at five, and the population size was set at 10. The efficiency analysis was conducted over key sensitivity, cost function, Euclidean distance, known-plaintext attack (KPA), harmonic mean, known ciphertext attack (KCA), arithmetic mean, CCA, and CPA. The efficiency of the developed model has been compared through various heuristic algorithms such as Harris hawks optimizer (HHO) [28], entity framework Harris hawks optimizer (EF-HHO) [29], BCO [26], and ROA [27]. _6.2. Effectiveness Analysis Using Euclidean Distance_ The overall analysis of the recommended ABC-ROA-based privacy preservation model over the Ethereum network with three datasets in terms of Euclidean distance is given in Figure 5. From the analysis, dataset 2 gives a very low Euclidean distance than dataset 1 and 2. While using dataset 2, the developed ABC-ROA-based privacy preservation model gives improved Euclidean distance of 36.03%, 8.62%, 2.54%, and 8.77% over HHO, EFHHO, BCO, and ROA. In the graph analysis, the developed ABC-ROA method is utilized to show effective performance. Here, the existing EF-HHO algorithm attains second-best performance. While considering all three datasets, the Euclidean distance of the proposed method shows better performance in dataset 2. Thus, the developed ABC-ROA-based data privacy preservation model over the Ethereum network gives higher effectiveness than the other heuristic algorithms. ----- ##### ROA-based data privacy preservation model over the Ethereum network gives higher ef-ROA-based data privacy preservation model over the Ethereum network gives higher ef _Electronics 2023, 12, 1404_ 18 of 29 ##### fectiveness than the other heuristic algorithms. fectiveness than the other heuristic algorithms. **Figure 5. Effectiveness analysis on the offered blockchain-based privacy preservation model using** Euclidean distance. **Figure 5. Figure 5. Effectiveness analysis on the offered blockchain-based privacy preservation model usingEffectiveness analysis on the offered blockchain-based privacy preservation model using** Euclidean distance. Euclidean distance.6.3. Performance Analysis Using the Harmonic Mean _6.3. Performance Analysis Using the Harmonic MeanThe harmonic mean analysis in terms of Euclidean distance, Pearson correlation, and_ ##### 6.3. Performance Analysis Using the Harmonic Mean the spearman correlation on the developed privacy preservation system over the The harmonic mean analysis in terms of Euclidean distance, Pearson correlation, and Ethereum network among various datasets is given in Figure 6. In dataset 2, the developed The harmonic mean analysis in terms of Euclidean distance, Pearson correlation, and the spearman correlation on the developed privacy preservation system over the Ethereum ##### the spearman correlation on the developed privacy preservation system over thenetwork among various datasets is given in FigureABC-ROA-based privacy preservation system over the Ethereum network provides en- 6. In dataset 2, the developed ABC- Ethereum network among various datasets is given in Figure 6. In dataset 2, the developedROA-based privacy preservation system over the Ethereum network provides enhancedhanced harmonic means of25.71%, 27.77%, 35%, and 27.1% than HHO, EF-HHO, BCO, and ROA. From the given graph analysis, the Pearson correlation can be analyzed to meas ##### ABC-ROA-based privacy preservation system over the Ethereum network provides en-harmonic means of25.71%, 27.77%, 35%, and 27.1% than HHO, EF-HHO, BCO, and ROA. ure the strength and direction between the two variables. Here, the Pearson correlation ##### hanced harmonic means of25.71%, 27.77%, 35%, and 27.1% than HHO, EF-HHO, BCO,From the given graph analysis, the Pearson correlation can be analyzed to measure the lies in the range of [−1 to 1]. Moreover, the negative correlation is denoted as “−1” as well ##### and ROA. From the given graph analysis, the Pearson correlation can be analyzed to meas-strength and direction between the two variables. Here, the Pearson correlation lies in the as the positive correlation can be represented as “1”. However, Spearman’s correlation range of [ 1 to 1]. Moreover, the negative correlation is denoted as “ 1” as well as the ##### ure the strength and direction between the two variables. Here, the Pearson correlation− − positive correlation can be represented as “1”. However, Spearman’s correlation can be usedcan be used to measure the association between the variables. As a result, the analysis of ##### lies in the range of [−1 to 1]. Moreover, the negative correlation is denoted as “−1” as well to measure the association between the variables. As a result, the analysis of the designedthe designed ABC-ROA-based privacy preservation system is superior to the other heu ##### as the positive correlation can be represented as “1”. However, Spearman’s correlationABC-ROA-based privacy preservation system is superior to the other heuristic approaches.ristic approaches. can be used to measure the association between the variables. As a result, the analysis of the designed ABC-ROA-based privacy preservation system is superior to the other heu- ristic approaches. **Figure 5. Effectiveness analysis on the offered blockchain-based privacy preservation model using** ##### the designed ABC-ROA-based privacy preservation system is superior to the other heu (a) (b) **Figure 6. Cont.** ##### (a) (b) ----- _Electronics 2023, 12, x FOR PEER REVIEW_ 20 of 30 _Electronics 2023, 12, 1404_ 19 of 29 (c) (c) **Figure 6. Effectiveness analysis of the developed blockchain-based privacy preservation system to** **Figure 6.Figure 6. (a) Euclidean distance, ( Effectiveness analysis of the developed blockchain-based privacy preservation system toEffectiveness analysis of the developed blockchain-based privacy preservation system to b) Pearson correlation and (c) Spearman correlation.** (a) Euclidean distance, (b) Pearson correlation and (c) Spearman correlation. (a) Euclidean distance, (b) Pearson correlation and (c) Spearman correlation. _6.4. Effectiveness Analysis Using the Arithmetic Mean_ _6.4. Effectiveness Analysis Using the Arithmetic Mean6.4. Effectiveness Analysis Using the Arithmetic Mean_ Comparison of the Pearson and Spearman correlations of the designed ABC-ROA Comparison of the Pearson and Spearman correlations of the designed ABC-ROA-Comparison of the Pearson and Spearman correlations of the designed ABC-ROA based privacy preservation system among various heuristic algorithms is shown in Figure based privacy preservation system among various heuristic algorithms is shown in Figurebased privacy preservation system among various heuristic algorithms is shown in Figure 7. 7. The sum of the numerical values of each observation divided by the total number of The sum of the numerical values of each observation divided by the total number of7. The sum of the numerical values of each observation divided by the total number of observations is known as the arithmetic mean. From the analysis, the developed ABC observations is known as the arithmetic mean. From the analysis, the developed ABC-ROA-observations is known as the arithmetic mean. From the analysis, the developed ABC ROA-based privacy preservation achieves secured data transfer when dataset 3 shows a based privacy preservation achieves secured data transfer when dataset 3 shows a lowROA-based privacy preservation achieves secured data transfer when dataset 3 shows a low value. Regarding the spearman correlation value in dataset 2, the developed ABC value. Regarding the spearman correlation value in dataset 2, the developed ABC-ROA-low value. Regarding the spearman correlation value in dataset 2, the developed ABC ROA-based privacy preservation model has high arithmetic means of 40%, 20%, 20%, and based privacy preservation model has high arithmetic means of 40%, 20%, 20%, and 34.4%,ROA-based privacy preservation model has high arithmetic means of 40%, 20%, 20%, and 34.4%, better than HHO, EF-HHO, BCO, and ROA. As a result, the designed ABC-ROA better than HHO, EF-HHO, BCO, and ROA. As a result, the designed ABC-ROA-based34.4%, better than HHO, EF-HHO, BCO, and ROA. As a result, the designed ABC-ROA based privacy preservation system over the Ethereum network shows higher effectiveness privacy preservation system over the Ethereum network shows higher effectiveness thanbased privacy preservation system over the Ethereum network shows higher effectiveness than other heuristic algorithms. other heuristic algorithms. than other heuristic algorithms. (a) (b) (a) (b) **Figure 7.** Effectiveness analysis of the designed blockchain-based privacy preservation model in **Figure 7.Figure 7. Effectiveness analysis of the designed blockchain-based privacy preservation model inEffectiveness analysis of the designed blockchain-based privacy preservation model in** terms of (a) Pearson correlation and (b) Spearman correlation. terms of (terms of (a) Pearson correlation and (a) Pearson correlation and (b) Spearman correlation.b) Spearman correlation. _6.5. Cost Function Analysis on the Proposed Model6.5. Cost Function Analysis on the Proposed Model_ _6.5. Cost Function Analysis on the Proposed Model_ Comparison of the proposed ABC-ROA-based privacy preservation system conver-Comparison of the proposed ABC-ROA-based privacy preservation system conver Comparison of the proposed ABC-ROA-based privacy preservation system conver gence rate to existing meta-heuristic algorithms with various datasets is shown in Figuregence rate to existing meta-heuristic algorithms with various datasets is shown in Figure 8. gence rate to existing meta-heuristic algorithms with various datasets is shown in Figure Compared to HHO, EF-HHO, BCO, and ROA, the performance of the ABC-ROA-based8. Compared to HHO, EF-HHO, BCO, and ROA, the performance of the ABC-ROA-based 8. Compared to HHO, EF-HHO, BCO, and ROA, the performance of the ABC-ROA-based privacy preservation system, the cost function is highly improved by 2.98%, 3.07%, 3.56%,privacy preservation system, the cost function is highly improved by 2.98%, 3.07%, 3.56%, privacy preservation system, the cost function is highly improved by 2.98%, 3.07%, 3.56%, and 4.12%, respectively, in dataset 3 at iteration value of 15. If the iteration increases, thenand 4.12%, respectively, in dataset 3 at iteration value of 15. If the iteration increases, then and 4.12%, respectively, in dataset 3 at iteration value of 15. If the iteration increases, then the cost function of the designed ABC-ROA method gets decreased. Hence, the graph ----- _Electronics 2023, 12, 1404_ 20 of 29 the cost function of the designed ABC-ROA method gets decreased. Hence, the graph analysis shows better performance in the recommended method. The existing EF-HHO analysis shows better performance in the recommended method. The existing EF-HHO algorithm achieves second-best performance. As a result, the developed ABC-ROA-based algorithm achieves second-best performance. As a result, the developed ABC-ROA-based privacy preservation model using blockchain technology performs more effectively than privacy preservation model using blockchain technology performs more effectively than other algorithms. other algorithms. (a) (b) (c) **Figure 8.** Convergence analysis on developed blockchain-based privacy preservation model in **Figure 8. Convergence analysis on developed blockchain-based privacy preservation model in terms** terms (a) dataset 1, (b) dataset 2, and (c) dataset 3. (a) dataset 1, (b) dataset 2, and (c) dataset 3. _6.6. Effectiveness Analysis Using Key Sensitivity6.6. Effectiveness Analysis Using Key Sensitivity_ The key sensitivity of the obtained optimum key in the ABC-ROA-based privacyThe key sensitivity of the obtained optimum key in the ABC-ROA-based privacy preservation system for three datasets with various existing meta-heuristic algorithms atpreservation system for three datasets with various existing meta-heuristic algorithms at various percentage levels is shown in Figurevarious percentage levels is shown in Figure 9. The key sensitivity of the proposed system 9. The key sensitivity of the proposed system is noticed as a lower value while increasing the percentage of the key for all the threeis noticed as a lower value while increasing the percentage of the key for all the three datasets. From the analysis, the developed model shows key sensitivity improvements ofdatasets. From the analysis, the developed model shows key sensitivity improvements of 11%, 15%, 20%, and 19% to heuristic algorithms such as HHO, EF-HHO, BCO, and ROA,11%, 15%, 20%, and 19% to heuristic algorithms such as HHO, EF-HHO, BCO, and ROA, respectively. In the given graph analysis, it shows the equivalence performance. Basedrespectively. In the given graph analysis, it shows the equivalence performance. Based on on the key sensitivity value, the ROA algorithm is not effective to secure the data in thethe key sensitivity value, the ROA algorithm is not effective to secure the data in the blockblockchain. At learning percentage 30, the key sensitivity of the existing EF-HHO algorithmchain. At learning percentage 30, the key sensitivity of the existing EF-HHO algorithm secures second-best performance. As a result, the suggested blockchain-based privacysecures second-best performance. As a result, the suggested blockchain-based privacy preservation model executes more effectively than other heuristic algorithms.preservation model executes more effectively than other heuristic algorithms. ----- _Electronics 2023, 12, x FOR PEER REVIEW_ 22 of 30 _Electronics 2023, 12, 1404_ 21 of 29 (a) (b) (c) **Figure 9. Performance analysis of the designed blockchain-based privacy preservation model con-** **Figure 9. Performance analysis of the designed blockchain-based privacy preservation model con-** cerning (a) dataset 1, (b) dataset 2, and (c) dataset 3. cerning (a) dataset 1, (b) dataset 2, and (c) dataset 3. _6.7. Performance Analysis Using CPA and CCA6.7. Performance Analysis Using CPA and CCA_ Comparison of performance analysis of the ABC-ROA-based privacy preservationComparison of performance analysis of the ABC-ROA-based privacy preservation system over the existing heuristic algorithms in terms of chosen ciphertext attacks (CCA)system over the existing heuristic algorithms in terms of chosen ciphertext attacks (CCA) and chosen plaintext attacks (CPA) with three datasets is given in Tablesand chosen plaintext attacks (CPA) with three datasets is given in Tables 2 and 3, respec- 2 and 3, respectively. In CPA, the attacker can be used to encrypt the message. Hence, the goal of the CPA attacktively. In CPA, the attacker can be used to encrypt the message. Hence, the goal of the is to reduce the security of the encryption scheme. Here, symmetric and asymmetricCPA attack is to reduce the security of the encryption scheme. Here, symmetric and asymcryptography can be used. However, the CPA is often feasible in the diverse applications.metric cryptography can be used. However, the CPA is often feasible in the diverse appliSince the CPA is essential for public key cryptography where the encryption key is publiccations. Since the CPA is essential for public key cryptography where the encryption key and so attackers can encrypt any plaintext they choose. Moreover, the CCA can able tois public and so attackers can encrypt any plaintext they choose. Moreover, the CCA can decrypt the ciphertext message. Here, the CCA can be widely used in cryptanalysis toable to decrypt the ciphertext message. Here, the CCA can be widely used in cryptanalysis collect information by obtaining the decryptions of the chosen ciphertext, since it needs toto collect information by obtaining the decryptions of the chosen ciphertext, since it needs recover the hidden secret key which is used for the decryption. The proposed model hasto recover the hidden secret key which is used for the decryption. The proposed model more secure shared data in the transmission while analyzing the results. The CCA value ofhas more secure shared data in the transmission while analyzing the results. The CCA the developed system is guaranteed with a lower value on raising the proportion of keyvalue of the developed system is guaranteed with a lower value on raising the proportion variations for all three datasets. From the dataset 2 analysis of CPA, the ABC-ROA-basedof key variations for all three datasets. From the dataset 2 analysis of CPA, the ABC-ROAprivacy preservation model has high performance of 7.58%, 5.93%, 4.15%, and 7.64% thanbased privacy preservation model has high performance of 7.58%, 5.93%, 4.15%, and HHO, EF-HHO, BCO, and ROA for the key variations of 50. The proposed blockchain-based7.64% than HHO, EF-HHO, BCO, and ROA for the key variations of 50. The proposed blockprivacy preservation model is more effective than other heuristic algorithms.chain-based privacy preservation model is more effective than other heuristic algorithms. ----- _Electronics 2023, 12, 1404_ 22 of 29 **Table 2. Effectiveness analysis using CCA with three datasets for the developed blockchain-based** data privacy preservation system over Ethereum network. **Key Variations in the** **EF-HHO** **HHO [29]** **BCO [27]** **ROA [28]** **ABC-ROA** **Percentage** **[30]** Dataset 1 10 99.911 91.762 95.836 99.319 87.762 20 99.877 93.113 96.495 99.396 89.113 30 99.845 94.312 97.079 99.673 90.312 40 99.873 95.739 97.806 99.787 91.739 50 99.927 96.462 98.194 99.861 92.462 Dataset 2 10 99.909 91.626 95.767 99.31 87.626 20 99.876 92.894 96.385 99.389 88.894 30 99.843 94.071 96.957 99.669 90.071 40 99.871 95.452 97.661 99.784 91.452 50 99.926 96.283 98.105 99.859 92.283 Dataset 3 10 99.926 92.981 96.454 99.429 88.981 20 99.899 93.969 96.934 99.494 89.969 30 99.872 94.909 97.391 99.729 90.909 40 99.895 96.057 97.976 99.823 92.057 50 99.94 96.858 98.399 99.884 92.858 **Table 3. Effectiveness analysis using CPA with three datasets for the developed blockchain-based** data privacy preservation system over Ethereum network. **Key Variations in the** **EF-HHO** **HHO [29]** **BCO [27]** **ROA [28]** **ABC-ROA** **Percentage** **[30]** Dataset 1 10 58.784 52.465 55.624 58.236 43.465 20 62.388 59.351 60.87 61.866 50.351 30 65.623 63.651 64.637 65.13 54.651 40 68.534 65.454 66.994 68.075 56.454 50 71.114 65.723 68.418 70.684 56.723 Dataset 2 10 57.875 49.222 53.548 57.313 40.222 20 61.469 55.313 58.391 60.935 46.313 30 64.722 60.534 62.628 64.219 51.534 40 67.695 63.39 65.542 67.223 54.39 50 70.332 64.698 67.515 69.891 55.698 Dataset 3 10 68.646 61.093 64.87 68.191 52.093 20 71.919 66.26 69.089 71.501 57.26 30 74.772 70.154 72.463 74.389 61.154 40 77.269 72.702 74.986 76.918 63.702 50 79.448 74.553 77 79.127 65.553 ----- _Electronics 2023, 12, 1404_ 23 of 29 _6.8. Statistical Analysis of the Designed Method_ The statistical analysis of the designed blockchain-based data privacy preservation system over the Ethereum network is shown in Table 4. The designed ABC-ROA method attains 14.9%, 0.7%, 11.4%, and 4.1% better performance than HHO, EF-HHO, BCO, and _Electronics 2023, 12, x FOR PEER REVIEW ROA regarding dataset 1. Throughout the analysis, the experimental outcome has attained24 of 30_ superior performance when compared to other traditional approaches. attains 14.9%, 0.7%, 11.4%, and 4.1% better performance than HHO, EF-HHO, BCO, and **Table 4. Statistical analysis on developed data privacy preservation model over the Ethereum network.** ROA regarding dataset 1. Throughout the analysis, the experimental outcome has attained superior performance when compared to other traditional approaches. Terms **HHO [29]** **EF-HHO [30]** **BCO [27]** **ROA [28]** **ABC-ROA** Dataset 1 **Table 4. Statistical analysis on developed data privacy preservation model over the Ethereum net-** Best 6.5506 5.5954 6.2903 5.8088 5.5696 work. Worst 6.6609 6.6609 6.6609 6.6609 6.125 **Terms** **HHO [29]** **EF-HHO [30]** **BCO [27]** **ROA [28]** **ABC-ROA** Mean 6.5815 5.6644 6.3546 5.9459 5.6203 Dataset 1 Median Best 6.5506 6.5506 5.5963 5.5954 6.29276.2903 5.8088 5.8587 5.5696 5.5836 StandardWorst 0.050579 6.6609 0.212396.6609 0.120526.6609 6.6609 0.18825 6.125 0.11126 Deviation Mean 6.5815 5.6644 6.3546 5.9459 5.6203 Median 6.5506 5.5963 Dataset 2 6.2927 5.8587 5.5836 Standard Deviation Best 6.5078 0.050579 5.02810.21239 6.28310.12052 0.18825 5.2618 0.11126 4.8695 Worst 6.6822 6.6822Dataset 2 6.6716 6.6822 5.7822 Mean Best 6.5701 6.5078 5.1008 5.0281 6.36626.2831 5.2618 5.4352 4.8695 4.9401 Median Worst 6.5303 6.6822 5.0323 6.6822 6.30726.6716 6.6822 5.3044 5.7822 4.8807 Mean 6.5701 5.1008 6.3662 5.4352 4.9401 Standard deviationMedian 0.071683 6.5303 0.329515.0323 0.118696.3072 5.3044 0.3084 4.8807 0.18276 Standard deviation 0.071683 0.32951 0.11869 0.3084 0.18276 Dataset 3 Dataset 3 Best 6.552 6.2897 6.5597 6.3459 6.2438 Best 6.552 6.2897 6.5597 6.3459 6.2438 Worst 6.7356 6.7356 6.7356 6.7356 6.5356 Worst 6.7356 6.7356 6.7356 6.7356 6.5356 Mean 6.6597 6.3163 6.625 6.4021 6.3077 Mean 6.6597 6.3163 6.625 6.4021 6.3077 MedianMedian 6.6337 6.6337 6.2905 6.2905 6.60296.6029 6.3933 6.3933 6.2847 6.2847 StandardStandard deviation 0.080236 0.088813 0.054733 0.082943 0.063698 0.080236 0.088813 0.054733 0.082943 0.063698 deviation _6.9. ANOVA Test for the Developed Data Privacy Preservation Model over the Ethereum Net-_ _6.9. ANOVA Test for the Developed Data Privacy Preservation Model over the Ethereum Networkwork_ The validation of the ANOVA test for the designed ABC-ROA method regardingThe validation of the ANOVA test for the designed ABC-ROA method regarding fit ness function is shown in Figure 10. Thus, the experimental result of the developed fitness function is shown in Figure 10. Thus, the experimental result of the developed method attains superior performance compared to other traditional approaches. method attains superior performance compared to other traditional approaches. (a) (b) **Figure 10. Cont.** ----- _Electronics 2023, 12, x FOR PEER REVIEW_ 25 of 30 _Electronics 2023, 12, 1404_ 24 of 29 (c) (c) **Figure 10. ANOVA of the designed method for privacy preservation over the Ethereum network** **Figure 10. ANOVA of the designed method for privacy preservation over the Ethereum network** **Figure 10. ANOVA of the designed method for privacy preservation over the Ethereum network** regarding fitness function (regarding fitness function (aa) dataset 1, () dataset 1, (bb) dataset 2, and () dataset 2, and (c) dataset 3. c) dataset 3. regarding fitness function (a) dataset 1, (b) dataset 2, and (c) dataset 3. _6.10.6.10.Validation of Control for Parameters of Different Algorithms Using the Designed MethodValidation of Control for Parameters of Different Algorithms Using the Designed Method_ _6.10. Validation of Control for Parameters of Different Algorithms Using the Designed Method_ The validation of control for parameters of different existing algorithms regarding The validation of control for parameters of different existing algorithms regarding The validation of control for parameters of different existing algorithms regarding Euclidean distance and Pearson and Spearman correlations is shown in Figure 11. Here, Euclidean distance and Pearson and Spearman correlations is shown in Figure 11. Here, Euclidean distance and Pearson and Spearman correlations is shown in Figure 11. Here, the evaluation of the parameter for the proposed ABC-ROA method is taken as ird= 0.06. the evaluation of the parameter for the proposed ABC-ROA method is taken as ird= 0.06. the evaluation of the parameter for the proposed ABC-ROA method is taken as ird = 0.06. Throughout the analysis, the developed method achieves enhanced performance com Throughout the analysis, the developed method achieves enhanced performance com Throughout the analysis, the developed method achieves enhanced performance comparedpared to the other existing methods. pared to the other existing methods. to the other existing methods. (a) (b) (a) (b) (c) (c) **Figure 11. Analysis of controlling the parameters of the designed method using (a) Euclidean distance,** (b) Pearson correlation, and (c) Spearman correlation. (b) (c) ----- _Electronics 2023, 12, 1404_ **Figure 11. Analysis of controlling the parameters of the designed method using (a) Euclidean dis-25 of 29** tance, (b) Pearson correlation, and (c) Spearman correlation. **7. Security Analysis** **7. Security Analysis** The developed ABC-ROA-based privacy preservation system model is evaluated The developed ABC-ROA-based privacy preservation system model is evaluated with various attacks such as KCA, KPA, adaptive chosen-plaintext analysis (ACPA), and with various attacks such as KCA, KPA, adaptive chosen-plaintext analysis (ACPA), and Ciphertext-Only Analysis (COA) assessed based on three datasets by comparing with re Ciphertext-Only Analysis (COA) assessed based on three datasets by comparing with cently used algorithms, as shown in Figures 12–15, respectively. Correlating one original recently used algorithms, as shown in Figures 12–15, respectively. Correlating one original datum with all original data and one sanitized datum with all sanitized data defines KPA datum with all original data and one sanitized datum with all sanitized data defines KPA analysis. KCA analysis is described as correlating each sanitized data with its data re analysis. KCA analysis is described as correlating each sanitized data with its data restored stored data. The ACPA attack is similar to the CPA attack. It selects the plaintext and data. The ACPA attack is similar to the CPA attack. It selects the plaintext and ciphertext ciphertext that are learned from past encryptions. In a COA attack, it uses known data that are learned from past encryptions. In a COA attack, it uses known data collection. collection. In the ABC-ROA-based privacy preservation system, the ACPA, COA, KPA, In the ABC-ROA-based privacy preservation system, the ACPA, COA, KPA, and KCA and KCA analysis shows the lowest value and indicates the minimum error. While ana analysis shows the lowest value and indicates the minimum error. While analyzing the lyzing the evaluation of different attacks for the designed ABC-ROA method, it is revealed evaluation of different attacks for the designed ABC-ROA method, it is revealed that that the designed ABC-ROA based privacy preservation over the Ethereum network at the designed ABC-ROA based privacy preservation over the Ethereum network attains tains effective performance. effective performance. (a) (b) (c) **Figure 12. Effectiveness analysis of the implemented ABC-ROA based privacy preservation model** **Figure 12. Effectiveness analysis of the implemented ABC-ROA based privacy preservation model** using KCA in terms of (a) dataset 1, (b) dataset 2, and (c) dataset 3 using KCA in terms of (a) dataset 1, (b) dataset 2, and (c) dataset 3. ----- _Electronics Electronics 20232023,, 12 12, x FOR PEER REVIEW, x FOR PEER REVIEW_ 27 of 30 27 of 30 _Electronics 2023, 12, 1404_ 26 of 29 ((aa) ) ((bb) ) ((cc) ) **Figure 13. Figure 13. Performance analysis of the developed ABC-ROA based privacy preservation model us-Performance analysis of the developed ABC-ROA based privacy preservation model us-** **Figure 13. Performance analysis of the developed ABC-ROA based privacy preservation model using** ing KPA regarding (ing KPA regarding (aa) dataset 1, () dataset 1, (bb) dataset 2, and () dataset 2, and (cc) dataset 3. ) dataset 3. KPA regarding (a) dataset 1, (b) dataset 2, and (c) dataset 3. ((aa) ) ((bb) ) **Figure 14. Cont.** ----- _Electronics Electronics 20232023,, 12 12, x FOR PEER REVIEW, x FOR PEER REVIEW_ 28 of 30 28 of 30 _Electronics 2023, 12, 1404_ 27 of 29 ((cc) ) **Figure 14. Figure 14. Performance analysis of the developed ABC-ROA based privacy preservation model us-Performance analysis of the developed ABC-ROA based privacy preservation model us-** **Figure 14. Performance analysis of the developed ABC-ROA based privacy preservation model using** ing ACPA regarding (ing ACPA regarding (aa) dataset 1, () dataset 1, (bb) dataset 2, and () dataset 2, and (cc) dataset 3. ) dataset 3. ACPA regarding (a) dataset 1, (b) dataset 2, and (c) dataset 3. ((aa) ) ((bb) ) ((cc) ) **Figure 15.Figure 15. Figure 15. Performance analysis of the developed ABC-ROA based privacy preservation model usingPerformance analysis of the developed ABC-ROA based privacy preservation model us-Performance analysis of the developed ABC-ROA based privacy preservation model us-** ing COA regarding (ing COA regarding (aa) dataset 1, () dataset 1, (bb) dataset 2, and () dataset 2, and (cc) dataset 3. ) dataset 3. COA regarding (a) dataset 1, (b) dataset 2, and (c) dataset 3. **8. Conclusions8. Conclusions8. Conclusions** A new blockchain-based privacy preservation model over the Ethereum networkA new blockchain-based privacy preservation model over the Ethereum network A new blockchain-based privacy preservation model over the Ethereum network was developed for preserving data privacy using blockchain technology. The data werewas developed for preserving data privacy using blockchain technology. The data were was developed for preserving data privacy using blockchain technology. The data were collected from standard databases. Initially, the data were sanitized, and an optimalcollected from standard databases. Initially, the data were sanitized, and an optimal key collected from standard databases. Initially, the data were sanitized, and an optimal key key developed using the ABC-ROA algorithm. The optimal key generation followed ----- _Electronics 2023, 12, 1404_ 28 of 29 the objective functions HF rate, IP rate, FR, and DM. The sanitized data progressed to the data restoration process, which restored the data in the database. These data were formed as subchains, known as the supply chain framework. The developed blockchain framework gave better privacy for the data over the supply chain network with the help of the generated optimal key. The effectiveness of the proposed blockchain-based privacy preservation model was compared with the existing privacy preservation models. The proposed ABC-ROA-based privacy preservation model performed 20.2% better than HHO, 17.4% better than EF-HHO, 13.7% better than BCO, and 20.7% better than ROA while considering dataset 2 with the key variation of 50. Therefore, compared to other privacy preservation approaches, the developed ABC-ROA-based privacy preservation model performs better for all key variations than other heuristic algorithms. One of the most important challenges in the existing privacy preservation model over the Ethereum network is scalability. Due to the scalability issues, it cannot provide the optimal solution, and also it generates issues such as inefficiency and limited block size. In this research, the developed ABC-ROA method was utilized to solve these issues. The estimation of convergence and optimization of deep-structure architectures were utilized to resolve the scalability issues. Moreover, implementing standard machine learning and deep learning approaches provides the ability to solve these issues. **Author Contributions: Conceptualization, Y.V.R.S.V.; Methodology, Y.V.R.S.V.; Data curation, Y.V.R.S.V.;** Writing—Original draft preparation, Y.V.R.S.V.; Visualization, K.J.; Investigation, K.J.; Validation, Y.V.R.S.V.; Reviewing and Editing, K.J. All authors have read and agreed to the published version of the manuscript. **Funding: This research did not receive any specific funding.** **Data Availability Statement: The data underlying this article are available in DataCo Smart Supply** [Chain for Big Data Analysis database, at https://www.kaggle.com/shivkp/customer-behaviour](https://www.kaggle.com/shivkp/customer-behaviour) (accessed on 10 January 2023). **Conflicts of Interest: The authors declare no conflict of interest.** **References** 1. Weng, J.; Weng, J.; Zhang, J.; Li, M.; Zhang, Y.; Luo, W. DeepChain: Auditable and Privacy-Preserving Deep Learning with [Blockchain-Based Incentive. IEEE Trans. Dependable Secur. Comput. 2021, 18, 2438–2455. [CrossRef]](http://doi.org/10.1109/TDSC.2019.2952332) 2. Tahir, S.; Tahir, H.; Sajjad, A.; Rajarajan, M.; Khan, F. Privacy-preserving COVID-19 contact tracing using blockchain. J. Commun. _[Netw. 2021, 23, 360–373. [CrossRef]](http://doi.org/10.23919/JCN.2021.000031)_ 3. Huang, C.; Zhao, Y.; Chen, H.; Wang, X.; Zhang, Q.; Chen, Y.; Wang, H.; Lam, K.-Y. ZkRep: A Privacy-Preserving Scheme for [Reputation-Based Blockchain System. IEEE Internet Things J. 2021, 9, 4330–4342. [CrossRef]](http://doi.org/10.1109/JIOT.2021.3105273) 4. Wu, G.; Wang, S.; Ning, Z.; Zhu, B. Privacy-Preserved Electronic Medical Record Exchanging and Sharing: A Blockchain-Based [Smart Healthcare System. IEEE J. Biomed. Health Inform. 2021, 26, 1917–1927. [CrossRef] [PubMed]](http://doi.org/10.1109/JBHI.2021.3123643) 5. Yang, Y.; Wu, J.; Long, C.; Liang, W.; Lin, Y.-B. Blockchain-Enabled Multiparty Computation for Privacy Preserving and Public [Audit in Industrial IoT. IEEE Trans. Ind. Inform. 2022, 18, 9259–9267. [CrossRef]](http://doi.org/10.1109/TII.2022.3177630) 6. Du, R.; Ma, C.; Li, M. Privacy-Preserving Searchable Encryption Scheme Based on Public and Private Blockchains. Tsinghua Sci. _[Technol. 2023, 28, 13–26. [CrossRef]](http://doi.org/10.26599/TST.2021.9010070)_ 7. Yang, Q.; Wang, H. Privacy-Preserving Transactive Energy Management for IoT-Aided Smart Homes via Blockchain. IEEE Internet _[Things J. 2021, 8, 11463–11475. [CrossRef]](http://doi.org/10.1109/JIOT.2021.3051323)_ 8. Tran, Q.N.; Turnbull, B.P.; Wu, H.-T.; de Silva, A.J.S.; Kormusheva, K.; Hu, J. A Survey on Privacy-Preserving Blockchain Systems [(PPBS) and a Novel PPBS-Based Framework for Smart Agriculture. IEEE Open J. Comput. Soc. 2021, 2, 72–84. [CrossRef]](http://doi.org/10.1109/OJCS.2021.3053032) 9. Yang, Y.; Wei, L.; Wu, J.; Long, C.; Li, B. A Blockchain-Based Multidomain Authentication Scheme for Conditional Privacy [Preserving in Vehicular Ad-Hoc Network. IEEE Internet Things J. 2021, 9, 8078–8090. [CrossRef]](http://doi.org/10.1109/JIOT.2021.3107443) 10. Baza, M.; Sherif, A.; Mahmoud, M.M.E.A.; Bakiras, S.; Alasmary, W.; Abdallah, M.; Lin, X. Privacy-Preserving Blockchain-Based [Energy Trading Schemes for Electric Vehicles. IEEE Trans. Veh. Technol. 2021, 70, 9369–9384. [CrossRef]](http://doi.org/10.1109/TVT.2021.3098188) 11. Zhang, X.; Jiang, S.; Liu, Y.; Jiang, T.; Zhou, Y. Privacy-Preserving Scheme with Account-Mapping and Noise-Adding for Energy [Trading Based on Consortium Blockchain. IEEE Trans. Netw. Serv. Manag. 2021, 19, 569–581. [CrossRef]](http://doi.org/10.1109/TNSM.2021.3110980) 12. Wu, Y.; Tang, S.; Zhao, B.; Peng, Z. BPTM: Blockchain-Based Privacy-Preserving Task Matching in Crowdsourcing. IEEE Access **[2019, 7, 45605–45617. [CrossRef]](http://doi.org/10.1109/ACCESS.2019.2908265)** 13. Abdelsalam, H.A.; Srivastava, A.K.; Eldosouky, A. Blockchain-Based Privacy Preserving and Energy Saving Mechanism for [Electricity Prosumers. IEEE Trans. Sustain. Energy 2021, 13, 302–314. [CrossRef]](http://doi.org/10.1109/TSTE.2021.3109482) ----- _Electronics 2023, 12, 1404_ 29 of 29 14. Zou, S.; Xi, J.; Xu, G.; Zhang, M.; Lu, Y. CrowdHB: A Decentralized Location Privacy-Preserving Crowdsensing System Based on [a Hybrid Blockchain Network. IEEE Internet Things J. 2021, 9, 14803–14817. [CrossRef]](http://doi.org/10.1109/JIOT.2021.3084937) 15. Rahman, M.S.; Khalil, I.; Moustafa, N.; Kalapaaking, A.P.; Bouras, A. A Blockchain-Enabled Privacy-Preserving Verifiable Query Framework for Securing Cloud-Assisted Industrial Internet of Things Systems. IEEE Trans. Ind. Inform. 2021, 18, 5007–5017. [[CrossRef]](http://doi.org/10.1109/TII.2021.3105527) 16. Zhang, C.; Zhu, L.; Xu, C.; Sharif, K. PRVB: Achieving Privacy-Preserving and Reliable Vehicular Crowdsensing via Blockchain [Oracle. IEEE Trans. Veh. Technol. 2020, 70, 831–843. [CrossRef]](http://doi.org/10.1109/TVT.2020.3046027) 17. Chulerttiyawong, D.; Jamalipour, A. A Blockchain Assisted Vehicular Pseudonym Issuance and Management System for [Conditional Privacy Enhancement. IEEE Access 2021, 9, 127305–127319. [CrossRef]](http://doi.org/10.1109/ACCESS.2021.3112013) 18. Lin, C.; He, D.; Huang, X.; Xie, X.; Choo, K.-K.R. PPChain: A Privacy-Preserving Permissioned Blockchain Architecture for [Cryptocurrency and Other Regulated Applications. IEEE Syst. J. 2021, 15, 4367–4378. [CrossRef]](http://doi.org/10.1109/JSYST.2020.3019923) 19. Rahmadika, S.; Astillo, P.V.; Choudhary, G.; Duguma, D.G.; Sharma, V.; You, I. Blockchain-Based Privacy Preservation Scheme for [Misbehavior Detection in Lightweight IoMT Devices. IEEE J. Biomed. Health Inform. 2022, 27, 710–721. [CrossRef]](http://doi.org/10.1109/JBHI.2022.3187037) 20. Xiong, T.; Zhang, R.; Liu, J.; Huang, T.; Liu, Y.; Yu, F.R. A blockchain-based and privacy-preserved authentication scheme for [inter-constellation collaboration in Space-Ground Integrated Networks. Comput. Netw. 2022, 206, 108793. [CrossRef]](http://doi.org/10.1016/j.comnet.2022.108793) 21. Singh, S.; Rathore, S.; Alfarraj, O.; Tolba, A.; Yoon, B. A framework for privacy-preservation of IoT healthcare data using Federated [Learning and blockchain technology. Futur. Gener. Comput. Syst. 2021, 129, 380–388. [CrossRef]](http://doi.org/10.1016/j.future.2021.11.028) 22. Guo, L.; Xie, H.; Li, Y. Data encryption based blockchain and privacy preserving mechanisms towards big data. J. Vis. Commun. _[Image Represent. 2019, 70, 102741. [CrossRef]](http://doi.org/10.1016/j.jvcir.2019.102741)_ 23. Mohan, D.; Alwin, L.; Neeraja, P.; Lawrence, K.D.; Pathari, V. A private Ethereum blockchain implementation for secure data [handling in Internet of Medical Things. J. Reliab. Intell. Environ. 2021, 8, 379–396. [CrossRef]](http://doi.org/10.1007/s40860-021-00153-2) 24. Elisa, N.; Yang, L.; Chao, F.; Cao, Y. A framework of blockchain-based secure and privacy-preserving E-government system. Wirel. _[Netw. 2018, 1–11. [CrossRef]](http://doi.org/10.1007/s11276-018-1883-0)_ 25. Dewangan, N.K.; Chandrakar, P.; Kumari, S.; Rodrigues, J.J. Enhanced privacy-preserving in student certificate management in [blockchain and interplanetary file system. Multimedia Tools Appl. 2022, 1–20. [CrossRef]](http://doi.org/10.1007/s11042-022-13915-8) 26. Abidi, M.H.; Alkhalefah, H.; Umer, U.; Mohammed, M.K. Blockchain-based secure information sharing for supply chain [management: Optimization assisted data sanitization process. Int. J. Intell. Syst. 2020, 36, 260–290. [CrossRef]](http://doi.org/10.1002/int.22299) 27. [Dutta, T.; Bhattacharyya, S.; Dey, S.; Platos, J. Border Collie Optimization. IEEE Access 2020, 8, 109177–109197. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.2999540) 28. Moazzeni, A.R.; Khamehchi, E. Rain optimization algorithm (ROA): A new metaheuristic method for drilling optimization [solutions. J. Pet. Sci. Eng. 2020, 195, 107512. [CrossRef]](http://doi.org/10.1016/j.petrol.2020.107512) 29. Elgamal, Z.M.; Yasin, N.B.M.; Tubishat, M.; Alswaitti, M.; Mirjalili, S. An Improved Harris Hawks Optimization Algorithm with [Simulated Annealing for Feature Selection in the Medical Field. IEEE Access 2020, 8, 186638–186652. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.3029728) 30. Hu, H.; Ao, Y.; Bai, Y.; Cheng, R.; Xu, T. An Improved Harris’s Hawks Optimization for SAR Target Recognition and Stock Market [Index Prediction. IEEE Access 2020, 8, 65891–65910. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.2985596) **Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual** author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/electronics12061404?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/electronics12061404, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/2079-9292/12/6/1404/pdf?version=1678939714" }
2,023
[]
true
2023-03-15T00:00:00
[ { "paperId": "9156c07c27c6739408bcfceae55d9bbd1cc14df3", "title": "Privacy-Preserving Searchable Encryption Scheme Based on Public and Private Blockchains" }, { "paperId": "d7fad1aa2bac0e7368d2abe21d4409d5c15af221", "title": "Blockchain-Enabled Multiparty Computation for Privacy Preserving and Public Audit in Industrial IoT" }, { "paperId": "2f8fbae457ea0a597f57342d77505aebc14f09b5", "title": "Enhanced privacy-preserving in student certificate management in blockchain and interplanetary file system" }, { "paperId": "7dd5e5efa92765c909cd8eaa9c9949c82ea794c8", "title": "CrowdHB: A Decentralized Location Privacy-Preserving Crowdsensing System Based on a Hybrid Blockchain Network" }, { "paperId": "acb772e615373b3856abcc9411ff4778a61e683f", "title": "A Blockchain-Enabled Privacy-Preserving Verifiable Query Framework for Securing Cloud-Assisted Industrial Internet of Things Systems" }, { "paperId": "2e85d94f5a6b541e11bb7a06c7e3bbde57268c32", "title": "Blockchain-Based Privacy Preservation Scheme for Misbehavior Detection in Lightweight IoMT Devices" }, { "paperId": "ef89aac4a4cce36f2986d173a10dcfdc3d973493", "title": "A Blockchain-Based Multidomain Authentication Scheme for Conditional Privacy Preserving in Vehicular Ad-Hoc Network" }, { "paperId": "0b21645d8e7f84ddc97734a597b9c5f222d760a3", "title": "Privacy-Preserving Scheme With Account-Mapping and Noise-Adding for Energy Trading Based on Consortium Blockchain" }, { "paperId": "8776224a0a099597d5cf841776f72c0b775afe11", "title": "A blockchain-based and privacy-preserved authentication scheme for inter-constellation collaboration in Space-Ground Integrated Networks" }, { "paperId": "210ac509ad7310226c3b4ceb5099d2361f888dd1", "title": "Blockchain-Based Privacy Preserving and Energy Saving Mechanism for Electricity Prosumers" }, { "paperId": "d055ec0254351d709a59ecb8f20eecf3e85370ef", "title": "A framework for privacy-preservation of IoT healthcare data using Federated Learning and blockchain technology" }, { "paperId": "f6f63f46c3955113a850c22f4c40fa6333ffe416", "title": "Privacy-Preserved Electronic Medical Record Exchanging and Sharing: A Blockchain-Based Smart Healthcare System" }, { "paperId": "0cd88c83357ec0aaf9dbac8c79237f66a0d5ec99", "title": "Privacy-Preserving Blockchain-Based Energy Trading Schemes for Electric Vehicles" }, { "paperId": "4d2cad9f22c590a474dbe58869955648a044faa0", "title": "ZkRep: A Privacy-Preserving Scheme for Reputation-Based Blockchain System" }, { "paperId": "d45d23c874b6831cd22787362079c48adf318049", "title": "A private Ethereum blockchain implementation for secure data handling in Internet of Medical Things" }, { "paperId": "33dac2d110a0cf088e145b12a1184af7628a248c", "title": "Privacy-Preserving Transactive Energy Management for IoT-Aided Smart Homes via Blockchain" }, { "paperId": "daaaebbe1910f41da860db2e90105d84372a4691", "title": "PRVB: Achieving Privacy-Preserving and Reliable Vehicular Crowdsensing via Blockchain Oracle" }, { "paperId": "1a8c6b39d67103fe0e772a63535abb0f4a16110a", "title": "Rain optimization algorithm (ROA): A new metaheuristic method for drilling optimization solutions" }, { "paperId": "85be5b16c9749a4eeef3568426fd6385bbe4b29f", "title": "Blockchain‐based secure information sharing for supply chain management: Optimization assisted data sanitization process" }, { "paperId": "e4de4928dfd42a68ce602d075f2d98b1f88a6bb3", "title": "PPChain: A Privacy-Preserving Permissioned Blockchain Architecture for Cryptocurrency and Other Regulated Applications" }, { "paperId": "80f614b891b30f2a86aabc00c3c5d25e6d0b8f82", "title": "Data encryption based blockchain and privacy preserving mechanisms towards big data" }, { "paperId": "07f229cc6e5b80fc8ffbbb3e4db85142466f55a9", "title": "DeepChain: Auditable and Privacy-Preserving Deep Learning with Blockchain-Based Incentive" }, { "paperId": "fa1bbc1cd49a9872b993feb0d33a405ec2f3c94d", "title": "BPTM: Blockchain-Based Privacy-Preserving Task Matching in Crowdsourcing" }, { "paperId": "4a4587a43228c9b7e0784190656ab10d583cebc1", "title": "A framework of blockchain-based secure and privacy-preserving E-government system" }, { "paperId": "71e3b0b2fb888dc5aa0ff6b2d4659dff32498023", "title": "A Survey on Privacy-Preserving Blockchain Systems (PPBS) and a Novel PPBS-Based Framework for Smart Agriculture" }, { "paperId": "52fba58cd69d0f71878004dd2c616d88a0799479", "title": "Privacy-preserving COVID-19 contact tracing using blockchain" }, { "paperId": "c4dc2fe1d1189d5c8215b0fcc73b7708bb3d8146", "title": "A Blockchain Assisted Vehicular Pseudonym Issuance and Management System for Conditional Privacy Enhancement" }, { "paperId": "b6a934727f869f620e50f84fa7f46d6159164fdc", "title": "An Improved Harris’s Hawks Optimization for SAR Target Recognition and Stock Market Index Prediction" }, { "paperId": "97eabc4d6bfd27cd2f33c1c3de5cb073cd645b59", "title": "Border Collie Optimization" }, { "paperId": "b384e0796db848d7b14d214d886798c1900eb09b", "title": "An Improved Harris Hawks Optimization Algorithm With Simulated Annealing for Feature Selection in the Medical Field" } ]
25,772
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0106a389bab04617737f3a786e08e31fa4afe7e5
[ "Computer Science" ]
0.895569
Why and how informatics and applied computing can still create structural changes and competitive advantage
0106a389bab04617737f3a786e08e31fa4afe7e5
Applied Computing and Informatics
[ { "authorId": "2022247", "name": "S. Mitropoulos" }, { "authorId": "1728886", "name": "C. Douligeris" } ]
{ "alternate_issns": [ "2634-1964" ], "alternate_names": [ "Appl Comput Informatics" ], "alternate_urls": [ "https://www.journals.elsevier.com/applied-computing-and-informatics/", "https://www.emeraldgrouppublishing.com/journal/ajms" ], "id": "256a5738-77cc-4cee-ad61-ef18264f639c", "issn": "2210-8327", "name": "Applied Computing and Informatics", "type": "journal", "url": "http://www.elsevier.com/wps/find/journaldescription.cws_home/724375/description#description" }
PurposeIn the new digital age, enterprises are facing an increasing global competition. In this paper, we first examine how Information Technology (IT) can play an important role in giving significant competitive advantage in the modern enterprises. The business value of IT is examined, as well as the limitations and the trade-offs that its applicability faces. Next, we present the basic principles for a successful IT strategy, considering the development of a long-term IT renovation plan, the strategic alignment of IT with the business strategy, and the adoption of an integrated, distributed, and interoperable IT platform. Finally, we examine how a highly functional and efficient IT organization can be developed.Design/methodology/approachOur methodological approach was based to the answers of the following questions: 1. Does IT still matter? 2. What is the business value created by IT along with the corresponding limitations and trade-offs? 3. How could a successful IT Strategy be build up? 4. How could an effective? T planning aligned with the business strategy be build up? 5. How could a homogenized and distributed corporate IT platform be developed? and finally, 6. How could a high-performance IT-enabled enterprise be build up?FindingsThe enterprises in order to succeed in the new digital area need to: 1. synchronize their IT strategy with their business strategy, 2. formulate a long-term IT strategy, 3. adopt IT systems and solutions that are implemented with elasticity, interoperability, distribution, and service-orientation. 4. keep a strategic direction towards the creation of an exceptional organization based on IT.Originality/valueThis paper is original with respect to the integrated approach the overall problem is examined. There is a prototype combined investigation of all perspectives for an effective enforcement of IT in a way that causes acceleration in competitive advantage when conducting business.
ERROR: type should be string, got "https://www.emerald.com/insight/2210-8327.htm\n\n# Why and how informatics and applied computing can still create structural changes and competitive advantage\n\n## Sarandis Mitropoulos\n### Regional Development, Ionian University, Lefkada, Greece, and\n## Christos Douligeris\n### Informatics, University of Piraeus, Piraeus, Greece\n\nAbstract\nPurpose – In the new digital age, enterprises are facing an increasing global competition. In this paper, we first\nexamine how Information Technology (IT) can play an important role in giving significant competitive\nadvantage in the modern enterprises. The business value of IT is examined, as well as the limitations and the\ntrade-offs that its applicability faces. Next, we present the basic principles for a successful IT strategy,\nconsidering the development of a long-term IT renovation plan, the strategic alignment of IT with the business\nstrategy, and the adoption of an integrated, distributed, and interoperable IT platform. Finally, we examine\nhow a highly functional and efficient IT organization can be developed.\nDesign/methodology/approach – Our methodological approach was based to the answers of the following\nquestions: 1. Does IT still matter? 2. What is the business value created by IT along with the corresponding\nlimitations and trade-offs? 3. How could a successful IT Strategy be build up? 4. How could an effective?\nT planning aligned with the business strategy be build up? 5. How could a homogenized and distributed\ncorporate IT platform be developed? and finally, 6. How could a high-performance IT-enabled enterprise be\nbuild up?\nFindings – The enterprises in order to succeed in the new digital area need to: 1. synchronize their IT strategy\nwith their business strategy, 2. formulate a long-term IT strategy, 3. adopt IT systems and solutions that are\nimplemented with elasticity, interoperability, distribution, and service-orientation. 4. keep a strategic direction\ntowards the creation of an exceptional organization based on IT.\nOriginality/value – This paper is original with respect to the integrated approach the overall problem is\nexamined. There is a prototype combined investigation of all perspectives for an effective enforcement of IT in\na way that causes acceleration in competitive advantage when conducting business.\nKeywords New technologies, IT strategy, Strategic alignment, Business value, SOA\nPaper type Research paper\n\n1. Introduction\nThe steadily increasing global market competition is significantly reducing the turnover time\nof innovation products. Information Technology (IT) can provide a significant competitive\nmarket advantage to an enterprise because it dynamically promotes relationships with its\ncustomers and suppliers, strengthens strategic alliances, promotes innovative products and\nservices, adapts existing solutions, and achieves lower operating costs, while it improves the\ninternal business processes. It is true that new technologies, such as the cloud, mobile\ncomputing, the Internet of Things (IoT), and the use of artificial intelligence (AI) have brought\n\n© Sarandis Mitropoulos and Christos Douligeris. Published in Applied Computing and Informatics.\nPublished by Emerald Publishing Limited. This article is published under the Creative Commons\nAttribution (CC BY 4.0) licence. Anyone may reproduce, distribute, translate and create derivative works\nof this article (for both commercial and non-commercial purposes), subject to full attribution to the\n[original publication and authors. The full terms of this licence may be seen at http://creativecommons.](http://creativecommons.org/licences/by/4.0/legalcode)\n[org/licences/by/4.0/legalcode](http://creativecommons.org/licences/by/4.0/legalcode)\n\n\n## Informatics and applied computing\n\nReceived 7 June 2021\nRevised 19 July 2021\n16 September 2021\nAccepted 17 September 2021\n\nApplied Computing and\nInformatics\nEmerald Publishing Limited\ne-ISSN: 2210-8327\np-ISSN: 2634-1964\n[DOI 10.1108/ACI-06-2021-0149](https://doi.org/10.1108/ACI-06-2021-0149)\n\n\n-----\n\n## ACI\n\n\naround structural changes in the way businesses operate as well as in the way markets\nfunction. For example, the internet has created hyperarchies that influence the creation of\nnew business models [1], by replacing the traditional enterprise hierarchies, while on the other\nhand it has necessitated a comprehensive review of the traditional business strategies. This\nrevision of the companies’ business strategies, as well as the new supply value chain have\nsignificantly increased the productivity, the quality of services offered to customers and the\noverall value proposition of the enterprises.\nNevertheless, there are many enterprises, which, despite investing a significant amount of\nmoney and human resources in applied computing, as well as in relevant research, have failed\nin having the initially expected advantages. This happened because these enterprises failed\nto adopt and then successfully implement the new e-business and e-governance models [2]. In\nthe context of rapidly responding to market needs and to their fierce competition, many\nenterprises have only pursued short-term benefits from applied computing and informatics\nby trying only to accelerate the development of new services and products at the lowest\npossible operating costs. But ignoring the fundamentals of strategy formulation has led to a\nconvergence of business practices based on cost leadership [3], something we have seen\nhappening in the past with the dotcom enterprises. This problem still exists with the latest\ninformatics technologies, where we see that enterprises still do not incorporate the\ninformatics and computing applications in their processes in the right way [2, 4, 5]. The main\nfaults that enterprises make regarding the adoption of informatics are that they do not:\n\n(1) synchronize their IT strategy with the goals of their business strategy,\n\n(2) formulate a long-term IT strategy,\n\n(3) keep a strategic direction towards the creation of an exceptional organization based\non IT,\n\n(4) adopt effective IT systems, and integrated IΤ solutions; instead these are\nimplemented hastily, inelastically and without interoperability.\n\nThe exponential growth of ubiquitous computing drives the need for new business models,\nwhich must serve an effective IT-enabled business strategy directly related to the Digital\nTransformation. The problem space here is how the effective adoption of new IT technologies\ncan be achieved which in turn drives the requirements of renewed business strategy and\nprocesses, and of culture change [29]. The main problem which modern enterprises face\nconcerns the right IT implementation along with the adaptation of the right strategy, process,\nand culture. This paper tries to answer all these requirements in a consistent and\nmethodological way. Even though a considerable body of research towards this direction\nexists, there is still a gap for the development of a perfect alignment between the strategy the\nenterprises must follow and an intelligent and effective way the new technologies are\nimplemented so that they create a competitive advantage [5]. Internet, cloud computing,\nmobile computing, software-oriented architectures (SOA), Internet of things (IoT), blockchain,\nserver virtualization and other modern technologies provide the enterprises with the\nopportunity to migrate from the traditional business models to new ones [30, 31, 35, 36].\nAlthough there exist theoretical frameworks in the literature [2–5], there is still a lack of of\nframeworks that combine the characteristics of hybrid, both research and practical methods\napproaches. The outcomes of the existing approaches show that the use of IT has not been\nproven fully effective, except from some exceptional cases.\nTo achieve an effective IT incorporation, this paper proposes a framework with very\nspecific suggestions. This framework can be used as a roadmap for future readiness and\ngrowth. The evaluation discussion shows the effectiveness of the proposed framework and\nmethodology. This framework in short consists of the following steps:\n\n\n-----\n\n(1) Examination of the expected business value to be created by IT along with any\nexpected limitations and necessary trade-offs.\n\n(2) Formulation of an effective IT strategy aligned with the overall business strategy.\n\n(3) Development of an effective IT planning.\n\n(4) Development of a homogenized, interoperable, and distributed IT platform.\n\n(5) Development of a high IT-enabled organization in all its dimensions.\n\n(6) Establishment of an on-going evaluation system for the continuous improvement of\nthe IT strategy and the corresponding infrastructure.\n\nThe issues and steps mentioned above are thoroughly analysed and potential solutions are\nproposed. A thorough and comprehensive investigation of whether IT can still make\nstructural changes in the operation of businesses is performed. In addition, it is examined how\nthis can be achieved with respect to the creation of business value considering the\ncorresponding limitations and necessary trade-offs. But since creating value without\ndeveloping a rational strategy is impossible the principles of a sound IT strategy are\nexamined, as well as how this can be effectively developed. Finally, implementation issues\nand evaluation results are discussed, as well as potential future work.\n\n2. IT still does matter\nMany researchers argue that IΤ no longer offers any innovation over the competition and\nhave, therefore, IΤ has reached the stage of maturing as a service [6]. This viewpoint has been\nexpressed without considering its high and disruptive evolution. Τhere is obviously a part of\nIT that is common to almost all enterprises, making this utility approach workable.\nNevertheless, this is only one side of the coin because IT is creating new situations that\naccelerate the developments in the operation of markets and enterprises. The concept of\nubiquitous computing, for example, is a new trend which will surely cause structural changes,\nwhile new ideas in virtual communities can create significant changes in corporate\ncollaboration by forming on-demand virtual business partner hyperarchies. In addition, new\nintelligent algorithms for machine learning, artificial intelligence and biotechnology are being\ndeveloped, thus facilitating research into new products (e.g. drugs and crystal solid materials)\nand services (e.g. telemedicine) with significant business and social benefits [7]. Thus, the new\ninformation technologies can leverage the capacity for innovation. Enterprises through the\nsmart and targeted use of IT will be able to create growth, and innovative products and\nservices at the right time.\nIn addition, enterprises need to focus on another perspective of IT, that of the growing\nbusiness value through digital culture. A well-designed business strategy, synchronized\nwith that of IT, creates an innovation-oriented corporate culture. Such a culture is not easy\nto be copied from the competition, thus, giving the enterprises that adopt it a clear lead.\nThere are many important examples that prove this to be true, such as the ones of Toyota\nand Dell – their supply chain and production chain management practices make them\nstand out [6].\nThe innovative and smart use of IT (e.g. green informatics) can create an additional\ncompetitive advantage which, most of the times, is closely connected with a change of culture\n\n[8]. However, the technology is not a panacea. In practice, many enterprises do not effectively\nenforce the new IT-enabled business models because their adoption of IT has not been\ndirected to develop a highly IT-enabled corporate culture, they have not integrated their\ndigital strategy with the overall business strategy, or they have not adequately understood\nthe new technology trends while developing a corresponding strategy.\n\n\n## Informatics and applied computing\n\n\n-----\n\n## ACI\n\n\n3. Business value creation: limitations and trade-offs\nNowadays, the relationship between business strategy and IT-based innovation is a strongly\ninteracting one. A successful business strategy must define an IT infrastructure driven by\ninnovation. Obviously, all of the expected benefits offer improved business performance and\nare, therefore, critical to the business success [5, 9].\nIT enables an increased control of the operating costs as well as of productivity. Indeed,\nthe assembly of products e.g. through the Computer Integrated Manufacturing (CIM) and the\nRobotics systems of the industry 4.0 era, increases productivity and quality, while reducing\ncosts. We also notice that information technology significantly increases the knowledge of\nthe business environment through monitoring systems, such as the ones used in the product\ndelivery status or in the stock levels of warehouses. IT systems can help develop staff skills\nthrough e-learning and on-the-job on-line assistance.\nNevertheless, there exist several criteria which need to be fulfilled for enterprises to be able\nto differentiate from competitors. Such criteria include among others the brand name visibility,\nthe quick-to-market response, and the service quality [10]. Despite these advantages, there are\nimportant limitations that need to be considered, such as the existence of isolated and\nproprietary software applications and databases. Outdated information (legacy) systems,\nwhich require high-cost adapters to bridge the newer applications, and the use of a wide range\nof different and heterogeneous technologies, are two other examples. The low utilization of\nthe existing IT resources is also a problem. All these limitations significantly reduce the\nefficiency of IT, the process support, the access and dissemination of information, the\nimplementation of new projects, and the staff adaptation to the new conditions, while they\nfurther increase the operating costs.\nThere are solutions to such problems, which enterprises are called upon to adopt, such as\ne.g. the Enterprise Application Integration (EAI), the rapid development of new applications,\nand the development of business process management systems By gradually adopting such\nsolutions an enterprise can in a reasonable time develop a much more integrated business\nenvironment, launch new services in a timely fashion, achieve world-wide access, improve its\nbusiness operations control, acquire resources on demand, work on a flexible infrastructure\nand lower costs [2, 10].\nMost of the times, the enforcement of these IT solutions requires the consideration of a\nvariety of trade-offs, while the main question is: “is there adequate capacity, qualities and\nstrategic direction for a change from the old economy environment to the new one?”. The\nanswer to this question includes: (1) the adoption of an innovation culture, which in turn\nrequires management of change and capability building up for the human resources, and (2)\nan elastic and interoperable infrastructure [11, 12].\nIn addition, the changing of an operational business model can drive the requirements\nfor higher adaptation in the application software portfolio creating another trade-off\namong efficiency, innovation, experimentation, and conformance to the relevant standards\n\n[11–13].\n\n4. The proposed methodological approach\nGiven that IT is still considered to be able to bring about significant structural changes and\nprovide a business competitive advantage in the modern markets along with all the related\noperational benefits, the question is how this goal can be achieved methodologically. This\ndevelopment can be considered to have the following four main dimensions:\n\n(1) Development of a successful IT strategy, whose goals should be synchronized with\nthe business strategy, e.g. there can be no strategy to penetrate global markets, while\nthe company’s information system cannot support the internationalization of the\ninformation it offers and manages.\n\n\n-----\n\n(2) The IT development plan which should consider the organizational structure of\nenterprise, the demanded quality of service to its customers, the firm’s human\nresources and the existing IT infrastructure and systems. Everything may need to be\nchanged partially or to a very large extent.\n\n(3) Development of the IT Infrastructure as a homogenized and distributed service\nplatform provided either to the internal customers of the company or to the external\nones.\n\n(4) Development of services and policies for management, research and innovation,\ntraining for the creation of a technologically high-performance enterprise,\n\nAll these key dimensions of development of innovation and IT solutions are shown in\nFigure 1 as phases.\n\n5. Building up a successful IT strategy\nA business strategy expresses the vision, the mission and the main business goals of a\nenterprise or organization. The business goals must be interpreted into subgoals that are\nenforced on specific high-level business domains, e.g. sales, customer relationships,\nproduction, and logistics. In fact, these domains, according to the balanced scorecard\n(BSC) approach [5], are influenced by the corporate strategy regarding the perspectives of the\nfuture readiness and the innovation, the internal process improvement, the customer\norientation, the cost control, and the financial goals. It is obvious that enterprises need to\neffectively approach these perspectives if they want to achieve a transition from the oldfashion business models to the new ones. Thus, the strategic components incorporated in\nthese perspectives, must be refined and improved.\nThese components, in fact, construct the high-level business domains upon which the\ngoals of the business strategy must be enforced. Of course, these domains must contain\nsubdomains or, in other words, lower-level domains that express the implementation\ncomponents of an enterprise or organization. Such components include among others\ntechnical processes, business data, operations, and informatics implementations. This topdown approach is very useful for the executives because it provides them with a tool for the\nsuccessful enforcement of a business strategy considering all its dimensions.\nThe IT potentiality in terms of new technologies and solutions, makes the IT strategy to be\nthe determinant while the business strategy to be the weak area which needs improvement\n\n[5]. This fact is expressed by the Strategic Alignment (SA) model [14] for the business and IT\nstrategies and the organisational and IT infrastructures.\nFor example, ubiquitous computing brings new capabilities to the enterprises and, thus, it\npushes for a revision of the corresponding business strategies. The goal is to achieve a\n\n\n## Informatics and applied computing\n\nFigure 1.\nThe basic dimensions\nof a high-performance\nenterprise based on\ninnovation and new\ntechnologies\n\n\n-----\n\n## ACI\n\n\ncompetitive advantage over the competition through the development of new products and\nthe appropriate modification of the business scope, the distinct competencies, and the\nbusiness governance, along with the improvement of the organizational infrastructure which\nconcerns the business processes, the human resource capability, and the administrative\nstructure. As mentioned, this approach helps enterprises to conform to the new requirements\nthat arise due to the new technological solutions and the potentialities created by them.\nTowards this direction, the following are proposed:\n\n(1) IT and business strategy alignment,\n\n(2) dimensioning of IT resource requirements,\n\n(3) adoption of new IT architectures and their management, based on innovation and\nadaptability, and\n\n(4) selection of Key Performance Indicators (KPI) for evaluation reasons.\n\nThe first point imposes an IT strategic plan fully aligned with the business strategy. The next\npoint asks for an open, interoperable, scalable, and distributed IT infrastructure, while the\nthird point relates to a high-performance IT-enabled enterprise [4, 5, 15]. Key performance\nmetrics are always required so that the evaluation of the adopted overall approach to be\npossible.\n\n6. Developing an effective IT planning aligned with the business strategy\nIT planning must incorporate the current as well as the future enterprise needs and\ntechnological trends. Thus, it must cover long-term issues as well as whatever is necessary\nfor future readiness. A Strategic Alignment (SA) between the business and IT strategies\naddresses four main areas: business strategy, IT strategy, organizational infrastructure, and\nIT infrastructure [5, 16]. These areas need to be aligned with each other according to the\nbusiness requirements and the type of enterprises. Namely, the requirement for alignment\nvaries from enterprise to enterprise and while it is generally very helpful, it may, however,\nlimit the degrees of freedom in some business cases. In fact, trade-offs and equilibria between\nstrategic alignment and flexibility must be addressed successfully. For example, the\nproduction and other business operations, like the customer relationship management (CRM),\nare positively affected by the Strategic Alignment. On the other hand, the business planning,\nthe marketing, and the sales are less influenced by the SA. Specifically, enterprises, whose\ncritical operations do not focus on IT, do not require such a strict SA. The more supportive\nand functional the role of IT is, the more SA it requires. On the contrary, where IT moves\nwithin a strategic role, the need for a strict SA decreases [17–19]. The following perspectives\ncan be considered in an IT environment [20]:\n\n(1) IT infrastructure: emerging IT infrastructure technologies and solutions call for the\nreformulation of the IT strategy, while the business strategy is implicitly impacted.\n\n(2) IT organization infrastructure: emerging IT calls for the reformulation of the\norganizational infrastructure, while the business strategy is implicitly impacted.\n\n(3) competitive potentiality: emerging IT capabilities, as well as new IT governance\npatterns call for the reformulation of the business strategy, while the organizational\ninfrastructure is implicitly impacted.\n\n(4) service level: the IT resources use, as well as the orientation to the customer calls for\nthe reformulation of the IT infrastructure, while the organizational infrastructure is\nimplicitly impacted.\n\n\n-----\n\n7. Developing a homogenized and distributed corporate IT platform\nThe following services can be offered in an enterprise network in an open and interoperable\nmanner [21, 22]:\n\n(1) electronic services for channel-management, as far as all the involved parties, like\nenterprises, clients, and suppliers, are concerned,\n\n(2) security, that concerns the IT resources protection,\n\n(3) communication mechanisms needed for the communication and interworking\nbetween all the internal/external business entities,\n\n(4) database and file management services, for the purpose of making the required data\nand files available over the enterprise network,\n\n(5) application services, for the purpose of making the required applications, like EPR,\nSCM and HRM, available over the enterprise network, and\n\n(6) management of IT facilities, needed for integration and synchronization of the\ninfrastructure layers and for provision of servers and platforms.\n\nNew trends in IT technologies call for the distribution, homogenization, integration, and\ninteroperability of systems and services [23]. Open IT standards and SOA are among the\ncurrent information systems technologies that enterprises need to move towards because\nthey offer increased interoperability, transfer of application services to heterogeneous\nenvironments, enterprise application integration, service reusability, high operational control\nand monitoring, and flexible service configuration and measurability in service\nperformance [24].\nA SOA helps to create networks of services for the purpose of their common management.\nService orientation here is mentioned as an architectural approach of the new IT systems\nwithout restricting its implementation to specific offered solutions, as it is the Enterprise\nService Bus (ESB). In most cases, service orientation must be followed due to the benefits\nmentioned above. Figure 2 introduces the conceptual approach of a service grid, where\napplication services are interacting each other through several underlying platform services,\nlike routing, message exchange and message queuing, data and knowledge management,\ndistributed API’s, security, accounting, QoS assurance, system monitoring and control.\nIndeed, the adoption of IT service grids can prove to be a key factor for enterprises in their\nefforts to gain a significant competitive advantage. New informatics and applied services\nneed to be distributed, scalable, open, and reliable in a low cost and quick-to-market\n\n**_Services_**\n\n**_Services_** **_Services_**\n\n**_Protocols:_**\n**_Rest, XML/SOAP, WSDL, UDDI_** **_Transportation:_**\n\n**_Routing, Queuing, MOM_**\n\n**_Utilities:_**\n\n**_Services_**\n\n**_Services_** **_Accounting, Security_**\n\n**_Data and Knowledge Management:_**\n\n**_Management:_** **_Directories, Discovery, DBMS, NFS_**\n**_QoS, SLA, Monitoring_**\n\n**_Services_** **_Services_**\n\n**_Services_**\n\n\n## Informatics and applied computing\n\nFigure 2.\nThe conceptual\napproach of SOA and\nthe Service Grid\n\n\n**_Services_**\n\n**_Services_**\n\n**_Protocols:_**\n**_Rest, XML/SOAP, WSDL, UDDI_**\n\n**_Utilities:_**\n\n**_Services_** **_Accounting, Security_**\n\n**_Management:_**\n**_QoS, SLA, Monitoring_**\n\n\n**_Services_** **_Services_**\n\n**_Services_**\n\n\n-----\n\n## ACI\n\nFigure 3.\nThe supply chain\n\n\ndeployment fashion. Furthermore, this adoption will cause channel enhancements, like\ndisintermediation, mitigation of information asymmetry, world-wide access, virtualization,\ncost reduction and control, efficient management information, economies of scale, and\nimproved strategic positioning.\nService grids will be enhanced with the new ubiquitous computing capabilities. For\nexample, mobile computing can effectively leverage the quality of service, as well as the\ncollaboration between employers, employees, customers, strategic partners, and third parties\n\n[25]. Furthermore, smartphones can provide a variety of functionalities, like user interaction,\ntask management, user online help, blogging, wiki, chatting, and remote access from trusteed\nparties or customers through appropriate authentication and authorizations, that can\nsignificantly leverage the operational effectiveness of modern enterprises. The dissemination\nof notifications (e.g. Google Cloud Messages) in mobile apps is another example which can\nfacilitate business operations.\n\n8. Developing a high-performance IT-enabled enterprise\nThe IT services implemented with technologies like SOA and ubiquitous computing, must be\nsupplemented by IT management services. Towards this direction [21, 24], the following\nmanagement services can be identified:\n\n(1) IT administration services, for the platforms, the IT systems planning and the project\nmanagement, the SLAs, and the negotiations with the IT suppliers,\n\n(2) IT architecture services, which need the incorporation of system management policies\nfor the effective management of IT resources,\n\n(3) IT R&D services, concerning new products, services, processes, and operations using\nIT, and\n\n(4) training services in the use of IT, strongly needed for the capacity building of\nenterprise’s staff.\n\nAlong with the management, the overall business processes as well as any activity involved\nin a process need to be designed effectively. It has been well-documented that business\nprocesses which are supported by IT, along with the right organizational and management\nservices, are significantly more efficient than those of the old economy [26, 27]. The supply\nchain is a classic example that proves the truth of the matter. Figure 3 illustrates the process\nof supply chain. In brief, creating a performance-oriented IT organization requires process\ninnovation and efficiency, effective communication with the customers, the partners, and the\nsuppliers, inventory management, development of new products and services, and agile\nadaptation of the existing ones.\nIT-enabled enterprises need to be supported by several key functionalities and\ntechnologies [22, 28, 33, 34]. Bridges between the different information systems via\nappropriate interfaces by using e.g. web services (XML or json-based approaches), are also\nrequired. Of course, the adoption of these technologies may raise critical security and risk\nmanagement issues., All the vulnerabilities and security holes must be eliminated in an way.\n\n\n-----\n\n9. Analysis discussion and evaluation\nHereafter, some examples are provided regarding the right way to employ new technologies\nin modern enterprises. As mentioned in the Accenture report “Improving Business ROI with\nDigital Technologies”[1], “to become a more efficient finance organization, considerable\ninvestments must be made to improve processes and technology”. According to this survey, the\nreasons of fail are that (1) there is no clear strategy and vision, (2) they do not deal effectively\nwith the legacy systems, and (3) they do not understand adequately the existing digital\ncapabilities. The outcomes of this survey are fully aligned with the assumptions of our\nresearch presented above.\nIn the report entitled “The ROI of IoT: The 7 benefits it can bring to your business”[2], it is\nstated that “if you know how your key equipment and assets are behaving, and how people are\ninteracting with them, you can unlock the value of that data through an IoT software program\nthat delivers actionable insights and improves business processes in future”. This is a key issue\nfor the development of an improved long-term planning and strategy, which in turn it is\nnecessary for the success of an IT investment project, according to our research.\nAccording to [32], 88% of companies reported that they already use cloud services, while\n50% of companies expect to have all their data stored in the cloud for a period of two years.\nThe same survey revealed that most enterprises already have applications in production\nbased on cloud at a level of about 47%, while a significant percentage intend to develop such\napplications. However, these enterprises are almost entirely concerned with the security of\ntheir data, which requires a high-performance IT-enabled enterprise accordingto our research.\nFinally, according to the study conducted in [22], by adopting virtualization technology\nfor a computer centre, the cost of investing in it gradually yielded positive profits over the\nannual investment cost over a period of five years. Also, the return on investment (ROI)\nbecame positive after the second year, while during the fourth year, the total investment cost\nbecame significantly lower than the projected profits.\nItshould benotedthat fora continuousevaluation of theITstrategy and its implementation,\nit is necessary to define critical success factors of the organizational and the IT infrastructures\nthat are linked to certain KPIs. These indicators will assess not only the effectiveness of\ninternal processes, customer satisfaction or financial gains, but also the achievement of the\nstructural changes necessary to adapt the organization to an ever-evolving environment.\n\n10. Conclusions and future work\nThis paper presented in a concise and detailed way, the very important role that IT still plays\nnowadays in modern enterprises, giving them a competitive advantage in modern globalized\nmarkets. It was posited that enterprises that adopt IT in a smart and innovative way have a\nsignificant capacity building capability, improved future readiness, and better strategic\npositioning.\nIn the future, we intend to evaluate our approach in an enterprise. The process of analysis\nwill include the business strategy, processes and practices, organizational structures, IT\ninfrastructure, etc. The whole task faces the restrictions of the confidentiality of business\ninformation making access to all this information rather difficult, especially when this\ninformation may be made public. An organization before the application of our methodology\nrequires an adequate mapping (a master plan), while then a road map should be developed\nwith all the technical details of integration of practices, plans, technological solutions,\nevaluations, etc. Evaluations could be based among others in the implementation of a\nbalanced scorecard, where the results will be expressed via Key Performance Indicators\nconcerning crucial success factors (CSF) in all, the internal processes, the quality of services to\nthe customers, and the financial profit which is essentially the main motivation of every\nenterprise.\n\n\n## Informatics and applied computing\n\n\n-----\n\n## ACI\n\n\nNotes\n[1. https://www.accenture.com/nl-en/blogs/insights/how-digital-technologies-improve-business-roi](https://www.accenture.com/nl-en/blogs/insights/how-digital-technologies-improve-business-roi)\n\n(last access: 14/7/2021).\n\n[2. https://blog.worldsensing.com/critical-infrastructure/roi-iot/ (last access 14/7/2021).](https://blog.worldsensing.com/critical-infrastructure/roi-iot/)\n\nReferences\n\n1. Naved K, Tabassum F. Reinventing business organizations: the information culture framework.\nSingapore Management Rev. 2005; 27(2): 37-63.\n\n2. Porter ME, Heppelmann JE. How smart, connected products are transforming competition. Harv\nBusiness Rev Spotlight Managing Internet Things. 2014, November 2014.\n\n3. Teece DJ. A capability theory of the firm: an economics and (strategic) management perspective.\nTaylor & Francis Online. 2017, New Zealand Economic Papers; 53(1) 2019.\n\n4. Tiwana A. IT strategy for non-IT managers. MIT Press. 2017, London, England.\n\n5. Mitropoulos S. An integrated model for formulation, alignment, execution and evaluation of\nbusiness and IT strategies. Int J Business Syst Res. 2021; 15(1): 90.\n\n6. Vandenbosch B, Lyytinen K. Much ado about IT: a response to “the corrosion of IT advantage”\nby Carr’ NG, J Business Strategy. 2004; 25(6): 10-12.\n\n7. Schmidt J, Marques MRG, Botti S, Marques MAL. 2019. Recent advances and applications of\n[machine learning in solid-state materials science. Available at: https://www.nature.com/articles/](https://www.nature.com/articles/s41524-019-0221-0)\n[s41524-019-0221-0.](https://www.nature.com/articles/s41524-019-0221-0)\n\n8. Benlamri R, Sparer M. Leadership, innovation and Entrepreneurship as driving forces of the global\neconomy. Proceedings of the 2016 ICLIE, Springer Proceedings in Business and Economics. 2016.\n\n9. Koi-Akrofi GW. Justification for IT investments: evaluation methods, frameworks, and models.\nTexila Int J Management. 2017; 3(2).\n\n10. Melarkode A, From-Poulsen M and Warnakulasuriya S. Delivery agility through IT. Business\nStrategy Rev. 2004, Autumn 2004.\n\n11. Wadhwa M, Harper A. Technology, innovation and enterprise transformation. IGI Glob book Ser\nAdv Business Inf Syst Analytics (Abisa). 2015, Hershey, USA.\n\n12. Schrage M, Kiron D, Hancock B, Breschi R. Performance management’s digital shift. MIT Sloan\n[Management Rev. February 26, 2019. Available at: https://sloanreview.mit.edu/projects/](https://sloanreview.mit.edu/projects/performance-managements-digital-shift/)\n[performance-managements-digital-shift/.](https://sloanreview.mit.edu/projects/performance-managements-digital-shift/)\n\n13. Dong J and Yang CH. Business value of big data analytics: a systems-theoretic approach and\nempirical test. Inf Management. 2020; 57(1): 103124.\n\n14. Ilmudeen, A, Bao, Y, Alharbi, IM. How does strategic alignment affect firm performance? the roles\nof information technology investment and environmental uncertainty, J Enterprise Inf\nManagement. 2019; 32(3): 457-76.\n\n15. Davies P. Strategic objectives and principles. 2015. Version V1.0, University of Sussex, Date of\n[Issue: 9th April 2015 Available at: http://www.sussex.ac.uk/its/pdfs/IT_Strategy_2015.pdf.](http://www.sussex.ac.uk/its/pdfs/IT_Strategy_2015.pdf)\n\n16. Chugunov A, Misnikov Y, Roshchin E, Trutnev D. Electronic governance and open society:\nchallenges in Eurasia. 5th International Conference. 2018, EGOSE 2018.\n\n17. Tallon, P. How information technology infrastructure flexibility shapes strategic alignment: a\ncase study investigation with implications for strategic IS planning. Plann Inf Syst. 2015; 15,\nAcemap: 425-55.\n\n18. Tallon P, Kraemer K. Investigating the relationship between strategic alignment and IT business\nvalue: the discovery of a paradox, In: Shin N (Eds), Creating business value with IT: challenges\nand solutions, Hershey, Pennsylvania, PA: Idea Group Publishing, 2003.\n\n19. Kaleka A, Morgan NA. How marketing capabilities and current performance drive strategic\nintentions in international markets. Ind Marketing Management. 2017; 78: 108-121.\n\n\n-----\n\n20. Coleman P, Papp R. Strategic alignment: analysis of perspectives. Proceedings of the 2006\nSouthern Association for Information Systems Conference, March 11–12, 2006, Florida, FL. 2006.\n\n21. Weill P, Subramani M, Broadbent M. Building IT infrastructure for strategic agility. MIT Sloan\nManagement Rev. 2002, Fall 2002.\n\n22. Lambropoulos G., Mitropoulos S., Douligeris C. Improving business performance by employing\nvirtualization technology: a case study in the financial sector. Computers. 2021; 10(4): 52.\n\n23. Razis M., Mitropoulos S. An integrated approach for the banking intranet/extranet information\nsystems: the interoperability case. Publ Int J Business Syst Res. 2021, Inderscience Publishers.\nforthcoming paper.\n\n24. Katsikogiannis G., Kallergis D., Garofalaki Z., Mitropoulos S., Douligeris C. A policy-aware service\noriented architecture for secure machine-to-machine communications. J Ad Hoc Networks. 2018;\n80, November 2018, Elsevier Science Publishers.\n\n25. Chassapis P, Mitropoulos S, Douligeris C. A prototype mobile application for the athens\nnumismatic museum. J Appl Comput Inform. 2020; ahead-of-print(ahead-of-print), Emerald\nPublishers.\n\n26. Mitropoulos S. A simulation-based approach for IT and business strategy alignment and\nevaluation. Int J Business Inf Syst. 2012; 10(4): 369-396, Inderscience Publishers.\n\n27. Mitropoulos S, Giannakos K, Achlioptas J, Douligeris C. A prototype workflow MIS for supply\nchain management: architecture, implementation and business evaluation. Publ Int J Business Inf\nSyst. 2021, Inderscience Publishers. forthcoming paper.\n\n28. Mitropoulos S, Mitsis C, Valacheas P, Douligeris C. An online Emergency medical management\ninformation system using mobile computing. J Appl Comput Inform. 2020. online version,\nEmerald Publishers.\n\n29. Sayabek, Z, Suieubayeva, S, Utegenova, A. Digital transformation in business, ISCDTE 2019:\ndigital age: chances, challenges and future lecture notes in networks and systems. 2020; 84:\n408-415, Springer, Cham.\n\n30. Gimpel H, R€oglinger M. Digital Transformation Changes and chances- Insights based on an\n[empirical study. 2015. Available at: https://fim-rc.de/wp-content/uploads/2020/02/Fraunhofer-](https://fim-rc.de/wp-content/uploads/2020/02/Fraunhofer-Studie_Digitale-Transformation.pdf)\n[Studie_Digitale-Transformation.pdf.](https://fim-rc.de/wp-content/uploads/2020/02/Fraunhofer-Studie_Digitale-Transformation.pdf)\n\n31. IDC. Exploring the Impact of infrastructure virtualization on digital transformation strategies and\n[carbon emissions, white paper. 2019, Available at: https://www.vmware.com/content/dam/](https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/company/vmware-exploring-impact-of-infrastructure-virtualization-on-digital-transformation-strategies-and-carbon-emissions-whitepaper.pdf)\n[digitalmarketing/vmware/en/pdf/company/vmware-exploring-impact-of-infrastructure-](https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/company/vmware-exploring-impact-of-infrastructure-virtualization-on-digital-transformation-strategies-and-carbon-emissions-whitepaper.pdf)\n[virtualization-on-digital-transformation-strategies-and-carbon-emissions-whitepaper.pdf.](https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/company/vmware-exploring-impact-of-infrastructure-virtualization-on-digital-transformation-strategies-and-carbon-emissions-whitepaper.pdf)\n\n[32. Oracle KPMG. Oracle and KPMG cloud threat report. 2020, Available at: https://www.oracle.com/](https://www.oracle.com/a/ocom/docs/cloud/oracle-cloud-threat-report-2020.pdf)\n[a/ocom/docs/cloud/oracle-cloud-threat-report-2020.pdf, [Accessed 12 June 2021].](https://www.oracle.com/a/ocom/docs/cloud/oracle-cloud-threat-report-2020.pdf)\n\n33. Huang M, et al.. An effective service-oriented networking management architecture for 5Genabled internet of things. Compuert Networks. 2020; 173: 107208.\n\n34. Niu Y, et al. Exploiting device-to-device communications in joint scheduling of access and\nbackhaul for mmWave small cells. IEEE JSAC. 2015; 33(10): 2052-2069.\n\n35. Ahmad Qadri Y, et al.. The future of healthcare internet of things: a survey of emerging\ntechnologies. IEEE Commun Surv Tutorials. 2020; 22(2): 1121-1167.\n\n36. Bera B, et al.. Designing blockchain-based access control protocol in IoT-enabled smart-grid\nsystem. IEEE Internet Things J. 2021; 8(7): 5744-5761.\n\nCorresponding author\n[Sarandis Mitropoulos can be contacted at: smitropoulos@ionio.gr](mailto:smitropoulos@ionio.gr)\n\nFor instructions on how to order reprints of this article, please visit our website:\nwww.emeraldgrouppublishing.com/licensing/reprints.htm\nOr contact us for further details: permissions@emeraldinsight.com\n\n\n## Informatics and applied computing\n\n\n-----\n\n"
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1108/aci-06-2021-0149?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1108/aci-06-2021-0149, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.emerald.com/insight/content/doi/10.1108/ACI-06-2021-0149/full/pdf?title=why-and-how-informatics-and-applied-computing-can-still-create-structural-changes-and-competitive-advantage" }
2,021
[]
true
2021-10-01T00:00:00
[ { "paperId": "f5d9cdc4e20b80e455d1016d797f626788edbc07", "title": "An online emergency medical management information system using mobile computing" }, { "paperId": "943df55f3385793fbbee9ca23597baa1cde334a2", "title": "Improving Business Performance by Employing Virtualization Technology: A Case Study in the Financial Sector" }, { "paperId": "3456d7b76609193aec77ed861b6c40c4736763d8", "title": "Designing Blockchain-Based Access Control Protocol in IoT-Enabled Smart-Grid System" }, { "paperId": "67023ff6f414784305e0ddf5aa7af8e2f167a5ab", "title": "A prototype mobile application for the Athens Numismatic Museum" }, { "paperId": "d6ea75b325737824d9232f95a227ad2d04fe9c6a", "title": "An effective service-oriented networking management architecture for 5G-enabled internet of things" }, { "paperId": "efcc84bc5542453001483bc9c81c7d4c6c409b80", "title": "The Future of Healthcare Internet of Things: A Survey of Emerging Technologies" }, { "paperId": "b0aea93214fc501217f5091e8886520ff6a8b794", "title": "Business value of big data analytics: A systems-theoretic approach and empirical test" }, { "paperId": "0273507eb05f1135f3a05f9c7adc9a56f12c7c5c", "title": "Recent advances and applications of machine learning in solid-state materials science" }, { "paperId": "3150c9cde9ba3bc220c87971e59fd8bb5979318f", "title": "How does business-IT strategic alignment dimension impact on organizational performance measures" }, { "paperId": "5b9d8e9a622dad3f0550b51ac1641295cc4497b2", "title": "A capability theory of the firm: an economics and (Strategic) management perspective" }, { "paperId": "758153bebcd9e4e90cac7cb1166d5cf8f303d842", "title": "How Does Strategic Alignment Affect Firm Performance? The Roles of Information Technology Investment and Environmental Uncertainty" }, { "paperId": "35b5c808a8e82dba941252d7447acf3679edf5c6", "title": "A policy-aware Service Oriented Architecture for secure machine-to-machine communications" }, { "paperId": "4c9885c4a071b864732ef7f89b89bd8d130bd516", "title": "Justification for IT Investments: Evaluation Methods, Frameworks, and Models" }, { "paperId": "22cfeafc9fd60144cc8e74fbab10c3a6701bd198", "title": "How marketing capabilities and current performance drive strategic intentions in international markets" }, { "paperId": "06765dd985e923bb50787f2679af61e45c5a2f88", "title": "How Information Technology Infrastructure Flexibility Shapes Strategic Alignment: A Case Study Investigation with Implications for Strategic IS Planning" }, { "paperId": "76116b7fa6ace076fc82a9b0508a990ad6624070", "title": "Exploiting Device-to-Device Communications in Joint Scheduling of Access and Backhaul for mmWave Small Cells" }, { "paperId": "8119a80c6059bfda198e1f6e5b52cf7351b0962d", "title": "How Smart, Connected Products Are Transforming Competition" }, { "paperId": "4d632c0828decf67cba6b58a05ed305a65dd4b54", "title": "Technology, Innovation, and Enterprise Transformation" }, { "paperId": "a8be48fa449bb6db40ac7ea062d433a0add5681f", "title": "A simulation-based approach for IT and business strategy alignment and evaluation" }, { "paperId": "194edbd38b15ee3b1c4d72d98ce109ff966385fe", "title": "Much ado about IT: a response to “the corrosion of IT advantage” by Nicholas G. Carr" }, { "paperId": "959a535954985bcc2ec55b9f76ec9cbbab47fab4", "title": "Building IT Infrastructure for Strategic Agility" }, { "paperId": "c3d9ef3c9c664559bb7e7bb63f6beefd3a43e86b", "title": "A prototype workflow MIS for supply chain management: architecture, implementation and business evaluation" }, { "paperId": "605a7299c12396b64ec2b5bb1b0f0cb175a5025c", "title": "An integrated approach for the banking intranet/extranet information systems: the interoperability case" }, { "paperId": "044750a44d4accc4724beaeb73eb2c4c1e133086", "title": "An integrated model for formulation, alignment, execution and evaluation of business and IT strategies" }, { "paperId": "35cbc9b8d5250b99416f80b764c5ac2eb17f9d93", "title": "Electronic Governance and Open Society: Challenges in Eurasia" }, { "paperId": "6fcba108774e4023a43649930d55c612780e8730", "title": "IT Strategy for Non-IT Managers: Becoming an Engaged Contributor to Corporate IT Decisions" }, { "paperId": "b7d6826b49c6f152dd3c223e13216e377dd55bc2", "title": "Leadership, Innovation and Entrepreneurship as Driving Forces of the Global Economy" }, { "paperId": "e8f3656f673e7edaf1b6ada2366df07710c5f801", "title": "Digital Transformation : Changes and Chances – Insights based on an Empirical Study" }, { "paperId": "06865e2dd2ee73aeacf83e3daceb30b86a0550a4", "title": "STRATEGIC ALIGNMENT: ANALYSIS OF PERSPECTIVES" }, { "paperId": "a8dd24d8f87cc7d3d8d69542be9bed166c85c031", "title": "Investigating the Relationship between Strategic Alignment and Information Technology Business Value: The Discovery of a Paradox" }, { "paperId": "4df9d9574ebd255ca8cf1fb8cc1d78adab605618", "title": "Investigating the relationship between strategic alignment and IT business value: the discovery of a paradox" }, { "paperId": null, "title": "Strategic objectives and principles" }, { "paperId": null, "title": "Digital transformation in business, ISCDTE 2019: digital age: chances, challenges and future lecture notes in networks and systems" }, { "paperId": null, "title": "Performance management's digital shift. MIT Sloan Management Rev" }, { "paperId": null, "title": "Reinventing business organizations: the information culture framework" }, { "paperId": null, "title": "Exploring the Impact of infrastructure virtualization on digital transformation strategies and carbon emissions" }, { "paperId": null, "title": "Delivery agility through IT" }, { "paperId": null, "title": "Oracle and KPMG cloud threat report" } ]
9,334
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01079c5c50c9b17fe9c05eccc3764f25df77393c
[ "Computer Science" ]
0.906166
A Link-Layer Virtual Networking Solution for Cloud-Native Network Function Virtualisation Ecosystems: L2S-M
01079c5c50c9b17fe9c05eccc3764f25df77393c
Future Internet
[ { "authorId": "2072871840", "name": "Luis F. Gonzalez" }, { "authorId": "143917913", "name": "I. Vidal" }, { "authorId": "145764713", "name": "F. Valera" }, { "authorId": "2232383964", "name": "Raul Martin" }, { "authorId": "2232320867", "name": "Dulce Artalejo" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-156830", "https://www.mdpi.com/journal/futureinternet" ], "id": "c3e5f1c8-9ba7-47e5-acde-53063a69d483", "issn": "1999-5903", "name": "Future Internet", "type": "journal", "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-156830" }
Microservices have become promising candidates for the deployment of network and vertical functions in the fifth generation of mobile networks. However, microservice platforms like Kubernetes use a flat networking approach towards the connectivity of virtualised workloads, which prevents the deployment of network functions on isolated network segments (for example, the components of an IP Telephony system or a content distribution network). This paper presents L2S-M, a solution that enables the connectivity of Kubernetes microservices over isolated link-layer virtual networks, regardless of the compute nodes where workloads are actually deployed. L2S-M uses software-defined networking (SDN) to fulfil this purpose. Furthermore, the L2S-M design is flexible to support the connectivity of Kubernetes workloads across different Kubernetes clusters. We validate the functional behaviour of our solution in a moderately complex Smart Campus scenario, where L2S-M is used to deploy a content distribution network, showing its potential for the deployment of network services in distributed and heterogeneous environments.
## future internet _Article_ # A Link-Layer Virtual Networking Solution for Cloud-Native Network Function Virtualisation Ecosystems: L2S-M **Luis F. Gonzalez *** **, Ivan Vidal *** **, Francisco Valera** **, Raul Martin and Dulce Artalejo** Telematic Engineering Department, Universidad Carlos III de Madrid, Avda. Universidad, 30, 28911 Leganés, Spain; fvalera@it.uc3m.es (F.V.); 100384060@alumnos.uc3m.es (R.M.); 100384053@alumnos.uc3m.es (D.A.) *** Correspondence: luisfgon@it.uc3m.es (L.F.G.); ividal@it.uc3m.es (I.V.)** **Abstract:** Microservices have become promising candidates for the deployment of network and vertical functions in the fifth generation of mobile networks. However, microservice platforms like Kubernetes use a flat networking approach towards the connectivity of virtualised workloads, which prevents the deployment of network functions on isolated network segments (for example, the components of an IP Telephony system or a content distribution network). This paper presents L2S-M, a solution that enables the connectivity of Kubernetes microservices over isolated link-layer virtual networks, regardless of the compute nodes where workloads are actually deployed. L2S-M uses software-defined networking (SDN) to fulfil this purpose. Furthermore, the L2S-M design is flexible to support the connectivity of Kubernetes workloads across different Kubernetes clusters. We validate the functional behaviour of our solution in a moderately complex Smart Campus scenario, where L2S-M is used to deploy a content distribution network, showing its potential for the deployment of network services in distributed and heterogeneous environments. **Keywords: microservices; cloud computing; virtual networks** **1. Introduction** **Citation: Gonzalez, L.F.; Vidal, I.;** Valera, F.; Martin, R.; Artalejo, D. A Link-Layer Virtual Networking Solution for Cloud-Native Network Function Virtualisation Ecosystems: L2S-M. Future Internet 2023, 15, 274. [https://doi.org/10.3390/fi15080274](https://doi.org/10.3390/fi15080274) Academic Editor: Izzat Alsmadi Received: 14 July 2023 Revised: 7 August 2023 Accepted: 15 August 2023 Published: 17 August 2023 **Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). In the last couple of years, the continuous development of the Internet has led to an unprecedented increase in the demand for telecommunication services from users. This increase has brought new challenges for operators and service providers, which have been forced to adapt new models and disruptive paradigms to accommodate the everincreasing demand. These challenges include, among others, reducing development cycles and decreasing the speed needed to launch new services to the market; supporting their continuous update transparently to users to satisfy the constant demand for innovation; supporting the scalable operation of services by taking into account a high potential number of users, where an appropriate quality of experience must be delivered; etc. In response to these challenges, cloud technologies, particularly the cloud-native model, have received great interest from the involved actors in the provision of Internet services. According to the Cloud Native Computing Foundation (CNCF) [1], “Cloud native _technologies empower organisations to build and run scalable applications in modern, dynamic_ _environments such as public, private, and hybrid clouds”. This model favours an application_ design based on microservice architectures [2], in contrast to traditional approaches based on monolithic designs. By following a microservices approach, an application is developed as a set of small services communicated through simple, well-defined mechanisms, for example, based on HTTP. This model allows to overcome the inherent limitations of the monolithic design, where every application is developed as a single indivisible block, which requires more coordination between development teams, introduces further complexity in the updating processes, and does not allow to independently scale parts of an application. In a cloud-native model, microservices can be executed in virtualisation containers [3]. This allows a high degree of flexibility when deploying an application since containers are ----- _Future Internet 2023, 15, 274_ 2 of 28 lightweight in comparison with traditional virtualisation platforms based on hypervisors. Containers can be exported to other virtualisation platforms with enough computational, networking and storage capacity to run them. Furthermore, they are able to pack the required software to execute every microservice independently from the rest since their own containers have all the software needed to run the service without the need for emulating an entire operating system. In the same fashion, containers offer a scalable solution that allows to flexibly adapt a service to its demand. For example, if an application experiences a sharp increase in traffic, new containers can be quickly deployed to provision for this demand, minimising service cut-offs. Given the rise in the popularity of container technology, there are several platforms that allow the management and orchestration of containers, both open source, such as Kubernetes (K8s) [4], Docker Swarm [5] or OpenStack [6], as well as solutions offered by cloud providers, like Google Kubernetes Engine [7] or Amazon Elastic Kubernetes Service [8]. It is worth mentioning that K8s has become the most popular tool in the service cloud-native market. According to the 2021 survey of CNCF [9], performed over the global cloud-native community (including software and technology, financial services, consultancy and telecommunication organisations), 96% of the surveyed reported the use of K8s in their organisation. The rise of the cloud-native model has provided multiple benefits for the deployment of Internet services. However, microservice technologies have also been regarded as excellent candidates for the deployment of network functions in cloud-native environments for the next generation of mobile networks (5G and 6G). The network function virtualisation (NFV) paradigm has greatly assisted in the agile deployment and development of network services (NSes) in both cloud and edge environments. NFV aims at the softwarisation of network services and functionalities through the use of virtualisation technologies, such as virtual machines (VMs), reducing the deployment and development costs since it is not necessary to develop and maintain the dedicated hardware involved in the provision of some network functions. Naturally, containers have been regarded as the next step for the deployment of NSes under the NFV umbrella since their lightweight nature and easier management can enhance the provision of NSes in comparison with more computationally demanding solutions. Naturally, the provision of network services involve certain degree of orchestration between containers to ensure its proper functionality. In this regard, microservice platforms like K8s are excellent candidates for this purpose, thanks to their orchestration and abstraction tools (automated rollouts and rollbacks, self-healing properties, service discovery . . .) that assist in the swift and efficient deployment of NSes in data centre environments, allowing microservices to properly interact with each other to offer a complex application. Some examples of NSes that could be deployed using NFV technologies include load balancing, service discovery and routing functionalities. All these services must be connected to one or several virtual networks able to isolate each VNF in different local area network (LAN) domains. This behaviour provides these functions with a finer control of their networking aspects within the platform where they are deployed, regardless of their location. These virtual networks are of the utmost importance for any NFV deployments since they enable isolation between different VNF instances at the network level to operate independently and securely, thus allowing to implement the VNF chaining necessary for the deployment of complex NSes. Virtual networks are necessary for the deployment of NSes in cloud-native environments. However, microservice platforms, like K8s, usually take a hands-off approach towards the connectivity of container networking: the flat networking approach. In this model, all microservices are visible to each other at the network (IP) layer through the use of networking agents deployed over a platform. These communications are especially useful when the applications and workloads must communicate through application-based communications (i.e., using APIs to communicate with one another) since these mechanisms dissociate the network configuration from the application itself. This approach in turn can provide high availability and higher resilience to failures, and its implementation ----- _Future Internet 2023, 15, 274_ 3 of 28 can benefit microservice-based applications in multiple ways (high availability, automatic service discovery, etc.). Unfortunately, there is an important downside that could limit the deployment of NSes in microservice platforms, as this flat networking approach prevents the creation and management of virtual networks, which are necessary to interconnect all the VNFs that compose a NS. Since all microservices are able to “see” each other at the network layer, there is no isolation between them. In consequence, the VNF chaining needed cannot be performed. Therefore, it is impossible to effectively deploy NSes in microservice platforms that only implement a flat networking approach towards their connectivity, as they lack the necessary tools to create and manage the required virtual networks used in NFV deployments. In order to address this limitation, this paper presents a networking solution that enables the link-layer connectivity of microservice platforms using software-defined networking (SDN) technology. More concretely, link-layer secure connectivity for microservice platforms (L2S-M) provides a programmable data plane that virtualised workloads can use to connect among each other at the link-layer level, enabling the establishment of point-topoint or multi-point links between workloads in a microservice environment. Furthermore, this paper also explores the potential of L2S-M to provide link-layer communications between workloads located in different clusters if they are managed by a microservice platform, or sites, which can be managed by other virtualisation platforms based on VMs. As a validation use case example, this paper presents a Smart Campus scenario, where L2SM is deployed to communicate different campuses located in geographically distributed scenarios and managed using distinct virtualisation technologies/orchestration functions, implementing a content delivery network (CDN) service to provide multimedia content in a university environment. **2. Background** The world of telecommunications and services has recently experienced an unprecedented demand for more efficient, resilient and robust applications able to support the ever-increasing demand of consumers, thanks to the new services and solutions that have flourished under the umbrella of the 5th Generation of mobile networks (5G). In this regard, the traditional monolithic design used in the development of telecommunication services falls short in many aspects. Under this model, the functionality of a complex application is wrapped as a single program or process, usually running inside a single host able to have all the necessary resources to execute its modules and functionalities. This model has significant drawbacks that come with this design architecture: higher complexity, less adaptability (a change in one module can have effects in the entirety of the code), lower scalability and long-term support [10]. To combat all these challenges that prevent the effective development of new telecommunication applications and services, microservices architectures have rose as a key enabler towards the development and deployment of scalable, resilient and cost-efficient applications. In this model, each application is split into individual modules that can be distributed among several hosts and architectures. In order to build a complex functionality, each module is able to communicate with each other regardless of their physical location, and they can operate independently of the rest of the services, as it will usually perform a single task [11]. This model assists with all of the problems that come with monolithic applications since complexity can be alleviated by focusing on each module (i.e., not having to modify an entire program), and scalability is increased through the deployment of multiple copies of each module used in the provision of the service (as well as being distributed in multiple architectures). Precisely due to this paradigm shift in application development, there is a conscious effort in applying novel virtualisation technologies to enable this transition since traditional virtualisation technologies (i.e., hypervisor-based solutions) use more resources than container-based technologies [3]. Virtual machines rely on hypervisors, which are software programs that operate at the hardware level within a host: its purpose is the emulation ----- _Future Internet 2023, 15, 274_ 4 of 28 of an entire operating system isolated from the host that they are running, including its kernel. Containers, on the other hand, use the core operating system in the host to execute some of their functionalities, while having a file system isolated from the host [12]. Since containers require far fewer resources than a virtual machine, a single host is able to deploy a wider array of functionalities and applications, which in turn provides higher scalability, resiliency and efficiency in comparison with virtual machines [3,13,14]. This performance improvement makes containers the perfect candidate for the implementation of distributed architectures for the implementation and deployment of applications within the 5th generation of networks. With this idea in mind, there have been conscious efforts to apply container technology for the deployment of network functions and verticals, for instance, in [14], where the authors define an application/service development kit (SDK) for the deployment of NFV-based applications for both hypervisor-based and containerised approaches. In [15], the authors build a lightweight docker-based architecture for virtualising network functions to provide IoT security (although VNFs were connected using standard Docker network capabilities). One fundamental aspect of container-based environments is guaranteeing that these containers are provided with a functional network interface, which enables the communication between microservices and other elements or devices outside (for example, devices connected to the Internet). In this regard, CNCF provides the container network interface (CNI) solution [16], a specification and a set of libraries to provide a reference framework to develop plugins that allow the configuration of network interfaces for containers. Currently, this reference framework has been adopted by multiple container management solutions, such as the case of K8s, which supports multiple plugins for a wide arrange of networking plugins for microservice platforms: Flannel [17], Calico [18] or Multus [19] are some of the most widely known (and utilised) in K8s. In consequence, the CNI model eases the addition of network interfaces in containers, allowing the connectivity between containers through their respective network interfaces. This connectivity model is clearly appropriate for microservice-based applications, where all of them must be able to communicate between each other (due to their intrinsic design, described in previous paragraphs). Naturally, microservice platforms tend to use this CNI solutions to define how the different microservices interact between each other. In this regard, K8s explicitly state that every pod, the minimal computation unit that can be deployed in K8s composed of one or more containers (sharing a common network namespace), must have its own IP address, so it is not necessary to explicitly build links between pods. Moreover, K8s imposes the restriction that all pods deployed in any networking implementation must be able to communicate with all other pods on any other node without the use of network address translation (NAT) functionalities. Regarding the logical interactivity of the services running within a microservice platform, although CNI plugins enable the network connectivity of the containers that are deployed over a cluster (as it was explained beforehand), they do not define how each of the microservices interact with each other using these communications, i.e., they do not implement the networking logic that a module (or set of modules) must have in order to provide the functionality of a complex application, or its relation with each module. For this reason, microservice platforms usually rely on the service mesh concept to connect different services (networking abstractions that can integrate one or several microservices, usually of the same type) and define how each microservice communicates within an infrastructure. Service mesh solutions like istio [20] and Envoy [21] use proxy functionalities that build a data plane between the microservices over a cluster, and use these functions to filter traffic and apply policies to the data sent to/from the proxies. This service mesh has its limitations, however, as it does not provide isolation between microservices since it only modifies routing information in a host, so the flat network implemented with the CNIs is still present regardless (i.e., all microservices can still see each other at the network (IP) level). ----- _Future Internet 2023, 15, 274_ 5 of 28 However, although this connectivity model is appropriate for applications, it is important to mention that it presents some limitations when deploying NFV services. In the NFV paradigm, services are deployed as sets of VNFs interconnected through virtual networks. These virtual networks provide the abstraction of point-to-point or multi-access links to the VNFs: they allow two or more VNFs to effectively connect to a single link-layer network, sharing a single broadcast domain, where all connected VNFs can be seen as ‘neighbours’ at a single IP hop distance. Furthermore, the traffic transmitted over a virtual network is not accessible to VNFs and entities outside the virtual network. Under the previous considerations, the connectivity model between cloud-native platforms presents some difficulties to support the abstraction offered by virtual networks commonly used in NFV ecosystems. Despite these limitations, the integration of cloud-native technologies in the NFV ecosystem can present important advantages. Among others, we can mention the following: (a) the use of lightweight containers, portable and scalable, and the use of continuous integration and continuous deployment (CI/CD) methodologies, offering a solution to tackle the development and deployment of NFV services in microservice platforms; (b) the immense popularity of the cloud-native model and its adoption state, both in development and production environments, opens up new opportunities to the incorporation of developers, manufacturers and cloud service providers into the NFV market, which will positively impact the innovation process and the flexibility of options for service deployment—additionally, the access to a vast catalogue of virtual functions, developed through the cloud-native model, would be enabled for the provision of NFV services of aggregated value; and (c) the current initiatives to translate cloud-native technologies to edge environments, like the cases of KubeEdge [22], OpenYurt [23] or K3s [24], centred around the use of K8s. These initiatives represent a promising alternative to have potentially limitless computation, storage and networking resources for the automatic deployment of operator services and verticals in the future. Even though the flat networking approach has been the de facto standard in microservice platforms, some CNI solutions have tried to go beyond this model to provide networking functionalities similar to the ones implemented in other virtual infrastructure manager (VIM) solutions. For example, the OpenShift SDN network plugin [25] allows the isolation of pods at the project level and the application of traffic policies, which can help with isolating workloads in a cluster (although they will still see each other at the network level, and it is only available for OpenShift clusters). The Nodus network controller [26] enables the creation of subnetworks that pods can use to enable their connectivity in a K8s cluster. However, this subnetting is limited since the subnetworks are all located in the same IP range, so pods are not completely isolated within a K8s cluster. The Kube-OVN CNI plugin [27] implements an OVN-based network virtualisation with K8s, which allows the creation of virtual subnetworks (not dependent on a single IP address range, contrary to the Nodus solution) that pods are able to attach to. This solution has its limitations however, as it does not allow the implementation of traffic engineering policies to route traffic between pods, it is not compatible with physical networking devices in a K8s clusters, and its inter-cluster capabilities is limited (it can only connect workloads at the network level). The open-source software community is taking the first steps towards the evolution of networking models in cloud-native technologies, as well as adapting them to the NFV ecosystem. As an example, the ETSI Open Source MANO (OSM) project [28], whose main objective is developing a management and orchestration platform for NFV environments in accordance to the ETSI standards, has supported the deployment of virtualised functions in K8s clusters since its SEVEN release. Nevertheless, this support is limited since it does not enable the creation of virtual links (VLs) to enable the isolated connectivity of different Kubernetes-based VNFs (KNFs) in a NS, as K8s does not natively provide a networking solution for creating virtual networks (i.e., OSM only deploys KNFs but does not define their connectivity, and all KNFs can communicate with each other). This is an important step towards the integration of NFV in microservice platforms since the management of ----- _Future Internet 2023, 15, 274_ 6 of 28 VLs (usually represented as virtual networks) is a fundamental aspect for the effective deployment of NSes, as seen in works like [29], where the authors perform a comprehensive analysis of NFV systems and deployments, or [30], where the authors explain in detail the use of virtual networking and its performance impact depending on the virtual link types. In this regard, the interest of the research community in closing the gap between NFV and microservice platforms can be seen in such works as [31], where the authors propose an architecture based on the monitoring, analysis, planning, and execution (MAPE) paradigm for the optimisation of network service performance to guarantee quality of service (QoS) in NFV environments, including container-based solutions, such as Docker and Kubernetes. Similarly, another example of this effort can be seen in the work [32], where authors propose an implementation of a novel framework that enables the optimal deployment of contained-based VNFs, using a multi-layer approach based on monitoring and analysing data from the physical, virtual and service layers. There have also been proposals to enhance the networking of microservice platforms in the NFV context. One prominent example is the network service mesh (NSM) [33]. NSM offers a set of application programming interfaces (APIs) that allow the definition of network services (for example, an IP router, a firewall or a VPN tunnel), and establishes virtual point-to-point links between pods that want to access determined network services and pods that implement such services. NSM is designed to provide its connectivity service in scenarios with multiple clusters or cloud infrastructures, keeping the CNI solution used in every cluster. NSM presents a promising approach for exploiting the potential of cloudnative technologies in an NFV ecosystem. In such an ecosystem, NSM would provide the abstraction of a network service. For example, if a VNF offers an IP routing service, NSM would allow the establishment of virtual point-to-point links among this VNF and the remaining VNFs that must have IP connectivity with the former, depending on the NFV service to be deployed. However, this connectivity service does not provide the versatility of a virtual network. On the one hand, NSM does not allow to connect multiple VNFs into a single link-layer network in such a way that they can share a single broadcast domain (i.e., NSM does not offer the multi-access link abstraction). This aspect can be a limiting factor to deploy telecommunication or verticals services in an NFV ecosystem. On the other hand, the NSM APIs do not offer an open interface that allows the cloud infrastructure administrator to flexibly manage the existing virtual links. Following the previous example, it could be desirable to change the configuration of a point-to-point link in such a way that it terminates in another IP router instance in order to support load balancing; mirroring port configurations could be required to monitor data traffic transmitted over a link; or the temporary shutdown of certain links and their subsequent activation could be needed when managing a security incident; etc. **3. Virtual Networking for Microservice Platforms: L2S-M** _3.1. Problem Statement_ Due to the intrinsic nature of cloud-native environments, applications are developed and deployed with the following principles of architectural patterns: scalability since applications should be able to increase or decrease their performance based on demand and cost effectiveness; elasticity, as applications should be able to dynamically adapt their workloads to react against the possible changes in the infrastructure; and agility, as applications should be deployed as fast as possible to minimise service down times. In principle, an application should be able to be deployed over a distributed cloud in such a way that it can provide its service without any, or minimal, disruptions, while also adapting its performance (and distribution) to the demand and available resources in an infrastructure [2]. One prominent case that focuses on this philosophy is microservice platforms: each function is composed of several modules (usually implemented as separate containers, as described in previous sections) that interact with each other to execute a complex functionality for one (or sometimes several) applications. These modules might not be deployed in the same infrastructure, physical equipment or virtual machine: it is ----- _Future Internet 2023, 15, 274_ 7 of 28 usual that they are distributed across multiple geographical locations, or even spread across multiple clouds managed by different service providers. This model is an antithesis of the monolithic design (i.e., a single application with all its components embedded in its code), and it provides several advantages over its counterpart: high availability since a module might have several copies distributed over the cloud; resilience since a failure in one module does not compromise the entire functionality of the application; and shorter development and deployment cycles. Taking into account the main characteristics of cloud-native environments, it is essential to define the networking between applications and modules to allow for seamless connectivity without compromising any of the benefits of the microservice model. In order to preserve such advantages, most solutions rely on a flat networking approach that facilitates communication among deployed functions and modules across different clouds, independently of the physical location and network configurations [17,18]. In this model, each microservice is able to reach at the IP level all of the containers that are deployed within an infrastructure. By using this approach, high-level APIs, such as RESTful APIs, can be implemented to enable the exchange information between functions (since all containers are connected between each other, their proper communication can only be guaranteed through application-layer mechanisms). Naturally, in order to effectively implement a complex application, it is essential that networking solution enables external connectivity for some of its modules (containers). As it usually is not possible to directly send data to a particular container from outside a microservice platforms, platforms like K8s rely on networking abstractions to enable its connectivity from the outside. In this regard, K8s relies on K8s Services to expose its own pods to the exterior. For this purpose, K8s uses a Service API to define a logical set of endpoints (usually these endpoints are pods) to make those pods accessible within the platform. Then, the own K8s distributes this incoming traffic to each pod. Furthermore, some CNI plugins (like Calico [18]) also implement some mechanisms to filter traffic for/to the pods, for instance, defining network policies that filter out undesired traffic (at the application layer). It is clear that this model has its advantages in cloud-native environments: as applications might not be permanently deployed in a single node or cloud (either due to its unavailability or to optimise resources), a flat networking approach allows communication with/to an application regardless of its distribution and IP addressing. Furthermore, this model allows application developers to pass over the inner networking being performed in an infrastructure since it is completely transparent to the applications themselves. In other words, developers can assume that other modules of an application will always have connectivity, and that no further configurations should be required in the modules to be used in the infrastructure. Despite its advantages, this model might not be suitable for all the services that could be provided through cloud-native platforms. In a previous work, we realised the potential of microservice platforms [34] as enablers for the deployment network and vertical services in the context of the 5th and 6th generations of mobile networks (5G/6G) in cloud-native environments (particularly, microservice platforms, which are a subset of functions in the cloud-native model). Microservice platforms, like Kubernetes (K8s), usually employ container technology to build applications in their managed infrastructures. Due to their lightweight nature, in some environments (e.g., resource constrained scenarios), these platforms have multiple benefits over traditional virtualisation solutions like OpenStack, which usually require more resources for the management and deployment of NSes. However, due to the intrinsic nature of NSes, it is necessary to ensure that communication at the lower layers is available (and not just at the application layer) for the functions that build the service. One prominent example of the necessities of virtual networking can be seen in the implementation of a router functionality. A traditional router must analyse incoming packets at the IP level from one of its network interfaces in order to check the destination ----- _Future Internet 2023, 15, 274_ 8 of 28 that will be forwarded. However, this functionality can only be performed if each one of its interfaces is located in a different LAN. The same situation is present when dealing with virtual networks: in virtual networks, each function is located at different LANs (despite their geographical location) and a router functionality would have one interface in each virtual network, enabling the analysis of the incoming/outgoing packets to be forwarded to the corresponding functions located in different (isolated) virtual networks. In consequence, functions are isolated between each other at the network level and can only communicate through the corresponding router, which enables the secure deployment and isolation of network functionalities, as well as their chaining (necessary for NS implementations). However, this behaviour is impossible to achieve in flat networking approaches: if all the functionalities are located in the same LAN (like they are in the flat networking approach), then there is no isolation between functions since they are in the same network domain, so the router cannot “decide” the routing/forwarding of the incoming/outgoing packets to each function. Hence, this problem heavily hinders the implementation of networking services in cloud native to host NSes in one, or several, clouds. To further illustrate the issue that microservice platforms have for the implementation of NSes in cloud environments, Figure 1 showcases a logic implementation of a multimedia content delivery service, in particular, a simple content delivery network (CDN) service. This CDN has two HTTP proxies: the first one is utilised to cache the content sent from a multimedia server, which allows to make the content closer to the end users (reducing download times and bandwidth used in the network) while the second one is a firewall that filters undesired requests from a function outside the CDN. Ideally, a service of such characteristics must enable sending the requests from/to the corresponding proxies and servers involved in the CDN, following the schema depicted in the upper side of the figure. Furthermore, the HTTP proxy must be able to analyse incoming packets at the network layer to appropriately filter undesired requests. Implementing services like the one depicted in this example is often performed in cloud environments through virtual links (i.e., virtual networks) that enable the isolated communication of different modules, which can only reach the corresponding peers of a particular virtual network. This enables the creation of links between functions similar to the ones that are used in NSes in the context of NFV, enabling the connectivity of microservices in schemas like the one shown in the upper part of the figure. **Figure 1. CDN service in a traditional networking approach vs. flat networking.** ----- _Future Internet 2023, 15, 274_ 9 of 28 Unfortunately, due to the intrinsic behaviour of the networking model used in microservice platforms, an implementation of such characteristics is impossible to achieve. As it is depicted in the bottom part of the figure, microservice platforms with flat networking approaches do not isolate these components: all of them are able to see each other at the network level, usually through a set of networking agents in charge of forwarding traffic to the rest of the components deployed over the platform. Although these agents can usually discriminate traffic based on some protocols (normally application-layer based), they do not completely isolate these components between them since they all are reachable at the IP layer. With the drawbacks of flat networking approaches in microservice platforms in mind, this paper presents a solution that enables the use of virtual networking for the deployment of network services and verticals in cloud-native environments: L2S-M. L2S-M aims at the provision of secure link-layer connectivity to NSes in cloud-native environments. Instead of being developed as a full networking solution to replace established connectivity solutions in different microservice platforms, L2S-M provides a flexible complementary approach to allow containers present in a cloud to attach into a programmable data plane that enables point-to-point or multi-point link-layer connectivity with any other container managed by the platform, regardless of its placement inside the infrastructure. This programmable data plane relies on software-defined networking (SDN) techniques to ensure the isolation of the link-layer traffic exchanged between containers. Moreover, SDN allows the application of traffic engineering mechanisms over the programmable data plane based on several factors like traffic priority or delay. _3.2. Functional Design of L2S-M_ In our previous work [35], L2S-M was first introduced as a complementary networking service to enable the deployment of NSes in NFV infrastructures (NFVIs) composed of resource-constrained environments (particularly, aerial microservice platforms based on K8s, as it is the de facto microservice platform and the one that could provide significant advantages in comparison with other VIMs [34] in UAV networks). In this regard, L2S-M was created to address the limitations that NFV Orchestration could have in such scenarios as seen in our previous works [36,37]. Particularly, L2S-M enables the creation of virtual networks that could connect different VNFs at the link-layer level, which is essential to ensure the deployment of NSes for aerial networks (as seen in our previous works like [38]). Moreover, the use of SDN could allow modifying the paths that traffic could use in aerial ad hoc network scenarios in response to sudden cut-offs, instead of leaving this task to the routing protocol in an aerial network. However, it is clear that there is a need for bringing virtual networking solutions in cloud and edge microservice-based platforms, which is necessary to ensure the proper provision of applications and services in the form of cloud network functions (CNFs), particularly in K8s platforms. The flexibility of its design allows L2S-M to not interfere with standard networking models already implemented in microservice platforms, such as K8s, bringing a complementary solution instead that can be exploited by any developers/platform owners interested in deploying CNFs in microservice platforms. Moreover, L2S-M also has the potential to enable secure link-layer connectivity between several cloud and edge solutions, effectively enabling inter-site communications for network functions and verticals deployed over multiple infrastructures. With these ideas in mind, this paper showcases the design of L2S-M as a cloud-native solution that enables the management of virtual networks in microservice platforms. Particularly, this paper presents a full architectural design of this solution, envisioned as a K8s operator used in data centre infrastructures that, due to its flexibility, can be exported to other kind of scenarios (e.g., edge environments). This work presents an implementation of L2S-M as a K8s operator in order to detail its functionality in the well-known microservice platform (although it can be exported to any microservice platform). ----- _Future Internet 2023, 15, 274_ 10 of 28 Figure 2 showcases the design of L2S-M in a cloud infrastructure. L2S-M delivers a programmable data plane that applications and services can use to establish point-to-point or multi-point links between each network function on demand. This objective is achieved through the creation and management of virtual networks, to which applications are able to attach, sharing the same broadcast domain between all the containers that joined one of these virtual networks, i.e., all containers will see each other in the same fashion as if they were in the same local network, regardless of their physical location within a cluster (set of machines governed by the same cloud-native management platform). This behaviour enables the direct point-to-point and multi-point link-layer communication between each container, isolating the traffic from each network to avoid unnecessary traffic filtering (for example, by having to implement multiple traffic policies for each application) and to ensure their secured operation. **Figure 2. L2S-M design in a cloud infrastructure.** The way that L2S-M is able to introduce this virtual networking model is through a set of programmable link-layer switches spread across the infrastructure as seen in Figure 2. These switches can either be physical equipment (labelled as Switch in the figure (such as the ones that can be found in traditional data centre infrastructures)) or virtual switches (labelled as V-SW in the figure), which can take advantage of the benefits of container virtualisation technology. In order to establish the point-to-point links between the switches to allow their communications and enable the desired in the cluster, IP tunnelling mechanisms are used, for instance virtual extensible LANs (VXLANs) [39] or generic routing encapsulation (GRE) [40]. That way, the basis of the L2S-M programmable data plane is established through this infrastructure of switches. Figure 2 showcases an infrastructure that divides three different availability zones with different characteristics. For example, one has physical switches that could be used to deploy hardware-accelerated functionalities, while another zone is an edge with resource-constrained devices like UAVs. It is worth mentioning that most networking solutions in cloud-native environments also rely on IP tunnelling mechanisms to build their communications ‘backend’. However, there are noticeable differences with respect to the approach used in L2S-M: first of all, these solutions build the IP tunnels to interconnect their own networking agents, which perform routing tasks in host IP tables themselves, and can interfere with the networking of some machines and/or functions (and cannot be easily modified) in turn. Furthermore, these tunnels are built between all members of a cluster to build a mesh, while L2S-M has ----- _Future Internet 2023, 15, 274_ 11 of 28 the flexibility to allow the use of any kind of topology and can be dynamically adapted depending on the necessities of the platform owners. This overlay of programmable link-layer switches serves as the basis for the creation of the virtual networks. In order to provide the full programmable aspect of the overlay, L2S-M uses an SDN controller (SDN-C in the figure) to inject the traffic rules in each one of the switches, specifying which ports must be used in order to forward, or to block, the corresponding traffic coming from the containers attached to the switches and/or other members of the overlay. This SDN controller can also be embedded into the own virtualisation infrastructure as shown in Figure 2. The use of this SDN approach can also enable the application of traffic engineering mechanisms to the traffic distributed across the programmable data plane. For instance, priority mechanisms could be implemented in certain services that are sensitive to latency constrains. Service mesh solutions like Istio [20] could be seen as alternatives that would enable a similar behaviour as the one provided by L2S-M. However, the service mesh was developed with the same ideas and concepts as the flat networking approach: instead of managing the network interfaces directly, service mesh solutions use proxy functionalities to forward/block traffic based on networking services (network abstractions) in a separate data plane from the one presently used in the cluster, basing its routing/forwarding in their logical definition (i.e., the user defines which services must communicate with each other). Although this is a favourable approach to provide high availability and keep the abstraction models present in the microservice platforms, this solution still does not address the isolation aspects needed for NFV deployments since all their containers are still located in the same LAN domain (through its CNI Plugin agents) and can be directly reached by the rest of the containers. Therefore, L2S-M provides a behaviour that service mesh cannot provide, as it does not have the proper tools to enable the use of virtual networks needed in NFV deployments. It is true that other solutions have explored similar virtual networking concepts in microservice platforms, highlighting Nodus [26] and Kube-OVN [27]. However, all these solutions have tried to implement a substitute for current CNI plugins, while L2S-M provides this behaviour as a complementary solution for those applications that may require such a degree of networking control. Furthermore, L2S-M has been designed to enable its seamless use with physical switching infrastructures (commonly found in data centre networks) through single root input/output virtualisation (SR-IOV) interfaces [41], which can greatly extend its use for multiple use cases in the NFV space (e.g., network acceleration, and 5G CORE deployments). Finally, L2S-M has a higher degree of flexibility to accommodate different SDN applications to introduce traffic engineering mechanisms based on the required scenario (which cannot be performed with the previous solutions, as they rely on an internal SDN mechanism that cannot be easily modified to implement new algorithms and applications). _3.3. Inter-Cluster Communication through L2S-M_ The previous section explained how L2S-M allows establishing link-layer virtual networks that connect CNFs executed in the same cluster through the combined use of network-layer tunnelling and SDN technologies. We refer to this type of connectivity as intra-cluster communications. Nonetheless, this idea can be extended to the inter-domain scope to provide link-layer connectivity between CNFs that run in different clusters. We will refer to the latter as inter-cluster communications. For this section, clusters may include any kind of cloud-native environments (not only K8s), since the L2S-M design is flexible enough to accommodate any kind of infrastructure. At this point, it is necessary to introduce two new elements in the L2S-M design to enable the inter-cluster communications: the network edge devices (NEDs) and the inter domain connectivity orchestrator (IDCO). The NEDs are programmable switches similar to the ones shown in the previous subsections: they can be either implemented as software, or be physical hardware present in ----- _Future Internet 2023, 15, 274_ 12 of 28 every site. Each cluster can be connected to one or more NEDs to constitute an inter-domain programmable switch infrastructure. Each NED must have network-layer connectivity with at least another NED. Following a similar approach to the one used for the intra-cluster communications, an overlay network is created by connecting the NEDs through secure network-layer tunnels that encapsulate the link-layer frames (e.g., VXLAN over IPSec). This overlay can be manually created when deploying the NEDs, although a new overlay manager can be present in order to manage the establishment of these tunnels. Each frame that is transmitted from a certain cluster A to a cluster B travels from one of the NEDs of cluster A to one of the NEDs of cluster B, traversing the overlay network and possibly going through other NEDs in other clusters. The interconnection of the cloudnative platforms with the NEDs will vary depending on its nature: for instance, a K8s cluster could deploy a NED as a pod in one of the nodes of the cluster and attach several ports into an L2S-M switch so that the communication in the cluster is managed through several predefined virtual networks in the cluster (to indicate to L2S-M to which ports the traffic should be sent for inter-domain communications); in OpenStack environments, a NED can be a VM attached to a provider network, which can be relied on to distribute and/or send traffic accordingly. Nevertheless, this setup provides the link-layer communications between elements in different clusters (although it does not isolate traffic between them yet). The IDCO element is in charge of managing the inter-cluster virtual networks. It has both northbound and southbound interfaces. The northbound interface is implemented as an HTTP REST API that allows external authorised entities to create, modify or delete the virtual networks. The southbound interface is used by the IDCO to obtain information from the NEDs and to inject the switching rules in them through a SDN southbound protocol (e.g., OpenFlow [42] or P4Runtime [43]). The IDCO decides how the frames that belong to each inter-cluster virtual network should traverse the overlay network and injects several rules in the NEDs to create the needed paths and accomplish the network isolation between them. **4. Implementing L2S-M in a Cloud-Native Platform for Intra-Site Connectivity:** **K8s Case** _4.1. L2S-M Implementation as a K8s Operator_ This subsection introduces an implementation of L2S-M as a K8s operator, enabling the creation and management of a virtual network in a distributed K8s infrastructure to securely communicate workloads at the link-layer level. Although this subsection focuses on the intra-site implementation of L2S-M, the validation present in this paper will showcase the basic functionality of inter-cluster communications (described in the previous section) to demonstrate its functionality in those scenarios. Figure 3 depicts the detailed implementation of the full L2S-M solution in a K8s cluster. This implementation contemplates the deployment of all the components depicted in Figure 2 with their respective particularities to allow functionality inside a K8s cluster since the implementation of this solution is not a straightforward task due to the complexity that pod namespace isolation and API model introduce in K8s. First of all, L2S-M requires the deployment of a set of L2 switches over the K8s infrastructure, as it is showcased in Figure 2. These switches are necessary components to enable the establishment of the L2 overlay required to exchange data between pods. However, instead of directly installing the switch in the node itself, L2S-M relies on the advantages that K8s provides (containerisation, automatic deployment, life-cycle management, etc.) to deploy each switch as a pod on every node of the cluster. Particularly, L2S-M uses a “daemonset” (a K8s resource that deploys one pod per node in the cluster) that installs an open virtual switch (OVS) in the node. Although any L2 switch solution can be used for this purpose, L2S-M uses OVS due to its compatibility with multiple OS and distributions, as well as its simplicity for its installation and configuration. ----- _Future Internet 2023, 15, 274_ 13 of 28 **Figure 3. L2S-M implementation in a K8s infrastructure.** Once the switch infrastructure is available in the K8s cluster, it is necessary to deploy point-to-point links between the desired neighbouring nodes to enable link-layer connectivity between each other through IP tunnelling mechanisms, as showcased in the design and in Figure 2. Instead of building a mesh with all the members of a K8s cluster (a common practice in CNI plugin solutions), each programmable switch is only interconnected to the desired peer, which can either be a virtual switch or a physical one. The figure depicts these connections (thick blue links between V-SWs) performed using VXLAN tunnels, although any IP tunnelling mechanism can be used (for instance, GRE [40]). L2S-M must be able to create this overlay as well. However, having the switches containerised introduces a problem, as it can be seen in Figure 3: the pods are not able to reach the other nodes directly since they are located behind their own namespace inside the node, so directly building an IP tunnel with the node would not work, or it would be necessary to use the CNI plugin for the standard K8s networking (which follows the flat networking approach and is incompatible with the concept of the solution). To avoid this struggle, L2S-M builds the VXLAN tunnels beforehand in the host namespace (since they must be able to see other at the IP layer, as a requirement of the K8s cluster), either using a dedicated interface for this purpose or its main interface. Afterwards, L2S-M “moves” these tunnel interfaces into the switch pod using the Multus CNI plugin since this plugin enables to bring the pre-created VXLAN interfaces into the OVS pod without losing the link-layer configuration. Regardless of the containerisation of the switches, it is necessary to enable the attachment of the pods with the switches of the overlay deployed in every node in order to perform the data exchange between other pods. However, every pod is located in its own namespace, so connectivity between the pods and their (virtualised) switch cannot be directly established (as it can be seen in Figure 3). To overcome this problem, a virtual Ethernet (vEth) element can be used to exchange messages between each namespace, as it mimics a “real” Ethernet cable, where packets sent at one end of the vEth appear at the other end, regardless of the namespaces at which they are located. L2S-M builds a set of vEth pairs in the host namespace, and then L2S-M attaches one extreme to the switch, leaving the remaining one in the host namespace. Once a pod desires to connect to a virtual network, L2S-M uses the Multus plugin to attach the other end into the pod, effectively connecting these two elements (just as it is done in a physical switch): ----- _Future Internet 2023, 15, 274_ 14 of 28 when a packet is generated inside the pod, the vEth will forward the packet into the Linux Kernel, and the packet will be forwarded to the switch. However, K8s does not have the tools to deal with the definition and management of virtual networks on demand, or allowing the assignment of the corresponding vEth pairs in each one of the hosts to the workloads deployed in the cluster. This is where a key element in the L2S-M design comes into play: the L2S-M K8s operator. A K8s operator [44] is a software extension to Kubernetes that allows the management of custom resources in a K8s cluster, which might contain any information and can only be used by coordinating the K8s API events with the operator to perform certain actions or events in the cluster. The L2S-M K8s operator takes advantage of a pre-existing CustomResourceDefinition (i.e., resources that are not native to the K8s environments) (CRD) from Multus (the wellknown NetworkAttachmentDefinition (NAD)) to define the virtual networks that we want to create in a cluster. If we want a pod to be added into one of these virtual networks, it will be written in its metadata in the same fashion as it would be done in a standard Multus definition (a common standard in real K8s deployments). However, these are not mere Multus annotations since they must be managed by an element that is able to identify which pods want to belong to which virtual networks, as well as to perform the necessary actions to attach the ports of each switch into the pod. Therefore, L2S-M includes the definition of the L2S-M operator, which is an agent (deployed as a pod) that is deployed in a controller node of a cluster as it can be seen in Figure 3. This pod will constantly monitor the calls between the K8s API, picking up several events that occur within the cluster. Depending on the type of event that is picked up by the operator, it will perform a different action. In this fashion, the L2S-M operator will be triggered when a “creation event” with a NAD is registered from the K8s API to check if the corresponding resource is a virtual network or a standard Multus definition (in such a case, the operator will not perform additional actions). If it is a virtual network, it will register its creation in the cluster, writing this network in its database (L2S-M DB). After the creation of a virtual network, once a pod starts its deployment in a cluster, it will generate a creation event and, if the pod being generated includes a NAD annotation in its metadata, the operator will begin to process this annotation prior to its deployment. The operator will then identify each one of these annotations to see if the pods express their desire to be attached into one, or several, virtual network(s), checking if any of the networks created in the cluster are present. If not, the operator will let Multus handle the deployment. Otherwise, the operator will retrieve the node where the pod is going to be deployed, and will check if there are available interfaces (i.e., free vEth ends) in the host namespace. Once a vEth is selected, the operator will assign that interface to the pod and register that it belongs to a particular virtual network. In case the Kubernetes API schedules the pod’s deletion, the operator will remove the interface from the virtual network in the corresponding node, and the vEth will be returned to the host namespace to be available for future workloads. During all these events, the operator will be modifying its DB depending on their actions. In order to provide the mechanisms to isolate traffic between virtual networks, L2S-M contemplates the deployment of a software-defined networking (SDN) controller in the K8s cluster as seen in Figure 3. The L2S-M operator and the controller interact through a common API, which allows the operator to communicate the interfaces (ports) where each pod is attached since the operator knows which virtual networks the pod belongs to. The SDN controller will use this information to send the appropriate traffic rules to each one of the programmable switches to ensure that the traffic generated in each network will only be sent to the proper ports (either forwarding the information to the corresponding neighbour in the overlay or to one of the ports in the switch). This mechanism ensures that the traffic in each virtual networks becomes isolated between each other since traffic will not be forwarded between workloads unless they belong to the same virtual network (i.e., they are treated as if they were in the same LAN). The current version of L2S-M [45] does not implement the entire isolation mechanism: it is able to isolate most of the traffic within a ----- _Future Internet 2023, 15, 274_ 15 of 28 virtual network since it can interact with an ONOS controller [46] to enable the forwarding only with the appropriate ports depending on the network. However, traffic destined to multiple hosts/pods (e.g., broadcast traffic in ARP Requests) must be forwarded to all elements in the overlay since ONOS does not natively implement a way to isolate this traffic. Future versions of L2S-M will fully isolate traffic in their respective virtual networks, regardless of their nature, using a specific SDN application used with ONOS. _4.2. Virtual Network Management Flow_ Figures 4–6 showcase the communications that are established between all the components of the L2S-M solution, divided into four main steps: the creation of a virtual network in the K8s cluster, the attachment of a pod into a virtual network, the deletion of a pod in the cluster and the deletion of a virtual network in a K8s cluster. **Figure 4. Creation of a virtual network in L2S-M.** **Figure 5. Attachment of a pod into a virtual network in L2S-M.** ----- _Future Internet 2023, 15, 274_ 16 of 28 **Figure 6. Deleting a pod from a virtual network in L2S-M.** 4.2.1. Virtual Network Creation First of all, when a user wants to create a virtual network, as seen in Figure 4, the user will instruct K8s through its command-line interface (kubectl) to create the resource inside the cluster (i.e., a NAD with the definition of the virtual network). Once this creation is performed, the L2S-M operator picks up the K8s event and checks that this NAD definition corresponds to a virtual network. The way the L2S-M knows is that the NAD includes a virtual interface (i.e., an interface that is not physically defined in the host) called “l2sm-vNet”, informing L2S-M that the annotation is a virtual network. Once the event is picked up, the operator registers the creation inside its database, completing the creation of the virtual network. 4.2.2. Attachment of a Pod into a Virtual Network The attachment of a pod into one, or more, virtual networks using L2S-M follows this structure as seen in Figure 5: 1. When a pod wants to be deployed in the cluster associated with one (or several) virtual network, it will introduce the corresponding annotation in its descriptor, using the standard Multus annotation format. The user will then use kubectl to deploy the pod, generating a creation event in the K8s cluster. 2. The L2S-M will pick up the event and check whether the pod has the corresponding annotation and if so, it will check each annotation element to see if it corresponds to a virtual network NAD from its database. Once it matches, L2S-M checks in its database the free vEth in the node where the deployment is being performed (these data are retrieved using the K8s API), writing an entry in the database for that interface with the name of the pod and the virtual network that this interface is associated with. 3. The L2S-M operator updates the deployment with the new interface annotation, instructing the Multus agent in the node of the vEth pair interface that will be aggregated to the pod. Once this operation is completed, the pod finishes its deployment phase attached to the OVS switch of the node. 4. After the deployment, the L2S-M operator sends the SDN controller the new attachment of the pod, notifying the controller that the new port of the switch is associated to a new virtual network. With this information, the SDN controller can configure all the switches of the overlay with the corresponding rules to exclusively forward packets between the members of the virtual network. This behaviour is up to the application running in the SDN controller, which is in charge of finding the appropriate path ----- _Future Internet 2023, 15, 274_ 17 of 28 between the pods and configuring the forwarding rules of the switches. One way to perform this could be using intent-based connectivity in such a way that L2S-M provides the MAC address of the members of the virtual network to the SDN controller using intents so that the controller can properly configure the paths between them. 4.2.3. Detachment of a Pod from a Virtual Network The procedure to delete a pod from a virtual network is very similar to the one for its deployment as seen in Figure 6: 1. Once the pod is scheduled to be deleted from the cluster (either from a deletion event or a failure in the pod/node), L2S-M picks up the event generated from the K8s API and realises that the pod being removed is attached into a virtual network. 2. L2S-M removes the pod entry from its database, marking the interface as idle. 3. Simultaneously, L2S-M sends to the SDN controller the instruction to remove the pod from the virtual network (e.g., removing the previous intent(s) generated in the attachment phase). 4. The SDN controller will configure the forwarding tables from the switches to remove the entries related to the pod that have been deleted, effectively removing the pod from the virtual network. 4.2.4. Virtual Network Deletion Finally, once a user wants to create a virtual network, as seen in Figure 7, the user will instruct K8s through kubectl to delete the resource inside the cluster. Similarly to the creation of networks, L2S-M picks up the K8s event and removes the virtual network entry from its database. This action will only be possible if all pods have been detached from the network; otherwise the operator will throw an error and prevent the deletion of the virtual network. **Figure 7. Deleting a virtual network in L2S-M.** _4.3. L2S-M Information and Uses_ L2S-M has been released as a publicly available open-source code that can be used in most K8s distributions [45]. This solution has received interest from the research community, as it has been used in multiple European research projects in different contexts: LABYRINTH [47], focused on providing security functions using UAVs, using L2S-M to enable communications between the aircrafts; and the FISHY project [48], focused on providing a coordinates framework for cyber-resilient supply chain systems, where L2S-M is used as the basis of the main platform built to deploy the functionalities used in the project, the FISHY reference framework (FRF). As it was described in previous paragraphs, current NFV orchestrators, like the wellknown Open Source MANO (OSM) [28], have limited support for the deployment of NFV cloud functions using K8s since there is no native way to create virtual networks able to interconnect several VNFs in the clusters. However, our ongoing work includes the definition of a new feature to be included in the codebase of OSM to enable the deployment of network functions in K8s clusters. This feature, named “Connectivity among CNFs ----- _Future Internet 2023, 15, 274_ 18 of 28 using SDN”, has already been approved and it is currently in the design phase in direct collaboration with the OSM community. The details of this feature can be seen in the official OSM site [49]. In this regard, we will briefly describe the steps that will be performed to add a K8s cluster in the OSM ecosystem using L2S-M, as well as the deployment of a NS using virtual networks within the cluster: - At the cluster (VIM) registration time, the OSM user selects that the data plane used in the CNFs communications is provided by L2S-M. Then, the user defines the resource definitions (i.e., .yaml templates or Helm [50] charts) in order to tell the orchestrator how to build and manage these networking resources within the cluster. - When a new network is deployed using the orchestrator, the resource orchestrator component (RO) of OSM takes the values and configuration parameters of each VL in the descriptor and translates them into the parameters used in the L2S-M virtual networks. After this process, the orchestrator contacts the K8s cluster and follows the flow seen in Figure 4. - Once the VLs have been processed, then the orchestrator proceeds to add the corresponding K8s annotations to each VNF to add the VLs associated with them, and start their deployment within the K8s cluster (as seen in Figure 5), finalising the deployment of the NS using L2S-M as the data plane networking solution in the cluster. **5. Practical Experience with L2S-M** _5.1. Description of the Testbed_ This section describes the testbed that is considered to validate the implementation of the design introduced in Section 3. This testbed mimics the infrastructure that could be deployed in a university Smart Campus environment. Universities, due to the nature of the academic and research activities that they perform on a daily basis, must have a powerful, reliable and secure infrastructure that allows them to flexibly deploy various applications used by the members of the university. These applications, ideally, should be able to effectively use the resources provided by the university infrastructure, as well as having good scalability properties (to adapt the service to the possible demands, which can dynamically be modified) and be resilient to temporary failures and/or service cut-offs. In this regard, microservices are able to provide most of these characteristics. Unfortunately, some of these services require networking capabilities that solutions like K8s are not able to provide. One prominent example of this situation is the use case presented in this paper: the implementation of a content-delivery network (CDN) for the distribution of academic content in a Smart Campus scenario. Generally, universities are composed of several campuses spread in distributed geographical locations. Each campus has its own size and importance inside the structure of the university, especially regarding the resources that they are able to provide for their own cloud environments for content distribution. This can potentially be an issue since a pure centralised model might impact the performance needed to effectively send content to remote campuses, possibly being a more desirable solution to move the content closer to the users to avoid overloading the main infrastructure of the network, which in turn may reduce latency as well. A centralised model can also introduce a single point of failure if the main infrastructure is down and/or link disruptions occur. Finally, it is important to mention that the network infrastructures of these campuses should be able to be dynamically modified to accommodate the demand that each campus may require at each moment, without interfering with the functionality. Microservice solutions like K8s are not usually able to provide the necessary tools to implement a distributed scenario due to its flat networking approach and the limitations of inter-cluster communications. However, L2S-M will be used in this paper to provide a CDN to distribute content across different campuses of a wide variety of characteristics located across geographically distributed regions, while also allowing to easily accommodate new infrastructures and members to fit the necessities of the university (for instance, setting up a temporary network for an event). ----- _Future Internet 2023, 15, 274_ 19 of 28 This use case includes four different sites distributed along two campuses to prove the effectiveness of the solution in an intra-campus scenario and of the NEDs for the inter-cluster communications. Each campus is designed to include different resources to showcase the interaction between heterogeneous infrastructures. Accordingly, there are two different sites inside each campus, resembling the cloud and edge of each campus, where each site can have its own infrastructure and implementation, to validate the initial statements. The scenario presented in this paper can be seen in Figure 8. The first campus in the scenario (on the right) is composed of two different sites: one central campus environment, and one temporary edge site, which is set up only if an event is performed in the university facilities (e.g., a conference or a workshop). The central infrastructure of this campus is regarded as the main cloud of the whole organisation, holding most of the software and teaching resources available to the students and teachers. In consequence, the representation of this site is conducted through two VMs (since these are considered “heavier” machines in terms of available resources). One of the VMs is used as a general-purpose server to host multimedia and software contents. The remaining VM hosts the corresponding NED used to connect this site with the rest of the sites of the university, both the ones in other campuses (inter-campus communications) and the ones in its own campus (intra-campus communications), as long as they are in different sites or clusters. The second site is the edge environment of the campus, and it is meant to represent the temporary devices deployed for an opportunistic event that members of the university will use to retrieve the content, as well as providing a cache server for the university’s CDN. It comprises two Raspberry Pi 4 Model B computers that act as nodes of a K8s cluster where L2S-M is installed. The NED of this site provides the connection of this site with the general cloud, which is the other site in this same campus, and with the other cloud present in the second campus. The second campus has two different sites as well. The first one is the designated campus’ cloud, which is used as a proxy for the connectivity of the edge environment deployed in the remaining site. This edge environment provides the infrastructure that students will use on a regular basis to download the university content. Naturally, the site offers a proxy as part of the CDN in order to store content in its premises, closing the information to be closer to the students. The NED of the cloud allows for the connection to the two sites in Campus 1 and to the other site in the same campus’ edge. The second edge NED connects this site with the same campus’ cloud. The structure of both edges is symmetrical in our deployment, although each site may have a different infrastructure and configuration, validating the initial premise of the benefits of L2S-M. The use case that is implemented to validate this research aims to simulate the previously described CDN, where a content server is located in the general cloud, to store all the desired data, one proxy server is located in the edge of the first campus, and two proxy servers are located on the second campus, one on its cloud and one on the edge, all aiming to cache data closer to the user. They will be deployed as different pods running a nginx [51] web server with the functionality of an HTTP reverse caching proxy. In the cloud of the first campus, the server was installed inside a VM to act as the main content server (i.e., where the information is permanently stored). In order to protect the access to this main cloud from external sources, an HTTP proxy was deployed in the cloud of the neighbouring campus. Regarding the edge sites (present in both campuses), an additional HTTP proxy was deployed to cache the content coming from the remote server. Both edges were designed in a symmetrical way. Apart from the proxy, one access point (AP) was deployed as a pod on each edge, giving the user the possibility to download the content from the CDN by connecting to the proxy available in the edge. To effectively enable this connection, a domain name system (DNS) provides the user with the IP address of the edge HTTP proxy when introducing the URL corresponding to that content. To avoid reaching the HTTP server directly without connecting to the enabled AP service, a firewall service (developed as a Linux router) was introduced in both edges, and it is in charge of forwarding the traffic from the AP into the nginx proxy and vice versa. ----- _Future Internet 2023, 15, 274_ 20 of 28 **Figure 8. Use case high-level design.** For this scenario, some virtual networks must be created to attach the different components, once again in a symmetrical way between campuses 1 and 2, connecting, in the first network, the router and firewall with the DNS and the AP. This can be seen in Figure 9, where Net1 corresponds to the virtual network in the first campus and Net2 to the one in the second campus. Another virtual network is deployed among the different campuses, connecting the content server with the different proxies and routers. This stands in Figure 8 as Net3. All these deployments allowed for the connection between elements of the network to be established and for the behaviour of the scenario to be as expected. **Figure 9. Use case detailed implementation.** ----- _Future Internet 2023, 15, 274_ 21 of 28 _5.2. Experimental Environment_ Figure 9 showcases the detailed implementation of the scenario, with the different pods that were deployed on each site and the connections between them. Starting with the first campus, the edge site is represented using a rack of four Raspberry Pi4 Model B computers, all of them having 8 GB of RAM and running an Ubuntu Server 20.04 installation. All these RPis were connected using gigabit Ethernet connectivity (since these are considered fixed devices that could be placed in classrooms all over the campus). All RPis are part of the same K8s cluster, using the L2S-M operator for intracluster communications. In order to enable the communications between pods and avoid loops over the L2S overlay created within the cluster, a RYU SDN controller was deployed (running as a pod) using the spanning tree protocol (STP). Each RPi hosts a different functionality, all of them deployed as K8s pods, which can be seen in more detail in Figure 9. The first RPi, the closest one to the users, hosts the AP functionality (enabling the requests of content downloads) and the DNS service within the edge cluster. The next hop of the CDN service, once a download has been requested from the AP, is the firewall (implemented as a Linux router) that will redirect the HTTP requests to the proxy, located in the second RPi. This proxy also provides a cache that allows hosting some of the requested content inside the cluster (allowing to have the information closer to the users). The cache can be dynamically modified depending on the demand and the status of the network. If the content is not found in this proxy, it will redirect the request to the proxy located in the campus cloud through the NED present in the remaining RPi, which oversees these inter-cluster communications. This cloud is composed of a single VM (Ubuntu Cloud 20.04) that hosts a K8s cluster. Two functionalities were deployed (as pods): one proxy HTTP, in charge of redirecting the requests from the edge campus to the main cloud of the university, and the SDN controller used for inter-cluster communication. This last component is essential for the whole functionality of the university since it allows the configuration of all the NEDs present in each cluster/site. Similar to the intra-site communications, this controller was implemented using the RYU SDN controller running a SPT protocol to avoid network loops. All of the previously described K8s clusters were installed using the 1.26 version of kubeadm [52], running the K8s 1.26 release. using containerd as the container runtime. For the default networking CNI Plugin, we selected Flannel [17] since it is one of the most-used CNI plugin solutions in production clusters. In the case of the main campus, the content server is directly installed and configured in one VM within an OpenStack cluster, and it is connected to the cloud of the second campus and to the other site in the campus premises through a NED, installed inside another VM. Both VMs run an Ubuntu Cloud 20.04 image, using 2 CPUs and 8 GBs of RAM. This content server stores all the multimedia files, Linux images, Debian packages and many other types of files that could be used daily in a university environment. The combination of all the functionalities and elements of both campuses build the CDN service that is implemented in this work. Due to the nature of the activities that could be present in a university, the edge environment of the main campus is considered a temporary infrastructure that is aggregated into the Smart Campus infrastructure. This new edge site, built as a K8s cluster using L2S-M for intra-cluster communications, has the same functionalities as the edge of the second campus, with the exception that its proxy will directly request the content to the main content server, rather than using an intermediate proxy. All of the configuration, deployment and network files used in this validation section can be found in the corresponding repository [53]. _5.3. Functional Validation and Results_ 5.3.1. Throughput Performance This first set of validation tests aimed at showcasing the possible impact that L2S-M could have in the available throughput for the virtual functions deployed over a K8s cluster. ----- _Future Internet 2023, 15, 274_ 22 of 28 To test this impact, we used the well-known traffic-generation tool iperf3 [54] in order to test the total available bandwidth between the two pods deployed in the scenario, testing both the standard K8s networking (flannel [17] for intra-cluster networking, and K8s NodePort for inter-cluster networking) and L2S-M in each scenario. In particular, the pods were deployed using the following configurations: - Two pods deployed in the same node of Campus 2 edge cluster (RPi1). - Two pods deployed in different nodes of Campus 2 edge cluster (RPi1 and RPi3). - Two pods deployed in different clusters (RPi1 of Campus 2 edge cluster and Campus 2 cloud cluster). For each configuration, an iperf3 flow of 180 s was established between each pod in both directions, using the standard K8s networking and the L2S-M in every iteration of the test. In this regard, each iteration was run 30 times, and then the average of each run was used to calculate the values shown in Table 1. **Table 1. Throughput comparison between Flannel and L2S-M in all the possible scenario configurations.** **Test** **Flannel (Mb/s)** **L2S-M (Mb/s)** intra-node 4860 5350 intra-cluster 870 847 inter-cluster 915 869 As it can be seen in Table 1, the performance of L2S-M does not introduce any significant performance degradation in comparison with the standard K8s networking approaches (Flannel and service abstraction). L2S-M improves the throughput between the pods when they are co-located in the same node since the traffic generated between them does not need to pass through the Flannel agent deployed in the node, which introduces some performance degradation. For the connectivity between pods that are located inside the K8s cluster, but in different nodes, L2S-M and Flannel exhibit quite similar performance, with a slight throughput decrease in the L2S-M case. Overall, the performance between both solutions is very similar, and showcases that L2S-M does not harm the traffic performance over the K8s cluster. Regarding the inter-site connectivity scenario, L2S-M provides approximately 50 Mb/s less throughput than its counterpart as expected since K8s services establish the connection directly without the use of IP tunnelling mechanisms, while L2S-M still requires the use of VXLAN tunnels between the infrastructures, which in turn introduces some overhead in the packets exchanged between the pods/VMs. Nevertheless, NodePort communications do not isolate the exchanged traffic between pods, unlike L2S-M, which only distributes traffic to the pods located in the same virtual network. These tests showcase that L2S-M does not introduce significant performance degradation in comparison with standard CNI plugin communications in K8s. 5.3.2. CDN Download Test The purpose of this test is to show the capacity of L2S-M to implement complex NSes, like the CDN proposed in this paper, over heterogeneous infrastructures implemented with different management and orchestration solutions. For this round of tests, we performed the downloading of a fairly large video file (1 GB) from Campus 2 edge and measure the average throughput that the CDN has when a user tries to download some media content in the Smart Campus scenario. For this test, a pod located in the edge cluster of Campus 2 downloads a video using [a specific URL that represents the video content “https://university-content/edu.mp4”.](https://university-content/edu.mp4) When the pod tries to download the video content (using the well-known wget program), the DNS service present in the campus edge will translate the URL into the IP address of the local (i.e., campus) nginx server, sending the HTTP request in the process, following the process described in the previous subsection. ----- _Future Internet 2023, 15, 274_ 23 of 28 In order to test the efficiency of the CDN in realistic scenarios, these tests emulate the congestion of one of the links used for the download, using iperf3 to send a TCP download at the maximum available rate possible. In this case, the link that is congested during the tests is the one interconnecting both campus clouds since in a realistic scenario it is expected to be the link that exchanges the highest amount of data between campuses. The cache will be set in two different ways: firstly by disabling the whole cache (making the server a simple http proxy), and secondly, allowing the cache to store the whole video. These modes will clearly showcase the impact of having a cache inside the scenario by reflecting if there is any significant improvement to the available throughput and/or the download speed. With all these considerations, the performed tests were the following: one set of tests, where the cache in the edge was disabled, performing the download when the link between campuses was idle and then congested. Each set was performed 30 times. Afterwards, the cache was enabled to hosts the whole video, repeating the aforementioned tests (idle and congested links). Figure 10 showcases the average throughput results, including the 95% confidence intervals. **Figure 10. Throughput with cache enabled and disabled.** As it can be seen in the figure, the presence of a cache obviously improves performance in terms of throughput in both scenarios: since the content is closer to the user, it traverses fewer functions and infrastructures, which in turn makes it easier for the nginx to send the content from one site to the other. When the cache is disabled, there is also a significant decrease between the idle and congested scenario. This is the expected behaviour since the content must traverse (and be processed) by an additional nginx server (the one located in the second campus). This behaviour can be further seen in Figure 11, where the traffic on each nginx element can be seen for two runs: one where the cache is disabled, and one where it is enabled (in both cases, the link was not congested). As it can be seen in all figures, traffic is present in all pods/VMs when the cache is disabled since the traffic is generated from the video-source VM in the main campus cloud and traverses the middle nginx pod. Since these entities must process this traffic, the overall throughput is lower, and the download **Figure 10. Throughput with cache enabled and disabled.** ----- _Future Internet 2023, 15, 274_ 24 of 28 takes significantly more time to finish. On the other hand, when the cache is enabled, the traffic is only generated from the edge nginx in Campus 2, which in turn requires less packet processing in the infrastructure (fewer nginx servers are involved), so the overall bandwidth is higher, and the download is significantly shorter. **Figure 11. Traffic capture in every CDN element of the cluster.** 5.3.3. Cluster Addition and Wi-Fi Download Test The last test of this validation section will provide some insights about the capabilities of L2S-M to incorporate a new infrastructure into a complex scenario such as the Smart Campus. It is common that universities hold events with many participants that require access to the Internet or/and retrieve content from the university to perform some activity. Some examples might include congresses, practical/lab sessions, etc. In all these cases, it is important to be able to set up an adequate infrastructure able to effectively host the network services required for each activity. Nevertheless, this set-up must be quick and dynamic since these events frequently change, so the requirements and equipment needed will vary depending on the type of activity being performed. In this regard, the set-up of a K8s cluster is well known to be simple and fast to deploy (a cluster can be set up in 20 min using administration tools such as Kubeadm [52]). In a similar fashion, L2S-M (and its inter-cluster functions) can be easily deployed within a K8s cluster as well since its configuration and set-up can be performed in a short period of time (approximately 10 min). For this test, the Campus 1 edge was deployed using two RPis (which can act as AP) and a Mini-ITx compute node to provide the connectivity with the main university cloud. This last battery of tests uses the RPis Wi-Fi module to provide the channel for the downloads, emulating a real scenario, where other clients would connect into an AP and download the content from there (rather than downloading it in the equipment, unlike in previous tests). In this case, the audiovisual content was downloaded from an external PC connected into the virtualised AP of an RPI, enabling and disabling the cache to test the CDN function **Figure 11. Traffic capture in every CDN element of the cluster.** ----- _Future Internet 2023, 15, 274_ 25 of 28 ality. This process was repeated 30 times for every mode, obtaining the download values that can be seen in Table 2. **Table 2. Download throughput and time used for retrieving content using Wi-Fi.** **Cache Enabled** **Throughput (Mb/s)** **Download Time** Off 4.207 3:20 On 4.780 3:05 As it can be seen in the table, the download speeds are lower than the ones in previous tests. These were due to the use of an unstable medium, such as the Wi-Fi connectivity of the RPis. Nevertheless, both download speeds and average throughput were improved when the cache was enabled in the site, proving again the effectiveness of the CDN. Beyond the particular results (throughput, etc.) obtained for the different featured scenarios, the main objective accomplished was to show that L2S-M can be used to provide complex network services that might involve different kind of network interfaces like wireless interfaces. This is also a starting point for the exploration of this concept since L2SM could be used to enable the connectivity with private networks over an infrastructure (e.g., the same way in which VIMs like OpenStack connect to provider networks). **6. Conclusions and Future Work** The rise of new paradigms like NFV in the context of the 5th generation of mobile networks has provided new ways to enable the development and deployment of network services. In this regard, microservice platforms assist in the optimisation and orchestration of network functions in distributed infrastructures. However, these platforms have some limitations due to the flat networking approach that they implement for the communication of their workloads (containers) in order to build a complex application. Furthermore, these solutions can also have some limitations for connectivity with other infrastructures, requiring networking abstractions that could prevent communicating network functions between clusters or other platforms. To address these issues, this paper has presented L2S-M as a solution that enables linklayer connectivity as a service in cloud-native ecosystems. L2S-M provides a programmable data plane that microservice platforms (like K8s) can use to create virtual networks that can be used by the containers to communicate at the link-layer level, since they will see the rest of the containers as if they were located in the same local area network (even though they might be distributed in different locations, depending on the cluster and the underlying infrastructure). Using these virtual networks, L2S-M provides the necessary network isolation required to deploy network and vertical specific functions in microservice platforms, which current solutions cannot easily provide. Furthermore, since L2S-M uses SDN technology to establish the paths between pods in a cluster, other SDN applications can be flexibly deployed to support traffic engineering and optimise traffic distribution, using an alternative network path across the L2S-M overlay. This paper also provides a first exploration of the potential of L2S-M to provide intercluster communications between containers using virtual networks, enabling the direct communication of network functions between heterogeneous infrastructures managed by different platforms, which can implement different virtualisation techniques, or even run bare-metal functions. This paper also presented the use implementation of L2S-M in a complex Smart Campus scenario, deploying a CDN to distribute multimedia content in a complex, distributed and heterogeneous scenario. The tests performed in the validation of this paper showed that L2S-M is suitable to deploy complex NSes based in microservices that require the use of multiple isolated virtual networks for their proper functionality, interconnecting workloads located in different infrastructures over geographically distributed locations. ----- _Future Internet 2023, 15, 274_ 26 of 28 Moreover, these tests depicted the flexibility of L2S-M to incorporate new infrastructures, like Wi-FI access points, to extend the functionality of the NSes in the use cases. Our future work for the advancement and further development of L2S-M includes the implementation of the overlay manager figure in the solution to dynamically modify the overlay network. Furthermore, we will also explore the implementation of L2S-M with SR-IOV interfaces to enable its direct use with the physical switching equipment commonly present in data centre infrastructures. This future work will also involve the exploration of alternative SDN controllers to increase the functionality and isolation aspects of L2S-M, as well as the development and application of SDN algorithms to apply traffic engineering mechanisms. Finally, we want to contribute to the relevant open-source communities with L2S-M. In this regard, we are working on a feature in OSM to support the creation of virtual networks in K8s clusters, using L2S-M as the reference operator [49]. **Author Contributions: Funding acquisition, F.V.; Investigation, L.F.G., I.V. and F.V.; Supervision, I.V.** and F.V.; Validation, L.F.G., R.M. and D.A.; Writing—original draft, L.F.G., R.M. and D.A.; Writing— review and editing, L.F.G., I.V. and F.V. All authors have read and agreed to the published version of the manuscript. **Funding: This article has partially been supported by the H2020 FISHY Project (Grant agreement** ID: 952644) and by the TRUE5G project (PID2019-108713RB681) funded by the Spanish National Research Agency (MCIN/AEI/10.13039/5011000110). **Institutional Review Board Statement: Not applicable.** **Informed Consent Statement: Not applicable.** **Data Availability Statement: Data sharing not applicable.** **Conflicts of Interest: The authors declare no conflict of interest.** **References** 1. [Cloud Native Computing Foundation. Building Sustainable Ecosystems for Cloud Native Software. Available online: https:](https://www.cncf.io) [//www.cncf.io (accessed on 11 June 2023).](https://www.cncf.io) 2. Liu, G.; Huang, B.; Liang, Z.; Qin, M.; Zhou, H.; Li, Z. Microservices: Architecture, container, and challenges. In Proceedings of the 2020 IEEE 20th International Conference on Software Quality, Reliability and Security Companion (QRS-C), Macau, China, [11–14 December 2020; pp. 629–635. [CrossRef]](http://doi.org/10.1109/QRS-C51114.2020.00107) 3. Morabito, R.; Kjällman, J.; Komu, M. Hypervisors vs. Lightweight Virtualization: A Performance Comparison. In Proceedings of [the 2015 IEEE International Conference on Cloud Engineering, Tempe, AZ, USA, 9–13 March 2015; pp. 386–393. [CrossRef]](http://dx.doi.org/10.1109/IC2E.2015.74) 4. [The Linux Foundation. Kubernetes: Production-Grade Container Orchestration. Available online: https://kubernetes.io (accessed](https://kubernetes.io) on 11 July 2023). 5. [Docker. Swarm Mode Overview. Available online: https://docs.docker.com/engine/swarm/ (accessed on 11 July 2023).](https://docs.docker.com/engine/swarm/) 6. [The OpenStack Project. OpenStack: Open Source Cloud Computing Infrastructure. Available online: https://www.openstack.org](https://www.openstack.org) (accessed on 11 July 2023). 7. [Google Kubernetes Engine. Available online: https://cloud.google.com/kubernetes-engine (accessed on 11 July 2023).](https://cloud.google.com/kubernetes-engine) 8. [Amazon Elastic Kubernetes Service (EKS). Available online: https://aws.amazon.com/es/eks/ (accessed on 11 July 2023).](https://aws.amazon.com/es/eks/) 9. [Cloud Native Computing Foundation. CNCF Survey 2021. Available online: https://github.com/cncf/surveys (accessed on](https://github.com/cncf/surveys) 11 June 2023). 10. Ponce, F.; Márquez, G.; Astudillo, H. Migrating from monolithic architecture to microservices: A Rapid Review. In Proceedings of the 2019 38th International Conference of the Chilean Computer Science Society (SCCC), Concepcion, Chile, 4–9 November 2019; [pp. 1–7. [CrossRef]](http://dx.doi.org/10.1109/SCCC49216.2019.8966423) 11. Ren, Z.; Wang, W.; Wu, G.; Gao, C.; Chen, W.; Wei, J.; Huang, T. Migrating Web Applications from Monolithic Structure to Microservices Architecture. In Proceedings of the 10th Asia-Pacific Symposium on Internetware, Beijing, China, 16 September 2018; [Association for Computing Machinery: New York, NY, USA, 2018. [CrossRef]](http://dx.doi.org/10.1145/3275219.3275230) 12. Joy, A.M. Performance comparison between Linux containers and virtual machines. In Proceedings of the 2015 International Conference on Advances in Computer Engineering and Applications, Ghaziabad, India, 19–20 March 2015; pp. 342–346. [[CrossRef]](http://dx.doi.org/10.1109/ICACEA.2015.7164727) 13. Moravcik, M.; Segec, P.; Kontsek, M.; Uramova, J.; Papan, J. Comparison of LXC and Docker Technologies. In Proceedings of the 2020 18th International Conference on Emerging eLearning Technologies and Applications (ICETA), Kosice, Slovenia, [12–13 November 2020; pp. 481–486. [CrossRef]](http://dx.doi.org/10.1109/ICETA51985.2020.9379212) ----- _Future Internet 2023, 15, 274_ 27 of 28 14. Acar, U.; Ustok, R.F.; Keskin, S.; Breitgand, D.; Weit, A. Programming Tools for Rapid NFV-Based Media Application Development in 5G Networks. In Proceedings of the 2018 IEEE Conference on Network Function Virtualization and Software Defined Networks [(NFV-SDN), Verona, Italy, 27–29 November 2018; pp. 1–5. [CrossRef]](http://dx.doi.org/10.1109/NFV-SDN.2018.8725610) 15. Sairam, R.; Bhunia, S.S.; Thangavelu, V.; Gurusamy, M. NETRA: Enhancing IoT security using NFV-based edge traffic analysis. _[IEEE Sens. J. 2019, 19, 4660–4671. [CrossRef]](http://dx.doi.org/10.1109/JSEN.2019.2900097)_ 16. [Cloud Native Computing Foundation. CNI: The Container Network Interface. Available online: https://www.cni.dev (accessed](https://www.cni.dev) on 11 July 2023). 17. [Flannel: A Simple and Easy Way to Configure a Layer 3 Network Fabric Designed for Kubernetes. Available online: https:](https://github.com/flannel-io/flannel) [//github.com/flannel-io/flannel (accessed on 11 July 2023).](https://github.com/flannel-io/flannel) 18. [Tigera. Calico Open Source. Available online: https://www.tigera.io/project-calico/ (accessed on 11 July 2023).](https://www.tigera.io/project-calico/) 19. Multus: A Container Network Interface (CNI) Plugin for Kubernetes That Enables Attaching Multiple Network Interfaces to [Pods. Available online: https://github.com/k8snetworkplumbingwg/multus-cni (accessed on 11 July 2023).](https://github.com/k8snetworkplumbingwg/multus-cni) 20. Istio Authors. Istio: Simplify Observability, Traffic Management, Security, and Policy with the Leading Service Mesh. Available [online: https://istio.io (accessed on 4 August 2023).](https://istio.io) 21. Envoy Project Authors. ENVOY: An Open Source Edge and Service Proxy, Designed for Cloud-Native Applications. Available [online: https://www.envoyproxy.io (accessed on 4 August 2023).](https://www.envoyproxy.io) 22. [KubeEdge: A Kubernetes Native Edge Computing Framework. Available online: https://kubeedge.io (accessed on 11 July 2023).](https://kubeedge.io) 23. [OpenYurt: An Open Platform That Extends Upstream Kubernetes to Edge. Available online: https://openyurt.io (accessed on](https://openyurt.io) 11 July 2023). 24. [Lightweight Kubernetes: The Certified Kubernetes Distribution Built for IoT & Edge Computing. Available online: https://k3s.io](https://k3s.io) (accessed on 11 July 2023). 25. [Red Hat, Inc. About the OpenShift SDN Network Plugin. Available online: https://docs.openshift.com/container-platform/4.](https://docs.openshift.com/container-platform/4.13/networking/openshift_sdn/about-openshift-sdn.html#about-openshift-sdn) [13/networking/openshift_sdn/about-openshift-sdn.html#about-openshift-sdn (accessed on 5 June 2023).](https://docs.openshift.com/container-platform/4.13/networking/openshift_sdn/about-openshift-sdn.html#about-openshift-sdn) 26. [Akraino. Nodus. Available online: https://github.com/akraino-edge-stack/icn-nodus/tree/master (accessed on 5 June 2023).](https://github.com/akraino-edge-stack/icn-nodus/tree/master) 27. The Linux Foundation. Kube-OVN: The Most Advanced Kubernetes Network Fabric for Enterprises. Available online: [https://www.kube-ovn.io (accessed on 5 June 2023).](https://www.kube-ovn.io) 28. [European Telecommunications Standards Institute (ETSI). Open Source MANO (OSM). Available online: https://osm.etsi.org](https://osm.etsi.org) (accessed on 11 July 2023). 29. Mijumbi, R.; Serrat, J.; Gorricho, J.L.; Bouten, N.; De Turck, F.; Boutaba, R. Network Function Virtualization: State-of-the-Art and [Research Challenges. IEEE Commun. Surv. Tutor. 2016, 18, 236–262. [CrossRef]](http://dx.doi.org/10.1109/COMST.2015.2477041) 30. Lai, W.P.; Wang, Y.H. On the performance impact of virtual link types to 5G networking. In Proceedings of the 2017 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), Kuala Lumpur, Malaysia, [12–15 December 2017; pp. 1470–1474. [CrossRef]](http://dx.doi.org/10.1109/APSIPA.2017.8282265) 31. Karkazis, P.A.; Railis, K.; Prekas, S.; Trakadas, P.; Leligou, H.C. Intelligent Network Service Optimization in the Context of [5G/NFV. Signals 2022, 3, 587–610. [CrossRef]](http://dx.doi.org/10.3390/signals3030036) 32. Uzunidis, D.; Karkazis, P.; Roussou, C.; Patrikakis, C.; Leligou, H.C. Intelligent Performance Prediction: The Use Case of a [Hadoop Cluster. Electronics 2021, 10, 2690. [CrossRef]](http://dx.doi.org/10.3390/electronics10212690) 33. [Network Service Mesh: The Hybrid/Multi-Cloud IP Service Mesh. Available online: https://networkservicemesh.io (accessed](https://networkservicemesh.io) on 11 July 2023). 34. Gonzalez, L.F.; Vidal, I.; Valera, F.; Sanchez-Aguero, V. A Comparative Study of Virtual Infrastructure Management Solutions for UAV Networks. In Proceedings of the 7th Workshop on Micro Aerial Vehicle Networks, Systems, and Applications, Virtual, 24 June–2 July 2021; Association for Computing Machinery: New York, NY, USA, 2021; pp. 13–18. 35. Gonzalez, L.F.; Vidal, I.; Valera, F.; Lopez, D.R. Link Layer Connectivity as a Service for Ad-Hoc Microservice Platforms. IEEE _[Netw. 2022, 36, 10–17. [CrossRef]](http://dx.doi.org/10.1109/MNET.001.2100363)_ 36. Gonzalez, L.F.; Vidal, I.; Valera, F.; Sanchez-Aguero, V.; Nogales, B.; Lopez, D.R. NFV orchestration on intermittently available SUAV platforms: Challenges and hurdles. In Proceedings of the IEEE INFOCOM 2019—IEEE Conference on Computer [Communications Workshops (INFOCOM WKSHPS), Paris, France, 29 April–2 May 2019; pp. 301–306. [CrossRef]](http://dx.doi.org/10.1109/INFCOMW.2019.8845040) 37. Gonzalez, L.F.; Vidal, I.; Valera, F.; Nogales, B.; Sanchez-Aguero, V.; Lopez, D.R. Transport-Layer Limitations for NFV Orchestra[tion in Resource-Constrained Aerial Networks. Sensors 2019, 19, 5220. [CrossRef] [PubMed]](http://dx.doi.org/10.3390/s19235220) 38. Nogales, B.; Sanchez-Aguero, V.; Vidal, I.; Valera, F.; Garcia-Reinoso, J. A NFV System to Support Configurable and Automated Multi-UAV Service Deployments. In Proceedings of the 4th ACM Workshop on Micro Aerial Vehicle Networks, Systems, and Applications, Munich, Germany, 10–15 June 2018; Association for Computing Machinery: New York, NY, USA, 2018; pp. 39–44. [[CrossRef]](http://dx.doi.org/10.1145/3213526.3213534) 39. Mahalingam, M.; Dutt, D.; Duda, K.; Agarwal, P.; Kreeger, L.; Sridhar, T.; Bursell, M.; Wright, C. Virtual eXtensible Local Area Network (VXLAN): A Framework for Overlaying Virtualized Layer 2 Networks over Layer 3 Networks. RFC 7348, 2014. [Available online: https://www.rfc-editor.org/rfc/rfc7348.txt (accessed on 16 August 2023).](https://www.rfc-editor.org/rfc/rfc7348.txt) 40. Li, T.; Farinacci, D.; Hanks, S.P.; Meyer, D.; Traina, P.S. Generic Routing Encapsulation (GRE). RFC 2784, 2000. Available online: [https://www.rfc-editor.org/rfc/rfc2784.txt (accessed on 16 August 2023).](https://www.rfc-editor.org/rfc/rfc2784.txt) ----- _Future Internet 2023, 15, 274_ 28 of 28 41. [VMWare. What is Single Root I/O Virtualization (SR-IOV). Available online: https://docs.vmware.com/en/VMware-vSphere/](https://docs.vmware.com/en/VMware-vSphere/8.0/vsphere-networking/GUID-CC021803-30EA-444D-BCBE-618E0D836B9F.html) [8.0/vsphere-networking/GUID-CC021803-30EA-444D-BCBE-618E0D836B9F.html (accessed on 11 June 2023).](https://docs.vmware.com/en/VMware-vSphere/8.0/vsphere-networking/GUID-CC021803-30EA-444D-BCBE-618E0D836B9F.html) 42. [Open Networking Foundation (ONF). OpenFlow Switch Specification v1.0–v1.5. Available online: https://opennetworking.org/](https://opennetworking.org/software-defined-standards/specifications/) [software-defined-standards/specifications/ (accessed on 18 April 2023).](https://opennetworking.org/software-defined-standards/specifications/) 43. [Open Networking Foundation (ONF). P4 Open Source Programming Language. Available online: https://p4.org (accessed on](https://p4.org) 18 April 2023). 44. [The Linux Foundation. Operator Pattern. Available online: https://kubernetes.io/docs/concepts/extend-kubernetes/operator/](https://kubernetes.io/docs/concepts/extend-kubernetes/operator/) (accessed on 28 June 2023). 45. Gonzalez, L.F.; Vidal, I.; Valera, F.; Lopez, D.R. Link-Layer Secure Connectivity for Microservice Platforms (L2S-M). Available [online: http://l2sm.io (accessed on 11 July 2023).](http://l2sm.io) 46. Open Networking Foundation. Open Network Operating System (ONOS), Open Source SDN Controller for Building Next[Generation SDN/NFV Solutions. Available online: https://opennetworking.org/onos (accessed on 11 July 2023).](https://opennetworking.org/onos) 47. [European H2020 LABYRINTH Project. Ensuring Drone Traffic Control and Safety. Available online: https://labyrinth2020.eu/](https://labyrinth2020.eu/) (accessed on 11 July 2023). 48. European H2020 FISHY Project. A Coordinated Framework for Cyber Resilient Supply Chain Systems Over Complex ICT [Infrastructures. Available online: https://fishy-project.eu/ (accessed on 11 July 2023).](https://fishy-project.eu/) 49. Gonzalez, L.F.; Vidal, I.; Valera, F.; Nogales, B.; Lopez, D.R. Feature 10921: Connectivity among CNFs Using SDN. Available [online: https://osm.etsi.org/gitlab/osm/features/-/issues/10921 (accessed on 11 July 2023).](https://osm.etsi.org/gitlab/osm/features/-/issues/10921) 50. [Helm Authors. HELM: The Package Manager for Kubernetes. Available online: https://helm.sh (accessed on 6 August 2023).](https://helm.sh) 51. [F5 Inc. Nginx Documentation. Available online: https://nginx.org/en/docs/ (accessed on 18 April 2023).](https://nginx.org/en/docs/) 52. [The Linux Foundation. Creating a Cluster with Kubeadm. Available online: https://kubernetes.io/docs/setup/production-](https://kubernetes.io/docs/setup/production-environment/tools/kubeadm/create-cluster-kubeadm/) [environment/tools/kubeadm/create-cluster-kubeadm/ (accessed on 18 April 2023).](https://kubernetes.io/docs/setup/production-environment/tools/kubeadm/create-cluster-kubeadm/) 53. [Gonzalez, L.F.; Vidal, I.; Valera, F.; Artin, R.M.; Artalejo, D. Smart Campus Scenario. Available online: https://github.com/](https://github.com/Networks-it-uc3m/Smart-Campus-Scenario) [Networks-it-uc3m/Smart-Campus-Scenario (accessed on 6 August 2023).](https://github.com/Networks-it-uc3m/Smart-Campus-Scenario) 54. [Dugan, J.; Elliott, S.; Mah, B.A.; Poskanzer, J.; Prabhu, K. What Is iPerf/iPerf3? Available online: https://iperf.fr/ (accessed on](https://iperf.fr/) 11 June 2023). **Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual** author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/fi15080274?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/fi15080274, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/1999-5903/15/8/274/pdf?version=1692267150" }
2,023
[ "JournalArticle" ]
true
2023-08-17T00:00:00
[ { "paperId": "3a545977e316c1f52926a2d2b17d7167f1828b95", "title": "Intelligent Network Service Optimization in the Context of 5G/NFV" }, { "paperId": "42db85b2dc52997efa9a6c8882298e8835e16d40", "title": "Link Layer Connectivity as a Service for Ad-Hoc Microservice Platforms" }, { "paperId": "26c83e011c74d869ba715b859974fb45d498df91", "title": "Intelligent Performance Prediction: The Use Case of a Hadoop Cluster" }, { "paperId": "0107278d781099481d35f2ed0207632144cbb614", "title": "Transport-Layer Limitations for NFV Orchestration in Resource-Constrained Aerial Networks" }, { "paperId": "c87f0010784bf6e34f4713163b8ec8f0f4645d0b", "title": "Migrating Web Applications from Monolithic Structure to Microservices Architecture" }, { "paperId": "47c199393b70fb09d169a3c74ef6c4b57c22d3a1", "title": "A NFV system to support configurable and automated multi-UAV service deployments" }, { "paperId": "dba65f8c9d379dad42583c3bbbec8685de9fd0a7", "title": "NETRA: Enhancing IoT Security Using NFV-Based Edge Traffic Analysis" }, { "paperId": "340c6b4f57b8125f644c0e937934b0c6b23603ac", "title": "Network Function Virtualization: State-of-the-Art and Research Challenges" }, { "paperId": "5bd0abaa8c3fe1b4cff21b93db62bcf7a53fee44", "title": "Virtual eXtensible Local Area Network (VXLAN): A Framework for Overlaying Virtualized Layer 2 Networks over Layer 3 Networks" }, { "paperId": "0312d85340352f014ec44ae34898fff58df4dbea", "title": "Generic Routing Encapsulation (GRE)" } ]
25,151
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0109b47dabb716f4886a9c4317735459dbf5b13b
[ "Computer Science" ]
0.878876
MONSOON: A Coevolutionary Multiobjective Adaptation Framework for Dynamic Wireless Sensor Networks
0109b47dabb716f4886a9c4317735459dbf5b13b
Proceedings of the 41st Annual Hawaii International Conference on System Sciences (HICSS 2008)
[ { "authorId": "2055747", "name": "P. Boonma" }, { "authorId": "122573052", "name": "J. Suzuki" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
# MONSOON: A Coevolutionary Multiobjective Adaptation Framework for Dynamic Wireless Sensor Networks ## Pruet Boonma and Junichi Suzuki Department of Computer Science University of Massachusetts, Boston pruet, jxs @cs.umb.edu { } ## Abstract _Wireless sensor applications (WSNs) are often required_ _to simultaneously satisfy conflicting operational objectives_ _(e.g., latency and power consumption). Based on an obser-_ _vation that various biological systems have developed the_ _mechanisms to overcome this issue, this paper proposes a_ _biologically-inspired adaptation mechanism, called MON-_ _SOON. MONSOON is designed to support data collection_ _applications, event detection applications and hybrid appli-_ _cations. Each application is implemented as a decentralized_ _group of software agents, analogous to a bee colony (appli-_ _cation) consisting of bees (agents). Agents collect sensor_ _data and/or detect an event (a significant change in sen-_ _sor reading) on individual nodes, and carry sensor data to_ _base stations. They perform these data collection and event_ _detection functionalities by sensing their surrounding en-_ _vironment conditions and adaptively invoking biologically-_ _inspired behaviors such as pheromone emission, reproduc-_ _tion and migration. Each agent has its own behavior pol-_ _icy, as a gene, which defines how to invoke its behaviors._ _MONSOON allows agents to evolve their behavior policies_ _(genes) and adapt their operations to given objectives. Sim-_ _ulation results show that MONSOON allows agents (WSN_ _applications) to simultaneously satisfy conflicting objec-_ _tives by adapting to dynamics of physical operational envi-_ _ronments and network environments (e.g., sensor readings_ _and node/link failures) through evolution._ ## 1. Introduction Autonomous adaptability is a key challenge in wireless sensor networks (WSNs) [1–4]. With minimal intervention to/from human operators, WSN applications are required to adapt their operations to dynamic changes in physical operational environments (e.g., sensor readings) and network environments (e.g., network traffic and node/link failures). A critical issue in this challenge is that each WSN application tends to have conflicting operational objectives. For example, the success rate of data transmissions from individual nodes to base stations is an important objective because higher success rate ensures that base stations have more data for operators to better understand a physical oper ational environment and make better informed decisions. At the same time, the latency of data transmissions from individual nodes to base stations is another important objective. Lower latency ensures that base stations can collect sensor data for operators to understand a physical operational environment more quickly and make more timely decisions. Success rate and latency conflict with each other. For improving success rate, hop-by-hop recovery is often applied; however, this can degrade latency. For improving latency, nodes may transmit data to base stations with the shortest paths; however, success rate can degrade because of traffic congestion on the paths. In order to address this adaptability issue, the authors of the paper envision autonomous WSN applications that understand their operational objectives and simultaneously satisfy them against the dynamics of network environments. Toward this vision, the authors observe that various biological systems have developed the mechanisms to overcome the above adaptability issue. For example, each bee colony autonomously satisfies conflicting objectives to maintain its well-being [5]. Those objectives include maximizing the amount of collected honey, maintaining temperature inside a nest and minimizing the number of dead drones. If bees focus only on foraging, they fail to ventilate their nest and remove dead drones. Given this observation, the proposed application architecture, called BiSNET/e (Biologically-inspired architecture for Sensor NETworks, evolutionary edition), applies key biological mechanisms to implement adaptive WSN applications. Figure 1 shows the BiSNET/e runtime architecture. The BiSNET/e runtime operates atop TinyOS on each node. It consists of two software components: agents and middle_ware platforms, which are modeled after bees and flowers,_ respectively. Each WSN application is designed as a decentralized group of agents. This is analogous to a bee colony (application) consisting of bees (agents). Agents collect sensor data and/or detect an event (a significant change in sensor reading) on platforms (flowers) atop individual nodes. Then, they carry sensor data to base stations, in turn, to a backend server (the MONSOON server in Figure 1), which is modeled after a nest of bees. Agents perform ----- |Col1|Col2| |---|---| |EAs|DA| these data collection and event detection functionalities by autonomously sensing their surrounding environment conditions and adaptively performing biological behaviors such as pheromone emission, reproduction, migration, swarming and death. A middleware platform runs on each node, and hosts an arbitrary number of agents (Figure 1). It provides a series of runtime services that agents use to perform their functionalities and behaviors. This paper describes a key mechanism in BiSNET/e, called MONSOON[1], which is an co-evolutionary adaptation framework for agents. Each agent possesses its own behavior policy, as a gene, which defines how to invoke its behaviors. MONSOON allows agents to evolve their behavior policies via genetic operations (mutation and crossover) across generations and simultaneously adapt the behavior policies to conflicting objectives in dynamic physical operational environments and network environments. Currently, MONSOON considers three objectives: success rate, latency and power consumption. The evolution process in MONSOON frees application designers from anticipating all possible environment conditions and tuning their agents’s behavior policies to the conditions at design time. Instead, agents can autonomously evolve and tune their behavior policies. This significantly simplifies the implementation and maintenance of agents (i.e., WSN applications). MONSOON supports data collection applications, event detection applications and hybrid applications. Different types of applications are implemented with different types of agents. Data collection and event detection applications are implemented with data collection agents (DAs) and _event detection agents (EAs), respectively. Both DAs and_ EAs are used to implement hybrid applications, which perform both data collection and event detection. In hybrid applications, DAs and EAs coevolve and adapt their behavior policies (genes) in a symbiotic manner. EAs helps DAs improve their behavior policies, and vice versa. This paper is organized as follows. Section 2 overviews the BiSNET/e runtime, and Section 3 describes the design of MONSOON. Section 4 evaluates MONSOON with a series of simulation results. Simulation results show that MONSOON allows agents (WSN applications) to simul 1Multiobjective Optimization for Network of Sensors using a cOevOlutionary mechaNism taneously satisfy conflicting objectives by adapting to dynamics of physical operational environments and network environments (e.g., sensor readings and node/link failures) through evolution. Sections 5 and 6 conclude with some discussion on related work. ## 2. The BiSNET/e Runtime At the beginning of a WSN’s operation, one DA and one EA are deployed on each node. They have randomlygenerated behavior policies. A DA collects sensor data on each node periodically (i.e., at each data collection cycle) and carry the data to a base station on a hop-by-hop basis. An EA collects sensor data on each node periodically, and if it detects an event (i.e., a significant change in sensor data), carries the data to a base station on a hop-by-hop basis. If an event is not detected, the EA discards the data. ### 2.1. Agent Structure and Behaviors Each agent consists of attributes, body and behaviors. _Attributes carry descriptive information on an agent. They_ include agent type (i.e., EA or DA), behavior policy (gene), sensor data to be reported to a base station, the data’s time stamp, and the ID of a node where the data is collected. _Body implements the functionalities of an agent: collect-_ ing and processing sensor data (e.g., discarding it or reporting it to a base station). _Behaviors implement actions inherent to all agents. Sim-_ ilar to biological entities (e.g., bees), agents sense their surrounding environment conditions and behave according to the sensed conditions without any intervention from/to other agents, platforms, base stations and human operators. This paper focuses on the following seven behaviors. **(1) Food gathering and consumption: Biological enti-** ties strive to seek food for living. For example, bees gather nectar to produce honey. Similarly, each agent periodically reads sensor data (as nectar) to gain energy (as honey)[2], and consumes a constant amount of energy for living. **(2) Pheromone emission: Agents may emit different** types of pheromones: migration and alert pheromones. They emit migration pheromones on their local nodes when they migrate to neighboring nodes. Each migration pheromone references the destination node an agent has migrated to. Agents also emit alert pheromones when they fail migrations within a timeout period. Each alert pheromone references a possibly failed node that an agent could not migrate to. Each pheromone has its own concentration, which decays by half at every data collection cycle. A pheromone disapears when its concentration becomes zero. **(3) Replication: EAs may make a copy of themselves** in response to the abundance of stored energy, while DAs always make a copy of themselves in each data collection 2The concept of energy in BiSNET/e does not represent the amount of physical battery in a node. It is logically affects agent behaviors. ----- cycle. A replicated (child) agent is placed on the node that its parent resides on, and it inherits the parent’s agent type and behavior policy (gene). Replicated agents are intended to move toward base stations to report collected sensor data. **(4) Migration: Agents may move from one node to an-** other. Migration is used to transmit agents (sensor data) to base stations. Each agent chooses a migration destination node by sensing three types of pheromones available on the local node: base station, migration and alert pheromones. Each base station periodically propagates base station _pheromones to individual nodes in the network. Their con-_ centration decays on a hop-by-hop basis. Using base station pheromones, agents can sense where base stations exist approximately, and move toward the base stations by climbing pheromone’s concentration gradient[3]. An agent may move to a base station by following a migration pheromone trace on which many other agents have traveled. The trace can be the shortest path to the base station. Conversely, an agent may goes off a migration pheromone trace and follows another path to a base station when the concentration of migration pheromones is too high on the trace (i.e., when too many agents have followed the trace). This avoids separating the network into islands. The network can be separated with the migration paths that too many agents follow, because the nodes on the paths consume more power and go down earlier than the others. An agent may also avoid moving to a node referenced by an alert pheromone. This allows agents to reach base stations by bypassing link/node failures. **(5) Swarming: Agents may swarm (or merge) with oth-** ers on their ways to base stations. Multiple agents become a single agent. (A DA can merge with both DAs and EAs, and an EA can merge with both EAs and DAs.) The resulting agent (swarm) aggregates sensor data contained in other agents, and uses the behavioral policy of the best agent in the swarm in terms of latency and power consumption. This data aggregation saves power consumption of nodes because in-node data processing requires much less power consumption than data transmission does. **(6) Reproduction: Once agents arrive at the MON-** SOON server ( Figure 1), they are evaluated according to their objectives. Then, MONSOON selects best-performing (or elite) agents, and propagates them to individual nodes. An agent running on each node performs reproduction with one of the propagated agents. A reproduced agent inherits a behavior policy (gene) from its parents via crossover, and mutation may occur on the inherited behavior policy. Reproduced agents perform a generation change by taking over existing agents running on individual nodes. Reproduction is intended to evolve agents so that the agents that fit better to the environment become more abun 3Base station pheromones are designed after the Nasonov gland pheromone, which guides bees to move toward their nest [6]. dant. It retains the agents whose fitness to the current network conditions is high (i.e., the agents that have effective behavior policies, such as moving toward a base station in a short latency), and eliminates the agents whose fitness is low (i.e., the agents that have ineffective behavior policies, such as consuming too much power to reach a base station). Through successive generations, effective behavior policies become abundant in agent population while ineffective ones become dormant or extinct. This allows agents to adapt to dynamic network conditions. **(7) Death: Agents periodically consume energy for liv-** ing, and expend energy to invoke their behaviors. (The energy costs to invoke behaviors are constant for all agents.) Agents die due to lack of energy when they cannot balance energy gain and expenditure. The death behavior is intended to eliminate the agents that have ineffective behavior policies. For example, an agent would die before arriving at a base station if it follows a too long migration path. When an agent dies, the local platform removes the agent and releases all resources allocated to the agent. ### 2.2. Behavior Sequences for DAs and EAs Figures 2 and 3 show a sequence of behaviors that each DA and EA perform on a node in each data collection cycle. A DA reads sensor data (as nectar) with the underlying sensor device and gains a constant amount of energy (as honey). Given the energy intake (EF), each agent updates its energy level as follows. _E(t) = E(t −_ 1) + EF (1) _E(t) is the current energy level of the DA, and E(t −_ 1) is the DA’s energy level in the previous data collection cycle. _t is incremented by one at each data collection cycle._ If a DA’s (E(t)) becomes very low (below the death threshold: TD), the DA dies due to starvation[4]. A DA replicates itself in each data collection cycle. A replicating (parent) agent splits its energy units to halves ( _[E][(][t][)]2[−][E][R]_ ), gives a half to its child agent, and keeps the other half. ER is the energy cost for an agent to perform the replication behavior. A child agent contains the sensor data that its parent collected, and carries it to a base station. Each replicated DA migrates toward a base station on a hop by hop basis. On each intermediate node, it examines Equation 2 to determine which next node it migrates to. An DA calculates this weighted sum (WS j) for each neighboring node j, and moves to a node that generates the highest weighted sum. t denotes pheromone type; P1 _j,_ 4If all agents are dying on a node at the same time, a randomly selected agent for each type (i.e., EA and DA) will survive. At least one agent of each type runs on each node. _WS j =_ �t=31 _wt_ _PPtmaxt,_ _j − −PPtmintmin_ (2) ----- _P2 j and P3 j represent the concentrations of base station,_ migration and alert pheromones on the node j. Ptmax and _Ptmin denote the maximum and minimum concentration of_ _Pt among neighboring nodes._ When a DA is migrating to a neighboring node, it emits a migration pheromone on the local node. If the DA’s migration fails, it emits an alert pheromone. Each alert pheromone spreads to one-hop away neighboring nodes. **for each data collection cycle** Read sensor data and gain energy (EF ). Update energy level (E(t)). **if E(t) < the death threshold (TD)** **then Invoke the death behavior.** Invoke the replication behavior to make a child agent. Give the half of the current energy level to a replicated (child) agent. **do** **for each migrating agent** Determine the destination node of migration. Emit a migration pheromone on the local node. Migrate to a neighboring node. **do** **if Migration fails** �Emit an alert pheromone on the local node. **then**   Propagate it to neighboring nodes. **Figure 2. A Sequence of DA Behaviors in** **Each Data Collection Cycle** is used to smooth out short-term minor oscillations in the data series of E. It places more emphasis on the long-term transition trend of E; only significant changes in E have the effects to change TR. The α value is a constant to control the responsiveness of EWMA against the changes of E. Similar to DAs, a parent EA splits its energy units to halves, gives a half to its child agent, and keeps the other half. The EA keeps replicating itself until its energy level becomes less than its TR. A child agent contains the sensor data that its parent collected, and carries it to a base station. EAs perform the migration behavior with Equation 2 in the same way as DAs do. ### 2.3 Agent Behavior Policy EAs and DAs have the same structure for behavior policies (genes). Each behavior policy contains a set of weight values in Equation 2 (wt, 1 ≤ _t ≤_ 3). w1 and w3 are non negative, and w2 can be negative. These weight values govern how agents perform the migration behavior. For example, if an agent has zero for w2 and w3, the agent ignores migration and alert pheromones, and moves toward the base stations by climbing the concentration gradient of base station pheromones. If an agent has a positive value for w2, it follows a migration pheromone trace on which many other agents have traveled. A negative w2 value allows an agent to go off a migration pheromone trace and follow another path toward a base station. If an agent has a positive w3, it moves to a base station by bypassing link/node failures. ## 3. MONSOON MONSOON is a coevolutionary multiobjective adaptation mechanism designed for agents in BiSNET/e. It allows agents to heuristically adapt to multiple objectives simultaneously. This adaptation process is performed through elite _selection and genetic operations. The elite selection process_ evaluates each type of agents (DAs and EAs) that arrive at base stations, based on given objectives, and chooses the best (or elite) ones. Elite agents are propagated to the network in order to perform genetic operations and reproduce an offspring (next generation) agent on each node. Elite selection is performed in the MONSOON server (see Figure 1), and genetic operations are performed in each node. ### 3.1. Operational Objectives Agents (DAs and EAs) consider three conflicting objectives: latency, cost and success rate of their migration (i.e., data transmission) from individual nodes to base stations. **(1) Latency represents the time required for an agent** (DA or EA) to travel to a base station from a node where the agent is born (replicated). As depicted below, latency is measured as a ratio of this agent travel time to the physical distance (PD) between a base station and a node where the **while true** Read sensor data and gain energy (EF ). Update energy level (E(t)). **if E(t) < the death threshold (TD)** **then Invoke the death behavior.** **while E(t) > the replication threshold (TR(t))** �Invoke the replication behavior to make a child agent. **do** Give the half of the current energy level to the child agent. **do** **for each migrating agent** Determine the destination node of migration. Emit a migration pheromone on the local node. Migrate to a neighboring node. **do** **if Migration fails** �Emit an alert pheromone on the local node. **then**   Propagate it to neighboring nodes. **Figure 3. A Sequence of EA Behaviors** When an EA reads sensor data (as nectar) with the underlying sensor device and gains energy (as honey), its current energy level (E(t)) is updated with Equation 3. _E(t) = E(t −_ 1) + _S · M_ (3) _S represents the absolute difference between the current_ and previous sensor data. M is metabolic rate, which is a constant value between 0 and 1. Each EA replicates itself if its energy level exceeds the replication threshold: TR(t) (Figure 3). The replication threshold is continuously adjusted as EWMA (Exponentially Weighted Moving Average) of each EA’s energy level: _TR(t) = (1_ − α)TR(t − 1) + αE(t) (4) _TR(t) is the current replication threshold, and TR(t −_ 1) is the one in the previous data collection period. EWMA ----- agent is born. The MONSOON server knows the location of each node with a certain localization mechanism. _Latency =_ _[Agent travel time][ (][sec][)]_ (5) _PD (meter)_ **(2) Cost represents the amount of power consumption** required for an agent (DA or EA) to travel to a base station from a node where the agent is born. It is measured with the total number of data transmissions, each node’s radio transmission range (radius), and PD. _Cost =_ _[Total][ #][ of data transmissions]_ (6) _Transmission range/PD_ The total number of data transmissions include successful and unsuccessful (failed) agent migrations as well as the transmissions of migration or alert pheromones. **(3) Success Rate is measured differently for DAs and** EAs. For DAs, it is measured as follows. _S uccess rateDA =_ [#][ ofagents that arrive at base stations] (7) _The total # of nodes_ For EAs, success rate is measured as follows. is divided into small cubes. Each non-nominated agent is plotted in this hypercube space based on their objective values. A single agent is randomly selected from each cube as an elite agent. This elite selection is designed to maintain the diversity of elite agents’ genes. The diversification of agent genes contribute to improve agents’ adaptation even to unanticipated network conditions. Figure 5 shows an example hypercube space. Each axis is divided into two ranges; therefore, eight cubes exist in total. Thus, the maximum number of elite agents is eight. In this example, six (A to F) non-dominated agents are plotted in the hypercube space. Three agents (B, C, and D) are plotted in the lower left cube, while the other three agents (A, E, and F) are plotted in three different cubes. From the lower left cube, only one agent is randomly selected as an elite agent. A, E, and F are selected as elite agents because they are in different cubes. Latency # of successful agent migrations _S uccess rateEA =_ (8) _The total # of attempts of agent migrations_ ### 3.2. Elite Selection Figure 4 shows how elite selection occurs at the MONSOON server in each data collection cycle. The MONSOON server performs the same selection process for EAs and DAs separately. The first step is to obtain three objective values (i.e., latency, cost and success rate) from each of the agents that reach the MONSOON server via base stations. Then, each agent is evaluated whether it is dominated by another agent. An agent is considered to be dominated if another agent outperforms it in all of three objectives. Empty the archive **for each data collection cycle** Empty the population pool. Collect agents from the network. Add collected agents to the population pool. Move agents from the archive to the population pool. Empty the archive **do** **for each agent of the ones in the population pool** **if not dominated by all other agents in** **do** the population pool  **then Add the agent to the archive.** Select elite agents from the archive. Propagate elite agents to the network. **Figure 4. Elite Selection in MONSOON** In the next step, a subset of non-dominated agents are selected as elite agents. This is performed with a hypercube space, which a three dimensional space whose axes represent three objectives (i.e., latency, cost and success rate). Each axis of the hypercube space is divided so that the space |A B C D uccess Rate (Maximize)|(Minimize) Non-dominated agent F E Cost (Minimize)| |---|---| **Figure 5. An Example Elite Selection** ### 3.3. Genetic Operations Once elite DAs and EAs are selected, the MONSOON server propagates them to each node in the network. They are propagated with base station pheromones. Based on a certain reproduction probability, an agent performs the reproduction behavior on each node through genetic operations (crossover and mutation) when elite agents arrive at the node. As a mating partner, the agent selects one of the elite agents that has the most similar gene. Gene similarity is measured with the Euclidean distance between the values of two genes. DAs can mate with elite EAs, and EAs can mate with elite DAs. This cross-mating allows DAs and EAs to coevolve their behavior policies; DAs can improve EAs’ genes, and vice versa. During reproduction, an agent inherits the half of its gene from its parent agent and the other half from its parent’s mating partner. Mutation occurs on the child agent’s gene with a certain mutation probability by randomly changing gene values within a predefined value range. ## 4. Simulation Results This section shows a set of simulation results to evaluate MONSOON. It is evaluated with a data collection applica Success Rate ----- 100 90 80 70 60 50 Cost 40 Latency Success Rate 30 20 10 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 **Simulation Ticks** 2 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 **Generations** **(a) Static Network** tion (Section 4.1), event detection application (Section 4.2) and hybrid application (Section 4.3). A simulated WSN consists of 100 nodes uniformly deployed in an observation area of 300x300 square meters. Each node’s communication range is 30 meters. A base station is deployed on the northwestern corner of the observation area. The base station links the MONSOON server via emulated serial port connection. All the software components in the BiSNET/e runtime are implemented in nesC, and the MONSOON server is implemented in Java. Simulation time is counted with ticks. Each tick represents five minutes. In genetic operations, the reproduction probability is 0.75, and the mutation probability is 0.025. ### 4.1 Data Collection Application A data collection application is implemented with DAs that perform the sequence of behaviors shown in Figure 2. No EAs are used in this application. The data collection cycle corresponds to a simulation tick (five minutes). Figure 6 (a) shows the average objective values produced by DAs at each simulation tick. Each objective value gradually improves and converges at the 17th tick. This simulation result shows that MONSOON allows DAs to simultaneously satisfy conflicting objectives by evolving their behavior policies. Figures 8, and 9 and 10 show the objective values that elite DAs produced at the 20th tick. Since each objective value’s change is less than 1% from the 17th to 20th tick, it is fair enough to say that the elite DAs are on the Pareto front at the 20th tick. Figures 8, and 9 and 10 plot the elite DAs in three different perspectives: latency-cost, costsuccess rate, and latency-success rate perspectives. Each gray dot represents an elite DA, and a black dot represents overlapping elite DAs. These figures demonstrate that elite agents are well diversified as intended by an elite selection process described in Section 3.2. Figure 6 (b) shows how the performance of DAs changes against a dynamic node addition. 25 nodes are added at random locations at the 20th tick. Upon this change in the network environment, objective values degrade dramatically because DAs have randomly-generated behavior policies on the new nodes. Those DAs cannot migrate efficiently toward the base station. Also, enough pheromones are not available on new nodes; DAs cannot make proper migration decisions when they move to the new nodes. However, DAs gradually improve their performance again, and objective values converge again at the 43th tick. MONSOON allows DAs to autonomously recover application performance despite dynamic node addition by evolving their behavior policies. Figure 6 (c) shows how the performance of DAs changes against dynamic node failures. 25 nodes randomly fail at the 20th tick. Objective values degrade because some DAs try to migrate to failed nodes referenced by migration 2 1.8 1.6 1.4 1.2 1 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0.8 0.6 0.4 0.2 0 100 90 80 70 60 50 40 30 20 10 0 100 90 80 70 60 50 40 30 20 10 0 **(a) Static Network** 2 100 90 80 70 60 50 40 30 20 10 0 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 **Simulation Ticks** 2 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 **Simulation Ticks** **(b) Node Addition** 1.8 1.6 1.4 1.2 1 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0.8 0.6 0.4 0.2 0 100 90 80 70 60 50 40 30 20 10 0 100 90 80 70 60 50 40 30 20 10 0 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 **Simulation Ticks** **(c) Random Node Failure** 2 **(b) Node Addition** 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 **Simulation Ticks** 0.8 0.6 0.4 0.2 0 2 1.8 1.6 1.4 1.2 1 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0.8 0.6 0.4 0.2 0 100 90 80 70 60 50 40 30 20 10 0 100 90 80 70 60 50 40 30 20 10 0 **(c) Random Node Failure** 2 2 1.8 1.6 1.4 1.2 1 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 **Simulaiton Ticks** **(d) Selective Node Failure** 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 **Simulation Ticks** **(d) Selective Node Failure** 2 100 90 80 70 60 50 40 30 20 10 0 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 1.8 1.6 1.4 1.2 1 100 90 80 70 60 50 40 30 20 10 0 0.8 0.6 0.4 0.2 0 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 **Simulation Ticks** 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 **Simulation Ticks** **(e) Base Station Failure** **Figure 6. Objective Values of DAs** **without EAs** **(e) Base Station Failure** **Figure 7. Objective Values of EAs** **without DAs** ----- 1.00 0.99 0.98 0.97 0.96 0.95 0.94 0.93 0.92 0.91 0.90 0.0 0.2 0.4 0.6 0.8 1.0 1.2 **Latency (Minimize)** ### Figure 10. Latency-Success Rate Objective Values on the Pareto Front 1.20 1.15 1.10 1.05 1.00 0.95 0.0 0.2 0.4 0.6 0.8 1.0 1.2 **Latency (Minimize)** ### Figure 8. Latency-Cost Objective Values on the Pareto Front 1.00 0.99 0.98 0.97 0.96 0.95 0.94 0.93 0.92 0.91 0.90 0.95 1.00 1.05 1.10 1.15 1.20 **Cost (Minimize)** ### Figure 9. Cost-Success Rate Objective Values on the Pareto Front pheromones. This increases the number of unsuccessful agent migrations. However, DAs gradually improve their performance again, and objective values converge again at the 45th tick. MONSOON allows DAs to autonomously recover application performance despite dynamic node failures by evolving their behavior policies. Figure 6 (d) shows how the performance of DAs changes when nodes selectively fail in a specific area. At the 20th tick, 20 nodes fail in the middle of WSN observation area. Hence, a WSN has a hole in its middle area. Compared with Figure 6 (c), it takes longer time for DAs to recover their performance. Objective values converge at 50th tick again. The converged cost and latency are worse than the ones at the 20th tick because DAs have to detour a hole (i.e., a set of failed nodes) and take longer migration paths to the base station. This simulation results shows that MONSOON allows DAs to survive selective node failures through evolution. Figure 6 (e) shows how the performance of DAs changes against base station failures. In this simulation scenario, two base stations are deployed at the northwestern and southeastern corners of WSN observation area. At the 20th tick, a base station at the southeastern corner fails. Objective values degrade because some DAs try to migrate toward the failed base station referenced by base station pheromones. This increases the number of unsuccessful agent migrations. However, DAs gradually improve their performance again, and objective values converge again at the 37th tick. MONSOON allows DAs to autonomously evolve and recover application performance despite dynamic base station failures. ### 4.2 Event Detection Application An event detection application is implemented with EAs that perform the sequence of behaviors shown in Figure 3 in every simulation tick. No DAs are used in this application. This simulation study simulates an event, which occurs in the middle of WSN observation area at the 50th tick and radially spreads over time. Figure 7 (a) shows the average objective values at each simulation tick. Upon an event detection, objective values are low because EAs use random behavior policies at first. However, each objective value gradually improves and converges at the 45th tick. This simulation result shows that MONSOON allows EAs to simultaneously satisfy conflicting objectives by evolving their behavior policies. Figure 7 (b) shows how the performance of EAs changes against a dynamic node addition. 25 nodes are added at random locations at the 50th tick. Upon this environmental change, objective values degrade slightly because EAs have randomly-generated behavior policies on the new nodes. Those EAs cannot migrate efficiently toward the base station. However, EAs gradually improve their performance immediately, and objective values converge again at the 70th tick. MONSOON allows EAs to autonomously recover application performance despite dynamic node addition by evolving their behavior policies. Figure 7 (c) shows how the performance of EAs changes against dynamic node failures. 25 nodes randomly fail at the 50th tick. Objective values degrade slightly because some EAs try to migrate to failed nodes referenced by migration pheromones. This increases the number of unsuccessful agent migrations. However, EAs gradually improve their performance again, and objective values converge again at the 72th tick. MONSOON allows EAs to autonomously recover application performance despite dynamic node failures by evolving their behavior policies. Figure 7 (d) shows the result of a simulation when 20 sensor nodes are selected in selective fashion, i.e. create a hole in the middle of network, to be deactivated at the 50th tick. Hence, the sensor network contains a hole in the middle of the network. Compared with the result in figure 7 (c), MONSOON takes longer time to improve the performance of the WSN. The success rate converges at about the 75th tick to approximately 38%. The cost and latency also show the similar trend. Particularly, after the 52nd tick, the av ----- 100 90 80 70 60 50 40 30 20 10 0 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 erage value of cost and latency are higher than the values just before the 20th tick because agents have to detour in a longer path to avoid the hole in the middle of the network. The simulation results shows that MONSOON allows WSN to survives a selective sensor nodes failure by adjusting the operational parameters of WSN to be suitable to the changes in network condition. Figure 7 (e) shows the result of a simulation which initially has two base stations deployed at the northwestern and southeastern corner of the observation area. Then, at the 50th tick, the base station at the southeastern corner is deactivated. From the figure, at the 51st tick, the success rate drops sharply to about 20% from around 50% in the 50th tick because more than a half of the agents still try to move to the base station at the southeastern corner. However, the success rate is improved successively and reach the same level as before the base station is deactivated at the 66th tick. Cost and latency show the same trend. MOSOON allows WSN to survives a base station failure by autonomously directing all agents to the remaining base station. ### 4.3 Hybrid Application This section represents simulation results from a sensor network with two application deployed simultaneously. Figure 11 shows the average objective values from collected DAs, i.e. for data collection application, in each simulation ticks. On the other hand, figure 12 shows the average objective values from collected EAs, i.e. for event collection application, in each simulation ticks. In the figure 12 (a), at 50th simulation ticks, oil spill happens and EAs start detecting and moving to the base station. The impact of EAs on DAs can be observed from the figure with the drop in success rate and the increase of cost and latency. However, within ten simulation ticks, MONOON allows DAs to adapt to the EAs and retain their performance. The simulation results shows that MONSOON allows a WSN application to adapt to the other application such that they can co-exist tranquilly in a same sensor network. Figure 12 (b), (c), (d) and (e) show the similar scenario as in figure 12 (b), (c), (d) and (e), respectively. The simulation result in the former set of the figures also show the similar trend as in the later set of the figures; therefore, MONSOON allows a WSN application to adapt to network changes, i.e. partial node failure or the base station failure, even when it has to work simultaneously with another application on the same network. Figure 12 (a) portraits the same scenario as in figure 7 (a). In the figure, 12 (a), sensor network hosts two applications, data collection and event detection. However, the objective values of event detection application, i.e. EAs, in figure 12 (a) are improved faster than in figure 7 (a). For example, the latency is reduced to lower than 0.05 at around the 28th tick in figure 12 (a) but it takes about the 38th tick 2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 1 6 11 16 21 26 31 36 **Simulation Ticks** **(a) Static Network** 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 **Generations** **(a) Static Network** 100 90 80 70 60 50 40 30 20 10 0 100 90 80 70 60 50 40 30 20 10 0 2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0.8 0.6 0.4 0.2 0 1 6 11 16 21 26 31 36 41 46 51 56 **Simulation Ticks** **(b) Node Addition** 100 90 80 70 60 50 40 30 20 10 0 100 90 80 70 60 50 40 30 20 10 0 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 **Simulation Ticks** **(b) Node Addition** 2 1.8 1.6 1.4 1.2 1 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 1 6 11 16 21 26 31 36 41 46 51 56 61 66 **Simulation Ticks** **(c) Random Node Failure** 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 **Simulation Ticks** **(c) Random Node Failure** 100 90 80 70 60 50 40 30 20 10 0 100 90 80 70 60 50 40 30 20 10 0 2 100 90 80 70 60 50 40 30 20 10 0 0.8 0.6 0.4 0.2 0 2 1.8 1.6 1.4 1.2 1 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 **Simulation Ticks** **(d) Selective Node Failure** 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 **Simulation Ticks** **(d) Selective Node Failure** 2 100 90 80 70 60 50 40 30 20 10 0 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 1.8 1.6 1.4 1.2 1 100 90 80 70 60 50 40 30 20 10 0 0.8 0.6 0.4 0.2 0 1 6 11 16 21 26 31 36 41 46 51 56 61 **Simulation Ticks** 1 6 11 16 21 26 31 36 41 46 51 56 61 66 **Simulation Ticks** **(e) Base Station Failure** **Figure 11. Objective Values of DAs** **with EAs** **(e) Base Station Failure** **Figure 12. Objective Values of EAs** **with DAs** ----- in figure 7 (a) to reduce to the same level. Thanks to crossmating (see Section 3.3), MONSOON allows event detection application, i.e., EAs, to improve its objective values by using information from the other application. Figure 12 (b), (c), (d) and (e) also show the similar results. ### 4.4 Power Consumption Figure 13 shows the impact of MONSOON and BiSNET/e on power consumption, and compare it with the power consumption by RUGGED [7, 8]. RUGGED is a gradient-based routing protocol. Figure 13 compares the average power consumption of nodes running BiSNET/e and RUGGED in the simulation scenario of Figure 6 (a). BiSNET/e consumes more power than RUGGED first because agents use random behavior policies. However, MONSOON allows agents to evolve their behavior policies and, in turn, reduce power consumption. After the 17th tick, power consumption is mostly same in BiSNET/e and RUGGED. Power consumption is nearly constant in RUGGED because it does not have dynamic adaptation mechanisms. 400 350 BiSNET/e RUGGED 300 250 200 150 100 50 0 1 3 5 7 9 11 13 15 17 19 **Simulation Ticks** **Figure 13. Average Power Consumption** ### 4.5 Memory Footprint Table 1 shows the memory footprint of the BiSNET/e runtime in a MICA2 mote, and compares it with the footprint of Blink (an example program in TinyOS), which periodically turns on and off an LED, RUGGED and Agilla, which is a mobile agent platform for WSNs [9]. The BiSNET/e runtime is lightweight in its footprint thanks to the simplicity of the biologically-inspired mechanisms in BiSNET/e. BiSNET/e can even run on a smaller-scale nodes, for example, TelosB, which has 48KB ROM. **Table 1. Memory Footprint in a MICA2 Mote** |Col1|RAM (KB)|ROM (KB)| |---|---|---| |BiSNET|2.5|30.0| |Blink|0.04|1.6| |RUGGED|0.84|20| |Agilla|3.59|41.6| ## 5. Related Work This work is an extension to the authors’ prior work, BiSNET [10]. BiSNET allows agents to autonomously adapt to dynamic network conditions. However, it does not investigate evolutionary adaptation (i.e., MONSOON); agent behavior policies are manually configured through trial-and-errors and fixed at runtime. Unlike BiSNET, BiSNET/e allows agents to dynamically adapt their behavior policies even to unanticipated network conditions. MONSOON is designed as an extension to an existing mutiobjective optimization algorithm, called PESA-II [11], which in turn extends the NSGA-II algorithm [12]. MONSOON executes elite selection and genetic operations at physically different locations (i.e., at the MONSOON server and individual nodes, respectively), while both PESA-II and NSGA-II execute the two processes at the same location. In MONSOON, an agent chooses a mate that has the closest gene to its own gene, in order to consider the agent’s performance stability. A mate is randomly chosen from the elite archive in PESA-II. In NSGA-II, a mate is selected with a binary tournament. Moreover, unlike PESA-II and NSGA-II, MONSOON considers coevolution between different types of agents (DAs and EAs). kOS is an operating system that applies biological mechanisms to implement adaptive WSN applications [13]. However, kOS has not implemented specific biologicallyinspired mechanisms yet. Also, [13] does not provide any evaluation results as well as the implementation details of kOS. In contrast, BiSNET/e implements specific biologically-inspired mechanisms such as pheromone emission, reproduction, genetic operations and migration. Moreover, this paper evaluates the impacts of those mechanisms on WSN applications’ (i.e., agents’) adaptability. Agilla proposes a programming language to implement mobile agents for WSNs, and provides a runtime system (interpreter) to operate agents on TinyOS [9]. On the other hand, BiSNET/e does not focus on investigating a new programming language for WSNs. BiSNET/e and Agilla provide a similar set of behaviors such as migration and replication. However, Agilla does not address the research issues that BiSNET/e focuses on: evolutionary adaptation to conflicting objectives. In addition, BiSNET/e focuses on its design simplicity and runtime lightweight. As shown in table 1, BiSNET/e is much more lightweight than Agilla. Several research efforts have applied genetic algorithms to WSNs, for example, to cluster-based routing [14–17], data processing [18], localization [19] and node placement [20,21]. Every work uses a fitness function that combines multiple objective values as a weighted sum, and uses the function to rank agents/genes in elite selection. Application designers need to manually configure the weight values in a fitness function through trial-and-errors. In BiSNET/e, no manually-confired parameters exist for elite selection because of a domination ranking mechanism. As a result, BiSNET/e requires much less configuration cost for application designers. Also, [14,15,17,19–21] do not consider dynamics in the network, but assumes the network is static. Evolutionary multiobjective optimization algorithms ----- have been used for node placement [22–24] and routing [25,26]. In each of these work, an optimization process is performed in a central server. This can lead to scalability issue as the network size increases. In contrast, MONSOON is carefully designed to perform its adaptation process in both the MONSOON server and individual nodes. ## 6. Conclusion This paper describes an evolutionary multiobjective adaptation framework, MONSOON, in a biologicallyinspired application architecture, called BiSNET/e. MONSOON allows WSN applications to simultaneously satisfy conflicting operational objectives by adapting to dynamics of physical operational environments and network environments (e.g., sensor readings and node/link failures) through evolution. Thanks to a set of simple biologicallyinspired mechanisms, the BiSNET/e runtime is implemented lightweight. ## References [1] K. Akkaya and M. Younis, “A survey of routing protocols in wireless sensor networks,” Elsevier Ad Hoc Networks, vol. 3, no. 3, pp. 325–349, 2005. [2] J. Blumenthal, M. Handy, F. Golatowski, M. Haase, and D. Timmermann, “Wireless sensor networks - new challenges in software engineering,” in Proc. of IEEE Emerging _Technologies and Factory Automation, September 2003._ [3] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless sensor networks: a survey,” Elsevier _J. of Computer Networks, vol. 38, pp. 393–422, 2002._ [4] P. Rentala, R. Musunuri, S. Gandham, and U. Sexena, “Survey on sensor networks,” in Proc. of the 7th ACM Int’l Conf. _on Mobile Computing and Networking, 2001._ [5] T. Seeley, The Wisdom of the Hive. Harvard University Press, 2005. [6] J. B. Free and I. H. Williams, “The role of the nasonov gland pheromone in crop communication by honey bees,” Int’l J. _of Behavioural Biology, Brill Publishing, vol. 41, 1972._ [7] J. Faruque, K. Psounis, and A. Helmy, “Analysis of gradientbased routing protocols in sensor networks,” in Proc. of _IEEE/ACM Int’l Conf. on Distributed Computing in Sensor_ _Systems, 2005._ [8] J. Faruque and A. Helmy, “RUGGED: Routing on fingerprint gradients in sensor networks,” in Proc. of IEEE Int’l _Conf. on Pervasive Services, 2004._ [9] C. L. Fok, G. C. Roman, and C. Lu, “Rapid development and flexible deployment of adaptive wireless sensor network applications,” in Proc. of 25th IEEE Int’l Conf. on Distributed _Computing Systems, June 2005._ [10] P. Boonma and J. Suzuki, “BiSNET: A biologicallyinspired middleware architecture for self-managing wireless sensor networks,” Elsevier J. of Computer Networks. doi:10.1016/j.comnet.2007.06.006, 2007. [11] D. W. Corne, N. R. Jerram, J. D. Knowles, and M. J. Oates, “PESA-II: Region-based selection in evolutionary multiobjective optimization,” in Proc. of Genetic and Evolutionary _Computation Conference, 2001._ [12] K. Deb, S. Agrawal, A. Pratab, and T. Meyarivan., “A fast elitist non-dominated sorting genetic algorithm for multiobjective optimization: NSGA-II,” in Proc. of INRIA Par_allel Problem Solving from Nature, 2000._ [13] M. Britton, V. Shum, L. Sacks, and H. Haddadi, “A biologically inspired approach to designing wireless sensor networks,” in Proc. of The 2nd European Workshop on Wireless _Sensor Networks, 2005._ [14] R. Khanna, H. Liu, and H. Chen, “Self-organisation of sensor networks using genetic algorithms,” Inderscience Int’l J. _of Sensor Networks, vol. 1, no. 3, pp. 241–252, 2006._ [15] S. Hussain and A. W. Matin, “Hierarchical cluster-based routing in wireless sensor networks,” in Proc. of the 5th Int’l _Conf. on Info. Processing in Sensor Nets, 2006._ [16] S. Jin, M. Zhou, and A. S. Wu, “Sensor network optimization using a genetic algorithm,” in Proc. of Multiconf. on _Systemics, Cybernetics and Informatics, 2003._ [17] K. P. Ferentinos and T. A. Tsiligiridis, “Adaptive design optimization of wireless sensor networks using genetic algorithms,” Elsevier J. of Computer Nets., vol. 51, no. 4, 2007. [18] J. Hauser and C. Purdy, “Sensor data processing using genetic algorithms,” in Proc. of the 43rd IEEE Midwest Sym_posium on Circuits and Systems, 2000._ [19] V. Tam, K. Y. Cheng, and K. S. Lui, “Using micro-genetic algorithms to improve localization in wireless sensor networks,” J. of Comm., Academy Publisher, vol. 1, no. 4, 2006. [20] H. Y. Guo, L. Zhang, L. L. Zhang, and J. X. Zhou, “Optimal placement of sensors for structural health monitoring using improved genetic algorithms,” Smart Materials and Struc_tures, vol. 13, no. 3, pp. 528–534, 2004._ [21] J. Zhao, Y. Wen, R. Shang, and G. Wang, “Optimizing sensor node distribution with genetic algorithm in wireless sensor network,” in Proc. of Int’l Symp. on Neural Nets., 2004. [22] D. B. Jourdan and O. L. de Weck, “Multi-objective genetic algorithm for the automated planning of a wireless sensor network to monitor a critical facility,” in Proc. of SPIE De_fense and Security Symposium, 2004._ [23] R. Rajagopalan, P. K. Varshney, C. K. Mohan, and K. G. Mehrotra, “Sensor placement for energy efficient target detection in wireless sensor networks: A multi-objective optimization approach,” in Proc. of Annual Conf. on Information _Sciences and Systems, 2005._ [24] A. M. Raich and T. R. Liszkai, “Multi-objective genetic algorithm methodology for optimizing sensor layouts to enhance structural damage identification,” in Proc. of the 4th _Int’l Workshop on Structural Health Monitoring, 2003._ [25] R. Rajagopalan, C. Mohan, P. Varshney, and K. Mehrotra, “Multi-objective mobile agent routing in wireless sensor networks,” in Proc. of IEEE Congress on Evolutionary _Computation, 2005._ [26] R. Rajagopalan, P. K. Varshney, K. G. Mehrotra, and C. K. Mohan, “Fault tolerant mobile agent routing in sensor networks: A multi-objective optimization approach,” in Proc. _of the 2nd IEEE Upstate New York Workshop on Communi-_ _cation and Networking, 2005._ -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/HICSS.2008.323?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/HICSS.2008.323, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "http://www.cs.umb.edu/~jxs/pub/hicss07-bisnet-e.pdf" }
2,008
[ "JournalArticle", "Conference" ]
true
2008-01-07T00:00:00
[]
14,728
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/010a74a07848f32bfc5c6155fa3977f24606075a
[ "Computer Science" ]
0.917127
Second Price Auctions - A Case Study of Secure Distributed Computating
010a74a07848f32bfc5c6155fa3977f24606075a
IFIP International Conference on Distributed Applications and Interoperable Systems
[ { "authorId": "1701915", "name": "B. Decker" }, { "authorId": "1788020", "name": "G. Neven" }, { "authorId": "1739936", "name": "Frank Piessens" }, { "authorId": "2891948", "name": "E. V. Hoeymissen" } ]
{ "alternate_issns": null, "alternate_names": [ "Distrib Appl Interoper Syst", "DAIS", "Distributed Applications and Interoperable Systems", "IFIP Int Conf Distrib Appl Interoper Syst" ], "alternate_urls": null, "id": "246ecad8-9680-4191-a119-7529c01ee421", "issn": null, "name": "IFIP International Conference on Distributed Applications and Interoperable Systems", "type": "conference", "url": "https://link.springer.com/conference/dais" }
null
# SECOND PRICE AUCTIONS ## A Case Study of Secure Distributed Computing Bart De Decker[1], Gregory Neven[2], Frank Piessens[3], Erik Van Hoeymissen[1] 1K. U. Leuven, Dept. Computer Science, Celestijnenlaan 200A, B-3001 Leuven, Belgium {Bart.DeDecker.Erik.VanHoeymissen}@cs.kuleuven.ac.be 2Research Assistant of the Fund for Scientific Research, Flanders, Belgium (F.W.O) Gregory.Neven@cs.kuleuven.ac.be _Postdoctoral Fellow of the Belgian National Fund for Scientific Research (F.W.O.)_ Frank.Piessens@cs.kuleuven.ac.be **Abstract** Secure distributed computing addresses the problem of performing a computation with a number of mutually distrustful participants, in such a way that each of the participants has only limited access to the information needed for doing the computation. Over the past two decades, a number of solutions requiring no _trusted third party have been developed using cryptographic techniques. The_ disadvantage of these cryptographic solutions is the excessive communication overhead they incur. In this paper, we use one of the SDC protocols for one particular application: second price auctions, in which the highest bidder acquires the item for sale at the price of the second highest bidder. The protocol assures that only the name of the highest bidder and the amount of the second highest bid are revealed. All other information is kept secret (the amount of the highest bid, the name of the second highest bidder, ...). Although second price auctions may not seem very important, small variations on this theme are used by many public institutions: e.g., a call for tenders, where contract is given to the lowest offer (or the second lowest). The case study serves two purposes: we show that SDC protocols can be used for these kind of applications, and secondly, we assess the network overhead and how well these applications scale. To overcome the communication overhead, we use mobile agents and semi-trusted hosts. **Keywords:** Secure distributed computing, SDC, mobile agents, second price auction, agents, semi-trusted execution platform #### 1. INTRODUCTION Secure distributed computing (SDC) addresses the problem of distributed computing where some of the algorithms and data that are used in the computa ----- 218 _MOBILE AGENTS_ tion must remain private. Usually, the problem is stated as follows, emphasizing privacy of data. Let f be a publicly known function taking n inputs, and suppose there are n parties (named ), each holding one private input The n parties want to compute the value without leaking any information about their private inputs (except of course the information about that is implicitly present in the function result) to the other parties. An ex ample is voting: the function f is addition, and the private inputs represent yes or no votes. In case an algorithm is to be kept private, instead of just data, one can make f an interpreter for some (simple) programming language, and let one of the be an encoding of a program. In descriptions of solutions to the secure distributed computing problem, the function f is usually encoded as a boolean circuit, and therefore secure distributed computing is also often referred to as secure circuit evaluation. It is easy to see that an efficient solution to the secure distributed computing problem would be an enabling technology for a large number of interesting distributed applications across the Internet. Some example applications are: auctions ([8]), charging for the use of algorithms on the basis of a usage count ([9, 10]), querying a secret database ([6]), various kinds of weighted voting, protecting mobile code integrity and privacy ([10, 5]), ... Secure distributed computing is trivial in the presence of a globally trusted third party(TTP): all participants send their data and code to the TTP (over a secure channel), the TTP performs the computation and broadcasts the results. The main drawback of this approach is the large amount of trust needed in the TTP. Solutions without a TTP are also possible. Over the past two decades, a fairly large variety of solutions to the problem has been proposed. An overview is given by Franklin [3] and more recently by Cramer [2] and Neven [7]. These solutions differ from each other in the cryptographic primitives that are used, and in the class of computations that can be performed (some of the solutions only allow for specific kinds of functions to be computed). The main drawback, however, of these solutions is the heavy communication overhead that they incur. In this paper, we investigate a case study: second price auctions. Here, the highest bidder wins but has to pay the second highest bid. The final outcome will only reveal the name of the winner and the amount of the second highest ##### bid. All other bids and even the name of the second highest bidder remain secret. We have chosen this application, because it illustrates the merits of SDC, and is somewhat exemplary for many other useful applications. For instance, the authority and many public institutions request for quotations before awarding the job/purchase to the lowest or second lowest offer. The reader can easily verify that determining the lowest (or second lowest) offer, without revealing the other quotations, is only a small variation on our case study. ----- _Second Price Auctions_ 219 In this case study, we try to be as specific as possible. We will show how SDC can be used in this application. Moreover, we will look at the performance. In ##### particular, we examine the communication overhead and the scalability of the application in terms of number of participants. Although the communication overhead seems prohibitively high, a reasonable remedy is proposed, using mobile agents and semi-trusted sites. Indeed, mobile agents employing SDC protocols can provide for a trade-off between communication overhead and trust. The communication overhead is alleviated if the communicating parties are brought close enough together. In our approach, every participant sends its representative agent to a trusted execution site. The agent contains a copy of the private data xi and is capable of running an SDC-protocol. Different participants may send their agents to different sites, as long as these sites are located closely to each other. Of course, a mobile agent needs to trust his execution platform, but we will show that the trust requirements in this case are much lower than for a classical TTP. Also, in contrast with protocols that use unconditionally TTPs, the trusted site is not involved directly. It simply offers ##### a secure execution platform: i.e. it executes the mobile code correctly, does not spy on it and does not leak information to other mobile agents. Moreover, the trusted host does not have to know the protocol used between the agents. In other words, the combination of mobile agent technology and secure distributed computing protocols makes it possible to use a generic TTP that, by offering a secure execution platform, can act as TTP for a wide variety of protocols in a uniform way. A detailed discussion of the use of mobile agent technology for advanced cryptographic protocols is given in [7]. The sequel of the paper is organized as follows: in section 2, we review one of the SDC protocols that will be used by the application; a design of the the application, second price auctions, is given in section 3; in this section, we also examine the communication overhead and tackle the scalability issue. In section 4, we introduce a modus operandi for the application. Finally, in section 5, we summarize the main outcomes of this paper. ### 2. SECURE DISTRIBUTED COMPUTING USING GROUP-ORIENTED CRYPTOGRAPHY In [4], Franklin and Haber propose a protocol that evaluates a boolean circuit on data encrypted with a homomorphic probabilistic encryption scheme for any number of participants. It resembles the protocol for two parties, proposed by Abadi and Feigenbaum ([1]). To extend the idea of [1] to the multi-party case, an encryption scheme is needed that allows anyone to encrypt, but needs the cooperation of all participants to perform a decryption. In a joint encryption scheme, all participants know the public key while each participant has his own pri ----- 220 _MOBILE AGENTS_ vate key Using the public key, anyone can create an encryption of some message m, where such that the private key of each participant in S is needed to decrypt. More formally, if de notes the decryption with private key, the relation between encryption and decryption is given by The plaintext m should be easily recoverable from In the joint encryption scheme used by Franklin and Haber, a bit b is encrypted as where _p and q are two primes such that_ mod 4, and The public key is given by where each represents the private (secret) key of participant This scheme has some additional properties that are used in the protocol: _XOR-Homomorphic. Anyone can compute a joint encryption of the XOR_ of two jointly encrypted bits. Indeed, if and then _Blindable. Given an encrypted bit, anyone can create a random ciphertext_ that decrypts to the same bit. Indeed, if and then is a joint encryption of the same bit. _Witnessable. Any participant can withdraw from a joint encryption by_ providing the other participants with a single value. Indeed, if it is easy to compute from First of all, the participants must agree on a value for N and g, choose a secret key and broadcast mod N to form the public key. To start the actual protocol, each participant broadcasts a joint encryption of his input bits. For an XOR-gate, everyone simply applies the XOR-homomorphism. The encrypted output of a NOT-gate can be found by applying the XOR-homomorphism with a default encryption of a one, e.g. [1,–1]. However, it is the AND-gate that causes some trouble. Suppose the encrypted input bits for the AND-gate are and To compute a joint encryption they proceed as follows: ----- _Second Price Auctions_ 221 1 Each participant chooses random bits and and broadcasts and 2 Each participant repeatedly applies the XOR-homomorphism to calculate and ##### Each participant broadcasts decryption witnesses and 3 Everyone can now decrypt and We have the following relation ##### between and Each participant is able to compute a joint encryption of he knows and (he chose them himself) and he received encryptions from the other participants, so he can compute as follows: If then so any default encryption for a zero will do, e.g. [1,1]. If then so is a valid substitution for and can be computed in an analogous way. He uses the XOR-homomorphism to combine all these terms, blinds the result and broadcasts this as 4 Each participant combines and again using the XOR-homomorphism, to form When all gates in the circuit have been evaluated, every participant has a joint encryption of the output bits. Finally, all participants broadcast decryption witnesses to reveal the output. #### 3. SECOND PRICE AUCTIONS In this section we consider second price auctions, where there is one item for sale and there are n bidders. The item will only be sold if the bid of one participant is strictly higher than the other bids. In all other cases there is no ----- 222 _MOBILE AGENTS_ winner. The clearing price is the second highest bid. The requirements for this type of auction are the following: if there is no winner, nothing is revealed; if there is a winner: – the identity of the highest bidder is revealed, but the highest bid remains secret; – the 2[nd] highest bid is revealed, but the identity of the 2[nd] highest bidder is kept secret; – no other information (other bids) are to be revealed. For three participants X, Y and Z, the boolean circuit is shown in Figure 1. The inputs to the circuit are 32-bit bids[1]. The output is the identity of the winner, represented by the bits and ( no winner, 01 winner is X, 10 winner is Y, 11 winner is Z), and the clearing price. If there is no winner, the clearing price is set to zero. To determine the winner, the circuit uses three comparators and a number of AND and OR gates. To determine the clearing price, four multiplexers are used. Consider the situation where X makes the highest bid. In this case and so the second input to the final multiplexer will be chosen. The input on this line is determined by the bids made by Y and Z. If then and Y will be selected as the clearing price. In the other cases or _Z will be the clearing price._ Our goal is to estimate the communication overhead of an implementation of secure distributed second price auctions with the protocol proposed by Franklin and Haber. The auction is designed as a boolean circuit and the communication overhead for secure circuit evaluation is estimated. The communication overhead is determined by the following steps in the protocol: broadcast of the encrypted input bits of each participant; evaluation of an AND gate: – broadcast of the encrypted bits – broadcast of the decryption witnesses – broadcast of the blinded broadcast of the output decryption witnesses. The associated communication overhead is: 1 In reality, fewer bits (e.g. 8 or 16) would suffice. ----- _Second Price Auctions_ 223 _Figure 1._ Boolean circuit implementation of second price auctions. for the broadcast of the input bits; for the evaluation of an AND gate; ##### for the decryption broadcast. where is the length of N in bits, which is the same as the number of bits needed to represent an element of is the number of input bits of participant i, n is the number of participants and out is the number of output bits of the circuit. In order to estimate the communication overhead, we need to ##### be able to determine the number of AND gates in the boolean circuit (note that each OR gate can be implemented with AND and NOT gates). Each comparator can be built with 374 AND-gates[2] 2The boolean function can be expressed as Hence if A and B are k-bit numbers, AND gates are needed. Both functions, and are needed for each comparator. ----- 224 _MOBILE AGENTS_ For participants, the circuit changes as follows. The number of comparators needed is now The final multiplexer will need to distinguish between different cases, i.e. n possible winners or no winner at all. The other n multiplexers are there to select the clearing price out of bids when there is a winner. The number of AND gates needed for each multiplexer as a function of the number of inputs m is shown in Figure 2. Besides the comparators and the multiplexers, some additional AND and OR gates are needed. However, the number of these gates is negligible compared to the number of gates needed for the comparators and multiplexers. In summary, the circuit has a total gate complexity of _Figure 2._ Number of AND gates needed in a mulitplexer The results of estimating the communication overhead for this circuit as a function of the number of participants n are summarized in Table 1[3]. Franklin and Haber’s protocol is linear in the number of broadcasts, so the total message complexity is However, it must be noted that this only holds on a network with broadcast or multicast functionality, such that the communication overhead of sending a message to all participants is the same as that of sending a message to a single participant. In absence of such infrastructure, the total message complexity is 3We choose to be 1024 bits. ----- _Second Price Auctions_ 225 #### 4. MODUS OPERANDI From the previous section, it should be clear that the design of the application has pros and cons: A major advantage is that our solution does not require a globally trusted third party that plays the role of the arbiter. The worst drawback is the immense communication overhead and the fact that the solution does not scale very well. There exists a trade-off between ’trust’ and ’communication overhead’ in both options, the first one using a TTP and the solution that uses SDC. In this section, we investigate this trade-off and present a nice remedy for the communication overhead. If a globally trusted third party is used, every participant, has to send ##### its private bid to that TTP who will select the highest bidder, determine the second highest bid, and disseminate its decision to the participants (see Figure 3). Of course, before sending its private data to the TTP, every _Figure 3._ 2nd Price Auction Using a TTP. must first authenticate the TTP, and then send through a safe channel. This can be accomplished via conventional cryptographic techniques. It is clear that this approach has a very low communication overhead: the data is only sent once to the TTP; later, every participant receives the result of the computation. However, every participant should uncondionally trust the TTP. It is not clear ----- 226 _MOBILE AGENTS_ whether n distrustful participants will easily agree on one single trustworthy site. If this site is compromised, all secrets, may be compromised! Also, the site needs the appropriate software for this particular application. Hence, for every new ‘application’ new software needs to be installed. Therefore, the participants not only need to trust the (security of) the site, but also the software for this application. In our approach (see Figure 4), the trust requirements are really minimal: every participant trusts its own execution site and expects that the other participants provide correct values for their own inputs. (Note that in this protocol, a participant cannot cheat, because of the use of witnesses.) Although our approach is very attractive, it suffers extensive communication overhead and does not scale well. _Figure 4._ 2nd Price Auction Using SDC. The communication overhead of SDC-techniques can be remedied by intro ducing semi-trusted execution sites and mobile agents (see Figure 5). Every participant sends its representative, agent to a trusted execution site The agent contains a copy of the private data and is capable of running a SDC-protocol. It is allowed that different participants send their agents to different sites. The only restriction being that the sites should be located closely to each other, i.e. should have high bandwidth communication between them. Of course, every execution site needs a mechanism to safely download an agent. However, that can be easily accomplished through convential cryptographic techniques. The amount of large distance communication is moderate: every participant sends its agent to a remote site, and receives the result from its ----- _Second Price Auctions_ 227 _Figure 5._ 2nd Price Auctions Using Agents (SDC) and Semi-Trusted Sites. agent. The agents use a SDC-protocol, which unfortunately involves a high communication overhead. However, since the agents are executing on sites that ##### are near each other, the overhead of the SDC-protocol is acceptable. No high bandwidth communication between the participants is necessary, and there is no longer a need for one single trusted execution site. The agents that participate in the secure computation are protected against malicious behaviour of other (non-trusted) execution sites by the SDC-protocols. That is sufficient to make this approach work. Moreover, in contrast with the approach where one uses ##### an unconditionally trusted third party, the trusted sites are not involved directly. They simply offer a secure execution platform: the trusted hosts do not have to know the protocol used between the agents. In other words, the combination ##### of mobile agent technology and secure distributed computing protocols makes it possible to use generic trusted third parties that, by offering a secure execution platform, can act as trusted third party for a wide variety of protocols in ##### a uniform way. Finally, the question remains whether it is realistic to assume that participants can find execution sites that are close enough to each other. Given the fact however that these execution sites can be generic, we believe that providing such execution sites could be a commercial occupation. Various deployment strategies are possible. Several service providers, each administering a set of geographically dispersed “secure hosts”, can propose their subscribers ##### an appropriate site for the secure computation. The site is chosen to be in the neighborhood of a secure site of the other service providers involved. Another ----- 228 _MOBILE AGENTS_ approach is to have execution parks, offering high bandwidth communication facilities, were companies can install their proprietary “secure site”. The park itself could be managed by a commercial or government agency. #### 5. CONCLUSIONS This paper demonstrates that second price auctions, and many other rele vant applications, can be implemented by using SDC protocols. That way, the participants can make sure that all confidential information is kept secret. The major disadvantage, the overwhelming communication overhead, can be remedied through the use of mobile agents and semi-trusted sites. There is no need for one generally trusted site, nor does the program code have to be endorsed by all participants. The trusted execution sites are generic and can be small (which might allow to draft a formal security for these sites). The communication overhead of secure distributed computing protocols is no longer prohibitive for their use since the execution sites are located closely to each other. #### References [1] M. Abadi and J. Feigenbaum, “Secure circuit evaluation, a protocol based on hiding infor mation from an oracle,” Journal of Cryptology, 2(1), p. 1–12, 1990 [2] R. Cramer. “An introduction to secure computation”, in LNCS 1561, pp 16–62, 1999. [3] M. Franklin, “Complexity and security of distributed protocols,” Ph. D. thesis, Computer Science Department of Columbia University, New York, 1993 [4] M. Franklin and S. Haber, “Joint encryption and message-efficient secure computation,” Journal of Cryptology, 9(4), p. 217–232, Autumn 1996 [5] S. Loureiro and R. Molva, “Privacy for Mobile Code”, Proceedings of the workshop on _Distributed Object Security, OOPSLA ’99, p. 37–42._ [6] G. Neven, F. Piessens, B. De Decker, “On the Practical Feasibility of Secure Distributed Computing: a Case Study”, Information Security for Global Information Infrastructures (S. Qing, J. Eloff, ed.), Kluwer Academic Publishers, 2000, pp. 361-370. [7] G. Neven, E. Van Hoeymissen, B. De Decker, F. Piessens, “Enabling Secure Distributed Computations : Semi-trusted Hosts and Mobile Agents”, to appear in Networking and Information Systems Journal 3 (2001). [8] N. Nisan, “Algorithms for selfish agents”, Proceedings of the 16th Annual Symposium on _Theoretical Aspects of Computer Science, Trier, Germany, March 1999, p. 1–15._ [9] T. Sander and C. Tschudin, “On software protection via function hiding”, Proceedings of _the second workshop on Information Hiding, Portland, Oregon, USA, April 1998._ [10] T. Sander and C. Tschudin, 'Towards mobile cryptography”, Proceedings of the 1998 IEEE _Symposium on Security and Privacy, Oakland, California, May 1998._ [11] T. Sander, A. Young, M. Yung, “Non-Interactive CryptoComputing for ”, preprint. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/0-306-47005-5_19?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/0-306-47005-5_19, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://link.springer.com/content/pdf/10.1007%2F0-306-47005-5_19.pdf" }
2,001
[ "JournalArticle" ]
true
2001-09-17T00:00:00
[]
5,508
en
[ { "category": "Political Science", "source": "s2-fos-model" }, { "category": "Law", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/010a9df80b73d0c589eaffcb3a9f4cb05b07d225
[]
0.907507
The Elections Clause Obligates Congress to Enact a Federal Plan to Secure U.S. Elections Against Foreign Cyberattacks
010a9df80b73d0c589eaffcb3a9f4cb05b07d225
[ { "authorId": "3964911", "name": "S. Malempati" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
While foreign adversaries continue to launch cyberattacks aimed at disrupting elections in the United States, Congress has been reluctant to take action. After Russia interfered in the 2016 election, cybersecurity experts articulated clear measures that must be taken to secure U.S. election systems against foreign interference. Yet the federal government has failed to act. Congress’s reticence is based on a misguided notion that greater federal involvement in the conduct of elections unconstitutionally infringes on states’ rights. Both state election officials and certain congressional leaders operate under the assumption that federalism principles grant states primacy in conducting federal elections. This Comment dispels the myth that Congress must defer to states to regulate federal elections. The text of the Elections Clause in Article I, Section 4 of the U.S. Constitution confers to Congress final authority in determining the “Times, Places and Manner” of federal elections. Therefore, the system of administering federal elections is based on decentralization rather than federalism. The risk of foreign interference in U.S. elections was a precise reason the founders bestowed on Congress ultimate control over federal elections. States and municipalities lack the capacity to effectively combat foreign cyber invasion. This Comment makes the case that Congress has a responsibility to exercise its power under the Elections Clause to create a federal plan to secure voter registration databases and voting mechanisms against cyberattacks in order to protect the integrity of American democracy. MALEMPATI_12.2.20 12/2/2020 2:44 PM 418 EMORY LAW JOURNAL [Vol. 70:417 INTRODUCTION ............................................................................................. 423 I. THE CURRENT CYBERSECURITY THREAT TO U.S. ELECTION INFRASTRUCTURE .............................................................................. 427 A. Russian Interference in the 2016 U.S. Election ........................ 428 B. Recommendations of Cybersecurity Experts to Strengthen U.S. Election Infrastructure ...................................................... 432 C. States Responded Inadequately and Ineffectively to Russian Cyberattacks ............................................................................. 437 II. THE LANDSCAPE OF CONGRESSIONAL AUTHORITY OVER FEDERAL ELECTIONS ......................................................................................... 440 A. Congressional Authority Under the Reconstruction Amendments and the Voting Rights Act .................................... 441 B. The Demise of the Voting Rights Act and Shifting StateFederal Authority to Regulate Elections .................................. 446 C. The Elections Clause Grants Congress Broad Authority to Regulate Federal Elections ...................................................... 448 1. Decentralization Versus Federalism .................................. 449 2. Congress Has Used Its Election Clause Authority to a Limited Degree ................................................................... 450 III. CONGRESS SHOULD ACT TO PROTECT U.S. ELECTION INFRASTRUCTURE .............................................................................. 454 A. Congress Has an Obligation Under the Elections Clause to Protect U.S. Democracy ........................................................... 454 B. Congress Has a Duty to Secure U.S. Elections Against Foreign Interference Because States Are Ill-Equipped and Reluctant to Do So ........................................................................................ 457 C. Congress Must Enact a Federal Plan to Preserve the Right of All Citizens to Vote ................................................................... 459 IV. A PROPOSED FEDERAL PLAN TO SECURE U.S. ELECTIONS ............... 462 A. Congress Should Establish Binding Federal Standards for States to Register Voters, Maintain Secure Voter Databases, and Check-in Voters at the Polls .............................................. 463 B. Congress Should Mandate Uniform Paper Ballots for All Federal Elections ...................................................................... 465 C. Congress Should Require All States to Submit to Federal Election Audits .......................................................................... 466 CONCLUSION ................................................................................................. 467 MALEMPATI_12.2.20 12/2/2020 2:44 PM 2020] THE ELECTIONS CLAUSE 419
# Emory Law Journal Emory Law Journal ### Volume 70 Issue 2 2020 # The Elections Clause Obligates Congress to Enact a Federal Plan The Elections Clause Obligates Congress to Enact a Federal Plan to Secure U.S. Elections Against Foreign Cyberattacks to Secure U.S. Elections Against Foreign Cyberattacks ### Suman Malempati [Follow this and additional works at: https://scholarlycommons.law.emory.edu/elj](https://scholarlycommons.law.emory.edu/elj?utm_source=scholarlycommons.law.emory.edu%2Felj%2Fvol70%2Fiss2%2F4&utm_medium=PDF&utm_campaign=PDFCoverPages) [Part of the Law Commons](http://network.bepress.com/hgg/discipline/578?utm_source=scholarlycommons.law.emory.edu%2Felj%2Fvol70%2Fiss2%2F4&utm_medium=PDF&utm_campaign=PDFCoverPages) ### Recommended Citation Recommended Citation Suman Malempati, The Elections Clause Obligates Congress to Enact a Federal Plan to Secure U.S. Elections Against Foreign Cyberattacks, 70 Emory L. J. 417 (2020). [Available at: https://scholarlycommons.law.emory.edu/elj/vol70/iss2/4](https://scholarlycommons.law.emory.edu/elj/vol70/iss2/4?utm_source=scholarlycommons.law.emory.edu%2Felj%2Fvol70%2Fiss2%2F4&utm_medium=PDF&utm_campaign=PDFCoverPages) This Comment is brought to you for free and open access by the Journals at Emory Law Scholarly Commons. It has been accepted for inclusion in Emory Law Journal by an authorized editor of Emory Law Scholarly Commons. For [more information, please contact law-scholarly-commons@emory.edu.](mailto:law-scholarly-commons@emory.edu) ----- ## THE ELECTIONS CLAUSE OBLIGATES CONGRESS TO ENACT A FEDERAL PLAN TO SECURE U.S. ELECTIONS AGAINST FOREIGN CYBERATTACKS ABSTRACT _While foreign adversaries continue to launch cyberattacks aimed at_ _disrupting elections in the United States, Congress has been reluctant to take_ _action. After Russia interfered in the 2016 election, cybersecurity experts_ _articulated clear measures that must be taken to secure U.S. election systems_ _against foreign interference. Yet the federal government has failed to act._ _Congress’s reticence is based on a misguided notion that greater federal_ _involvement in the conduct of elections unconstitutionally infringes on states’_ _rights. Both state election officials and certain congressional leaders operate_ _under the assumption that federalism principles grant states primacy in_ _conducting federal elections._ _This Comment dispels the myth that Congress must defer to states to regulate_ _federal elections. The text of the Elections Clause in Article I, Section 4 of the_ _U.S. Constitution confers to Congress final authority in determining the “Times,_ _Places and Manner” of federal elections. Therefore, the system of administering_ _federal elections is based on decentralization rather than federalism._ _The risk of foreign interference in U.S. elections was a precise reason the_ _founders bestowed on Congress ultimate control over federal elections. States_ _and municipalities lack the capacity to effectively combat foreign cyber_ _invasion. This Comment makes the case that Congress has a responsibility to_ _exercise its power under the Elections Clause to create a federal plan to secure_ _voter registration databases and voting mechanisms against cyberattacks in_ _order to protect the integrity of American democracy._ ----- INTRODUCTION ............................................................................................. 423 I. THE CURRENT CYBERSECURITY THREAT TO U.S. ELECTION INFRASTRUCTURE .............................................................................. 427 _A. Russian Interference in the 2016 U.S. Election ........................ 428_ _B. Recommendations of Cybersecurity Experts to Strengthen_ _U.S. Election Infrastructure ...................................................... 432_ _C. States Responded Inadequately and Ineffectively to Russian_ _Cyberattacks ............................................................................. 437_ II. THE LANDSCAPE OF CONGRESSIONAL AUTHORITY OVER FEDERAL ELECTIONS ......................................................................................... 440 _A. Congressional Authority Under the Reconstruction_ _Amendments and the Voting Rights Act .................................... 441_ _B. The Demise of the Voting Rights Act and Shifting State-_ _Federal Authority to Regulate Elections .................................. 446_ _C. The Elections Clause Grants Congress Broad Authority to_ _Regulate Federal Elections ...................................................... 448_ _1. Decentralization Versus Federalism .................................. 449_ _2. Congress Has Used Its Election Clause Authority to a_ _Limited Degree ................................................................... 450_ III. CONGRESS SHOULD ACT TO PROTECT U.S. ELECTION INFRASTRUCTURE .............................................................................. 454 _A. Congress Has an Obligation Under the Elections Clause to_ _Protect U.S. Democracy ........................................................... 454_ _B. Congress Has a Duty to Secure U.S. Elections Against Foreign_ _Interference Because States Are Ill-Equipped and Reluctant to_ _Do So ........................................................................................ 457_ _C. Congress Must Enact a Federal Plan to Preserve the Right of_ _All Citizens to Vote ................................................................... 459_ IV. A PROPOSED FEDERAL PLAN TO SECURE U.S. ELECTIONS ............... 462 _A. Congress Should Establish Binding Federal Standards for_ _States to Register Voters, Maintain Secure Voter Databases,_ _and Check-in Voters at the Polls .............................................. 463_ _B. Congress Should Mandate Uniform Paper Ballots for All_ _Federal Elections ...................................................................... 465_ _C. Congress Should Require All States to Submit to Federal_ _Election Audits .......................................................................... 466_ CONCLUSION ................................................................................................. 467 ----- INTRODUCTION Does federalism prevent Congress from taking action to secure U.S. elections against foreign cyberattacks? Since its founding, the United States has grappled with how to balance the authority of state governments against that of the federal government in managing elections.[1] Article I, Section 4 of the U.S. Constitution, often called the “Elections Clause,” grants each state the power to designate the “Times, Places and Manner” of federal elections, but it also states that “Congress _may at any time by Law make or alter such Regulations.”[2]_ Despite the seemingly sweeping power designated to Congress by the Elections Clause, scholars and the Supreme Court have traditionally viewed the regulation of elections and the voting process through the lens of state sovereignty.[3] Currently, U.S. election infrastructure consists of a heterogeneous array of voter registration procedures, registered voter databases, pollbooks, voting machines, and vote counting mechanisms that vary from state to state.[4] States are also inconsistent in the degree to which they delegate election management to counties and municipalities.[5] Two hundred and thirty years ago in the Federalist Papers, Alexander Hamilton explained the rationale for embedding Congress’s power to regulate elections into the Constitution.[6] Hamilton explained that leaving control of federal elections solely in the hands of state governments could create an existential risk to the nation.[7] With the Elections Clause, the drafters of the Constitution “reserved to the national authority a right to interpose, whenever _extraordinary circumstances might render that interposition necessary to its_ safety.”[8] Hamilton presciently recognized that the threat of foreign interference 1 Guy-Uriel E. Charles & Luis Fuentes-Rohwer, State’s Rights, Last Rites, and Voting Rights, 47 CONN. L. REV. 481, 514 (2014) (“This struggle between the states and the national government with respect to the apportionment of powers over elections has waxed and waned throughout American history.”). 2 U.S. CONST. art. I, § 4 (emphasis added). 3 _See, e.g., Shelby Cnty. v. Holder, 570 U.S. 529, 535 (2013) (stating the Voting Rights Act of 1965_ which granted federal oversight over the voting laws of certain states was “a drastic departure from the principle of federalism”); Justin Weinstein-Tull, Election Law Federalism, 114 MICH. L. REV. 747, 753 (2016) (describing “election law federalism” as consisting of “multiple sovereigns” at the federal, state, and local government levels). 4 Weinstein-Tull, supra note 3, at 754 (listing the differences in voting hours, funding schemes, absentee voting rules, and voter registration, or the “nuts and bolts of the election”). 5 _Id._ 6 THE FEDERALIST NO. 59 (Alexander Hamilton). 7 _Id. (“With so effectual a weapon in [state legislators’] hands as the exclusive power of regulating_ elections for the national government, a combination of few such men, in a few of the most considerable States, where the temptation will always be the strongest, might accomplish the destruction of the union.”). 8 _Id. (emphasis added)._ ----- in U.S. elections would be such an extraordinary circumstance.[9] He wrote in _Federalist 59 that “a firm union of this country, under an efficient government,_ will probably be an increasing object of jealousy to more than one nation of Europe; and that enterprises to subvert it will sometimes originate in the intrigues of foreign powers.”[10] In 2016, for the first time in the history of this nation, Hamilton’s prediction of foreign interference came true when Russia attempted to interfere with and influence the U.S. presidential election.[11] Along with a campaign of misinformation, Russia directly attacked U.S. election systems.[12] Beginning as early as 2014, the Russian government directed extensive activity against U.S. election infrastructure at the state and local levels.[13] A 2018 report by the Senate Intelligence Committee revealed that Russian operatives attempted to hack into the election systems of each of the fifty states.[14] Russia attacked a point of vulnerability in U.S. election infrastructure—states’ supposed primacy in conducting federal elections.[15] According to the Senate Intelligence Report, “[s]tate elections officials, who have primacy in running elections, were not sufficiently warned or prepared to handle an attack from a hostile nation-state actor.”[16] Hamilton’s interpretation of the Elections Clause suggests that Russian aggression is a clear reason for Congress to exert its constitutional authority to protect U.S. election infrastructure.[17] Despite the obvious risk that our democracy may be undermined by foreign interference, some members of Congress have expressed reluctance to take a greater role in protecting federal elections.[18] State officials have also pushed back and even rejected federal help in securing their state and local election 9 _Id._ 10 _Id._ 11 NAT’L ACAD. OF SCIS., ENG’G & MED., SECURING THE VOTE: PROTECTING AMERICAN DEMOCRACY 13 (2018) [hereinafter NAS REPORT]. 12 _Id. at 14._ 13 S. SELECT COMM. ON INTEL., S. REP. NO. 116-XX, REPORT ON RUSSIAN ACTIVE MEASURES CAMPAIGNS AND INTERFERENCE IN THE 2016 U.S. ELECTION 3 (2019) (partially redacted) [hereinafter SENATE INTELLIGENCE REPORT]. 14 _Id. at 12._ 15 _Id. at 4 (“Russian efforts exploited the seams between federal authorities and capabilities, and_ protections for the states.”). 16 _Id._ 17 _See_ _infra Part III._ 18 _See, e.g., Dean Dechiaro,_ _Election Officials Want Security Money, Flexible Standards, ROLL CALL_ (Aug. 15, 2019), https://www.rollcall.com/news/congress/election-officials-want-security-money-flexiblestandards (describing Senate Majority Leader Mitch McConnell’s reluctance to bring House-passed election security bills up for votes in the Senate). ----- systems out of concern for maintaining state sovereignty.[19] Although Congress has previously overridden the right of states to conduct elections by passing the Voting Rights Act of 1965 (VRA) under the Fifteenth Amendment, it has yet to invoke its full Elections Clause powers.[20] With its holding in Shelby County v. _Holder in 2013, the Supreme Court gutted the VRA, tilting the balance toward_ state autonomy in conducting elections.[21] Therefore, Congress can no longer rely solely on its power to enforce the Reconstruction Amendments to supersede state authority over elections.[22] This Comment argues that the threat of foreign attacks against U.S. election infrastructure requires Congress to exercise its power under the Elections Clause to enact legislation establishing a uniform system for federal elections.[23] This Comment takes the position that foreign cyber intrusion is the type of existential threat for which the Elections Clause gives Congress the authority to act. Because the Constitution grants Congress the ultimate authority to regulate federal elections, the creation of a federal system for elections does not intrude on state sovereignty. Part I describes the current cybersecurity threat to U.S. election infrastructure. A paucity of federal regulations poses significant risks in the face of such twenty-first century threats. This Part describes the scope of Russia’s attacks on state and local election systems during the 2016 election and catalogs the recommendations of cybersecurity experts in how best to secure election infrastructure against future attacks. By detailing how state and local election officials responded ineffectively to cyberattacks in 2016 and leading up to the 2018 election, this Comment predicts that without a comprehensive federal plan, Russia and other foreign actors may successfully disrupt future federal elections. 19 _See_ _infra Part III.B._ 20 Voting Rights Act of 1965, 52 U.S.C. § 10301; see South Carolina v. Katzenbach, 303 U.S. 301, 308 (1966) (upholding the invalidation of state laws restricting voter access to the polls as an appropriate means for carrying out Congress’s constitutional responsibilities under the Fifteenth Amendment). 21 570 U.S. 529, 557 (2013). 22 The Thirteenth, Fourteenth, and Fifteenth Amendments to the U.S. Constitution are often called the “Reconstruction Amendments.” The Thirteenth Amendment prohibited slavery. U.S. CONST. amend. XIII. The Fourteenth Amendment established birthright citizenship and created due process and equal protection rights against state action. U.S. CONST. amend. XIV. The Fifteenth Amendment guaranteed the right to vote regardless of color or condition of previous servitude. U.S. CONST. amend. XV. 23 This Comment does not address one aspect of Russia’s interference in the 2016 election—a social media campaign of disinformation aimed at influencing voters. For a summary of that issue and recommendations for confronting Russia’s efforts, see Alex Stamos, Sergey Sanovich, Andrew Grotto & Allison Berke, _Combatting State-Sponsored Disinformation Campaigns from State-aligned Actors,_ _in SECURING_ AMERICAN ELECTIONS: PRESCRIPTIONS FOR ENHANCING THE INTEGRITY AND INDEPENDENCE OF THE 2020 U.S. PRESIDENTIAL ELECTION AND BEYOND 43 (Michael McFaul ed., 2019). ----- Next, Part II explores the history of the Supreme Court’s interpretation of constitutional provisions that confer differential authority to states and the federal government to regulate federal elections. This Part describes how the Court’s recognition of congressional authority to control federal elections has waxed and waned over the past 150 years. The Court has previously granted relatively broad powers to Congress to invalidate state legislation that infringed on citizens’ right to vote under the enforcement provisions of the Fourteenth and Fifteenth Amendments.[24] The expansion of congressional authority under the Reconstruction Amendments was followed by a reversion to greater state sovereignty over elections with the Court’s holding in Shelby County.[25] This Part explains that Shelby County represents a shift in the Court’s view towards greater state autonomy in conducting elections. Therefore, Congress must find another source of authority to enact federal election legislation. Part II argues that such authority can be found in the Elections Clause, which provides an underrecognized source of power for Congress to regulate federal elections. Despite the Supreme Court’s reluctance to infringe on states’ purported sovereignty in conducting elections, the Elections Clause gives Congress the power to supersede any state action regarding elections. The text and purpose of the Elections Clause provide a system for U.S. elections based on decentralization rather than federalism. Part III contends that, for three main reasons, Congress has an obligation to use its Election Clause authority to enact a federal election plan. First, foreign attacks on U.S. election infrastructure fall within the category of “extraordinary circumstances” as described by Hamilton, which provides the impetus for Congress to regulate the “Times, Places and Manner” of federal elections.[26] Cyber invasion by Russia and potentially other nation-states is a matter of national security that requires a federal response. Second, state and local officials lacked the capacity to manage the attacks during the 2016 U.S. election. Cyberattacks will continue to intensify without a coordinated national response, and states cannot be left to defend election infrastructure from such attacks. Third, insecure voting systems in several states violate the rights of voters under the Fourteenth Amendment by preventing voters from confidently knowing that their votes will count.[27] Therefore, despite the Supreme Court’s holding in _Shelby County, Congress also has a responsibility to step in where states have_ 24 U.S. CONST. amend. XIV, § 5; U.S. CONST. amend. XV, § 2. 25 _Shelby County, 570 U.S. at 544; Charles & Fuentes-Rohwer, supra note 1, at 514–15, 518._ 26 U.S. CONST. Article I, § 4; THE FEDERALIST NO. 59 (Alexander Hamilton). 27 _See Curling v. Kemp, 334 F. Supp. 3d 1303, 1328 (N.D. Ga. 2018) (“A wound or reasonably threatened_ wound to the integrity of a state’s election system carries grave consequences beyond the results in any specific election, as it pierces citizens’ confidence in the electoral system and the value of voting.”). ----- failed in securing their election systems pursuant to the Fourteenth Amendment’s enforcement provision. Lastly, Part IV provides a prescriptive solution and suggests legislation that Congress may enact. Namely, Congress should enact a federal election plan that provides for federal oversight of uniform procedures and standards that each state must follow while maintaining the decentralized conduct of elections.[28] The plan should include federally mandated standards for maintaining registration databases and electronic pollbooks. The federal plan should also require that all states use the same mechanism to generate voter-verified paper ballots, which are read by federally certified optical scanners. Finally, a federal election plan should mandate that all states submit to federal post-election audits. I. THE CURRENT CYBERSECURITY THREAT TO U.S. ELECTION INFRASTRUCTURE Securing U.S. elections and citizens’ confidence in the election process is of paramount importance to maintain this nation’s republican form of government. After the 2016 presidential election, evidence is clear that foreign powers are capable of interfering with U.S. election systems to, at minimum, erode voter confidence and at worst, suppress voter turnout, manipulate vote tallies, and sway election results.[29] Along with hacking into the Democratic National Committee’s servers and launching a disinformation campaign on social media, Russia directly targeted U.S. election infrastructure and continues to do so.[30] Cybersecurity experts are fully aware of the vulnerability of U.S. election systems and have developed clear, consensus recommendations on how best to secure elections against cyberattacks.[31] The onus is now on the federal government to create a national plan that will implement these recommendations. While decentralization provides some protection from a single crippling attack, it also creates a barrier to generating a cohesive and uniform response to foreign cyberattacks.[32] Although states and municipalities play a critical administrative role in conducting elections, they are generally ill-prepared to 28 _See NAS_ REPORT, supra note 11, at 16 n.11 (noting decentralization of U.S. elections is one aspect of the current U.S. election system that protects against cyberattacks). 29 Kim Zetter, The Crisis of Election Security, N.Y. TIMES MAG. (Sept. 26, 2018), https://www.nytimes. com/2018/09/26/magazine/election-security-crisis-midterms.html. 30 _See generally SENATE INTELLIGENCE REPORT, supra note 13; NAS_ REPORT, supra note 11. 31 _See_ _infra Part I.B._ 32 Lawrence Norden, How to Secure Elections for 2020 and Beyond, BRENNAN CTR. FOR JUST. (Oct. 23, 2019), https://www.brennancenter.org/our-work/research-reports/how-secure-elections-2020-and-beyond. ----- confront a threat from a foreign nation-state.[33] States and municipalities have demonstrated an inability to handle attacks from a foreign nation-state and have still not taken adequate steps to secure election infrastructure at the local level.[34] Therefore, a foreign threat to U.S. elections requires a uniform federal response, and Congress must pass legislation to preserve the integrity of federal elections. _A. Russian Interference in the 2016 U.S. Election_ The 2016 U.S. election presented challenges that states, municipalities, and the nation had not previously faced. Russia made a concerted to effort to interfere with and disrupt many aspects of the election.[35] One line of attack was to launch cyberattacks against electronic components of state election systems.[36] Actors sponsored by the Russian government “obtained and maintained access to multiple U.S. state or local electoral boards.”[37] Although the Senate Intelligence Committee found no evidence that vote tallies were changed or that voter registration records were altered, the committee’s insight is limited in this regard because a full forensic analysis has not been done.[38] What is certain is that Russian government-affiliated actors “conducted an unprecedented level of activity” that targeted state election systems leading up to the 2016 election.[39] Russian hacking into U.S. election infrastructure was a “watershed moment” in the history of U.S. elections.[40] Protecting election infrastructure became a national security issue when Russia targeted cyberattacks against U.S. voter databases and election systems.[41] The Intelligence Community first detected evidence of hacking into state election systems in the summer of 2016.[42] In July 33 _Id. (“[I]t is not reasonable to expect each of these state and local election officials to independently_ defend against hostile nation-state actors.”) (statement of Bob Brehm, co-executive director of the New York State Board of Elections) (internal quotation marks omitted); see _infra Part III.B._ 34 _See infra Part III.B._ 35 SENATE INTELLIGENCE REPORT, supra note 13, at 3. 36 NAS REPORT, supra note 11, at 1. 37 _Id. (quoting OFF. OF THE DIR. OF NAT’L INTEL.,_ ASSESSING RUSSIAN ACTIVITIES AND INTENTIONS IN RECENT US ELECTIONS iii (2017), https://www.dni.gov/files/documents/ICA_2017_01.pdf.) 38 NAS REPORT, supra note 11, at 2 n.3. The NAS committee was not aware of any ongoing investigation into the possibility that vote tallies were changed. Deficiencies in “intelligence gathering, information sharing, and reporting” leave some uncertainty about the exact consequences of Russia’s attacks. _Id.; SENATE_ INTELLIGENCE REPORT, supra note 13, at 5; Zetter, supra note 29. 39 SENATE INTELLIGENCE REPORT, supra note 13, at 5. 40 NAS REPORT, supra note 11, at xii. 41 _Id. at 117._ 42 SENATE INTELLIGENCE REPORT, _supra note 13, at 6. The U.S. Intelligence Community consists of_ sixteen agencies working under the coordination of the Office of the Director of National Intelligence. The sixteen agencies are: Central Intelligence Agency, Defense Intelligence Agency, Federal Bureau of Investigation, National Geospatial-Intelligence Agency, National Reconnaissance Office, National Security ----- 2016, Illinois noticed unusual activity on the state’s Board of Elections voter registry website.[43] An FBI investigation discovered that the activity resulted in data being exfiltrated from the voter registration database.[44] Ultimately, the FBI determined that Russian actors successfully penetrated Illinois’s voter registration database, viewed multiple database tables, and eventually accessed up to 200,000 voter registration records.[45] Russian cyber actors were in a position to delete or change voter data, although there is no evidence that they did so.[46] Further, evidence shows that Russian operatives targeted several small jurisdictions around the country. In the summer of 2016, General Staff of the Russian Army (GRU) officers sought “access to state and local election computer networks by exploiting known software vulnerabilities” on state and local government websites.[47] By mid-August 2016, federal cybersecurity personnel became confident that Russian cyber actors were probing the election infrastructures and voter registration databases of several states.[48] By late September of that year, U.S. intelligence agencies identified twenty-one states that were targeted by Russian government cyber actors.[49] Eventually, intelligence officials concluded that Russia had attempted to invade the election systems of all fifty states.[50] In one line of attack, GRU officers sent spear-phishing emails to over 120 Florida county election officials.[51] The emails contained an attached Word document carrying a virus that would permit the GRU to access an infected computer.[52] The FBI believes, through this operation, the GRU was able to gain access to the network of at least one county government in Florida.[53] Eventually, Agency/Central Security Service, U.S. Department of Energy, U.S. Department of Homeland Security, U.S. Department of State, U.S. Department of the Treasury, Drug Enforcement Administration, U.S. Air Force, U.S. Army, U.S. Coast Guard, U.S. Marine Corps, and U.S. Navy. NAS REPORT, supra note 11, at 1 n.2. Russian activity began as early as 2014. SENATE INTELLIGENCE REPORT, supra note 13, at 3. 43 SENATE INTELLIGENCE REPORT, supra note 13, at 6. 44 _Id._ 45 _Id. at 22._ 46 _Id._ 47 Michael McFaul & Bronte Kass, _Understanding Putin’s Intentions and Actions in the 2016 U.S._ _Presidential Election, in SECURING AMERICAN ELECTIONS,_ _supra note_ 23, at 5, 14. 48 SENATE INTELLIGENCE REPORT, supra note 13, at 7. 49 _Id._ 50 _Id. at 12._ 51 Herbert Lin, Alex Stamos, Nate Persily & Andrew Grotto, Increasing the Security of U.S. Election _Infrastructure, in SECURING AMERICAN ELECTIONS,_ _supra note_ 23, at 17, 18. 52 ROBERT S. MUELLER, III, U.S. DEP’T OF JUST., REPORT ON THE INVESTIGATION INTO RUSSIAN INTERFERENCE IN THE 2016 PRESIDENTIAL ELECTION 51 (2019). 53 _Id._ ----- a Russian operative was indicted by Special Counsel Robert Mueller for probing election websites of certain rural counties in Georgia, Florida, and Iowa in October 2016.[54] Russia also targeted electronic pollbook systems in several states.[55] In one example of an attack on Election Day in 2016, registered voters in North Carolina were denied the right to vote when the local electronic pollbook systems could not locate their records.[56] Although hacking was never proven to be the cause of the electronic pollbook discrepancy, a forensic analysis was not conducted as county election officials in North Carolina declined the FBI’s offer to investigate.[57] The Intelligence Community understood the seriousness of the foreign attacks.[58] In October 2016, the Department of Homeland Security (DHS) and the Office of the Director on National Intelligence issued a joint statement on election security, which revealed that the probing of state election systems had originated from “servers operated by a Russian company.”[59] The statement also warned state and local governments about the cybersecurity threats and asked them to seek assistance from DHS.[60] In January 2017, then-DHS Secretary Jeh Johnson issued a statement designating U.S. election infrastructure as a part of the nation’s critical infrastructure, which made election systems an ongoing “priority for cybersecurity assistance and protections” from DHS.[61] Members of the Intelligence Community generally agreed that some of Russia’s motives for the cyberattack were to sow discord and undermine voters’ confidence in the 54 Indictment at 26, U.S. v. Netyksho, No. 18-cr-00215 (D.D.C. Jul. 13, 2018). 55 Benjamin Wofford, The Hacking Threat to the Midterms Is Huge. And Technology Won’t Protect Us, VOX (Oct. 25, 2018, 5:00 AM), https://www.vox.com/2018/10/25/18001684/2018-midterms-hacked-russiaelection-security-voting. 56 _Id. Electronic pollbooks are electronic voter check-in databases that are increasingly being used in_ place of paper voter rolls in precincts around the U.S. See _infra Part I.B._ 57 Wofford, supra note 55. 58 SENATE INTELLIGENCE REPORT, supra note 13, at 7–8. 59 Press Release, DHS & ODNI Election Sec., Joint Statement on Election Security (Oct. 7, 2016), https://www.dni.gov/index.php/newsroom/press-releases/press-releases-2016/item/1635-joint-dhs-and-odnielection-security (“We believe, based on the scope and sensitivity of these efforts, that only Russia’s senior-most officials could have authorized these activities.”). 60 _Id._ 61 Press Release, Jeh Johnson, DHS Sec’y, Statement on the Designation of Election Infrastructure as a Critical Infrastructure Subsector (Jan. 6, 2017), https://www.dhs.gov/news/2017/01/06/statement-secretaryjohnson-designation-election-infrastructure-critical. Election infrastructure is comprised of “storage facilities, polling places, and centralized vote tabulations locations used to support the election process, and information and communications technology to include voter registration databases, voting machines, and other systems to manage the election process and report and display results on behalf of state and local governments.” Id. ----- U.S. election system.[62] However, intelligence officials believed that the general public did not fully comprehend the threat and had a dim understanding of the vastness of Russia’s attack during the 2016 election.[63] The attacks did not subside after the 2016 election. Russia continued to attack U.S. election infrastructure for the purpose of interfering with the 2018 midterm elections.[64] The Intelligence Community was clearly aware of the ongoing threat from Russia.[65] As one U.S. cybersecurity expert noted before the 2018 midterm elections, “The Russians will attempt, with cyberattacks and with information operations, to go after us again. They’re doing it right now.”[66] An October 11, 2018, DHS Report stated, “We judge that numerous actors are regularly targeting election infrastructure, likely for different purposes, including to cause disruptive effects, steal sensitive data, and undermine confidence in the election. We are aware of a growing volume of malicious activity targeting election infrastructure in 2018[.]”[67] There is now abundant evidence that Russia targeted the campaigns of at least a dozen House and Senate candidates in the 2018 midterm elections.[68] The Intelligence Community also believes that Russia continued its activity against state and local election systems.[69] The extent to which Russia succeeded in its endeavors in 2018 is still not known.[70] Russia has demonstrated it has sufficient sophistication and knowledge of U.S. voting patterns to understand that cyberattacks on local election systems could cause significant disruption.[71] Although it may be difficult to change vote tallies across the country in national elections, cyber actors can access databases in particular districts, manipulate voter files, and cause enough voter suppression to impact the outcome.[72] Therefore, an attack on a few key battleground states 62 SENATE INTELLIGENCE REPORT, supra note 13, at 35–36. 63 Wofford, supra note 55. 64 _Id._ 65 _Id._ 66 _Id. (quoting Eric Rosenbach, former Pentagon Chief of Staff)._ 67 SENATE INTELLIGENCE REPORT, supra note 13, at 21. 68 Wofford, supra note 55. 69 _See SENATE_ INTELLIGENCE REPORT, _supra_ note 13, at 10 (stating that prior to the 2018 midterm election, DHS determined “numerous actors are regularly targeting election infrastructure, likely for different purposes, including to cause disruptive effects, steal sensitive data, and undermine confidence in the election”). 70 _See Lin et al., supra note 51, at 18–19 (“[T]here is no evidence that votes were actually changed and_ that no lasting damage was done to voter registration databases. Nonetheless, these incidents should be viewed as precursors or dress rehearsals for similar attacks against the 2020 U.S. presidential election.”). 71 Eric Manpearl, _Securing U.S. Election Systems: Designating U.S. Election Systems as Critical_ _Infrastructure and Instituting Election Security Reforms, 24 B.U._ J. SCI. & TECH. L. 168, 175 (2018). 72 _Id. at 173–74; Zetter, supra note 29._ ----- during a presidential race could swing the election.[73] Because small manipulations are easier to perpetrate without detection, the risk that cyberattacks may affect the result of an election is “greatest when the electorate is evenly divided and vote counts are close, as has been the case recently in a number of Presidential elections.”[74] Attacks on specific competitive districts during congressional elections could also substantially change the composition of the federal legislature.[75] No proof exists that such attacks have occurred, but they are certainly a risk for the future.[76] The consensus opinion among the Intelligence Community is that the threat of foreign cyberattacks on U.S. election systems persists.[77] And the risk is not just from Russia. Evidence shows that China, Iran, North Korea, and ISIS have all conducted cyber intrusions against U.S. election infrastructure.[78] ### B. Recommendations of Cybersecurity Experts to Strengthen U.S. Election Infrastructure Election cybersecurity experts generally agree that certain remedies would create a more secure U.S. election system. Because of long-standing concerns about insecure voting systems and the recent recognition of foreign cyberattacks, the National Academy of Sciences, Engineering, and Medicine (“NAS”) appointed an ad hoc committee to consider the future of voting in the United States.[79] The NAS committee determined that, due to the events of the 2016 election and the ongoing threat of cyberattacks, the current U.S. system of voting must evolve.[80] In its report, the NAS committee noted that because of the new 73 Manpearl, supra note 71, at 175; NAS REPORT, supra note 11, at 16 n.11; see Zetter, supra note 29 (describing how a few thousand missing votes and a 537-vote victory for George W. Bush in Florida determined the result of the 2000 presidential election). 74 Lin et al., supra note 51, at 19. 75 Manpearl, supra note 71, at 175; NAS REPORT, supra note 11, at 16 n.11. 76 Zetter, supra note 29. 77 SENATE INTELLIGENCE REPORT, supra note 13, at 43 (quoting Russian Interference in the 2016 U.S. _Elections: Open Hearing Before the S. Comm. on Intelligence, 115th Cong. 117 (2017) (statement of Alex_ Halderman, Professor of Computer Science and Engineering, University of Michigan)); see Jeremy Herb, Brian Fung, Jennifer Hansler & Zachary Cohen, Russian Hackers Targeting State and Local Governments Have Stolen _Data, US Officials Say, CNN, https://www.cnn.com/2020/10/22/politics/russian-hackers-election-data/index._ html (Oct. 23, 2020, 11:39 AM) (reporting that “Russian state-sponsored hackers” targeted state and local government and stole voter registration information in the weeks leading up to the 2020 election). 78 William Roberts, Election Security: The Fight to Secure the Vote, 33 WASH. LAW. 12, 14 (2018). 79 The committee was charged with: (1) documenting the current state of technology, standards, and resources for voting technologies; (2) examining the challenges arising out of the 2016 federal election; (3) evaluating advances in current and upcoming technology that can improve voting; and, (4) providing recommendations to make voting “easier, accessible, reliable, and verifiable.” NAS REPORT, supra note 11, at 3–4. 80 _Id. at 121._ ----- foreign threat, “[w]e must think strategically and creatively about the administration of U.S. elections” and must “seriously reexamine . . . the role of federal and state governments in securing our elections.”[81] While cybersecurity experts are not in a position to opine on the constitutionality of federal authority to regulate states in conducting federal elections, they have a strong, coherent, consensus opinion on how best to secure election infrastructure against cybersecurity threats. Experts recommend measures to secure two critical aspects of elections: voter registration databases and vote-casting mechanisms.[82] First, voter registration lists must be complete and accurate.[83] The Help America Vote Act of 2002 (HAVA) required each state to create a statewide voter database, rather than leave the maintenance of voter registration to counties and municipalities.[84] The administration of voter registration databases requires two main large scale tasks.[85] Election administrators must (1) maintain the correct status and relevant information of citizens who are properly registered to vote; and (2) deliver precinct-specific lists of registered voters to each precinct.[86] Because of the complexity and flexibility needed to maintain accurate, upto-date lists of registered voters, lists are by necessity kept electronically.[87] Electronic voter registration databases are easier than paper counterparts to manage and maintain but are vulnerable to cyberattacks.[88] And in many states, “databases containing voter registration lists are connected, directly or indirectly, to the Internet or to state computer networks.”[89] This connectivity creates a significant risk of cyber invasion and manipulation.[90] Manipulation of voter registration data would cause chaos when voters arrive at the polls and find their names have been removed from the rolls.[91] Removing or changing data for a small number of voters in contentious congressional races or in swing states 81 _Id._ 82 Lin et al., supra note 51, at 17. 83 NAS REPORT, supra note 11, at 59. 84 Help America Vote Act of 2002, Pub. L. No. 107-252, 116 Stat. 1666 (codified as amended at 42 U.S.C. §§ 15301–15545) (requiring “a single, uniform, official, centralized, interactive, computerized statewide voter registration list defined, maintained, and administered at the state level”). 85 Lin et al., supra note 51, at 17. 86 _Id._ 87 NAS REPORT, supra note 11, at 57–61. 88 _Id. at 61._ 89 _Id. at 57._ 90 _See_ _infra Part I.C. Russia breached online voter databases in Illinois and Arizona, obtaining personal_ information on tens of thousands of registered voters. SENATE INTELLIGENCE REPORT, supra note 13, at 22–24; NAS REPORT, supra note 11, at 25. 91 SENATE INTELLIGENCE REPORT, supra note 13, at 2. ----- for a presidential race could change the results of an election.[92] The NAS recommends that election administrators routinely assess the integrity of voter registration databases and put in place systems that detect evidence of probing or tampering with the system.[93] The Senate Intelligence Committee recommends updating software in state voter registration systems and maintaining paper backup copies of registration databases.[94] Managing statewide voter registration databases requires states to deliver precinct-specific lists, also known as pollbooks, to each precinct.[95] Pollbooks, which can either be paper-based or electronic, are used to verify voter eligibility and check-in voters.[96] Over 80% of jurisdictions use preprinted paper pollbooks to check-in voters, but the use of electronic pollbooks (e-pollbooks) is increasing.[97] Between 2012 and 2016, there was a 75% increase in use of epollbooks, and now almost half of voters are checked in electronically.[98] E-pollbooks, which may or may not be networked or connected to the internet, provide some advantages over paper pollbooks. E-pollbooks generally speed up the check-in process and can better track which voters have already cast ballots.[99] When networked, e-pollbooks allow polling places to send and receive real-time updates to voter registration data, which is critical for states that use same-day registration.[100] However, e-pollbooks are vulnerable to cyberattacks that could change voter data, disrupt check-in procedures, and manipulate information on who has and has not voted.[101] Alternatively, a “denial of service” attack could simply shut down operation of an e-pollbook, which would altogether disrupt voting at a particular precinct.[102] Currently no national security standards exist for e-pollbooks, and security practices vary by state.[103] The NAS recommends jurisdictions that use epollbooks have paper backup lists available to be used in the event of any 92 Manpearl, supra note 71, at 175. 93 NAS REPORT, supra note 11, at 63. 94 SENATE INTELLIGENCE REPORT, supra note 13, at 57 (noting that one state’s voter registration system is more than ten years old). 95 NAS REPORT, supra note 11, at 69. 96 _Id. at 69–70._ 97 _Id. at 70._ 98 _Id._ 99 _Id._ 100 _Id. at 71._ 101 _Id._ 102 _Id. at 72._ 103 _Id. at 71._ ----- disruption or compromise to the electronic version.[104] The NAS also recommends that Congress provide funds for the U.S. Election Assistance Commission to develop national security standards for the use of e-pollbooks.[105] Second, cybersecurity experts generally agree that cybersecurity risks are inherent when states rely entirely on computers for voters to cast ballots.[106] Currently, jurisdictions use a variety of types of ballots, including paper, card, and machine-only, and votes are cast by a variety of mechanisms.[107] In the majority of jurisdictions, voters mark their choices on paper ballots, either by hand or by using a ballot-marking device (BMD).[108] Paper ballots are either hand-counted or machine-counted, most commonly by optical scanners.[109] Several states use direct recording electronic (DRE) voting machines in at least some jurisdictions.[110] DREs are free-standing computer units that record selections voters make using a touchscreen.[111] States purchased DREs with funding from HAVA, which was passed as a response to the problems with lever machines and punch card ballots in the 2000 presidential election.[112] The advent of DREs introduced “new technical challenges,” such as touchscreen miscalibration, which causes a voter’s intended selection of one candidate to be misinterpreted as a vote for another candidate.[113] Almost immediately, several security risks with DREs were identified, leading some states to decertify and stop using the machines as early as 2007.[114] Cybersecurity experts now recognize the full extent of the cybersecurity risks with DREs. In its report on election security, the NAS noted that because they are completely paperless, DREs create a risk that a cyberattack on the 104 _Id. at 72._ 105 _Id._ 106 _Id. at 78._ 107 _Id. at 37, 39._ 108 When voting with a BMD, a voter uses a touchscreen or keypad to mark his or her choices, after which the BMD prints a paper copy of the selections. The paper printout is human-readable. The paper is then scanned and tabulated by a separate device. With some BMD printouts, an optical scanner records and tallies the humanreadable ballot. With other BMDs, the actual selections are recorded on a barcode, which is then read by the tabulating machine. Id. at 39. 109 _Id. at 80._ 110 Lawrence Norden & Andrea Cordova, Voting Machines at Risk: Where We Stand Today, BRENNAN CTR. FOR JUST. (Mar. 5, 2019), https://www.brennancenter.org/our-work/research-reports/voting-machinesrisk-where-we-stand-today. 111 NAS REPORT, supra note 11, at 78. 112 Zetter, supra note 29. 113 NAS REPORT, supra note 11, at 78. 114 Zetter, supra note 29. ----- machines will be undetectable.[115] A computer virus could steal votes from one candidate and assign them to another or could stop a machine from accepting votes altogether.[116] According to the Senate Intelligence Report, DRE voting machines “can be programmed to show one result to the voter while recording a different result in the tabulation.”[117] Therefore, the report called for states to discontinue using DREs, which “are now out of date.”[118] A cybersecurity expert actually demonstrated in a courtroom how a DRE machine could be infected with malware that could alter vote counts on the machine.[119] The same expert showed that malware could be introduced remotely and be spread from machine to machine.[120] The Senate Intelligence Report concluded that “[p]aper ballots and optical scanners are the least vulnerable to cyberattack.”[121] Secure voting systems must allow a voter to verify that the recorded ballot reflects his or her intent, which is not possible with paperless DRE machines.[122] Therefore, the NAS recommends that “[w]ell designed, voter-marked paper ballots” be the standard way for voters to cast their votes.[123] The consensus opinion from national cybersecurity experts is that an independent record of the voter’s physical ballot is essential as a reliable audit tool.[124] An auditable record can be achieved by using hand-marked paper ballots.[125] When voting machines are used to mark ballots, the machine must provide a physical, human-readable record of the voter’s selections.[126] National security experts also agree that the threat of foreign interference in U.S. elections persists.[127] In his testimony before Senate Intelligence Committee, former Assistant Attorney General for National Security John Carlin stated, I’m very concerned about . . . our actual voting apparatus, and the attendant structures around it . . . . We’ve literally seen it already, so 115 NAS REPORT, supra note 11, at 78. 116 SENATE INTELLIGENCE REPORT, supra note 13, at 42. 117 _Id._ 118 _Id._ at 59; _see also Zetter, supra note 29 (noting that as early as 2007, some states have decertified_ electronic voting machines after finding them to be susceptible to viruses and malicious software). 119 Curling v. Kemp, 334 F. Supp. 3d 1303, 1308 (N.D. Ga. 2018). 120 _Id. at 1309. Accordingly, a federal judge in Georgia ordered a permanent injunction against the use of_ DRE machines in the state after 2019. See infra Part III.C. 121 SENATE INTELLIGENCE REPORT, supra note 13, at 59. 122 NAS REPORT, supra note 11, at 79. 123 _Id._ 124 _Id. at 79–80._ 125 _Id. at 42._ 126 _Id. at 78._ 127 SENATE INTELLIGENCE REPORT, supra note 13, at 43. ----- shame on us if we can’t fix it heading into the next election cycles. And it’s the assessment of every key intel professional, which I share, that Russia’s going to do it again because they think it was successful. So we’re in a bit of a race against time heading up to the two-year election. Some of the election machinery that’s in place should not be.[128] Consequently, “[g]iven Russian intentions to undermine the credibility of the election process, states should take urgent steps to replace outdated and vulnerable voting systems.”[129] _C. States Responded Inadequately and Ineffectively to Russian Cyberattacks_ In the summer of 2016, after it became clear to the Intelligence Community that foreign actors were attacking state election infrastructure, intelligence officials began the process of reaching out to states to offer cybersecurity support.[130] During a call with state election officials on August 15, 2016, DHS Secretary Jeh Johnson offered to provide help to states by inspecting voting systems for viruses and other signs of cyber invasion.[131] DHS proposed conducting on-site risk and vulnerability assessments as well as remote “cyber hygiene scans” on internet-connected election management systems such as voter registration databases.[132] Several states rejected the offer for help. According to Secretary Johnson, the general response from state officials was “[t]his is our responsibility and there should not be a federal takeover of the election system.”[133] Then-Georgia Secretary of State Brian Kemp cited concerns about “federal overreach” and claimed that help from federal intelligence agencies would “subvert the [C]onstitution to achieve the goal of federalizing elections under the guise of security.”[134] Similarly, Louisiana Secretary of State Tom Schedler chided Congress for overemphasizing the extent of the risk and stated that election administration should be left to the states because “[t]hat’s 128 _Id. (quoting Interview by Senate Select Comm. on Intel. with John Carlin, Former Assistant Att’y Gen._ for Nat’l Sec. (Sept. 25, 2017)). 129 _Id. at 58._ 130 _Id. at 46–47._ 131 _Id. at 47–48; Aliya Sternstein, At Least One State Declines Offer for DHS Voting Security, NEXTGOV_ (Aug. 25, 2016), https://www.nextgov.com/cybersecurity/2016/08/some-swing-states-decline-dhs-votingsecurity-offer/131037/. 132 SENATE INTELLIGENCE REPORT, supra note 13, at 52. 133 _Id. at 47._ 134 Sternstein, supra note 131. ----- what the Constitution says.”[135] Republican legislators also blocked funds for election security in Minnesota and Arizona.[136] Even more concerning, many states failed to recognize the extent or seriousness of the threat and chose not to heed warnings from the Intelligence Community.[137] Several states also opposed the decision of Secretary Johnson to designate U.S. election systems as critical infrastructure.[138] DHS initially intended to make the designation in August 2016 but held off until January 2017 because of pushback from state election officials.[139] Again rejecting federal support, the National Association of Secretaries of State (NASS) expressed opposition to DHS’s critical infrastructure designation, mistakenly citing states’ primacy in regulating elections.[140] The NASS stated that DHS “has no authority to interfere with elections, even in the name of national security.”[141] Secretary Kemp declared that “[d]esignating voting systems or any other election system as critical infrastructure would be a vast federal overreach.”[142] Despite the dire warnings and offers to help from the Intelligence Community, states did little to respond to the ongoing threat of cyberattacks on election systems. Even after the breaches to databases in Illinois and Arizona were known, states continued to struggle to respond to security risks.[143] States have displayed widely varying degrees of concern about election security and efforts to address the security risks. For the most part, states relied on the same insecure infrastructure to conduct elections in 2018 as they did in 2016, despite the known risks.[144] But the attacks on local elections systems did not subside 135 Aliya Sternstein, _9 States Accept DHS’s Election Security Support, NEXTGOV_ (Sept. 21, 2016), https://www.nextgov.com/cybersecurity/2016/09/9-states-accept-dhss-election-security-support/131741/. 136 Gopal Ratnam, Democrats Target State Elections with Focus on Election Security, ROLL CALL (Aug. 22, 2019), https://www.rollcall.com/news/congress/democrats-target-state-elections-focus-election-security. 137 _See infra Part III.B._ 138 Manpearl, supra note 71, at 186. The purpose of a critical infrastructure designation is to allow the Federal Government to partner with and provide support to the identified sectors. The designation added U.S. election systems to the other critical infrastructure sectors: chemical; commercial facilities; communications; critical manufacturing; dams; defense industrial base; emergency services; energy; financial services; food and agriculture; government facilities; health care and public health; information technology; nuclear reactors, materials, and waste; transportation systems; and water and wastewater systems. Press Release, Off. of the Press Sec’y, Presidential Policy Directive—Critical Infrastructure Security and Resilience (Feb. 12, 2013), https://obamawhitehouse.archives.gov/the-press-office/2013/02/12/presidential-policy-directive-criticalinfrastructure-security-and-resil. 139 SENATE INTELLIGENCE REPORT, supra note 13, at 48–49. 140 Manpearl, supra note 71, at 187. 141 Nat’l Ass’n of Sec’ys of State, NASS Resolution Opposing the Designation of Elections as Critical _Infrastructure, at 21–22 (Feb. 18, 2017)._ 142 Sternstein, supra note 131. 143 SENATE INTELLIGENCE REPORT, supra note 13, at 39; Norden & Cordova, supra note 110. 144 Wofford, supra note 55. ----- after the 2016 election, and states continue to be ill-equipped to handle the attacks.[145] Georgia, for example, exhibited a grossly inadequate response to the cybersecurity challenges that came to light in the 2016 election. The Georgia Secretary of State’s Office left its registration database completely open to hackers with 6.5 million voter records exposed during a six-month period in 2016–17.[146] U.S. cybersecurity experts were able to access the database and even plant files during that time.[147] Malicious actors could have manipulated the data, including dropping voters from the database or changing their data.[148] But Georgia election officials claimed they saw no evidence that any election related data was compromised.[149] However, a forensic evaluation was not done initially because Georgia officials wiped the server that housed the data after the breach was discovered.[150] Evidence from an FBI image taken of the server before it was wiped shows that there may have been signs of tampering.[151] Georgia also knew of the substantial evidence that Russia was targeting election systems and that its paperless, internet-connected voting system was ripe for hacking.[152] Yet, it made no significant changes, and in the 2018 federal election, voters cast ballots on the same outdated, insecure system used in 2016.[153] Georgia election officials were reluctant to acknowledge the full extent of the vulnerability of Georgia’s electronic voting equipment even though security flaws in DRE machines had been known for over a decade and Georgia had not updated the software on its machines since 2005.[154] Therefore, Georgia voters used the same hackable and non-auditable voting machines in the 2018 145 _Id._ 146 NAS REPORT, supra note 11, at 58. 147 Frank Bajak, Georgia Election Server Wiped After Suit Filed, PBS NEWSHOUR (Oct. 26, 2017, 9:34 AM), https://www.pbs.org/newshour/politics/georgia-election-server-wiped-after-suit-filed. 148 NAS REPORT, supra note 11, at 57. 149 Frank Bajak, Georgia Election Server Showed Signs of Tampering, AP (Jan. 16, 2020), https://apnews. com/39dad9d39a7533efe06e0774615a6d05. 150 Kim Zetter, Georgia Election Systems Could Have Been Hacked Before 2016 Vote, POLITICO (Jan. 16, 2020, 11:07 PM), https://www.politico.com/news/2020/01/16/georgia-election-systems-could-have-beenhacked-before-2016-vote-100334. 151 _Id._ 152 _See Curling v. Kemp, 334 F. Supp. 3d 1303, 1327 (N.D. Ga. 2018) (“[Georgia] stood by for far too_ long, given the mounting tide of evidence of the inadequacy and security risks of Georgia’s DRE voting system and software.”). 153 _See_ Curling v. Raffensperger, 397 F. Supp. 3d 1334, 1382–92 (N.D. Ga. 2019) (summarizing the affidavits of 137 Georgia voters, 2 county pollworkers, and 15 pollwatchers, and concluding that the “same pattern of problems with Georgia’s voting systems and registration databases has persisted across multiple elections cycles”). 154 _Id. at 1339, 1348._ ----- midterm elections.[155] As a result, voters in Georgia experienced significant difficulty voting in 2018.[156] Problems reported by voters included long lines due to malfunctioning machines being taken out of service, machines selecting the wrong candidates when voters marked their choices on touchscreens, and checkin problems with e-pollbooks, including incorrect polling places or incorrect addresses listed for voters.[157] A federal court noted that Georgia state election officials had “stood by for far too long” and “buried their heads in the sand” rather than address the inadequacy and insecurity of Georgia’s voting system.[158] Similarly, North Carolina refused an offer from the FBI to investigate election irregularities in 2016.[159] A forensic analysis was never conducted after registered voters could not be located in local e-pollbook systems.[160] Although hacking was never proven as the cause of the e-pollbook discrepancy, it was discovered that Russia targeted e-pollbook systems in several states, including North Carolina.[161] Despite knowing that information, county election officials in North Carolina declined the FBI’s offer to investigate.[162] Given that some states and municipalities have demonstrated they are incapable and, in some instances, even unwilling to secure election infrastructure, the United States needs a national election infrastructure plan. Such a plan should follow the recommendations of national cybersecurity experts to provide uniformity and address vulnerabilities in many state and local election systems. II. THE LANDSCAPE OF CONGRESSIONAL AUTHORITY OVER FEDERAL ELECTIONS Many state election officials, scholars, and federal legislators consider primary authority over the conduct of federal elections to belong to the states. For example, the first recommendation in the Senate Intelligence Report on 155 _Id. at 1392; see Adam Levin & Beau Friedlander, Georgia’s Shaky Voting System, N.Y._ TIMES (Nov. 13, 2018), https://www.nytimes.com/2018/11/13/opinion/voting-machines-georgia-security.html (describing how Georgia, for its 2018 gubernatorial election, relied on the same voting system it used in 2016 despite the cybersecurity vulnerabilities that had been identified). 156 Mark Niesse, Long Lines and Equipment Problems Plague Election Day in Georgia, AJC (Nov. 6, 2018), https://www.ajc.com/news/state—regional-govt—politics/long-lines-and-equipment-problems-plagueelection-day-georgia/l7NUidWbMetr5OFdGcb5ZM/. 157 _Curling, 397 F. Supp. 3d at 1383._ 158 _Curling, 334 F. Supp. 3d at 1327._ 159 Wofford, supra note 55. 160 _Id._ 161 _Id._ 162 _Id._ ----- Russian interference in the 2016 election is to “reinforce states’ primacy in running elections.”[163] The Supreme Court’s view on whether the federal government or states have the ultimate right to prescribe the manner in which federal elections are conducted has been unclear. The pendulum of the Court’s interpretation of the differential authority between Congress and the states over federal elections has swung back and forth for two centuries. From the antebellum era to the Reconstruction Amendments to the VRA to the Court’s decision in Shelby County, the Court has expanded and contracted congressional authority relative to state sovereignty. But even this pendulum swing has remained in a somewhat narrow range because Congress has never attempted to exercise the full breadth of its authority under the Elections Clause. The vast majority of congressional action to regulate elections since the Civil War has been pursuant to the Reconstruction Amendments rather than the Elections Clause.[164] Even when congressional authority was at its peak under the VRA, Congress approached election legislation from a deferential framework. Congress only passed the VRA after the Civil Rights Movement’s expansive and concerted fight for voting rights in the South brought national attention and shifted public opinion on this issue.[165] The Supreme Court upheld this action by Congress under the Enforcement Clause of the Fifteenth Amendment because of the long-standing and pernicious evil of racial discrimination in voting.[166] But Congress has yet to exercise and the Court has yet to uphold the full extent of Congress’s power to enact federal election legislation under the Elections Clause, which extends beyond antidiscrimination. _A. Congressional Authority Under the Reconstruction Amendments and the_ _Voting Rights Act_ The end of the Civil War and the Reconstruction era brought a new paradigm to the balance of federal authority versus state autonomy. The Fourteenth Amendment provided an avenue for Congress to ensure that each state did not abridge or deny certain rights to its own citizens.[167] The Fifteenth Amendment 163 SENATE INTELLIGENCE REPORT, supra note 13, at 54. 164 Franita Tolson, The Spectrum of Congressional Authority Over Elections, 99 B.U. L. REV. 317, 341 (2019). 165 CAROL ANDERSON, ONE PERSON, NO VOTE: HOW VOTER SUPPRESSION IS DESTROYING OUR DEMOCRACY 21–22 (2018); see South Carolina v. Katzenbach, 303 U.S. 301, 315 (1966) (“The burden is too heavy—the wrong to citizens is too serious—the damage to our national conscience too great not to adopt more effective measures than exist today.”). 166 _Id. at 303–04._ 167 U.S. CONST. amend. XIV. ----- prohibited states from denying the right to vote “on account of race, color, or previous condition of servitude.”[168] Despite the Fifteenth Amendment guarantee, many former Confederate states still prevented African American citizens from exercising their new constitutional right to vote.[169] But embedded in the Reconstruction Amendments were enforcement provisions that established a role for Congress to protect the rights of all citizens against state action.[170] The constitutional enfranchisement of African American voters created a new framework for Congress to play a greater role in elections in order to protect the right to vote. While Congress had the power to enforce the Reconstruction Amendments to prevent states from infringing on their citizens’ right to vote, the Reconstruction-era framework preserved a concept of federalism and state sovereignty over the conduct of elections.[171] Congress attempted to exert broad authority to regulate elections through the Enforcement Acts of 1870 and 1871, which instituted a system of federal oversight for congressional elections.[172] However, despite Congress’s greater power to protect voters under the Reconstruction Amendments, the Supreme Court did not allow Congress full license to regulate elections. In United States v. Reese, the Court struck down provisions of the Enforcement Act of 1870 because they exceeded the scope of Congress’s mandate under the Fifteenth Amendment.[173] The Court held that section 4 of the statute was invalid because it created criminal penalties for state officials who denied citizens the right to vote.[174] According to the Court, the Fifteenth Amendment did not confer upon Congress expansive power to regulate elections and protect voters, but simply prevented states from discriminating based on race.[175] Similarly, the Court restrained Congress from using the Enforcement Act of 1870 to assert broad authority over states pursuant to the Fourteenth Amendment in United States v. Cruikshank.[176] In that case, election inspectors in Louisiana were criminally charged with conspiring to prevent two African American 168 U.S. CONST. amend. XV, § 1. 169 ANDERSON, supra note 165, at 2. 170 U.S. CONST. amend. XIII, § 2; U.S. CONST. amend. XIV, § 5; U.S. CONST. amend. XV, § 2. 171 Tolson, supra note 164, at 354. 172 Enforcement Act of 1870, ch. 114, 16 Stat. 140; Enforcement Act of 1871, ch. 99, 16 Stat. 433; Tolson, _supra note 164,_ at 358. 173 United States v. Reese, 92 U.S. 214, 220 (1875). 174 _Id. at 217–18, 220._ 175 _Id. at 217._ 176 United States v. Cruikshank, 92 U.S. 542, 555 (1875). ----- citizens from exercising their right to vote.[177] The Court dismissed the indictments, holding that the Louisiana officials did not intentionally discriminate based on race.[178] Importantly, the Court noted that the federal government had authority to prohibit discrimination under the Fourteenth Amendment, but the right to vote itself came from the states.[179] The Court, however, did not address Congress’s power to regulate elections and ensure the right to vote under the Elections Clause. The post-Reconstruction era, beginning with the federal government’s withdrawal of military troops in 1876, allowed Southern states to construct significant structural barriers to African American suffrage.[180] Discriminatory devices to prevent African Americans from voting were enacted into state laws and even embedded into the constitutions of several former Confederate states.[181] In addition to literacy tests, poll taxes, and good-morals requirements, the small percentage of African Americans who were able to cast ballots in the South often had to overcome outright violence.[182] During the Jim Crow era of renewed disenfranchisement, the Supreme Court invalidated several state laws designed to prevent African Americans from voting as violations of the Fourteenth and Fifteenth Amendment.[183] However, case-by-case litigation was essentially a game of whack-a-mole. Each time federal courts struck down a discriminatory state law that restricted the right of its citizens to vote, states found insidious, creative alternative ways to disenfranchise African American voters.[184] For example, after two Supreme Court decisions invalidated all-white primary elections, states such as South Carolina and Texas found ways to unofficially hold “pre-primaries” without such laws being on their books.[185] The Civil Rights Movement forced Congress 177 _Id. at 544–45._ 178 _Id. at 556–57._ 179 _Id. at 554–56 (holding that the Fourteenth Amendment only confers on Congress the power to ensure_ that states do not deny the equality of rights of their citizens, but states still assume the primary duty to guarantee these rights: “The power of the national government is limited to the enforcement of this guaranty.”). 180 ANDERSON, _supra note 165, at 2–3._ 181 Virginia E. Hench, The Death of Voting Rights: The Legal Disenfranchisement of Minority Voters, 48 CASE W. RES. L. REV. 727, 733–43 (1998). 182 ANDERSON, _supra note 165, at_ 14–18. 183 _See,_ _e.g., Schnell v. Davis, 336 U.S. 933, 933 (1949) (striking down, as a violation of the Equal_ Protection Clause, a provision of the Alabama state constitution that required citizens to understand and explain an article of the U.S. Constitution in order to exercise the right to vote). 184 ANDERSON, _supra note 165, at_ 13. 185 Smith v. Allwright, 321 U.S. 649, 656–57 (1944); ANDERSON, _supra note 165, at_ 13. ----- to enact a comprehensive plan to “banish the blight of racial discrimination in voting.”[186] Nearly a century after the Fourteenth and Fifteenth Amendments were ratified, Congress responded to the grassroots efforts of the Civil Rights Movement by passing the Voting Rights Act of 1965.[187] The VRA prescribed remedies for voting discrimination that it imposed on particular states that were known to have constructed the greatest barriers for African American voters.[188] By exercising its power under the Enforcement Clause of the Fifteenth Amendment, Congress supplanted the right of states to enact particular discriminatory voter qualification laws.[189] The VRA placed significant constraints on states’ autonomy in determining voter qualifications.[190] Section 5 of the VRA required states or counties that had a history of discriminating against African American voters, as defined in section 4(b), to submit to preclearance by the U.S. Attorney General of any new law that impacted voter qualifications or registration.[191] The Act also authorized federal examiners to directly place and remove voters from the registration lists of states and localities who fell under the VRA’s coverage formula.[192] When the Supreme Court upheld the VRA as “an appropriate means for carrying out Congress’ constitutional responsibility,” federal authority to regulate elections under the Reconstruction Amendments was at its zenith.[193] South Carolina challenged the VRA on the grounds it exceeded Congress’s powers and infringed on a function that had traditionally been left to states.[194] But the Court dismissed these concerns.[195] The Court held that “[a]s against the reserved powers of the States, Congress may use any rational means to effectuate the constitutional prohibition of racial discrimination in voting.”[196] The Court in 186 South Carolina v. Katzenbach, 383 U.S. 301, 308 (1966); ANDERSON, _supra note 165, at 21–22._ 187 The Voting Rights Act was signed into law by President Lyndon Johnson on August 6, 1965. _See_ Voting Rights Act of 1965, Pub. L. No. 89-110, §§ 1–19, 79 Stat. 437 (codified as amended in scattered sections of 52 U.S.C.); see Eric S. Lynch, Trusting the Federalism Process Under Unique Circumstances: United States _Election Administration and Cybersecurity, 60 WM._ & MARY L. REV. 1979, 1991–92 (2019) (noting that President Johnson introduced the voting rights bill to Congress three days after the “Bloody Sunday” Selma-toMontgomery march). 188 §§ 1–7, 79 Stat. at 437–41. 189 U.S. CONST. amend. XV, § 2 (“Congress shall have the power to enforce this provision through appropriate legislation.”); §§ 1–2, 79 Stat. at 437. 190 §§ 1–6, 79 Stat. at 437–40. 191 §§ 4(b)–5, 79 Stat. at 438–39. 192 § 7, 79 Stat. at 440–41. 193 South Carolina v. Katzenbach, 303 U.S. 301, 308 (1966). 194 _Id. at 323._ 195 _Id._ 196 _Id. at 324._ ----- _South Carolina v. Katzenbach stated that Congress’s authority relative to states’_ rights under the Enforcement Clause of the Fifteenth Amendment is just as broad as Congress’s power under the Necessary and Proper Clause.[197] Therefore, to prevent racial discrimination, the Supreme Court established that Congress had paramount authority to supersede state autonomy in determining who was eligible to cast a ballot. According to the Court, “[t]he Voting Rights Act was designed by Congress to banish the blight of racial discrimination in voting, which has infected the electoral process.”[198] The Court emphasized the “unique circumstances” that permitted Congress to exert such expansive powers to violate state sovereignty under the Fifteenth Amendment.[199] The unique circumstances to which the Court referred were the overt discriminatory actions of several former slave states that violated the Fifteenth Amendment.[200] In _Katzenbach, the Court’s_ ratification of Congress’s power to enact the VRA was specific to the era as well as the manner and degree to which the infringement on the rights of African Americans were being infringed.[201] Over the next almost fifty years, the Supreme Court continued to uphold the VRA as a legitimate exercise of Congress’s power to enforce the Fifteenth Amendment.[202] The Court recognized Congress’s authority to invalidate provisions that did not have a stated discriminatory purpose but had a disparate impact on the right of African Americans to vote. In _City of Rome v. United_ _States, the Court upheld the VRA’s ban on changes to a municipality’s voting_ provisions that would have had a discriminatory effect.[203] In that case, the city of Rome, Georgia challenged the VRA on federalism grounds.[204] But the Court made clear that the mandate embedded in the enforcement provisions of the Reconstruction Amendments trumped federalism concerns.[205] The Court stated that “principles of federalism that might otherwise be an obstacle to 197 _Id. at 325–26;_ _see Ex parte Virginia, 100 U.S. 339, 345–46 (1879) (“Whatever Legislation is_ appropriate, that is, adapted to . . . secure to all persons the enjoyment of perfect equality of civil rights and the equal protection of the laws against State denial or invasion, if not prohibited, is brought within the domain of _congressional power.”) (emphasis added)._ 198 _Katzenbach, 383 U.S. at 308._ 199 _Id. at 335 (“Under the compulsion of these unique circumstances, Congress responded in a permissibly_ decisive manner.”). 200 _Id._ 201 _Id. at 326–31._ 202 _See, e.g., Lopez v. Monterey Cnty., 525 U.S. 266, 287 (1999); City of Rome v. United States, 446 U.S._ 156, 173 (1980). 203 _City of Rome, 446 U.S. at 173._ 204 _Id. at 178._ 205 _Id. at 179._ ----- congressional authority are necessarily overridden by the power to enforce the Civil War Amendments ‘by appropriate legislation.’”[206] The Court held that Congress has the power to impose voting regulations on states and their political subdivisions because the “[Reconstruction] Amendments were specifically designed as an expansion of federal power and an intrusion on state sovereignty.”[207] The Supreme Court took its view of federal power over state regulations under the Fifteenth Amendment one step further in Lopez v. Monterey County.[208] In that case, Monterey County was subject to the coverage formula under section 4(b) of the VRA, but the State of California as a whole was not.[209] California passed a state law that determined the manner in which county judges were to be elected.[210] Voters alleged that the law was invalid as applied to Monterey County because any changes to existing law that applied to the county had to be precleared by the federal government.[211] The Court determined that the California law could not take effect in Monterey County until it received preclearance pursuant to section 5 of the VRA.[212] Therefore, the Court recognized that Congress’s authority to enforce the Reconstruction Amendments includes the power to supersede the rights of states to regulate their own counties. Accordingly, at end of the twentieth century, Congress had broad authority under the Fifteenth Amendment to regulate federal elections through the VRA. _B. The Demise of the Voting Rights Act and the Shifting State-Federal_ _Authority to Regulate Elections_ The twenty-first century brought a dramatic shift in the Supreme Court’s deference to Congress to enforce the Fifteenth Amendment through the VRA, which culminated in the Court’s gutting of the VRA in _Shelby County v._ _Holder.[213] Chief Justice John Roberts’s general ideology appears to limit_ congressional power in favor of state sovereignty through principles of federalism.[214] Relying on federalism, the Roberts Court has limited Congress’s 206 _Id. (quoting Fitzpatrick v. Bitzer, 427 U.S. 445, 456 (1976))._ 207 _Id. at 179–80._ 208 Lopez v. Monterey Cnty., 525 U.S. 266, 287 (1999). 209 _Id. at 269._ 210 _Id._ 211 _Id. at 271, 274._ 212 _Id. at 287._ 213 Shelby Cnty. v. Holder, 570 U.S. 529, 556–57 (2013). 214 Joshua A. Douglas, (Mis)Trusting States to Run Elections, 92 WASH. U. L. REV. 553, 580 (2015); see Adam B. Cox & Thomas J. Miles, Judging the Voting Rights Act, 108 COLUM. L. REV. 1, 3 (2008) (demonstrating ----- ability to oversee elections and has elevated the role of states in regulating various aspects of the voting process and election conduct.[215] In sharp contrast to the Civil Rights era that led to the VRA, the Court in recent years has more closely scrutinized Congressional regulation of voting and elections while affording more deference to election laws enacted by states.[216] In 2009, the Court foreshadowed its holding in Shelby _County by expressing_ outright hostility to the VRA in _Northwest Austin Municipal Utility District_ _Number One v. Holder.[217] In that case, a Texas municipal district challenged the_ VRA’s preclearance requirement.[218] The Court avoided the question of the VRA’s constitutionality by resolving the district’s claims on statutory grounds.[219] In dicta, however, the Court raised concerns about whether the VRA was constitutional.[220] In his majority opinion, Chief Justice Roberts noted that section 5 of the VRA “authorizes federal intrusion into . . . state and local policymaking” and “imposes substantial ‘federalism costs.’”[221] The Court also stated that section 5 exceeded Congress’s mandate under the Fifteenth Amendment by suspending all changes to election law in the jurisdictions falling under its coverage formula.[222] In the concluding paragraphs of the opinion, which foreshadowed _Shelby County, the Court claimed that the “exceptional_ conditions” that justified the VRA no longer exist as “we are now a very different Nation.”[223] Four years later, in Shelby County, the Supreme Court struck down section 4(b) of the VRA.[224] Section 4(b) had delineated the “coverage” formula that determined which states and localities were subject to federal preclearance before enacting new voting legislation.[225] In invalidating portions of the VRA, the Court described its rationale as a combination of federalism issues, concerns that judicial ideology impacts judicial decisions regarding voting rights). 215 Douglas, supra note 214, at 583. 216 _Id. at 579; see Franita Tolson, Election Law “Federalism” and the Limits of the Anti-Discrimination_ _Framework, 59 WM._ & MARY L. REV. 2211, 2215 (2018) (arguing that recent case law has limited the extent of Congress’s powers under the Fourteenth and Fifteenth Amendments due to federalism concerns and the Supreme Court now views states as having broad authority to regulate federal elections). 217 Nw. Austin Mun. Util. Dist. No. 1 v. Holder, 557 U.S. 193, 203 (2009). 218 _Id. at 196._ 219 _Id. at 205–06._ 220 _Id. at 204._ 221 _Id. at 202 (quoting Lopez v. Monterey Cnty., 525 U.S. 266, 282 (1999))._ 222 _Id._ 223 _Id. at 211._ 224 Shelby Cnty. v. Holder, 570 U.S. 529, 556–57 (2013). 225 Voting Rights Act of 1965, Pub. L. No. 89-110, § 4(b), 79 Stat. 437, 438 (codified as amended in scattered sections of 52 U.S.C.). ----- about equal sovereignty among states, and changed conditions regarding racial inequality in voting.[226] A concern for state sovereignty predominated Justice Roberts’s majority opinion.[227] The Court described the VRA’s requirement that certain states obtain federal permission before enacting voting laws as “a drastic departure from basic principles of federalism.”[228] Scholars and interested parties soon discovered that the _Shelby County_ decision definitively altered the Court’s view of the balance between state and federal government in regulating elections under the Reconstruction Amendments.[229] Prior to _Shelby County, the Court had generally recognized_ Congress’s authority to supersede state laws regulating elections in order to protect voters’ rights.[230] _Shelby County turned that assumption on its head._ Contrary to the prior understanding of the federal-state balance regarding elections, the Court stated that the original intent of the framers was for states to have primary authority to regulate federal elections.[231] The Court in _Shelby_ _County held that the VRA was only a legitimate exercise of Congress’s power_ when it was enacted because it was the product of a particular time in history.[232] However, the Court’s emphasis in _Shelby County on federalism and state_ sovereignty in conducting elections was misguided. The Court viewed the authority to regulate elections solely from an antidiscrimination perspective and, ignoring its City of Rome precedent, focused on overt discriminatory intent.[233] By only evaluating Congress’s power to protect the rights of minority voters under the Fourteenth and Fifteenth Amendments, the Court discounted Congress’s broad powers to contradict state laws and regulate elections under the Elections Clause. _C. The Elections Clause Grants Congress Broad Authority to Regulate_ _Federal Elections_ Congress’s authority to regulate federal elections under the Elections Clause 226 _Shelby County,_ 570 U.S. at 534–44, 547. 227 _Id. at 535 (stating that the VRA infringed on state sovereignty and section 4 violated “the principle that_ all states enjoy equal sovereignty”). 228 _Id._ 229 _See Charles & Fuentes-Rohwer, supra note 1, at 488, 522 (presenting the case against an “optimistic”_ reading of the Shelby County holding for voting rights advocates). 230 _Id. at 500–01, 516._ 231 _Id. at 517._ 232 _Id. at 495 (noting that Chief Justice Roberts’ majority opinion stated that the VRA was only acceptable_ in 1966 because “exceptional conditions can justify legislative measures not otherwise appropriate” (quoting South Carolina v. Katzenbach, 303 U.S. 301, 334 (1966))). 233 _See Shelby County, 570 U.S. at 551, 553, 556._ ----- is significantly broader than the Court has acknowledged since Shelby County.[234] In _Federalist 59, Alexander Hamilton explained that the Elections Clause_ invested ultimate authority to regulate federal elections in “the national legislature.”[235] Because of the clear mandate of the Elections Clause, the Supreme Court was remiss in Shelby County to overvalue state sovereignty in regard to the conduct of federal elections.[236] The Court mistakenly relied on what it called a “prevailing view that federalism best explains” the U.S. election system.[237] _1. Decentralization Versus Federalism_ The Elections Clause precludes viewing the balance of state-versus-federal authority to regulate elections through traditional notions of federalism.[238] The text and history of the Elections Clause demonstrate that the Constitution prescribed a system for federal elections based on decentralization rather than federalism.[239] Though often conflated, “federalism” and “decentralization” are distinct concepts.[240] Decentralization is a hierarchically organized “managerial concept” in which the leader at the top has plenary power over the subordinate units.[241] Federalism may be structurally similar to decentralization.[242] But as a political concept, federalism implies that the subordinate units retain certain rights and “areas of jurisdiction that cannot be invaded by the central authority[.]”[243] In the United States, federalism denotes separate sovereignty and a “system of parallel federal and state governance.”[244] Regarding federal elections, the Elections Clause prescribes a system of decentralization rather than federalism.[245] A traditional notion of federalism does not bar Congress from enacting broad legislation to dictate the manner in 234 Tolson, supra note 216, at 2217. 235 THE FEDERALIST NO. 59 (Alexander Hamilton). 236 Tolson, supra note 216, at 2214. 237 _Id. at 2216._ 238 _Id. at 2215–18; see Tolson, supra note 164, at 321–22._ 239 U.S. CONST. art. I, § 4; see Franita Tolson, Reinventing Sovereignty?: Federalism as a Constraint on _the Voting Rights Act, 65 VAND._ L. REV. 1195, 1247 (2012) (“The organizational structure of the [Elections] Clause itself is not really federalist, but reflects a decentralized organizational structure that is often confused with federalism.”); Weinstein-Tull, supra note 3, at 790 (noting that some scholars argue that federal election statutes do not implicate federalism, but demonstrate a form of “managerial decentralization”). 240 Edward L. Rubin & Malcolm Feeley, Federalism: Some Notes on a National Neurosis, 41 UCLA L. REV. 903, 910–11 (1994). 241 _Id._ 242 _Id. at 911._ 243 _Id._ 244 Weinstein-Tull, supra note 3, at 775. 245 Tolson, supra note 239, at 1202, 1247. ----- which federal elections will be conducted.[246] In contrast, states have no plenary power to regulate federal elections.[247] States can administer federal elections under direct grant from the Elections Clause but subject to Congress’s ultimate authority.[248] Pursuant to the Elections Clause, “the Constitution primarily treats states as election administrators rather than sovereign entities.”[249] Therefore, states may only regulate federal elections in a managerial sense.[250] Congress has the final say in how authority is delegated and has generally left states “to fill in . . . the blanks with respect to the nuts and bolts of federal elections[.]”[251] _2. Congress Has Used Its Election Clause Authority to a Limited Degree_ In addition to exercising federal authority over elections under the Fifteenth Amendment, Congress has, at times, used its Elections Clause power.[252] Two examples of statutes enacted under the Elections Clause that have been upheld by courts are the National Voter Registration Act of 1993 (NVRA) and the Help America Vote Act of 2002 (HAVA).[253] Congress enacted the NVRA to increase voter participation in elections by making voter registration easier for all eligible citizens.[254] The NVRA requires states to provide opportunities to register to vote when citizens interact with various state government offices, such as applying for driver’s licenses or applying for aid through public assistance and disability services offices.[255] The NVRA also authorizes the federal government to enforce its provisions through civil actions against states.[256] Federal courts have generally upheld the NVRA as a legitimate exercise of Congress’s Elections Clause authority.[257] Despite giving no weight to the 246 Tolson, supra note 216, at 2216 (“Congress and the courts can disregard state sovereignty in enacting, enforcing, and resolving the constitutionality of legislation passed pursuant to the Elections Clause.”). 247 Michael T. Morley, The Intratextual Independent “Legislature” and the Elections Clause, 109 NW. U. L. REV. 847, 849 (2015). 248 _Id._ 249 Harkless v. Bruner, 545 F. 3d 445, 454 (6th Cir. 2008). 250 _See Tolson, supra note 239, at 1197._ 251 Tolson, supra note 216, at 2218. 252 Franita Tolson, The Elections Clause and Underenforcement of Federal Law, 129 YALE L.J. F. 171, 173 (2019). 253 _See Help America Vote Act of 2002, Pub. L. No. 107-252, §§ 101–906, 116 Stat. 1666 (codified as_ amended at 52 U.S.C. §§ 20901–21145); see National Voter Registration Act of 1993, Pub. L. No. 103-31, §§ 1– 13, 107 Stat. 77 (codified as amended at 52 U.S.C. §§ 20501–20511). 254 § 2, 107 Stat. at 77. 255 §§ 4–5, 7, 107 Stat. at 78, 80–81. 256 § 11, 107 Stat. at 88. 257 _See Weinstein-Tull, supra note 3, at 762–63, 765._ ----- Elections Clause in Shelby County, the Supreme Court recognized Congress’s broad power to regulate voter qualification standards under the Elections Clause in Arizona v. Inter Tribal Council of Arizona, Inc.[258] In Inter Tribal Council, the Court held that the NVRA preempted an Arizona state law.[259] The Court noted that the Elections Clause grants Congress final policymaking authority over many aspects of federal elections.[260] The NVRA required states to accept a national mail registration form developed by the Federal Election Commission.[261] The Court held that the NVRA mandate that states “accept and use” a federal form to register voters superseded Arizona’s law that required voters to present proof of citizenship to register to vote.[262] In some cases, courts have noted that Congress’s right to disregard states’ autonomy under the Elections Clause is even broader than its powers under the Commerce Clause.[263] For example, “[i]f Congress determines that the voting requirements established by a state do not sufficiently protect the right to vote, it may force the state to alter its regulations.”[264] In ACORN v. Miller, the Sixth Circuit rejected Michigan’s challenge to the NVRA.[265] Michigan argued that “Congress overstepped its power to regulate federal elections by compelling state legislation to effectuate a federal program, directing states to legislate toward a federal purpose, and forcing states to bear the financial burden of enacting a federal scheme.”[266] However, the Sixth Circuit held that, unlike the Commerce Clause, the Elections Clause “specifically grants Congress the authority to force states to alter their regulations regarding federal elections.”[267] Congress’s power under the Elections Clause extends as far as commandeering state offices and state election officials to carry out federal 258 Arizona v. Inter Tribal Council of Arizona, Inc., 570 U.S. 1, 14–15 (2013). 259 _Id. at 14–15, 20._ 260 _Id. at 8–9._ 261 National Voter Registration Act of 1993, Pub. L. No. 103-31, § 6, 107 Stat. 77, 79–80 (codified as amended at 52 U.S.C. §§ 20501–20511). When HAVA was enacted, this function of the Federal Election Commission transferred to the Election Assistance Commission. See Help America Vote Act of 2002, Pub. L. No. 107-252, § 303, 116 Stat. 1666, 1713–14 (codified as amended at 52 U.S.C. §§ 20901–21145). 262 _Inter Tribal Council, 570 U.S. at 15._ 263 _See Harkless v. Bruner, 545 F. 3d 445, 454 (6th Cir. 2008) (“[U]nlike the Commerce Clause . . . Article_ I section 4 specifically grants Congress the authority to force states to alter their regulations regarding federal elections.” (quoting ACORN v. Miller, 129 F.3d 833, 836 (6th Cir. 1997))). Congress’s power to prescribe the details that state legislatures must adopt to hold federal elections stands in stark contrast to virtually all other provisions of the Constitution. Id. 264 _ACORN, 129 F.3d at 837._ 265 _Id. at 837–38._ 266 _Id. at 836._ 267 _Id._ ----- law.[268] For example, the NVRA imposes duties on state officials: each state must designate a particular state election official to be responsible for carrying out state obligations under the Act.[269] States have claimed that the NVRA violates the anticommandeering doctrine because it forces them to enact new legislation to administer a federal program.[270] The anticommandeering doctrine prohibits the federal government from compelling states to “implement, by legislation or executive action, federal regulatory programs.”[271] However, as it relates to commandeering, courts have distinguished the source of congressional power in upholding federal election legislation.[272] The prohibition on commandeering under Congress’s Commerce Clause authority does not extend to Congress’s authority under the Elections Clause.[273] In contrast to the Commerce Clause, the Elections Clause allows Congress to “conscript state agencies” to administer a federal election scheme.[274] Therefore, under the Elections Clause, Congress may “enact election legislation that forces a state to take action it might not otherwise take, without violating the anticommandeering doctrine.”[275] Despite this mandate, Congress has been reluctant to use the full extent of its Elections Clause authority because of “federalism” concerns.[276] Congress passed HAVA in response to the challenges encountered in the 2000 presidential election.[277] That election was plagued by unreliable voting systems that varied by jurisdiction, culminating in the “hanging chad” debacle in Florida.[278] HAVA provided federal funds for states to update their voting machines while placing several requirements on states.[279] HAVA’s mandatory provisions include allowing voters to review and verify votes before casting a 268 Tolson, supra note 216, at 2220 (noting that Congress’s primacy in regulating elections is embodied by “its independent authority to make legislation, alter state law, and commandeer state officials to implement federal law”). 269 National Voter Registration Act of 1993, Pub. L. No. 103-31, § 10, 107 Stat. 77, 87 (codified as amended at 52 U.S.C. §§ 20501–20511) 270 Voting Rts. Coal. v. Wilson 60 F.3d 1411, 1415–16 (9th Cir. 1995); see ACORN v. Edgar, 56 F.3d 791, 793 (7th Cir. 1995) (describing an argument by the state of Illinois that the NVRA would require it to change its state laws that govern voter registration). 271 Printz v. United States, 521 U.S. 898, 925 (1997). 272 Weinstein-Tull, supra note 3, at 782. 273 _Id._ 274 _Voting Rts. Coal., 60 F.3d at 1415._ 275 Weinstein-Tull, supra note 3, at 782. 276 _See infra Part III.A._ 277 Weinstein-Tull, supra note 3, at 757. 278 _Id._ 279 Help America Vote Act of 2002, Pub. L. No. 107-252, §§ 102, 301, 303, 116 Stat. 1666, 1670–71, 1704–05, 1708 (codified as amended at 52 U.S.C. §§ 20901–21145). ----- ballot, making voting accessible to people with disabilities, and centralizing voter registration databases at the state level.[280] But HAVA did not “fully nationalize election administration.”[281] Even after HAVA, states and municipalities remain relatively autonomous in conducting elections.[282] With HAVA, Congress used a carrot as much as a stick to coax states into making voting more secure and accessible.[283] HAVA required states to update voting machines and provided funds for the upgrades, but left states to determine which systems to use.[284] HAVA requires that elections be auditable, but stops short of requiring paper ballots.[285] In March 2018, the U.S. Election Assistance Commission announced that it would provide $380 million in election security grants to states, but it left states with discretion in how to use the funds.[286] Under the Elections Clause, Congress has much more authority than it exercised with HAVA. Congress can create a national plan for elections and force states to comply with and administer the plan.[287] Thus, unlike the antidiscrimination framework of the Fourteenth and Fifteenth Amendments, Congress is not constrained by federalism when it exerts its authority under the Elections Clause.[288] Courts can and should disregard claims of state sovereignty in resolving the constitutionality of legislation passed pursuant to the Elections Clause.[289] But Congress has exercised its Elections Clause power far less often than it has used its authority to enforce the Fourteenth and Fifteenth Amendments.[290] Because the Supreme Court’s decision in _Shelby County diminished Congress’s power to regulate elections_ under the Reconstruction Amendments, Congress must rely on its Elections Clause authority to enact legislation that protects U.S. election infrastructure.[291] 280 §§ 301, 303, 116 Stat. at 1704–05, 1708. 281 Weinstein-Tull, supra note 3, at 759. 282 _Id._ 283 _Cf. JAMES T._ BENNET, MANDATE MADNESS: HOW CONGRESS FORCES STATES AND LOCALITIES TO DO ITS BIDDING 211, 214–15 (2014) (describing and criticizing the “carrot and stick” approach of HAVA, which provided federal funds to help induce states to comply with the statute’s requirement that they update and modernize voting equipment). 284 §§ 102–305, 116 Stat. at 1670–71, 1714. 285 §§ 301, 116 Stat. at 1704–06. 286 _U.S. Election Assistance Commission to Administer $380 Million in 2018 HAVA Election Security_ _Funds, U.S._ ELECTION ASSISTANCE COMM’N NEWS (Mar. 29, 2018), https://www.eac.gov/news/2018/03/29/uselection-assistance-commission-to-administer-380-million-in-2018-hava-election-security-funds. 287 _See infra Part III.A._ 288 Tolson, supra note 252, at 173. 289 Tolson, supra note 216, at 2216. 290 Tolson, supra note 252, at 173. 291 Tolson, supra note 216, at 2215. ----- While Congress has not previously exercised the full extent of its power under the Elections Clause, it could do so to create a uniform federal election system. III. CONGRESS SHOULD ACT TO PROTECT U.S. ELECTION INFRASTRUCTURE Due to the threat of foreign interference in U.S. elections, Congress has both the authority and an obligation to act. The notion that Congress cannot create a federal plan for elections because such action would infringe on states’ rights misinterprets the Constitution. The Elections Clause gives Congress a definitive right to regulate federal elections.[292] The combination of multiple sources of constitutional authority—the Elections Clause and the Reconstruction Amendments—provides Congress with even greater power to act.[293] Congress is also duty-bound to protect the integrity of our democracy and to ensure the rights of all citizens to have their votes properly counted.[294] It has a responsibility to take action to protect U.S. election infrastructure in the face of cybersecurity threats because state and local election officials are incapable of doing so.[295] Therefore, to combat foreign interference, Congress must enact legislation to improve the security of election systems throughout the country. Congress should pass a federal plan for three main reasons. First, the structure and purpose the Elections Clause bestows upon Congress a duty to maintain the legitimacy of the federal government.[296] In other words, Congress must ensure that the result of federal elections reflects the will of voters. Second, states are illequipped and reticent to take the cybersecurity measures necessary to protect election infrastructure.[297] Third, the enforcement clauses of the Fourteenth and Fifteenth Amendments obligate Congress to protect the right of all citizens to vote.[298] _A. Congress Has an Obligation Under the Elections Clause to Protect U.S._ _Democracy_ The integrity of elections is critical to maintaining democracy in the United States. Almost 150 years ago, the Supreme Court analogized the power to 292 _See supra Part II.C._ 293 Tolson, supra note 164; see _infra Part III.C._ 294 _See_ United States v. Slone, 411 F.3d 643, 649 (2005) (“Under the Elections Clause, Congress is authorized to protect the integrity of federal elections.”). 295 _See infra Part III.B._ 296 _See U.S. CONST. art. I, § 4, cl. 1; Tolson, supra note 216, at 2218._ 297 _See infra Part III.B._ 298 U.S. CONST. amend. XIV, § 5; U.S. CONST. amend. XV, § 2; see infra Part III.C. ----- regulate federal elections to the right to defend the nation itself.[299] In Ex parte _Yarbrough, the Court stated “[t]hat a government whose essential character is_ republican . . . has no power by appropriate laws to secure this election from the influence of violence, of corruption, and of fraud, is a proposition so startling as to arrest consideration and demand the gravest consideration.”[300] Foreign interference in U.S. elections is not a necessary, but a sufficient, condition for Congress to exercise its authority under the Elections Clause. Congress has a constitutional responsibility to ensure the integrity of the U.S. election process and to protect the fundamental right of citizens to vote. The overarching purpose of the Elections Clause “is to ensure the continued existence and legitimacy of federal elections.”[301] Hamilton described the critical point of the Elections Clause: “every government ought to contain in itself the means of its own preservation.”[302] According to Hamilton, Congress must use its authority to assume from states the responsibility of regulating the manner of federal elections “whenever extraordinary circumstances might render that imposition necessary to its safety.”[303] Foreign interference in U.S. elections is one such extraordinary circumstance.[304] Therefore, for the safety of the nation and the preservation of confidence in federal elections, Congress has an obligation to invoke the Elections Clause to create a federal plan for election administration.[305] While Congress has occasionally exercised its broad powers to regulate elections under the Elections Clause, it has been reluctant to take full action against the threat of foreign interference. In response to Russia’s cyberattacks in 2016 and 2018, the Democratic-led House of Representatives attempted to take small steps to improve the security of federal elections. In 2018, Congress authorized $380 million under HAVA for states to bolster their election security.[306] While several states used the HAVA funds to strengthen cybersecurity and purchase new voting equipment, the amount of money is far 299 _Ex parte Yarbrough, 110 U.S. 651, 657–58 (1884)._ 300 _Id. at 657._ 301 Tolson, supra note 216, at 2218. 302 THE FEDERALIST NO. 59 (Alexander Hamilton). 303 _Id._ 304 Lynch, supra note 187, at 2008–11. 305 Tolson, supra note 216, at 2218. 306 Dustin Volz, U.S. Spending Bill to Provide $380 Million for Election Cyber Security, REUTERS (Mar. 21, 2018, 1:30 PM), https://www.reuters.com/article/us-usa-fiscal-congress-cyber/u-s-spending-bill-to-provide380-million-for-election-cyber-security-idUSKBN1GX2LC; Norden & Cordova, supra note 110. ----- from sufficient.[307] Congress has otherwise been reluctant to pass legislation that would be effective enough to prevent further cyberattacks.[308] Although the House passed three election security bills in 2019, predominantly along party-line votes, the bills have made no progress in the Senate.[309] Congressional Republicans have downplayed the extent of foreign interference in the 2016 and 2018 elections.[310] Objecting to the 2019 Securing America’s Federal Elections (SAFE) Act, Representative Rodney Davis (R-Ill.) stated that Congress should not force states to update voting technology because “there is no evidence of voting machines being hacked in 2016, 2018[,] or ever[.]”[311] Senate Majority Leader Mitch McConnell (R-Ky.), who has refused to bring any of the House bills up for a vote in the Senate, has also minimized the risk.[312] Senator McConnell even chided the media for fostering panic among voters and for not giving more credit to the current administration for preventing major security breaches in the 2018 election.[313] However, in objecting to the SAFE act, Congressional Republicans have primarily argued that the bill’s provisions interfere with the authority of states and localities to conduct elections.[314] Senator McConnell stated that while he believes Russian meddling to be real, he doesn’t believe that the federal government should tell states how to run elections.[315] The Republican sentiment, as expressed by Senator McConnell, misinterprets the authority granted to Congress under the Constitution. Because the Elections Clause gives Congress final policymaking authority over the times, places, and manners of federal elections, it “allows Congress to legislate independent of and without deference to state sovereignty.”[316] Therefore, the 307 Norden & Cordova, supra note 110. 308 _Id._ 309 For the People Act of 2019, H.R. 1, 116th Cong.; Stopping Harmful Interference in Elections for a Lasting Democracy (SHIELD) Act, H.R. 4617, 116th Cong.; Securing America’s Federal Elections (SAFE) Act, H.R. 2722, 116th Cong. 310 Maggie Miller & Julie G. Brufke, House Passes Sweeping Democratic-Backed Election Security Bill, HILL (Jun. 27, 2019, 5:00 PM), http://thehill.com/homenews/house/450737-house-passes-sweeping-democratbacked-election-security-bill; Hailey Fuchs & Karoun Demirjian, _Divided House Passes Election Security_ _Legislation over Republican Objection, WASH._ POST (Jun. 27, 2019, 4:45 PM), https://www.washingtonpost. com/powerpost/divided-house-passes-election-security-legislation-over-republican-objections/2019/06/27/a07 1c10c-98f1-11e9-8d0a-5edd7e2025b1_story.html. 311 Miller & Brufke, supra note 310. 312 Fuchs & Demirjian, supra note 310. 313 _Id._ 314 _Id._ 315 DeChiaro, supra note 18. 316 Tolson, supra note 164, at 324. ----- notion that Congress must cajole states to undertake security fixes to their election systems and abide by federal security standards is grossly misguided.[317] Congress has an obligation under the Elections Clause to preserve the legitimacy of the federal government by ensuring that federal elections reflect the will of the people.[318] A strong and uniform federal plan is needed to protect against efforts by foreign actors to disrupt U.S. elections. _B. Congress Has a Duty to Secure U.S. Elections Against Foreign_ _Interference Because States Are Ill-Equipped and Reluctant to Do So_ The United States is unique in that it currently has no nationwide election authority.[319] Conducting elections in the United States is a complex process “that involves multiple levels of government, personnel with a variety of skills and capabilities, and numerous electronic systems that interact in the performance of a multitude of tasks.”[320] State or local officials manage elections in accordance with state laws and local regulations.[321] Elections are administered by over 9,000 state and local jurisdictions containing over 114,000 polling places.[322] The thousands of jurisdictions vary widely in size, in funding available for election administration, and in the ability to detect and manage irregularities, particularly cyberattacks.[323] Several of the small elections offices “have few dedicated staff and little access to the latest information technology (IT) training or tools.”[324] A lack of cyber sophistication was evident in the 2016 election as states and municipalities were unequipped to deal with the severity of the threat. One state official said, “I don’t think any of us expected to be hacked by a foreign government.”[325] Another official stated, “If a nation-state is on the other side, it’s not a fair fight. You have to phone a friend.”[326] In most states, the decentralized structure means that counties and municipalities have varying 317 _See SENATE INTELLIGENCE REPORT, supra note 13, at 54 (stating in its recommendations that “[s]tates_ should remain firmly in the lead on running elections, and the federal government should ensure they receive the necessary resources and information”). 318 _See THE FEDERALIST NO. 59 (Alexander Hamilton) (“Every government ought to contain in itself the_ means of its own preservation.”). 319 NAS REPORT, supra note 11, at 31. 320 _Id. at 4._ 321 NAS REPORT, supra note 11, at 17. 322 Manpearl, supra note 71, at 169. 323 NAS REPORT, _supra note 11, at 17._ _See generally David C. Kimball & Brady Baybeck, Are All_ _Jurisdictions Equal? Size Disparity in Election Administration, 12 ELECTION L.J. 130 (2013) (discussing how_ size disparities lead to diverging experiences for election officials and voters in large versus small jurisdictions). 324 NAS REPORT, supra note 11, at 17. 325 SENATE INTELLIGENCE REPORT, supra note 13, at 39. 326 _Id._ ----- levels of resources to conduct elections.[327] County election officials, who are on the front lines of defending election equipment, often have very limited IT support.[328] A Wisconsin state election administrator noted that some counties’ election teams may only consist of “a county clerk and one more person working on elections.”[329] Many county officials have not received any cybersecurity training, even after the 2016 cyberattacks were made known. In Pennsylvania, election officials in three of the four largest counties had not received cybersecurity training as of August 2017.[330] In Michigan, officials in fewer than one-third of counties indicated that they received formal cybersecurity training.[331] And in Arizona, officials in only five of fifteen counties received such training.[332] States also vary widely in the level of security they maintain around voter registration databases. DHS analysis of state election systems found significant variance in the security of state voter registration databases, including lack of encryption and lack of backups in many states.[333] As of May 2017, forty-one states were still using voter registration systems that were created more than a decade prior.[334] Types of vote casting systems also vary dramatically from state to state. Forty-five states continue to use outdated voting machines that are no longer manufactured.[335] Some machines are at least fifteen years old and run on outdated software that is no longer supported, such as Windows XP.[336] In the November 2018 election, fourteen states did not use a voting mechanism that allowed for a voter-verified paper audit trail.[337] Many states understand the need for more secure voting equipment but lack sufficient financial resources. Although the 2018 HAVA funds were dispersed quickly, states did not have enough time to make major improvements to their 327 _See Norden & Cordova, supra note 110._ 328 _Id._ 329 _Id._ 330 Likhitha Butchireddygari, Many County Officials Still Lack Cybersecurity Training, NBC NEWS (Aug. 23, 2017, 5:20 AM), https://www.nbcnews.com/politics/national-security/voting-prep-n790256. 331 _Id._ 332 _Id._ 333 SENATE INTELLIGENCE REPORT, supra note 13, at 46. 334 Tim Lau, U.S. Elections Are Still Vulnerable to Foreign Hacking, BRENNAN CTR. FOR JUST. (Jul. 18, 2019), https://www.brennancenter.org/our-work/analysis-opinion/us-elections-are-still-vulnerable-foreignhacking. 335 Norden & Cordova, supra note 110. 336 _Id._ 337 Lin et al., supra note 51, at 22. ----- election systems before the 2018 midterm elections.[338] The funding has also been insufficient for states to overhaul their elections systems and replace outdated voting machines.[339] Most states recognized a need to purchase new equipment before the 2020 election, but two thirds of the state officials claimed that they lack the money to do so, even with the additional HAVA funds.[340] Consequently, states and municipalities cannot be relied on to successfully combat foreign cyberattacks against U.S. election systems. According to Senator Ron Wyden (D-Or.), If there was ever a moment when Congress needed to exercise its clear constitutional authorities, this is it. America is facing a direct assault on the heart of our democracy by a determined adversary. We would not ask a local sheriff to go to war against the missiles, planes and tanks of the Russian army. We shouldn’t ask a county IT employee to fight a war against the full capabilities and vast resources of Russia’s cyber army. That approach failed in 2016 and it will fail again.[341] Simply providing funding to states is also not enough. Congress must create a comprehensive plan to secure federal elections against foreign attacks. _C. Congress Must Enact a Federal Plan to Preserve the Right of All Citizens_ _to Vote_ Professor Franita Tolson has effectively described how Congress’s license to enact comprehensive federal election legislation may be even greater than its Elections Clause power alone because it derives from multiple sources of authority.[342] In addition to its obligation to preserve the integrity of federal elections under the Elections Clause, Congress has a responsibility to exercise its authority under the enforcement clauses of Fourteenth and Fifteenth Amendments to protect the right of all citizens to vote.[343] Multiple sources of authority confer even broader power when Congress acts to protect constitutional rights and may provide the impetus for the Supreme Court to find a federal statute valid where it would have considered it unconstitutional under a single source of authority.[344] Therefore, notwithstanding the Supreme Court’s 338 The EAC dispersed 96% of the HAVA funds by August 2018. Lynch, supra note 187, at 1999. 339 Norden & Cordova, supra note 110. 340 _Id._ 341 SENATE INTELLIGENCE REPORT, supra note 13, Minority Views of Senator Wyden, at 1. 342 Tolson, supra note 164, at 329. 343 _Id. at 324._ 344 _Id. at 329. The Supreme Court has been inconsistent in its recognition of a greater scope of authority_ when Congress acts pursuant to multiple sources of authority. Compare Tennessee v. Lane, 541 U.S. 509, 516 ----- holding in Shelby County, the Reconstruction Amendments provide additional power to Congress’s Elections Clause authority to establish a federal system for election infrastructure.[345] With this power comes a duty for Congress to act. Cyberattacks that disrupt the voting process and create risks that vote tallies will be manipulated infringe on the right of citizens to vote. The fundamental right to vote includes the right to be certain that one’s vote matters.[346] Courts have found that plaintiffs have standing to bring Fourteenth Amendment Due Process and Equal Protection claims where they allege that certain voting methods prohibit their votes from being properly counted.[347] In _Stewart v._ _Blackwell, the Sixth Circuit found that the increased probability that plaintiffs’_ votes would not be properly counted due to a faulty punch-card system was “neither speculative nor remote” and was therefore a justiciable claim.[348] Similarly, a Pennsylvania court found that voters had proper standing to bring a Fourteenth Amendment claim because the machines they used to vote did not allow them to know whether their votes had been cast or would be counted.[349] A recent lawsuit brought by voters in Georgia demonstrates how voting systems that are not secure against cyberattacks infringe on voters’ rights.[350] A federal court granted an injunction against using insecure DRE machines based on the merits of the plaintiffs’ Fourteenth Amendment Due Process and Equal Protection claims.[351] The plaintiffs in Curling claimed that the state had violated their Due Process rights by placing a “substantial burden” on their fundamental right to vote and had violated their Equal Protection rights by placing “more severe burdens” on their right to vote than voters who did not have to use DRE machines.[352] The court agreed and granted plaintiff’s relief in part because the (2004) (upholding Title II of the Americans with Disabilities Act (ADA) based on “the power to enforce the [F]ourteenth [A]mendment and to regulate commerce”), with Bd. of Trs. v. Garrett, 531 U.S. 356, 374 (2001) (ignoring the Congress’s Commerce Clause authority when invalidating the ADA in part as an improper exercise of the Fourteenth Amendment enforcement clause), _and Shelby Cnty. v. Holder, 570 U.S. 529, 553 (2013)_ (giving no weight to Congress’s additional authority for enacting the VRA under both the Fourteenth and the Fifteenth Amendments). 345 _See Tolson, supra note 164, at 330 (“[F]ar-reaching and potentially controversial legislation can gain_ substantial legitimacy from the fact that Congress can draw on multiple sources of power.”). 346 _See Curling v. Kemp, 334 F. Supp. 3d 1303, 1328 (N.D. Ga. 2018)._ 347 _E.g., id._ 348 Stewart v. Blackwell, 444 F.3d 843, 855 (6th Cir. 2006), superseded by Stewart v. Blackwell, 473 F.3d 692 (6th Cir. 2007). 349 Banfield v. Cortes, 922 A.2d 36, 44 (Pa. Commw. Ct. 2007). 350 _Curling, 334 F. Supp. 3d 1303; Curling v. Raffensperger, 397 F. Supp. 3d 1334 (N.D. Ga. 2019)._ 351 _Curling, 397 F. Supp. 3d at 1410._ 352 _Curling, 334 F. Supp. 3d at 1312._ ----- state’s ongoing use of an insecure voting method “pierce[d] citizens’ confidence in the electoral system and the value of voting.”[353] Therefore, in some instances, voting rights advocates can protect the right to vote against insecure voting systems through litigation.[354] Federal courts may be willing to recognize that an infringement on voters’ right to feel secure that their votes will count is an injury for which relief may be granted.[355] Insecure voting systems can also affect voters’ ability to merely cast a ballot. Long wait times to vote—resulting from erroneous registration data or voting equipment dysfunction—may impact minority voting districts to a greater degree than predominantly white precincts.[356] And as wait times increase, voter participation drops.[357] Consequently, the Equal Protection Clause and the Fifteenth Amendment may be implicated when citizens of color are disproportionately denied the right to vote when cyberattacks disrupt voting on election day. However, litigation is cumbersome and cannot always protect the rights of all voters or ensure the integrity of federal elections. Indeed, one impetus for the VRA in 1965 was that piecemeal litigation had failed to sustainably protect the African Americans’ right to vote in most jurisdictions in the Deep South.[358] With each hard fought victory in courts, state and local governments found ways to enact new restrictions.[359] Moreover, litigation only grants relief after harm has occurred. Courts can grant prospective relief to require security measures for future election cycles.[360] But there is no sufficient remedy for the harm to voters that has already occurred after they participated in an insecure election.[361] Thus, 353 _Curling, 397 F. Supp. 3d at 1411 (quoting Curling, 334 F. Supp. 3d at 1328)._ 354 _Id. at 1410._ 355 _Id.; see_ _Curling, 334 F. Supp. 3d at 1328 (“A wound or reasonably threatened wound to the integrity_ of a state’s election system carries grave consequences beyond the results in any specific election, as it pierces citizens’ confidence in the electoral system and the value of voting.”). Contra Heindel v. Andino, 359 F. Supp. 3d 341, 357 (D.S.C. 2019) (holding that plaintiffs failed to show a clearly impending injury that was traceable to state election officials because they “merely speculate and make assumptions about whether their votes will be inaccurately counted as the result of a potential hack” (quoting Clapper v. Amnesty Int’l, 568 U.S. 398, 411 (2013))). 356 Stephanie Mencimer, Even Without Voter ID Laws, Minority Voters Face More Hurdles to Casting _Ballots, MOTHER_ JONES (Nov. 3, 2014), https://www.motherjones.com/politics/2014/11/minority-voterselection-long-lines-id/; German Lopez, Minority Voters Are Six Times More Likely as White Voters to Wait More _Than an Hour to Vote, VOX (Nov. 8, 2016, 1:30 PM), https://www.vox.com/identities/2016/11/8/13564406/_ voting-lines-race-2016. 357 _Lopez, supra note 356._ 358 South Carolina v. Katzenbach, 303 U.S. 301, 314 (1966). 359 _Id.; ANDERSON, supra note 165, at 13; see supra Part II.A._ 360 _See_ _Curling, 397 F. Supp. 3d at 1412._ 361 _See_ _Curling, 334 F. Supp. 3d at 1315._ ----- the federal government must respond comprehensively to protect voters’ rights against cyberattacks from foreign actors. In sum, Congress must act to protect U.S. election infrastructure and to combat foreign interference in federal elections. Congress has the primary obligation to safeguard the legitimacy of the federal government, to protect the fundamental right of citizens to vote, and to ensure that the election results reflect the choice of the majority of voters. And Congress has the authority to act pursuant to the Elections Clause coupled with the enforcement provisions of the Reconstruction Amendments, which provide additional power to protect the right of all citizens to vote. IV. A PROPOSED FEDERAL PLAN TO SECURE U.S. ELECTIONS Congress has the power under the Elections Clause to enact legislation that establishes a federal plan to which state election authorities must adhere.[362] The Elections Clause authorizes Congress to designate the manner in which federal elections are conducted in order to protect the integrity of the federal government against a threat of foreign interference.[363] After Russian cyberattacks against state and local election systems in 2016 and 2018, and the anemic, ineffective response by state election officials, the need for a uniform federal election plan is evident.[364] Therefore, Congress has the obligation to enact a national plan that creates uniform standards across all election jurisdictions to ensure that federal elections are secure and that all citizens are able to exercise their right to vote and know their votes will count. A national plan for federal elections does not imply that the entirety of election administration should be conducted by the federal government. The decentralized approach to U.S. elections, which relies on states and localities to manage the nuts and bolts of elections, provides efficiency.[365] The cybersecurity benefit of a decentralized structure remains—it protects against the devastating impact of a single widespread cyberattack or technological breakdown.[366] But an ongoing role for states to conduct elections does not preclude implementing uniform rules and standards for federal elections. Measures to secure U.S. 362 _See supra Part III.A._ 363 _See id._ 364 _See supra Part I.C._ 365 _See THE FEDERALIST NO. 59 (Alexander Hamilton) (stating that regulation of federal elections is left_ to local administrations because “it may be more convenient and more satisfactory”). 366 Manpearl, supra note 71, at 182; NAS REPORT, supra note 11, at 119. ----- election infrastructure would be most effective if they are implemented at a national level.[367] Although Congress’s national plan for federal elections should be mandatory for states to follow, the Elections Clause does not grant Congress authority over state and local elections.[368] However, Congress can encourage states to follow a federal election plan for their own internal elections. First, because of logistics, efficiency, and cost, states would likely use federal election infrastructure to conduct state and local elections along with federal elections. Second, states’ inability to take appropriate cybersecurity measures for their own elections provides the impetus for Congress to act under the Fourteenth and Fifteenth Amendments to protect the right of all citizens to know that their votes with count.[369] Unlike the Elections Clause, the Fourteenth and Fifteenth Amendments apply to all elections: federal, state, and local.[370] Third, Congress could use its Spending Clause power to condition funding for election infrastructure on a state’s compliance with a federal plan for all elections conducted within the state.[371] A national election plan should have three main components. First, it should create uniform federal standards for securing voter registration databases and for transmitting voter information to polling places so that voters can be checked-in on election day. Second, Congress should require that all states implement a secure method of voting that uses a uniform ballot design. All voters should be allowed to mark and record their selections in the manner that is least susceptible to cyberattacks: hand-marked paper ballots read by secure, state-of-the-art optical scanners. Finally, to ensure the integrity of every federal election, states must be required to submit to federal post-election audits. 367 _See Mark Lanterman, Fair Elections and Cybersecurity, 75 BENCH &_ BAR MINN. 10, 10 (2018) (“[T]he sorts of measures that would most likely effect positive security outcomes are best implemented at a national level, where standardized procedures can provide a framework for ongoing improvement.”). 368 U.S. CONST. art. I, § 4, cl. 1. 369 _See supra Part III.C._ 370 U.S. CONST. amend. XV, § 1 (“The right of citizens of the United States to vote shall not be denied or abridged by the United States or by any State . . . .”) (emphasis added). 371 _See Art. I, § 8, cl. 1 (empowering Congress to “lay and collect Taxes, Duties, Imposts, and Excises, to_ pay the Debts and provide for the common Defence and general Welfare of the United States”); South Dakota v. Dole, 483 U.S. 203, 207 (1987) (“[O]bjectives not thought to be within Article I’s enumerated legislative fields may nevertheless be attained through the spending power and the conditional grant of federal funds.”) (internal quotation marks omitted). ----- _A. Congress Should Establish Binding Federal Standards for States to_ _Register Voters, Maintain Secure Voter Databases, and Check-in Voters at_ _the Polls_ Voter registration databases that are maintained electronically are particularly vulnerable to manipulation by malicious cyber actors.[372] Election administrators currently rely on county or state government IT departments to secure voter registration databases.[373] A DHS analysis found that the security of voter databases varied significantly by state, and many states lacked encryption and backups for their databases.[374] Federal intelligence and cybersecurity officials have made recommendations to states and have offered to provide cybersecurity measures to protect voter registration databases.[375] But many states have demonstrated a reluctance to receive help from the federal government or to follow recommendations.[376] Consequently, Congress must pass legislation that directs states to implement specific cybersecurity measures for voter registration databases, which include updating relevant software, creating paper back-ups, and instituting two-factor authentication for user access to the databases.[377] This action would not be novel—Congress has previously set mandatory requirements for state voter databases.[378] A federal plan should also require states to put in place standard security procedures for monitoring voter database integrity.[379] Such measures should include installing monitoring sensors on state registration systems to detect attempts to hack into the systems and reporting any identified compromises immediately to DHS.[380] A national plan must also create standards for transmitting voter data to polling places for voter verification and check-in. Because they are electronic, e-pollbooks are vulnerable to cyberattacks, particularly if they are locally 372 SENATE INTELLIGENCE REPORT, supra note 13, at 57. 373 NAS REPORT, supra note 11, at 58. 374 SENATE INTELLIGENCE REPORT, supra note 13, at 46. 375 _Id. at 52._ 376 _See supra Part III.B; SENATE INTELLIGENCE REPORT, supra note 13, at 48–49; see also_ _id., Minority_ Views of Senator Wyden, at 2 (“The Committee report describes a range of cybersecurity measures needed to protect voter registration databases, yet there are currently no mandatory rules that require that require states to implement even minimum security measures.”). 377 SENATE INTELLIGENCE REPORT, supra note 13, at 57. 378 Help America Vote Act of 2002, Pub. L. No. 107-252, 116 Stat. 1666 (codified as amended at 42 U.S.C. §§ 15301-15545) (requiring “a single, uniform, official, centralized, interactive, computerized statewide voter registration list defined, maintained, and administered at the state level”). 379 NAS REPORT, supra note 11, at 63. 380 SENATE INTELLIGENCE REPORT, supra note 13, at 57. ----- networked or connected to the internet.[381] Cyberattacks could change voter data, alter information on who has voted, or simply shut down operation of an epollbook through a “denial of service” attack.[382] Congress should, therefore, include national security standards for the use of e-pollbooks in its federal plan. Because e-pollbooks have advantages over paper and are easy to use, their use should not be discontinued.[383] Rather, the NAS recommends that Congress authorize and fund the National Institute of Standards and Technology to develop security standards along with verification and validation protocols for e-pollbooks.[384] In addition, each precinct should be required to maintain a paper copy of the precinct’s pollbook as a back-up in the event that voter data is manipulated or access to electronic data is disrupted.[385] _B. Congress Should Mandate Uniform Paper Ballots for All Federal Elections_ Voters across the country cast their ballots using methods that are subject to varying degrees of cyber risks, and many states are either unwilling or incapable of following the recommendations of cybersecurity experts.[386] Voting systems that do not provide human-readable printouts for voters to confirm their selections and do not maintain a voter-verified paper audit trail are most vulnerable to cyberattacks.[387] Experts have called for discontinuing the use of paperless DRE machines because they are vulnerable to hacking without detection and do not produce auditable paper trails.[388] Yet, in 2019, twelve states were still using paperless DRE machines in at least some jurisdictions, and four states still used them statewide.[389] Congress should pass legislation that prohibits states from using outdated, paperless voting machines and requires the use of a uniform method of voting that will provide an auditable paper trail. The Senate Intelligence Report concluded that “[p]aper ballots and optical scanners are the least vulnerable to cyberattack.”[390] The most secure and costeffective method for voting would be to use hand-marked paper ballots in all 381 See supra Part I.B. regarding the vulnerability of e-pollbooks. 382 NAS REPORT, supra note 11, at 71, 86. 383 _Id. at 72._ 384 _Id._ 385 _Id._ 386 _See supra Part III.B._ 387 SENATE INTELLIGENCE REPORT, supra note 13, at 42. 388 See supra Part I.B. for a detailed description of the security flaws associated with DRE voting machines. 389 Norden & Cordova, supra note 110. 390 SENATE INTELLIGENCE REPORT, supra note 13, at 59. ----- federal elections.[391] Using a uniform paper ballot for federal elections that voters mark by hand would also allow states to continue and expand the use of voteby-mail.[392] Alternatively, Congress could require and provide funding for uniform BMD machines to be used across all jurisdictions. The BMDs must produce a paper record of the voter’s choices, which each voter can review before casting their ballot. However, because BMD machines are potentially vulnerable to cyberattacks, the most secure election systems use hand-marked paper ballots as the primary method for voting.[393] Moving forward, Congress should mandate that all federal elections be conducted using human-readable paper ballots that are counted either by hand or by using federally certified optical scanners.[394] _C. Congress Should Require All States to Submit to Federal Election Audits_ As part of a federal election plan, Congress should require that all states submit to post-election audits. Audits require voter-verifiable paper ballots that provide a human-readable record of the voter’s selections.[395] Such audits provide assurance that the outcome of any election reflects the voters’ choices and is based on an accurate tabulation of the ballots cast.[396] NAS election cybersecurity experts recommend risk-limiting audits as the most efficient and effective means to ensure the reliability of an election.[397] Risk-limiting audits examine randomly selected, individual ballots until a predetermined level of statistical assurance is reached.[398] In 2017, risk-limiting audits were piloted statewide in Colorado, and several other states plan to conduct pilots in the next few years.[399] However, rather than leaving the requirement for audits to the discretion of states, Congress should pass 391 _Id.; Christopher Deluzio & Kevin Skoglund, Guess Which Ballot Costs Less and Is More Secure—_ _Paper or Electronic?,_ PATRIOT NEWS (Aug. 20, 2019), https://www.pennlive.com/opinion/2019/08/guesswhich-ballot-costs-less-and-is-more-secure-paper-or-electronic-opinion.html. 392 In 2016, Colorado, Oregon, and Washington used mail-only voting, and most ballots in California and Utah were cast by mail. NAS REPORT, supra note 11, at 48–50. 393 _See generally Andrew Appel, Richard A. DeMillo & Philip B. Stark, Ballot-Marking Devices (BMDs)_ _Cannot Assure the Will of the Voters,_ 19 ELECTION L.J. 432 (2020) (describing the vulnerability of BMD voting machines to hacking as well as risk that BMDs may not accurately record a vote as the voter had intended and arguing that the most secure method of voting is a system that uses hand-marked paper ballots). 394 NAS REPORT, supra note 11, at 80. 395 _Id. at 94._ 396 _Id._ 397 _Id. at 95._ 398 _Id._ 399 _Id._ ----- legislation to require all states to submit to federal risk-limiting audits after each federal election. The federal government’s response to ongoing Russian cyberattacks must extend beyond offers to provide resources to states.[400] To protect and defend U.S. elections, Congress must “establish mandatory nation-wide cybersecurity requirements.”[401] Such requirements must designate specific measures to ensure the security of voter registration databases and pollbooks and should compel the use of uniform paper ballots and post-election audits. CONCLUSION The right of citizens to freely choose who will represent them is the essence of our republican form of government. The founders understood that maintaining free and fair elections is a core tenet of this nation. Therefore, they placed in the Constitution the means for Congress to have final authority to regulate federal elections when the need arises. Russian cyberattacks on state and local election systems constitute a challenge to the core values of American democracy, which require a comprehensive, uniform federal response. To varying degrees over the past 150 years, Congress has imposed regulations on states to protect election integrity by ensuring that all citizens have the right to vote. The current threat requires an even greater response. This Comment describes a source of authority that authorizes Congress to prescribe cybersecurity measures to which states must adhere in conducting federal elections. The value implicit in the Elections Clause is that federal elections must be administered in a manner that produces a clear and legitimate outcome. Congress has the authority and an obligation under the Elections Clause to ensure the integrity of American democracy in the face of cyberattacks by a foreign adversary. Congress must exercise this power to create a comprehensive national plan for federal elections. SUMAN MALEMPATI[*] 400 SENATE INTELLIGENCE REPORT, supra note 13, Minority Views of Senator Wyden, at 1. 401 _Id._ - J.D. Candidate, Emory University School of Law, Class of 2021. I extend my deepest gratitude to Professor Robert Schapiro for his wisdom, guidance, and support throughout the writing process. Thank you to Natalie Baber and Connor Hees for providing insightful feedback. To the Emory Law Journal staff, particularly Brennan Mancil and Sam Reilly, thank you for the incredible work you have done make this Comment better and get it published. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.2139/ssrn.3590843?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.2139/ssrn.3590843, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "CLOSED", "url": "" }
2,020
[]
false
null
[]
32,311
en
[ { "category": "Economics", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/010d55f6ebe83e48bb83926c88b0d72be0c08538
[]
0.846713
Investment Portfolio Optimization in Indonesia (Study On: Lq-45 Stock Index, Government Bond, United States Dollar, Gold and Bitcoin)
010d55f6ebe83e48bb83926c88b0d72be0c08538
International Journal of Current Science Research and Review
[ { "authorId": "2241335791", "name": "I. Made" }, { "authorId": "2226312743", "name": "Gede Abandi Semeru" }, { "authorId": "119396810", "name": "Y. Nainggolan" } ]
{ "alternate_issns": null, "alternate_names": [ "Int J Curr Sci Res Rev" ], "alternate_urls": null, "id": "fd0a02a9-7a23-4ec4-ae7b-0a3d06990d81", "issn": "2581-8341", "name": "International Journal of Current Science Research and Review", "type": "journal", "url": null }
In forming their portfolios, investors should analyze the risk and return of each investment instrument. This is aimed at preventing investors from speculating and gambling with their investments. Conducting an investment portfolio optimization study on LQ-45 stock index, government bond, USD, gold, and Bitcoin can provide valuable insights due to unique market characteristics in Indonesia. This research analyzes the formation of investment instruments over the last 60 months, specifically from January 2018 to December 2022. The research method used in this study is quantitative research aimed at selecting several investment instruments for a portfolio in Indonesia. The portfolio aims to minimize risk and maximize return using the Markowitz method, also known as the optimal portfolio. To fulfill the objectives of this research, data on the prices of each instrument are required. An optimal portfolio can be obtained by combining two instruments: 18% bitcoin and 82% gold. This optimal portfolio can achieve an expected return of 1.29% with a risk level of 5.15%. Considering a risk-free rate of 0.375%, this portfolio forms a slope of 0.1775, which is the largest slope formed between the combination of risk-free instruments and risky portfolios. Investors should allocate their funds more wisely, considering not only the highest return but also the associated risk. High returns often come with high risks, so investors need to assess the risk-return trade-off before making investment decisions.
## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 ** ### **www.ijcsrr.org** # **Investment Portfolio Optimization in Indonesia (Study On: Lq-45 Stock ** **Index, Government Bond, United States Dollar, Gold and Bitcoin)** ### **I Made Gede Abandi Semeru [1], Yunieta A. Nainggolan [2]** 1,2 School of Business & Management, Institut Teknologi Bandung **ABSTRACT:** In forming their portfolios, investors should analyze the risk and return of each investment instrument. This is aimed at preventing investors from speculating and gambling with their investments. Conducting an investment portfolio optimization study on LQ-45 stock index, government bond, USD, gold, and Bitcoin can provide valuable insights due to unique market characteristics in Indonesia. This research analyzes the formation of investment instruments over the last 60 months, specifically from January 2018 to December 2022. The research method used in this study is quantitative research aimed at selecting several investment instruments for a portfolio in Indonesia. The portfolio aims to minimize risk and maximize return using the Markowitz method, also known as the optimal portfolio. To fulfill the objectives of this research, data on the prices of each instrument are required. An optimal portfolio can be obtained by combining two instruments: 18% bitcoin and 82% gold. This optimal portfolio can achieve an expected return of 1.29% with a risk level of 5.15%. Considering a risk-free rate of 0.375%, this portfolio forms a slope of 0.1775, which is the largest slope formed between the combination of risk-free instruments and risky portfolios. Investors should allocate their funds more wisely, considering not only the highest return but also the associated risk. High returns often come with high risks, so investors need to assess the risk-return trade-off before making investment decisions. **KEYWORDS:** Bitcoin, Government Bond, Gold, LQ-45, Portfolio Optimization, USD. **INTRODUCTION** The portfolio formed by an investor can provide high returns or, on the contrary, cause losses for the investor. In other words, risk is a deviation from the expected return. There is a positive relationship between return and risk in investing, known as high riskhigh return, which means the greater the risk that must be borne, the greater the resulting return. Return is the result obtained from an investment, which can be in the form of realized return or expected return that has not yet occurred but is expected to happen in the future. Meanwhile, portfolio risk consists of systematic and unsystematic risk. Both of these risks are often referred to as total risk. Some factors that influence this uncertainty include securities prices and interest rates, which can change at any time. The benefits of diversification are well-known through the principle that says "Don't put all your eggs in one basket", because if that basket falls, then all the eggs in it will break. In the context of investment, this proverb can be interpreted as a recommendation not to invest all the funds owned in only one asset, because if that asset fails, then all the invested funds will disappear. Investors expect to get maximum returns with minimum possible risk. However, the larger the profit obtained from an investment, the higher the associated risk. Therefore, investors need to consider the balance between risk and return in investing. Risk can be minimized by diversification or by combining several investment instruments into a portfolio. If one instrument experiences a loss while another instrument generates a profit, the profit from one instrument can offset the loss from the other investment instrument. Effective diversification of investment instruments yields efficient results in a portfolio, providing maximum expected returns with minimal variance for those expected returns. Such a portfolio is called a Markowitz Efficient Portfolio This study focuses on investment instruments in Indonesia, including the LQ-45 stock index, government bonds, United States dollar, gold, and Bitcoin can provide valuable insights due to unique market characteristics, diversification benefits, local investor perspectives, period-specific analysis, and the opportunity to contribute to existing knowledge. The selected instruments represent different asset classes, each with its own characteristics and potential benefits. By including a mix of equities (LQ-45 stock index), fixed income (bonds), currencies (US dollar), commodities (gold), and cryptocurrencies (bitcoin), this can analyze how diversification across these assets may impact portfolio performance and risk management. The LQ45 stock index is a widely recognized benchmark index for the Indonesian stock market, providing insights into the performance of the country's largest and most liquid stocks. Government bonds, on the other hand, represent fixed-income securities issued by ### 4922 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 ** ### **www.ijcsrr.org** the Indonesian government, offering income and potentially lower risk compared to equities. The US dollar is a commonly used global reserve currency and serves as a benchmark for many international transactions. Gold is a well-known precious metal and often considered a store of value. Bitcoin, as a cryptocurrency, represents a digital and decentralized form of currency with its own unique characteristics. By including a diverse set of instruments, this can conduct a comprehensive analysis that covers a broader range of investment options. This can enhance the understanding of portfolio optimization, risk management, and the potential for achieving better risk-adjusted returns. When forming a portfolio, investors seek to minimize risk and maximize returns. A portfolio that can achieve these goals is called an optimal portfolio. To form an optimal portfolio, several assumptions need to be made about investor behavior in making investment decisions. It is assumed that investors tend to avoid risk (risk averse). This type of investor would choose an investment with lower risk if presented with two investments with the same expected return but different levels of risk. **LITERATURE REVIEW** ***A.*** ***Portfolio*** Investing aims to generate profits with a certain level of risk. The purpose of creating an investment portfolio is to diversify risk so that the funds held have minimum risk. Investing in more than one investment instrument has lower risk compared to investing in only one instrument. The more investment instruments involved in the portfolio, the lower the risk. If there is a decrease in one investment instrument, then other instruments can offset or replace it. Therefore, investors must have diversity in their portfolio so that the funds held do not experience a decrease from their initial value (Markowitz, 1952) Markowitz assumed that investors would be able to create an efficient portfolio. He also stated that the portfolio should be diversified to achieve risk spreading. Such diversification will produce an efficient portfolio where it provides a higher level of return than other portfolios with the same risk and a lower risk than other portfolios with the same level of return. ***B.*** ***Optimal Portfolio*** The optimal portfolio is a portfolio that provides the highest expected return for a given level of risk or the lowest level of risk for a given level of expected return. In other words, it is the portfolio that offers the best risk-reward trade-off for an investor. The concept of the optimal portfolio was introduced by Harry Markowitz in his seminal paper "Portfolio Selection" in 1952. To find the optimal portfolio, an investor needs to consider the expected returns, standard deviations, and correlations of all the assets in the portfolio. The optimal portfolio can be identified by plotting the efficient frontier, which is a curve that represents the set of portfolios that offer the highest expected return for a given level of risk, or the lowest level of risk for a given level of expected return. The point on the efficient frontier that corresponds to the investor's risk tolerance and expected return is the optimal portfolio for that investor. The optimal portfolio is crucial for investors who want to maximize their returns while minimizing their risk. By diversifying their portfolio and selecting assets with low correlations, investors can reduce their portfolio's risk and increase their expected returns. The optimal portfolio is also useful for portfolio managers who want to construct a portfolio that meets the investment objectives of their clients while minimizing risk. ***C.*** ***Asset Allocation*** Asset allocation is more focused on placing funds in various investment instruments rather than emphasizing stock choices in the portfolio. From the study results, differences in performance are more due to asset allocation rather than investment choices. According to Markowitz (1952), asset allocation is one of the factors that determine the level of return and risk of the portfolio. Perrit and Lavine (1990) state that besides diversification, this asset allocation is a very important factor in investment, for practical reasons such as targeting long-term investments, determining the risk that investors can tolerate over time, and eliminating investment decision changes based on changes in financial conditions. ***D.*** ***Conceptual Framework*** This study presents a conceptual framework encompassing the key elements of modern portfolio theory (MPT), including the optimal portfolio, Sharpe ratio, portfolio variance and covariance, risk preference, and efficient frontier. Developed by Harry Markowitz in the 1950s, MPT offers a robust framework for constructing portfolios that strive to optimize the delicate balance between risk and return. ### 4923 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 ** ### **www.ijcsrr.org** **Figure 1.** Conceptual Framework **RESEARCH METHODOLOGY** ***A.*** ***Data Collection Method*** In this research, historical data was obtained by visiting websites that provide the required data. This research analyzes the formation of investment instruments over the last 60 months, specifically from January 2018 to December 2022. In selecting the instruments, several instruments were chosen to represent the entire range of instruments available in Indonesia. ***B.*** ***Data Analysis Method*** Due to the complexity of data processing, assistance from computer software, specifically Microsoft Excel, is required. Apart from being easy to operate, this software also offers the necessary functions and features for performing calculations. The functions in Microsoft Excel are highly useful for data processing, including stdevp (calculating standard deviation), average (calculating the mean), correl (calculating correlation), covar (calculating covariance), and varp (calculating variance). In addition to these functions, the additional features in Microsoft Excel, especially the Solver feature, are crucial for data processing. This feature allows for finding solution values in linear programming equations by setting value criteria and applying various constraints or objective function limitations. In addition to its usefulness, one of the advantages of Microsoft Excel is its ease of application in the portfolio calculation procedure using the Markowitz Method employed in this study. It is user-friendly and widely popular software in the community. ***C.*** ***Calculating Investment Instrument Returns and Market Value*** The historical data obtained consists of monthly instrument prices or given return values. For data that is still in the form of instrument prices, the initial step of calculation is to compute the monthly returns. ***D.*** ***Calculating Average Returns of Instruments and Market Value*** The next step is to calculate the average return and standard deviation. From the historical return data, the monthly average return and standard deviation are calculated. With a total of 60 records, the average return for each instrument and market value is calculated to obtain the monthly average return. ***E.*** ***Calculating Standard Deviation of Instruments and Market Value*** To simplify the calculation, the stdevp(argument) function is used, where the argument contains the return data of the instruments during the research period in Microsoft Excel software. ### 4924 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 www.ijcsrr.org** ***F.*** ***Calculating Correlation of Investment Instruments*** The next step is to calculate the correlation coefficient between instruments. The correlation coefficient is used to analyze whether a variable has a significant relationship with another variable. It helps determine the strength of the relationship, as well as how one variable influences the other. In this case, the variables are investment instruments. The correlation coefficient indicates the magnitude of the relationship between the movements of two variables relative to their respective deviations. In statistics, the correlation coefficient ranges between two extreme values: perfect positive correlation (+1), indicating a strong positive relationship, perfect negative correlation (-1), indicating a strong inverse relationship, and a correlation coefficient of zero (0), indicating no correlation. ***G.*** ***Calculating Covariance of Investments Instruments*** The next step is to calculate the covariance between instruments. Covariance is the average of the products of deviations between one instrument and another. ***H.*** ***Calculating Portfolio Variance*** The next step is to calculate the portfolio variance. In calculating the portfolio variance, the portfolio standard deviation is calculated first. The portfolio standard deviation is the square root of the portfolio variance. The portfolio variance is obtained by multiplying the covariances between instruments with the weights of each instrument in the portfolio. ***I.*** ***Calculating Portfolio Return and Standard Deviation*** The next step is to calculate the portfolio return. To calculate the portfolio return, we first calculate the average return per instrument per month over the research period. Then, the portfolio return can be calculated by accumulating the average returns per instrument in the research. The next step is to calculate the standard deviation of the portfolio. The next step is to find the portfolio return and portfolio standard deviation using the Solver feature in Microsoft Excel. To facilitate the calculation process, the solver feature in Microsoft Excel is used. In this feature, several variables need to be filled in order to obtain the instrument weights that minimize the variance. From filling in all the variables mentioned above, the spreadsheet calculation process is performed by clicking the solve button. The portfolio standard deviation and portfolio return resulting from the solver calculation process represent a combination of all instruments that minimize the variance, which is also known as the Global Minimum Variance (GMV) point. ***J.*** ***Constructing the Minimum Variance Frontier Curve*** The next step is to find the points that represent combinations of portfolio return and portfolio standard deviation, forming the minimum variance frontier curve using the solver feature in Microsoft Excel. Before finding these values, it is necessary to identify the instrument with the highest return and the instrument with the lowest return as individual instruments. If needed, the data should be plotted on a graph for easier search. Then, determine the number of points to be generated between the highest return and the lowest return, which will result in a return increment (delta return). To obtain these points, the solver feature is used with the objective function and constraints as described in section 3.10. The difference is that in the subject to constraints column, a constraint for the portfolio return is added. The lowest individual return is added to the delta return, resulting in a different standard deviation. Similarly, the other points are processed by adjusting the subject to constraints column with multiples of the delta return until reaching the highest individual return. From the generated points, a line can be drawn through all of them, forming a curve that opens to the right. This curve will also pass through the GMV point directly since this point represents the minimum point of the efficient frontier. ***K.*** ***Selecting the Efficient Frontier Curve*** The next step is to determine the efficient frontier, which is a part of the minimum variance frontier curve. By forming the points described in section 3.10, the minimum variance frontier curve can be created. From the data processing in section 3.9, the GMV point located on the minimum variance frontier curve is obtained. The curve below the GMV point on the minimum variance frontier is considered the non-efficient frontier. This is because, with the same standard deviation, portfolios on the minimum variance frontier curve above the GMV point can achieve higher returns. ### 4925 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 www.ijcsrr.org** ***L.*** ***Finding the Optimal Portfolio*** The next step is to find the optimal point in the risky asset portfolio. In order to obtain the optimal portfolio point, the risk-free rate needs to be determined. Once known, an equation to calculate the reward-to-variability ratio is created in specific cells, referencing other cells that contain portfolio returns, the risk-free rate, and portfolio standard deviation. ***M.*** ***Finding the Complete Optimal Portfolio*** The next step is to construct a portfolio that involves investment in a risk-free instrument. By combining the risk-free instrument with the optimal portfolio of risky instruments, a complete optimal portfolio can be formed. **RESULT AND DISCUSSION** ***A.*** ***Data Processing with Markowitz Method*** With the available historical data for several sample instruments including Bitcoin, gold, government bond, LQ45, and the US dollar from the period of 2018 to 2022, data processing is conducted. The goal is to form an optimal portfolio with measured portfolio performance. ***B.*** ***Instrument Return Analysis*** Investment Return Analysis begins with calculating the return of each instrument. According to Kritzman (1990, p.7) in his book titled "Asset Allocation for Institutional Portfolios," return is the income generated from an asset, adjusted for changes in prices that occur over a specific period, divided by the price of the asset at the beginning of the period. According to Levy (1999, p.198) in his book titled "Introduction to Investments," expected return represents the average of the potential rates. Expected return is also known as the mean return, simplified as the mean. Expected return has two components, namely the probability and the rate of return of an asset. The fluctuation in prices of each instrument makes it difficult for the author to estimate the probability distribution of each instrument. Therefore, to calculate the expected return per month, the researcher assumes that the probability distribution remains constant. This means that the denominator is the sum of monthly sample returns (closing price per month) for each instrument during the research period. In this study, there are sixty months from January 2018 to December 2022. ***C.*** ***Average Return and Risk*** The initial step in data processing, according to Bodie, Kane, and Marcus (2011, p.156), is to calculate the average return. From the historical returns of all instruments obtained, the average return can be calculated for each instrument over the entire research period. This is done by dividing the total return of each instrument during the research period by the number of months in the research period. The "average" function in Microsoft Excel can be used with the arguments of each return for the entire research period to obtain the expected return per instrument. The next step in data processing, according to Bodie, Kane, and Marcus (2011, p.156), is to calculate the risk (standard deviation) for each instrument over the entire research period. Risk is the square root of variance, so calculating risk is aligned with calculating variance. The "stdev" function in Microsoft Excel can be used with the arguments of each return for the entire research period to obtain the risk per instrument. **Table 1.** Standard Deviation and Monthl y Returns of Individual Instruments |Col1|No.|Instrument|Standar Deviation (σ)|Expected Return (E(r))|Col6| |---|---|---|---|---|---| ||1|Bitcoin (BTC)|21.67%|3.26%|| ||2|Gold|4.15%|0.86%|| ||3|Goverment Bond|5.59%|0.26%|| ||4|LQ45|5.27%|-0.14%|| ||5|US Dolar|2.77%|0.29%|| ### 4926 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 ** ### **www.ijcsrr.org** From Table 1. above, it can be seen that Bitcoin has the highest risk, with a standard deviation of 21.67% and an expected return of 3.26%. On the other hand, the US Dollar has the lowest risk, with a standard deviation of 2.77% and an expected return of 0.29%, confirming the concept of high risk high return. ***D.*** ***Correlation Coefficients*** The next step is to calculate the correlation coefficients for all the instruments. The correlation coefficient, or simply correlation, is a statistical measure used to assess the relationship between individual instrument returns or the tendency of two instruments to move together. The correlation coefficient of returns between two instruments is calculated using the statistical function "correl" in Microsoft Excel, with the arguments being the returns of the two instruments. **Table 2.** Correlation Coefficients amon g Instruments |Correlation|BTC|Emas|Goverment Bond|LQ45|US Dolar| |---|---|---|---|---|---| |BTC|1|0.00888|-0.03914|0.26831|-0.12025| |Gold|0.0088 8|1|-0.05084|-0.25173|0.41984| |Goverment Bond|-0.03914|-0.05084|1|-0.38037|0.55016| |LQ45|0.2683 1|-0.25173|-0.38037|1|-0.64063| |US Dolar|-0.12025|0.41984|0.55016|-0.64063|1| From Table 2. above, it can be observed that the correlations among instruments range between -0.64063 < ρ < 0.55016. No instrument exhibits positive correlation with all other instruments. For example, Bitcoin shows positive correlation with Gold and LQ45, but negative correlation with Government Bond and US Dollar, with coefficients of -0.03914 and -0.12025, respectively. On the other hand, no instrument exhibits negative correlation with all other instruments. Government Bond, for instance, shows negative correlation with Bitcoin, Gold, and LQ45, but positive correlation with US Dollar, with a coefficient of 0.55016. ***E.*** ***Covariance*** The next step is to calculate the covariance of all instruments. Covariance is a measure of how two different sets of data vary together. Covariance determines the extent to which two variables are related or how they vary together. Covariance is the average of the deviations from each data point to their respective means. By knowing the covariances and correlations among instruments, investors can determine the composition of available assets to achieve an optimal portfolio with minimal risk and maximum return. The covariance between two instruments is calculated using the covar statistical function in Microsoft Excel, with the arguments being the returns of the two instruments. |Table 3. Instruments Covariances|Col2|Col3|Col4|Col5|Col6| |---|---|---|---|---|---| |Covariance|BTC|XAU_ID R|Goverme nt Bond|LQ45|USD_ID R| |BTC|4.616 %|0.008%|-0.047%|0.301%|-0.071%| |XAU_IDR|0.008 %|0.169%|-0.012%|-0.054%|0.047%| |Goverment Bond|- 0.047%|-0.012%|0.307%|-0.110%|0.084%| |LQ45|0.301 %|-0.054%|-0.110%|0.273%|-0.092%| |USD_IDR|- 0.071%|0.047%|0.084%|-0.092%|0.075%| ### 4927 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 www.ijcsrr.org** ***F.*** ***Variance*** The portfolio variance is calculated using equation 3.9 in chapter 3. Due to the number of instruments used in this study being six, the equation becomes quite long and complex. The variance of each instrument is calculated using the multiplication function in Microsoft Excel. The portfolio variance is calculated in the spreadsheet with a matrix arrangement designed to facilitate the calculation of the long and complex equation. |le 4. Instrument Variances|Col2|Col3|Col4|Col5|Col6| |---|---|---|---|---|---| |Variance BTC||XAU_IDR|Goverment Bond|LQ45|USD_IDR| |[Individual] Weight (Wi)|0%|6%|2%|31%|61%| |[Portofolio] Total Weight (Wp)|100%||||| |[Individual] Variance|0.000%|0.001%|0.000%|0.007%|0.013%| |[Individual] Expected Return|3.26%|0.86%|0.26%|-0.14%|0.29%| |[Individual] Expected Return * (Wi)|0.000|0.001|0.000|0.000|0.002| |[Portofolio] Variance|0.022%||||| |[Portofolio] Standard Deviation|1.50%||||| |[Portofolio] Expected Return|0.19%||||| |Risk Free Rate|0.375%|4.5% annually|||| |CAL slope|-12.18%||||| ***G.*** ***Optimal Portfolio*** To obtain an optimal portfolio, several steps need to be taken, namely forming the minimum variance frontier curve, calculating the GMV (Global Minimum Variance) Portfolio point, selecting the efficient frontier curve, determining the optimal portfolio point, and forming several Capital Allocation Lines. The process of determining the optimal portfolio point will be detailed below. ***H.*** ***Forming the Minimum Variance Frontier Curve*** The minimum variance frontier curve is initially formed by the instruments that provide the highest return and the instruments with the lowest return. Once obtained, 20 other frontier points are formed that minimize variance. As a result, a curve is obtained that opens in the opposite direction to the Y-axis, which represents expected return. **Figure 2.** Minimum Variance Frontier Curve ### 4928 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 www.ijcsrr.org** ***I.*** ***Global Minimum Variance Portfolio*** The principle behind the frontier set of risky portfolios is to capture all levels of risk. However, investors are primarily interested in portfolios that provide the highest return. The entire range of portfolio compositions between risk levels and return levels is depicted in the arrangement of points on the efficient frontier of risky assets. From this arrangement, the Global Minimum Variance (GMV) Portfolio is determined, which minimizes variance while maximizing return. **Table 5.** Global Minimum Variance |Individual|W1|BTC|0%| |---|---|---|---| ||W2|XAU_IDR|6%| ||W3|Goverment Bond|2%| ||W4|LQ45|31%| ||W5|USD_IDR|61%| ||Total||100%| |Portofolio||Varian|0.022%| |||Std Dev|1.50%| |||Exp Return|0.19%| |||Risk Free Rate|0.375%| |||Slope|-12.18%| **Figure 3.** Global Minimum Variance Portfolio The GMV point represents the formation of the lowest-risk and efficient portfolio, obtained by minimizing the variance in the portfolio. Since minimizing variance in the portfolio corresponds to the points on the minimum frontier curve, the GMV point is guaranteed to lie on this minimum frontier curve. The GMV point is located on the curve with the smallest variance or standard deviation, so it lies at the end of the curvature of the minimum frontier curve. As this point is at the far end of the curve, it is ensured to be unique. If a line is drawn from the GMV point parallel to the X-axis (standard deviation), it forms the GMV line that serves to separate the efficient curve and the inefficient curve. The efficient curve is the minimum frontier curve located above the GMV line, while the inefficient curve consists of the minimum frontier below the GMV line. ### 4929 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 www.ijcsrr.org** ***J. Efficient Frontier Curve of Risky Assets*** The Efficient Frontier Curve of Risky Assets is a segment of the minimum variance frontier curve that provides efficient performance, aiming to achieve higher portfolio returns with the same level of risk. This curve is formed by a collection of portfolios that are located above the GMV portfolio line. **Figure 4.** Efficient Frontier Curve The connected curve above the GMV Portofolio line represents the efficient frontier of risky assets, while the disjointed curve below the GMV Portofolio line represents the inefficient frontier. This curve is a plot of the dominant efficient portfolios as they have higher returns compared to portfolios with the same standard deviation located below the GMV Portofolio. Assuming that investors are rational and risk-averse, they will choose portfolios with higher returns when faced with two portfolios that have the same level of risk. Therefore, portfolios located below the GMV portfolio do not need to be depicted in the graph above. In the efficient frontier curve, the portfolio with the lowest level of risk is the GMV portfolio with a standard deviation of 1.50% and a return of 0.29%. The curve will then bend parabolically, and the maximum return is achieved at a position of 3.26% with a standard deviation of 21.67%, where the entire portfolio is invested in bitcoin. **Figure 5.** Efficient Frontier Curve 2 As seen in Figure 4.5, the curve formed below the GMV portfolio line represents the inefficient frontier. This is evident in the case of the LQ45 instrument, which bears a risk of 5.27%. By diversifying and forming a portfolio, the expected return can be increased. By observing the intersection point of LQ45 with the efficient frontier curve (Ev), the expected return can be increased from -0.45% to 1.29% without increasing the risk. Similarly, as shown in Figure 4.5, in the case of Government Bond, which obtains ### 4930 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 ** ### **www.ijcsrr.org** a return of 0.26%, diversifying and forming a portfolio can reduce the risk. By observing the intersection point of Government Bond with the efficient frontier curve (Ep), the risk can be reduced from 5.59% to 1.54% without increasing the risk. This demonstrates that diversification in the form of a portfolio can reduce the level of risk in investments. In other words, investing in a single instrument alone is inefficient compared to investing in a portfolio. ***K. Optimal Portfolio*** From various combinations and allocations of instruments resulting in a portfolio, with the help of the solver function in Microsoft Excel, data on portfolio returns and standard deviations are obtained. These data are plotted on a graph to form the efficient frontier curve. **Figure 6.** Optimal Portfolio The Optimal Portfolio can be determined from one of the points on the efficient frontier curve. To determine which point is the optimal portfolio, another factor needs to be considered, which is the return rate of the risk-free asset. The return rate of the riskfree asset at the end of the research period or at the time of portfolio formation is 4.5% per year or 0.375% per month. As mentioned earlier, the best portfolio is the one that provides the best trade-off between the risk taken and the return obtained. The slope of the Capital Allocation Line (CAL) is a ratio that calculates the relationship between excess return and risk. It is referred to as the rewardto-variability ratio. **Table 6.** Optimal Portfolio |Individual|W1|BTC|18%| |---|---|---|---| ||W2|XAU_IDR|82%| ||W3|Govermen t Bond|0%| ||W4|LQ45|0%| ||W5|USD_IDR|0%| ||Total||100%| |Portofolio||Varian|0.27%| |||Std Dev|5.15%| |||Exp Return|1.29%| |||Risk Free Rate|0.375%| |||Slope|17.75%| ### 4931 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 ** ### **www.ijcsrr.org** From Table 6. it can be seen that the portfolio consists of only two instruments, namely bitcoin and gold. Gold has the largest weight with a composition of 82% and the weight of bitcoin is only 18%. From the types of assets obtained, the optimal portfolio is formed from a combination of the gold instrument, which has a correlation of 0.0088 or approximately 0.9%. This is in line with Markowitz's theory that in order to reduce risk, investors need to form a portfolio with the lowest possible correlation. This is to ensure that losses incurred from one or more instruments in the portfolio can be offset by other instruments with lower correlation. ***L. Capital Allocation Line and Efficient Frontier Curve*** In determining the previous portfolio, all the instruments used were risky assets. If we include an element or opportunity to invest in a risk-free asset, such as the interest rate of Bank Indonesia Certificates, a new portfolio will be obtained. The risk-free asset will be linked to a risky portfolio and form a straight line called the Capital Allocation Line (CAL). By finding the point where CAL intersects the Efficient Frontier curve, an optimal alternative portfolio can be obtained, known as the tangency portfolio, which represents the maximum slope (CAL slope) between the return of risky assets and the risk-free asset on the Efficient Frontier curve. The risk-free point (r_f) represents an instrument with a combination of standard deviation and expected return that is free from risk (standard deviation = 0), obtained from the Bank Indonesia interest rate instrument (Sertifikat Bank Indonesia - SBI). In this study, the average interest rate over the research period was taken, which is 4.5% per year or 0.375% per month. Thus, for the riskfree asset, the point (r_f) is obtained at the coordinates (0, 0.0375%). Therefore, CAL(A) can be formed by connecting the point (r_f) and the maximum expected return, which is the return of the bitcoin instrument. Bitcoin has the highest expected return among individual assets, which is 3.26% with a risk level of 21.67%. For the second asset allocation line, CAL(G) can be formed by connecting the point (r_f) and the global minimum variance portfolio (GMV portfolio) point. The GMV portfolio has an expected return of 0.19% with a risk level of 1.5%. As for CAL(P), it is formed from the point (r_f) to the tangency point between CAL and the efficient frontier curve. This point represents the optimal portfolio that provides the highest performance, with an expected return of 1.29% and a risk level of 5.15%, as shown in Figure 7. below: **Figure 7.** Capital Allocation Line If a line is drawn from the risk-free asset rate point (r_f) parallel to the Y-axis (standard deviation), it will intersect with the CAL lines. Among the capital allocation lines (CALs), CAL(P) forms the largest tangent angle with the risk-free asset line. This is considered the optimal portfolio according to Sharpe (1995) as it provides the highest value among the angles formed by the other CALs. **CONCLUSION AND RECOMMENDATION** The investment portfolio instruments have varying levels of return and risk. Bitcoin has the highest return among the instruments, with a difference of 3.26% compared to the others. However, it is also associated with a high level of risk, reaching 21.67%. Other instruments such as gold, government bonds, LQ45, and US dollars have much lower levels of return compared to bitcoin. ### 4932 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 ** ### **www.ijcsrr.org** Diversification in investment can help investors increase their investment returns while maintaining the same level of risk as individual assets. Additionally, the risk level of an asset can be reduced in a portfolio investment with the same return level as individual assets. Based on the calculation of returns from the five instruments and the average standard deviation, an optimal portfolio can be obtained by combining two instruments: 18% bitcoin and 82% gold. This optimal portfolio can achieve an expected return of 1.29% with a risk level of 5.15%. Considering a risk-free rate of 0.375%, this portfolio forms a slope of 0.1775, which is the largest slope formed between the combination of risk-free instruments and risky portfolios. Investors should allocate their funds more wisely, considering not only the highest return but also the associated risk. High returns often come with high risks, so investors need to assess the risk-return trade-off before making investment decisions. It is recommended for future research to use data from a period that is not a transitional phase. The data used in this study covers the years 2018 to 2022, which includes the period affected by the COVID-19 pandemic starting from early 2020. This global pandemic has significantly influenced all global economic movements, and it may be beneficial to analyze data from a more stable period for a more accurate assessment of investment performance. **REFERENCES** 1. Algarvio, H., Lopes, F., Sousa, J., & Lagarto, J. (2017). Multi-agent electricity markets: Retailer portfolio optimization using Markowitz theory. Electric Power Systems Research, 148, 282-294. 2. Bodie, Zvi, Alex Kane, & Alan J. Marcus. (2011). Investments. Singapore: Irwin/McGraww-Hill. 3. Dian, C. (2020). Pembentukan Portofolio Optimal Pada Beberapa Indeks Saham Menggunakan Model Markowizt. Jurnal Akuntansi Muhammadiyah (JAM), 10(2), 149-159. 4. Elton, Edwin J. dan Martin J. Gruber (1995). Modern Portfolio Theory And Invesment Analysis. John Wiley & Sons 5. Farkhati, F., Hoyyi, A., & Wilandari, Y. (2014). Analisis Pembentukan Portofolio Optimal Saham dengan Pendekatan Optimisasi Multiobjektif untuk Pengukuran Value at Risk. Jurnal Gaussian, 3(3), 371-380. 6. Fernández-Navarro, F., Martínez-Nieto, L., Carbonero-Ruz, M., & Montero-Romero, T. (2021). Mean Squared Variance Portfolio: A Mixed-Integer Linear Programming Formulation. Mathematics, 9(3), 223. 7. Fischer, E. Donald dan Jordan J. Ronald. (1995). Security Analysis And Portfolio Management. Prentice Hall Inc. 8. Grinold, Richard C. and Ronald N.Kahn. (1995). Active Portfolio Management: Quantitative Theory and Applications. Chicago: Probus Publishing. 9. Gurrib, I. (2014). Diversification in Portfolio Risk Management: The Case of UAE Financial Market. International Journal of Trade, Economic and Finance, 445-449. 10. Hali, N. A., & Yuliati, A. (2020). Markowitz Model Investment Portfolio Optimization: a Review Theory. International Journal of Research in Community Services, 1(3), 14-18. 11. Hanif, A., Hanun, N. R., & Febriansah, R. E. (2021). Optimization of Stock Portfolio Using the Markowitz Model in the Era of the COVID-19 Pandemic. The International Journal of Applied Business, 5(1), 37-50. 12. Ivanova, M., & Dospatliev, L. (2017). Application of Markowitz portfolio optimization on Bulgarian stock market from 2013 to 2016. International Journal of Pure and Applied Mathematics, 117(2), 291-307. 13. Jones, Charles P. (2000). Investment: Analysis and Management (7th Edition). USA: Wiley & Son, Inc. 14. Kamali, S. (2014). Portfolio optimization using particle swarm optimization and genetic algorithm. Journal of mathematics and computer science, 10(2), 85-90. 15. Konno, H., & Yamazaki, H. (1991). Mean-absolute deviation portfolio optimization model and its applications to Tokyo stock market. Management science, 37(5), 519-531. 16. Kritzman, Mark P. (1990). Asset allocation for institutional investors (2th Edition). USA: McGraw-Hill Companies. 17. Lee, H. S., Cheng, F. F., & Chong, S. C. (2016). Markowitz portfolio theory and capital asset pricing model for Kuala Lumpur stock exchange: A case revisited. International Journal of Economics and Financial Issues, 6(3S), 59-65. 18. Levy, Haim. (1998). Introductions to investments. South-Western Educational Publishing. 19. Lindblad, J. T. (2015). Foreign direct investment in Indonesia: Fifty years of discourse. Bulletin of Indonesian Economic Studies, 51(2), 217-237. ### 4933 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** ----- ## **International Journal of Current Science Research and Review ** ### **ISSN: 2581-8341 ** Volume 06 Issu e 07 July 2023 **DOI: 10.47191/ijcsrr/V6-i7-108, Impact Factor: 6.789 ** **IJCSRR @ 2023 ** ### **www.ijcsrr.org** 20. Manurung, Adler Haymans and C.Berlian. (2004). Portofolio investasi: Studi empiris 1996-2003 Majalah Usahawan, No.8 Th. XXXIII, 44-48 21. Markowitz, Harry M. (1952). Portfolio Selection. Journal of Finance, 7, 77-91. 22. Muis, M. A., & Adhitama, S. (2021). The Optimal Portofolio Creation Using Markowitz Model. Accounting and Financial Review, 4(1), 72-81. 23. Negara, I. N. W., Langi, Y. A., & Manurung, T. (2021). Analisis Portofolio Saham Model Mean–Variance Markowitz Menggunakan Metode Lagrange. D'cartesian, 9(2), 173-180. 24. Reilly, Frank K and Brown, Keith C. (2000). Investment analysis and portfolio management (6th Edition). USA: Harcourt, Inc. 25. Reilly, Frank K and Brown, Keith C. (2006). Investment analysis and portfolio management (8th Edition). USA: Tomson South-Western 26. Septyanto, E. D. (2019). Analisis Portofolio Optimal Menggunakan Metode Multi Objektif pada Saham Jakarta Islamic Index. UNP Journal of Mathematics, 2(1), 1-6. 27. Sharpe, William F; Gordon J. Alexander; Jeffrey. (1995). Invesment (5th Edition). Prentice Hall. 28. Siregar, B., & Pangruruk, F. A. (2021). A Portfolio Optimization Based on Clustering in Indonesia Stock Exchange: A Case Study of The Index LQ45. Indonesian Journal of Business Analytics, 1(1), 59-70. 29. Verdiyanto, R. (2020). An Empirical Implementation of Markowitz Modern Portfolio Theory on Indonesia Sharia Equity Fund: A Case of Bahana Icon Syariah Mutual Fund. Journal of Accounting and Finance in Emerging Economies, 6(4), 11591172 30. Xiao, Y., & Watson, M. (2019). Guidance on conducting a systematic literature review. Journal of Planning Education and Research, 39(1), 93-112. ***Cite this Article: I Made Gede Abandi Semeru, Yunieta A. Nainggolan (2023).*** ***Investment Portfolio Optimization in Indonesia*** ***(Study On: Lq-45 Stock Index, Government Bond, United States Dollar, Gold and Bitcoin). International Journal of Current*** ***Science Research and Review, 6(7), 4922-4934*** ### 4934 [*] Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 ** Available at: www.ijcsrr.org** ** Page No. 4922-4934** -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.47191/ijcsrr/v6-i7-108?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.47191/ijcsrr/v6-i7-108, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://ijcsrr.org/wp-content/uploads/2023/07/108-24-2023.pdf" }
2,023
[ "JournalArticle" ]
true
2023-07-24T00:00:00
[ { "paperId": "ebdbe428e77abe9440101d40b9071b0339cb790c", "title": "Portfolio Optimization Based on Clustering in Indonesia Stock Exchange: A Case Study of The Index LQ45" }, { "paperId": "80c2009d0d3cb61af902034b6655ab5cc2fb7314", "title": "The Optimal Portofolio Creation using Markowitz Model" }, { "paperId": "69ed95136cba84dc62f6a7e42c418df470f58b7a", "title": "Optimization of Stock Portfolio Using the Markowitz Model in the Era of the COVID-19 Pandemic" }, { "paperId": "8104adafaf320c5c5c2b161eceada70cb09e88d4", "title": "Mean Squared Variance Portfolio: A Mixed-Integer Linear Programming Formulation" }, { "paperId": "43afbb410455c8963227306a45df9b363cbf4040", "title": "Analisis Portofolio Saham Model Mean – Variance Markowitz Menggunakan Metode Lagrange" }, { "paperId": "42b798ae53824722daf8cfef1e17142114abf08e", "title": "An Empirical Implementation of Markowitz Modern Portfolio Theory on Indonesia Sharia Equity Fund: A Case of Bahana Icon Syariah Mutual Fund" }, { "paperId": "e4295ff08b2e20b9177fd58e42e5f1de2411353b", "title": "Markowitz Model Investment Portfolio Optimization: a Review Theory" }, { "paperId": "662bdd2a44672f09a3d77e00012e5a8c9430acfa", "title": "PEMBENTUKAN PORTOFOLIO OPTIMAL PADA BEBERAPA INDEKS SAHAM MENGGUNAKAN MODEL MARKOWIZT" }, { "paperId": "454ca515348a27984cc6fd2c6d5eef12ae9e23d5", "title": "APPLICATION OF MARKOWITZ PORTFOLIO OPTIMIZATION ON BULGARIAN STOCK MARKET FROM 2013 TO 2016" }, { "paperId": "5572fcc987d521aa4c0d244291536358b49cbd1a", "title": "Guidance on Conducting a Systematic Literature Review" }, { "paperId": "0ebe8309c30b2d0384b15c179da8bbb5aadbeac2", "title": "Multi-agent electricity markets: Retailer portfolio optimization using Markowitz theory" }, { "paperId": "bffda437077f0dabfb03d47c99c6ec78f895d17c", "title": "Markowitz Portfolio Theory and Capital Asset Pricing Model for Kuala Lumpur Stock Exchange: A Case Revisited" }, { "paperId": "92f49bf7fdd27f42b033daf90f1a1fcc802363e8", "title": "Investments" }, { "paperId": "ff5666f3caec461a4d8683d319073f79c67bda3b", "title": "ANALISIS PEMBENTUKAN PORTOFOLIO OPTIMAL SAHAM DENGAN PENDEKATAN OPTIMISASI MULTIOBJEKTIF UNTUK PENGUKURAN VALUE AT RISK" }, { "paperId": "8447623aec9be706bd7d517bfc700d91c8a8773a", "title": "Portfolio Optimization using Particle Swarm Optimization and Genetic Algorithm" }, { "paperId": "ba4c5aab9f535b63645b7f7947f7ace9c8b0af31", "title": "Diversification in Portfolio Risk Management: The Case of the UAE Financial Market" }, { "paperId": "16053126f5027144fe23feebde9c7cea23193644", "title": "Mean-absolute deviation portfolio optimization model and its applications to Tokyo stock market" }, { "paperId": "a8d5640416e8e7f0d60d62991c1ccedef10023a4", "title": "Security analysis and portfolio management" }, { "paperId": "39a0f436cbd7b98830b80b58f7e7b115673ef295", "title": "Portfolio Selection" }, { "paperId": null, "title": "Analisis Portofolio Optimal Menggunakan Metode Multi Objektif pada Saham Jakarta Islamic Index" }, { "paperId": "7806aa9cc9ab7a47c10bdea38f5b048588a4aa14", "title": "Investment Analysis And Portfolio Management 7th Edition" }, { "paperId": null, "title": "4933 * Corresponding Author: I Made Gede Abandi Semeru Volume 06 Issue 07 July 2023 19" }, { "paperId": null, "title": "Portofolio investasi: Studi empiris 1996-2003" }, { "paperId": "0b50e3fbffdf968df31965b3a521e6a22e239644", "title": "Active Portfolio Management: Quantitative Theory and Applications" }, { "paperId": null, "title": "Invesment (5th Edition)" }, { "paperId": null, "title": "Asset allocation for institutional investors (2th Edition)" }, { "paperId": "3019ae6463569cec9969ebad1b195f5996a333db", "title": "Modern portfolio theory and investment analysis" } ]
11,391
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" }, { "category": "Law", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/010d89793ad314e40acc52f726a8eb4440010419
[ "Computer Science" ]
0.85231
Blockchain as privacy and security solution for smart environments: A Survey
010d89793ad314e40acc52f726a8eb4440010419
arXiv.org
[ { "authorId": "31584949", "name": "Maad Ebrahim" }, { "authorId": "145556030", "name": "A. Hafid" }, { "authorId": "46256177", "name": "Etienne Elie" } ]
{ "alternate_issns": null, "alternate_names": [ "ArXiv" ], "alternate_urls": null, "id": "1901e811-ee72-4b20-8f7e-de08cd395a10", "issn": "2331-8422", "name": "arXiv.org", "type": null, "url": "https://arxiv.org" }
Blockchain was always associated with Bitcoin, cryptocurrencies, and digital asset trading. However, its benefits are far beyond that. It supports technologies like the Internet-of-Things (IoT) to pave the way for futuristic smart environments, like smart homes, smart transportation, smart energy trading, smart industries, smart supply chains, and more. To enable these environments, IoT devices, machines, appliances, and vehicles, need to intercommunicate without the need for centralized trusted parties. Blockchain replaces these trusted parties in such trustless environments. It provides security enforcement, privacy assurance, authentication, and other key features to IoT ecosystems. Besides IoT-Blockchain integration, other technologies add more benefits that attract the research community. Software-Defined Networking (SDN), Fog, Edge, and Cloud Computing technologies, for example, play a key role in enabling realistic IoT applications. Moreover, the integration of Artificial Intelligence (AI) provides smart, dynamic, and autonomous decision-making capabilities for IoT devices in smart environments. To push the research further in this domain, we provide in this paper a comprehensive survey that includes state-of-the-art technological integration, challenges, and solutions for smart environments, and the role of these technologies as the building blocks of such smart environments. We also demonstrate how the level of integration between these technologies has increased over the years, which brings us closer to the futuristic view of smart environments. We further discuss the current need to provide general-purpose Blockchain platforms that can adapt to unique design requirements of different applications and solutions. Finally, we provide a simplified architecture of futuristic smart environments that integrate these technologies, showing the advantage of such integration.
# **Blockchain as privacy and security** **solution for smart environments: A** **Survey** **MAAD EBRAHIM** **[1]** **, ABDELHAKIM HAFID** **[1]** **(Member, IEEE), and ETIENNE ELIE** **[2]** 1 Department of Computer Science and Operations Research (DIRO), University of Montreal, Montreal, QC H3T1J4 Canada 2 Intel Corporation, 2200 Mission College Blvd, Santa Clara, CA 95054 Corresponding author: Maad Ebrahim (e-mail: maad.ebrahim@umontreal.ca). **ABSTRACT** Blockchain was always associated with Bitcoin, cryptocurrencies, and digital asset trading. However, the benefits of Blockchain are far beyond that. It has been recently used to support and augment many other technologies, including the Internet-of-Things (IoT). IoT, with the help of Blockchain, paves the way for futuristic smart environments, like smart homes, smart transportation, smart energy trading, smart industries, smart supply chains, and more. To enable these smart environments, IoT devices, machines, appliances, and vehicles, will need to intercommunicate without the need for a centralized trusted party. Blockchain can replace third trusted parties by providing secure means of decentralization in such trustless environments. They also provide security enforcement, privacy assurance, authentication, and other important features to IoT ecosystems. Besides the benefits of Blockchain-IoT integration for smart environments, other technologies also have important features and benefits that attracted the research community. Software-Defined Networking (SDN), Fog, Edge, and Cloud Computing technologies, for example, play an important role in enabling realistic IoT applications. Moreover, the integration of Machine Learning and Artificial Intelligence (AI) algorithms provides smart, dynamic, and autonomous decisionmaking capabilities for IoT devices in smart environments. To push the research further in this domain, we provide in this paper a comprehensive survey that includes state-of-the-art technological integration, challenges, and solutions for smart environments, and the role of Blockchain and IoT technologies as the building blocks of such smart environments. We also demonstrate how the level of integration between these technologies has increased over the years, which brings us closer to the futuristic view of smart environments. We further discuss the current need to provide general-purpose Blockchain platforms that can adapt to different design requirements of different applications and solutions. Finally, we provide a simplified architecture of futuristic smart environments that integrates all these technologies, showing the advantage of such integration. **INDEX TERMS** Artificial Intelligence (AI), Blockchain, Cloud Computing, Edge Computing, Fog Computing, Internet-of-Things (IoT), Software-Defined Networking (SDN), Smart Environments **I. INTRODUCTION** Technology development is progressing rapidly, even faster than the expectations decades ago. The reason for such explosion in the technology is the huge effort that is being conducted by the research community, which aims to facilitate human life via developing a futuristic view of a smarter earth. There has been a lot of academic work and industrial adoption to create and implement prototypes of smart cities, which include smart homes, smart factories, smart cars, smart transportation, and various smart human gadgets. A lot of core technologies helped reaching this point of success for this futuristic human civilization. These technologies include, but not limited to, Internet-of-Things (IoT), Software-Defined Networking (SDN), Artificial In telligence (AI), and Cloud, Fog, and Edge Computing. In addition, Blockchain was able to augment those technologies with more features that are essential for the full automation that is needed in smart environments. The role of IoT and Blockchain in Smart environments can be understood from the increase in popularity of the web search terms that are shown in Fig. 1 over the last few years. IoT enabled every physical device to be connected to the internet in order to communicate with other physical devices and services. This can enable, for example, a fridge in the future to automatically detect missing items for its users and automatically order those items from the nearest grocery store. The system in the grocery store can respond to this order and automatically receive the required monetary value 1 ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey **FIGURE 1.** Google Trends search interest for IoT, Smart Home and Blockchain terms between 2004-2020. when accepting the transaction. The grocery items are then automatically collected and sent to the customer using a self driving vehicle. In this futuristic world, the owner of the house does not need to set his smart alarm clock, as it is automatically synchronized to wake him up for his next meeting. To reach his meeting on time, his self-driving car chooses the fastest and safest path using real-time information of the traffic in the city. The car is notified in real-time of nearby accidents in order to optimize its route. This car can communicate with other smart vehicles along the route to provide the safest driving experience for all vehicles on the road. Besides IoT, SDN enables for dynamic and programmable control and management of the underlying network in a smart way. SDN best suits IoT networks, since they change rapidly in terms of the number of devices, their locations, and the amount of data they send. Moreover, SDN enables for the integration of AI and machine learning into the decisionmaking process of load balancing, computational offloading, traffic control, and data flow in the network. SDN is considered one of the major factors in enabling IoT, and hence smart cities innovation [1]. However, it requires intensive computations and storage requirements that cannot be provided by the low-power and limited-resources IoT devices. That is why Cloud, Fog, and Edge Computing technologies were introduced to basically help providing storage and computation resources as paid services to manage such networks. Cloud Computing provides theoretically unlimited storage and computation resources as paid services. They are usually located in central data centers located in distant geographical locations. The distance can burden the core network by the huge amount of traffic created by IoT devices. This distance also increases the service response time (delay) for IoT devices, specially when they need a feedback on their requests. Such delay might not be acceptable for timesensitive IoT applications, such as self-driving cars, where a delay in milliseconds can cause catastrophic incidents. Hence, technologies such as Fog and Edge Computing provide solutions to these problems by bringing those resources closer to the IoT infrastructure. They minimize the delay and save the network bandwidth by performing preliminary preprocessing and analysis on IoT data before sending it to the cloud for heavier processing and permanent storage. There is one missing connection for all those technologies in order to enable the futuristic concept of "trustless" smart 2 **FIGURE 2.** The integration of IoT with Blockchain, SDN, AI, Cloud, Fog, and Edge Computing technologies. cities we talked about earlier. To enable the communica tion among multiple IoT devices that are usually manufactured/owned by different organizations, a third trusted party is usually needed to provide the trust among the devices performing the transactions. Blockchain can act here as that missing connection in order to provide this trust mechanism in a decentralized manner. Blockchain can also be used as a mechanism to permanently log the transactions executed by IoT devices, manage digital assets trading, and perform monetary transactions between them. Actually, Blockchains can do much more than that; they can support and enrich the development of the IoT industry, and they can mitigate many of the current limitations in SDN, Cloud, Fog, and Edge solutions for IoT applications. In this paper, we present a comprehensive survey that shows what Blockchain can provide by its integration with other technologies to build the foundation for future smart environments. This integration is oriented towards IoT applications and supported by different emerging technologies (see Fig. 2). We also show few applications that are brought to life with the help of Blockchain and its integration with those technologies. We further discuss some of the challenges and open research problems of Blockchain and its integration with those technologies. These challenges need to be addressed by the research community in order to provide a ready-to-go Blockchain-based decentralization platform for smart environments. The rest of the paper is organized as follows. Section II compares this work with existing surveys. We then start by introducing our definition for smart environments in Section III. Section IV briefly introduces Blockchain and some of its applications. In Section V, we present IoT-Blockchain integration and some applications for such integration. Section VI describes the benefits of integrating Cloud, Fog, and Edge Computing technologies into Blockchain-IoT ecosystems. Sections VII and VIII show how SDN and AI technologies, respectively, help supporting Blockchain-IoT infrastructures. ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey In Section IX, we discuss some challenges and open research problems that need to be addressed to successfully enable smoother technological integration. Finally, we present in Section X a general smart environment architecture that integrates the technologies presented in this survey to satisfy its requirements. **II. OUR WORK AND EXISTING SURVEYS** Presenting Blockchain integration into multiple technologies in the context of smart environments is what makes this work unique compared to previous surveys. We show how this integration is able to augment those technologies, and how it helps pave the way for smart environments of the future. Beside giving a brief introduction about Blockchain and its applications in smart environments, this work discusses Blockchain integration with IoT to allow for full automation in smart devices. We then study how Blockchain-based IoT solutions are augmented with SDN, Cloud, Fog, and Edge Computing to enhance the capabilities of IoT applications. Finally, we study the impact of AI and machine learning algorithms to make such solutions even smarter. Table 1 shows the level of technological integration to Blockchain-IoT solutions in reviews and surveys. The table shows that most existing surveys do not consider the inclusion of all these technologies to enhance BlockchainIoT integration. In addition, existing surveys do not elaborate on the direct impact such integration on smart environment applications. Most surveys focus on integrating Cloud Computing to mitigate the resource limitations of IoT devices while considering Fog and Edge technologies to provide privacy and minimize the delay. However, there is a lack of surveys covering the recent interest in using SDN and AI technologies for dynamic network management and complex optimization problems, respectively. Stojkoska and Trivodaliev [24], for example, focused on the role of IoT in smart home applications. In addition, Bhushan *et al.* [25] discussed the integration of Blockchain to support IoT applications in smart cities. Even though our work is oriented towards smart environments, we also study the effect of technological integration to establish such smart environments. We cover how other technologies help mitigate several problems in Blockchain-based IoT smart systems. Other surveys also discussed IoT integration with other technologies, like AI [26], or SDN [27]. However, these surveys do not consider smart environment applications, and do not cover the benefits of Blockchain decentralization properties. The majority of the surveys only focus on Blockchain and IoT integration, including the benefits and challenges of such integration [2], [5], [6], [9], [14]. Ali *et al.* [11], for example, reviewed Blockchain-based platforms and services that are used to augment IoT applications. Similarly, Lao *et al.* [17] covered the use of Blockchain to address IoT limitations and secure IoT networks. They also gave a comprehensive overview of IoT-Blockchain applications, including architectures, communication protocols, and traffic models for such **TABLE 1.** The integration of Cloud, Edge, Fog, SDN, and AI technologies in Blockchain-IoT solutions. |Authors|Year|Integration with Edge/Fog Cloud SDN AI| |---|---|---| |Conoscenti et al. [2]|2016|| |Christidis and Devetsikiotis [3]|2016|| |Reyna et al. [4]|2018|| |Ramachandran and Krishnamachar [5]|2018|| |Panarello et al. [6]|2018|| |Fernández-Caramés and Fraga-Lamas [7]|2018|  | |Banerjee et al. [8]|2018|   | |Atlam et al. [9]|2018|| |Zheng et al. [10]|2018|| |Ali et al. [11]|2019|   | |Dai et al. [12]|2019|   | |Ferrag et al. [13]|2019|  | |Makhdoom et al. [14]|2019| | |Salah et al. [15]|2019|  | |Yang et al. [16]|2019|   | |Lao et al. [17]|2020|| |Alharbi [18]|2020|  | |LI et al. [19]|2020|| |Luo et al. [20]|2020| | |Xie et al. [21]|2020|| |Mohanta et al. [22]|2020|  | |Chamola et al. [23]|2020| | applications. There are several other surveys that only focus on the effort to secure IoT networks using Blockchain [8], [28]. Riabi *et al.* [28], for example, covered contributions that use Blockchain to mitigate single-points-of-failures in centralized access control architectures for IoT devices. Even though the majority of the surveys only deal with Blockchain-IoT integration, some of them have different focus or interest. Reyna *et al.* [4] did cover BlockchainIoT integration, in particular, running Blockchain on IoT devices. Fernández-Caramés and Fraga-Lamas [7] reviewed Blockchain-IoT integration in healthcare, logistics, smart cities, and energy management systems. Likewise, Ferrag *et* *al.* [13] focused on Blockchain-IoT integration for applications in Internet-of-Vehicles (IoV), Internet-of-Energy (IoE), Internet-of-Cloud (IoC), and Edge Computing. In addition to those applications, Mohanta *et al.* [22] did overview existing security solutions for IoT networks using Blockchain and AI technologies. Garcia [29] reviewed the integration of AI, IoT, and Blockchain from the taxation, legal, and economical point of views. There are also other surveys that focus on the integration of Blockchain and other technologies outside the context of IoT. Ekramifard *et al.* [30], for example, produced a systematic literature review on the integration of Blockchain with AI; particularly identifying the applications that can benefit from such integration. Similarly, Salah *et al.* [15] surveyed the way Blockchain can enhance and solve AI limitations. They also reviewed the role of Blockchain in achieving decentralized AI schemes. Additionally, Akter *et al.* [31] investigated a diverse set of applications in the literature that 3 ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey are based on Blockchain, AI, and Cloud technologies. Xie *et al.* [21] surveyed Blockchain-based solutions to augment the Cloud Computing technology. They did study the role of Blockchain to provide decentralized Cloud Exchange services. Other surveys focused on Blockchain integration with Fog and/or Edge computing technologies. For instance, Baniata and Kertesz [32] covered contributions that integrate Blockchain with Fog Computing. likewise, Yang *et al.* [16] provided a survey on Blockchain integration with Edge computing, which can help decentralize network management. Moreover, SDN integration with Blockchain has been also presented in other surveys [18], [19]. Alharbi [18], for example, surveyed existing papers that secure SDN architectures from different attacks using Blockchain. Moreover, LI *et* *al.* [19] reviewed how Blockchain and SDN technologies complement each other when integrated together. Our work goes far beyond few technological integration between Blockchain and IoT to mitigate some of their limitations. We review existing work according to different levels of technological integration, which is oriented towards smart environment applications. We briefly discuss the benefits of Blockchain technology itself, and the benefits of BlockchainIoT solutions in smart environments. Then, we present the benefits of integrating Cloud, Fog, Edge, SDN, and AI technologies in Blockchain-IoT smart systems. We then present open research problems to be addressed to create smooth technological integration for smart futuristic environments. Finally, we provide a simplified smart environment architecture that shows how Blockchain help integrating the various technologies discussed in this paper. **III. SMART ENVIRONMENTS** Over history, humans have always created innovative solutions to make their lives easier. The recent technological inventions allowed us to live in an environment that was con sidered science fiction in the past. However, scientists always wanted to push this further by creating a smarter world where automation is included in every aspect of human lives. This was only possible through introducing IoT technology, where IoT sensors and actuators are embedded in physical devices and machines, making them able to interconnect through the internet. IoT, with the help of Big data analysis, AI, Machine Learning, and many other innovative technologies allowed for the realization of these smart environments [33], [34]. Governments, industries, and scientists are all racing towards creating and prototyping smart cities for boosting the life quality of their citizens. Smart homes, for example, provide a futuristic domestic environment that delivers a technologically advanced living experience for people [35]. While smart education includes smart campuses, smart universities, and smart classrooms for students in these smart cities [36]. Various innovative solutions were used to mitigate different challenges in realizing these environments, like using Blockchain to secure smart city applications [37]. The difficulty in realizing these applications increases as the 4 human interactions with the smart infrastructure get more complex, as in the case of smart transportation systems [38]. Therefore, the research in these complex smart infrastructures, including smart transportation systems, had become a dominant research topic in the context of smart environments [39]. The fourth industrial revolution, also called Industry 4.0, is another important component of smart environments, which could only be realized after introducing smart factories and smart supply chain systems [40]. To decrease the cost while increasing the quality of mass production, businesses stood up for shifting from traditional manufacturing to smart factories [41]. The introduction of Smart Industry allows for the automation of intelligent predictive maintenance strategies, which provide major cost saving over time-based preventive maintenance [42]. The fourth industrial revolution in smart cities also led to the development of distributed smart energy trading systems, which require Blockchain for security and reliability [43]. Smart farming is another component of smart environments, which was enabled by IoT, Wireless Sensor Networks (WSN), Cloud Computing, Fog Computing, as well as Big data analytics [44]. Big data analytic plays an important role in bringing real-time decision-making capabilities into smart farming environments by obtaining valuable information from the collected data [45]. For example, machine learning can automate decision-making in smart farming environments by predicting soil drought and crop productivity [46]. Smart vehicles also exist in almost all smart environments, including smart farming, smart factories, smart cities, and smart transportation systems, and we also see how Blockchain can secure the transactions between these smart vehicles [47]. IoT paved the way to pervasive computing, also called ubiquitous computing, which is the interconnection of sensors with computing capabilities through the internet. These smart devices are the main building blocks for smart environments, which aim at providing a comfortable living experience for humans via performing repetitive or risky tasks using these devices. Furthermore, other technologies are needed to get the full benefit out of these devices, including AI, machine learning, Big data analytics, as well as computer networks, parallel and distributed computing, and much more. The realization of smart environments needs a lot of work, so in this paper, we focus on how Blockchain provides security, privacy, and other features into smart environments with the help of other technologies. The outcomes of this research spot the light on what Blockchain can provide for smart environments. Besides, it proposes a simplified IoTbased smart environment architecture using Blockchain with Cloud/Fog computing, SDN, and AI technologies. **IV. BLOCKCHAIN AND ITS APPLICATIONS** Starting by Bitcoin in 2008, Blockchain went far beyond the world of cryptocurrencies as it was proposed by Nakamoto *et al.* [48]. Blockchain enabled many smart environment ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey ##### Previous Block Next Block ##### Tx1 Tx2 … Txn Tx1 Tx2 … Txn |Col1|Hash| |---|---| ||Other Info Nonce … Tx1 Tx2 Txn| **FIGURE 3.** The concept of Blockchain. |Hash|Col2| |---|---| |Other Info Nonce … Tx1 Tx2 Txn|| applications and solutions as has been clearly seen in recent studies. For example, Blockchain provides authentication and authorization for smart city applications [49], like managing real estate deals in smart cities [50]. Blockchain was also used as a core framework to secure smart vehicles [50] and smart grid systems [51]. Blockchain is essentially a distributed ledger that is made up of blocks (list of transactions) that are backwardconnected through hash pointers as shown in Fig. 3. Besides the transactions’ Merkel tree inside each block. These pointers guarantee the immutability of this ledger. Different consensus algorithms, such as Proof-of-Work (PoW) [48], have been proposed to securely update the ledger without the need to a centralized entity. In PoW, the nodes in the network compete to solve a puzzle that is solved only by trying different Nonce values. Blockchain can provide pseudonymity and traceability, which are very useful in dozens of domains. Those features are achieved with the help of different technologies, including cryptography, hashing, and digital signatures. In a public Blockchain, anyone can participate in the network to create and validate transactions and blocks. Con trarily, only authorized nodes can join a private Blockchain to read, create, or validate transactions and blocks. A consortium Blockchain has mixed features of both private and public Blockchain implementations, where only permissioned users can perform Blockchain transaction with different levels of restrictions. Bitcoin and Ethereum are examples of public Blockchain implementations, and Hyperledger Fabric [52] is an example of a consortium Blockchain. We can also have a private Blockchain using a private version of Hyperledger Fabric or Ethereum [53]. Zheng *et al.* [10] gave a taxonomy on different types and implementations of Blockchain, and the different consensus algorithms used in them. Beside PoW and Proof-of-Stake (PoS) consensus algorithms, Delegated-PoS (DPoS) [54] and Practical Byzantine Fault Tolerant (PBFT) [55] have been also considered in different implementations. However, PoW is still the most secure consensus algorithm despite its huge computational and energy consumption. The peer-to-peer (P2P) network of Blockchain should have decentralized management of the data that is synchronized among all peers in the network. To keep the data synchronized efficiently, two message transfer protocols are usually adopted between the nodes, i.e. Gossip [56] and Kademlia [57]. Bitcoin uses Gossip, which spreads information by communicating only with neighbors, mimicking the spread of epidemic diseases. Ethereum communication protocol on the other hand is inspired by Kademlia, which maintains a distributed hash table that specifies the communicating neighbors for each node. The peers in the network are usually of three types, i.e. core, full, and light nodes. All peers participate in validating and broadcasting transactions and blocks. Core nodes are responsible for network routing, whereas full nodes are responsible for storing the whole Blockchain. Light nodes are only responsible for maintaining users’ accounts in resource-constrained devices. Asymmetric cryptography and zero-knowledge proof [58] can be used to secure users’ data with the help of Blockchain [59]. Hence, Blockchain can be used as a secure distributeddatabase system, e.g. medical record system. The patients can ensure data integrity and privacy by giving data access only to specific medical firms. Records from different hospitals and clinics can be obtained in a secure manner without vulnerable central authorities. Peng *et al.* [60] implementation prevents falsified data retrieval to ensure authenticity, integrity, and efficiency of Blockchain data queries. With the help of smart contracts, automation in such systems mitigate the need for human centric auditing and revisioning. Smart contracts [61] were proposed in the 1994 by Nick Szabo [62] and rediscovered in the context of Blockchain with Ethereum [63]. They enforce rules and conditions inside transactions to lower the cost induced by third central parties, such as law firms. These contracts are automated and permanently stored in Blockchain immutable ledger. Running code inside Blockchain using smart contracts adds decentralization to applications in trustless environments. Smart contracts adds automation to network management, security services, and IoT applications. However, Blockchain still suffer many challenges and problems, like scalability [64] and privacy leakage [65]. Electricity consumption of PoW and the capitalism problem in PoS algorithm [66] are some consensus-related problems in Blockchain. Moreover, the financial use of Blockchain still needs much work from legal and law enforcement perspectives [67]. ***A. BLOCKCHAIN APPLICATIONS*** Zheng *et al.* [10] classified Blockchain applications to finance systems, reputation systems, public and social services, and security and privacy applications. In this survey however, we focus on security [68], AI [15], IoT [14], and healthcare [69] applications. We show how Blockchain has led to the development and enhancement of many applications in these domains. We give below brief descriptions of few Blockchain applications that demonstrate the power of Blockchain in enhancing and simplifying humans lives. Such applications and prototypes pave the way for the smart environments of the future we are looking for: *MedRec* [70]: Is a distributed Electronic Health Records (EHR) and medical research data. This decentralized medical claim system handles EHRs using Blockchain. It allows patients to easily and securely share their medical records 5 ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey across different health insurance companies, medical institutions, clinics, and pharmacies. It guarantees authentication, confidentiality, accountability for sharing patients’ data. Medical stakeholders can be incentivized to play the role of block miners in this system. *BitAV* [71]: Is a Blockchain-based fast anti-malware scanning application that secures entire networks. It can provide security services in a decentralized manner to computationallimited environments, like IoT networks, given enough RAM and storage. It is 1,400% faster than conventional antivirus software, and 500% less in terms of average update propagation flow. It achieves this performance using the P2P network maintenance mechanism inspired by Blockchain consensus. *OriginChain* [72]: Is an adaptable Blockchain-based traceability system. It is decentralized, transparent, and tamperproof; it traces the origin of products across complex supply chains. Private data, customer/product information, product certificates and photos are kept off-chain to increase the performance and save space. However, the hashes of that data are kept on-chain to ensure immutability. *E-Voting* : Are decentralized electronic voting systems that are usually required to scale well to large scale voting [73]. Yang *et al.* [74] used an Ethereum smart contract to give a prototype of a voting system that provides confidentiality using homomorphic encryption. The eligibility of the voters, and the integrity and validity of their votes can also be verified. Similarly, Khoury *et al.* [75] created transparent, consistent, and deterministic Ethereum smart contracts, which can be modified by voting organizers. Voters should pre-register with mobile phone numbers and can only vote once in that voting platform. Hjálmarsson *et al.* [76] used a smart contract in a private version of Ethereum to guarantee transparency and privacy. They used Blockchain-as-a-Service to host nationwide elections, but they still need additional measures to support countries with huge population. *Reputation Systems* : Are decentralized systems for rewards and educational records. Sharples and Domingue [77] democratise educational reputations beyond the academic community using a decentralized Blockchain-based framework. It creates a permanent distributed record of intellectual effort and associated reputational reward. It can be also used for crowd-sourced timestamped patenting, i.e. proof of academic, art, and scientific work. **V. BLOCKCHAIN & INTERNET OF THINGS (IOT)** IoT devices in smart environments should be digitally connected in order to share their data and automate their tasks. These devices are usually made up of sensors and actuators that connect through the internet. Blockchain can be used to increase IoT automation and solve a number of its limitations, including security, privacy, and scalability. That makes Blockchain one of the enabling technologies for IoT networks in smart environments. Using Blockchain for decentralized monetary transaction and digital asset trading is also an enabler for IoT devices in smart environments. Blockchain was also used as a distributed access management 6 system for the ever-growing IoT networks to mitigate the overhead of centralized architectures [78]. IoT is one of the biggest trends in today’s innovations [79]. It enables physical devices to communicate through the internet to send/receive data, and can perform actions. IoT has already emerged in humans life in different domains, such as smart home devices [24], smart cities [80], and smart transportation [81]. It has been proven to be well suited for E-business models, specially with the help of Blockchain and smart contracts [82], [83]. As an example, Zhang and Wen [82] proposed an E-business architecture to build systematic, highly efficient, flexible, reasonable, and low cost business-oriented IoT ecosystems. In addition, Zheng *et al.* [10] discussed IoT-Blockchain integration and its associated challenges. They identified scalability, ever-growing storage, privacy leakage, and selfish mining as critical problems. They pointed out that big-data analytics and AI can enhance Blockchain-IoT integration and their applications. Blockchain technology is a good solution to mitigate the problems of traditional central communication and management systems for large scale IoT devices. On one hand, resource limitations of IoT devices, and scalability issues in Blockchain create big problems for Blockchain-IoT integration. On the other hand, the continuous research effort has led to innovative solutions to these problems, such as IoTA [84]. IoTA is not an abbreviation; rather it comes from IoT and the word "iota", which means an extremely small amount. The name reflects its purpose to connect IoT devices through micro/zero value transactions. Shabandri and Maheshwari [85] developed this architecture as a protocol to provide trust in IoT networks. It eliminates transaction fees and the concept of mining to solve both of those problems. The main component of the IoTA is what they called the Tangle, which is a guided acyclic graph (DAG) for transaction storage. Shabandri and Maheshwari [85] demonstrated the performance of IoTA by implementing two IoT applications, namely a smart utility meter system and a smart car transaction system. ADEPT (Autonomous Decentralized Peer-to-Peer Telemetry) [86] is another example of a Blockchain-based databaselike framework for decentralized IoT networks. It is a proofof-concept that was produced by a collaboration between IBM and Samsung. ADEPT provides a secure and a lowcost interaction mechanism for IoT devices, where devices have the ability to make orders, pay for them, and confirm their shipment autonomously. The underlying technologies behind ADEPT are Ethereum smart contracts, BitTorrent file sharing and TeleHash peer-to-peer messaging. They used a mix of PoW and PoS consensus algorithms to provide secure decentralization for transaction approval. The huge number of IoT devices creates a burden on the network and raises problems such as data security, privacy, and integrity. Two of the most challenging issues in IoT security are heterogeneity and scalability of IoT devices that are distributed over the network. Blockchain can solve the security, privacy, and data integrity issues in a decentralized manner. Blockchain is also able to create traceable IoT ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey networks, where transaction data are recorded and verified without intermediary management and control [17]. Distributed Blockchain-based management of IoT devices can be also performed at the edge of the network to avoid using distant resources [87]–[89]. Blockchain can also provide a distributed digital payment system for IoT devices to reduce the cost induced by third parties. To protect IoT data in a Blockchain, a hybrid combination of private and public Blockchains is needed [90]. Data privacy is maintained by private management nodes, whereas the consensus algorithm is maintained by public nodes. In addition, traditional Blockchain consensus algorithms are not suitable for IoT-Blockchain solutions because of their computational and time requirements. To solve this problem, Samaniego and Deters [91] proposed a Blockchainas-a-service platform for IoT applications. They introduced structural improvements for Blockchain to fit IoT networks by improving consensus algorithms. Atlam *et al.* [9] listed several benefits for using Blockchain for IoT, like publicity, decentralization, resiliency, security, speed, cost saving, immutability, and anonymity. However, they also highlighted the challenges, such as scalability, processing power, time delay, storage requirements, lack of skills, legal and compliance, and naming and discovery. Similarly, Ramachandran and Krishnamachari [5] proposed Blockchain-based monetary exchange for data and compute in IoT networks. IoT transactions can also be recorded on Blockchain for future accounting and auditing. However, they were also concerned by Blockchain challenges, like latency, bandwidth consumption, transaction fees, transaction volumes, partition tolerance, and physical attacks on IoT devices. Dorri *et al.* [92] investigated the delay, expensive computations, and bandwidth overhead problems of Blockchain to better fit IoT applications. They proposed a secure, private, and lightweight hierarchical architecture for Blockchain-IoT applications. The hierarchical architecture is made up of three layers, namely local network (smart home), overlay network, and cloud storage (see Fig. 4). Likewise, Reyna *et al.* [4] argued the benefits of Blockchain-IoT integration to securely push code into IoT devices to speed up the deployment of new IoT ecosystems. They proposed Blockchain-based direct firmware update without the need to trust third-parties. Since IoT devices are manufactured by different vendors, they may not agree on sharing a common Blockchain. Hence, IoT devices should be able to send transactions across different Blockchain implementations with different consensus protocols. Dai *et al.* [12] named the integration of Blockchain and IoT as Blockchain-of-Things (BCoT). They stated that a successful integration requires interoperability, traceablitiy [72], reliability, and autonomicity [82]. They also stated that decentralization, heterogeneity, poor interoperability, privacy and security vulnerabilities [3] are critical issues for such integration. Beside those problems, the lack of publicly available IoT datasets for the research community is another |Col1|Local Storage|Col3| |---|---|---| |Smart IoT Devices||| |Centralized Private Blockchain||| **Smart Home** **FIGURE 4.** Hierarchical Blockchain-based IoT architecture for smart homes. problem to be addressed. Thus, Banerjee *et al.* [8] worked on standards for securely developing and sharing IoT datasets. They proposed two conceptual solutions to ensure IoT data integrity and privacy using Blockchain. Reyna *et al.* [4] proposed using three different communication approaches for the communication between IoT devices with the help of Blockchains (see Fig. 5). The first approach is IoT-IoT interactions, where IoT interactions take place offchain. This approach is the fastest among other approaches since only a part of IoT data is stored on-chain. The second approach is IoT-Blockchain, where all the interactions take place through Blockchain. With this approach, all IoT data are stored on-chain to ensure traceable interactions. The third approach is hybrid, where only part of the interactions/data goes through Blockchain, while the rest is done directly between IoT devices. The hybrid approach is better in terms of performance and security; however, it requires careful orchestration for those interactions. The impact of IoT on industry and enterprise systems asked for standardizing this technology to speed up its development and spread in this domain [93]. Christidis and Devetsikiotis [3] showed that Blockchain-IoT integration will cause transformations across industries, and open the door for new business models and distributed applications. It will also facilitate service and resource sharing, and automate time-consuming workflows. Blockchain smart contracts are the main source of such automation for complex multistep processes. They can reduce cost and time for future business models and applications in smart environments. 7 ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey **(b)** IoT-Blockchain **(c)** Hybrid Approach **(a)** IoT-IoT Interactions **FIGURE 5.** Three different Blockchain-IoT interaction architectures proposed by Reyna *et al.* [4]. ***A. BLOCKCHAIN-IOT APPLICATIONS*** A number of key applications were developed for smart environments with the help of Blockchain-IoT integration. Prototypes of these applications are necessary to study them and solve their limitations and issues. Here, we list few applications, case studies, and prototypes for smart environments that are brought to live using Blockchain-IoT integration: *Smart Homes* : Dorri *et al.* [92] proposed a smart-home case study that uses Blockchain to add security and privacy features to various IoT applications. They presented a lightweight secure system for IoT-based smart homes to minimize the overhead of consensus algorithms. Their hierarchical architecture consists of smart homes, an overlay network, and cloud storage. Other works later analyzed this architecture and the role of smart homes as miners in a private Blockchain [94], [95]. They used simulation results to show the insignificance of Blockchain overhead, including power consumption, compared to achieving confidentiality, integrity, availability, security, and privacy. *Energy Trading* : Sikorski *et al.* [96] presented a proof-ofconcept and a detailed implementation for energy trading using realistic data. It is a machine-to-machine electricity market for the chemical industry. They used a Blockchain smart contract for automatic confirmation of trading and payment commitments. They implemented a scenario of two electricity producers and one consumer that automatically trades energy using IoT. Producers publish energy trading offers for a given price, while consumers can read, analyse, and accept, or refuse those offers. Consumers can pick and accept the offer with the minimum cost using a smart contract execution as an atomic exchange of assets, i.e. currency for 8 energy. Each transaction is saved on the immutable ledger for future proofing. *Smart Things* : Panarello *et al.* [6] did a systematic survey of Blockchain-IoT integration. They covered various smartapplication domains, such as smart homes, smart properties, and smart cities. They also covered smart energy-trading, smart manufacturing, smart data-marketplaces, and other generic smart applications. They classified existing work based on the development levels, consensus algorithms, and technical challenges. They also identified the challenges that include confidentiality, authentication, integrity, availability, and nonrepudiation. Conoscenti *et al.* [2] gave a systematic literature review of Blockchain applications for IoT. They discussed 18 use cases of Blockchain, 4 of which are specifically designed for IoT. The rest of the use cases are applications for decentralized private data management systems that are inline with IoT applications. The four IoT-related Blockchain applications they discussed are: 1) *E-business models for IoT solutions* : Zhang and Wen [82] designed a methodology for transactions and payments between smart IoT devices using Blockchain smart contracts. 2) *IoT Data-Market* : Wörner and von Bomhard [97] proposed a prototype for a system where sensors can sell data directly to a data-market in exchange of Bitcoins. 3) *Public-Key Infrastructure* : Axon and Goldsmith [98] adapted what is called Certcoin [99] to a privacyaware Blockchain-based public-key infrastructure to avoid web certificate authorities, provide certificate transparency, and mitigate single points of failures. 4) *Enigma* : An autonomous Blockchain-based decentralized computation platform proposed by Shrobe *et al.* [100]. It allows different users to run computations on personal data with guaranteed privacy. In order to fully benefit from IoT applications and systems, we need to start by mitigating their current limitations and challenges. The power and resource limitations of IoT devices do not make them suitable to process heavy computations and store large data. Cloud Computing helped providing theoretically unlimited storage and computational resources for IoT devices. In addition, Fog and Edge Computing bring those resources closer to IoT devices to decrease network delay and bandwidth consumption. Cloud, Fog, and Edge resources allow for performing heavy analysis on IoT data and maintain real-time performance for time-sensitive IoT applications. These resources will extend the capabilities of Blockchain-IoT integration and mitigate many of its limitations and issues. **VI. BLOCKCHAIN & CLOUD, FOG, AND EDGE** **COMPUTING** The huge amount of IoT data generated in smart environments needs to be processed in large data centers that have enough computing and storage capacities. That is why Cloud Computing was proposed as the first solution for big data ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey analysis and storage for IoT-based applications in smart environments [101]. However, Fog Computing paradigm was evolved to support the Cloud in order to mitigate latency intolerance of real-time IoT applications in smart environments [102], such as autonomous vehicles. Similarly, Edge Computing utilizes available computational resources in smart environments, such as the resources in smart vehicles or smart phones, in order to reduce the latency even further [103]. Blockchain was again able to support these technologies by securing and protecting the privacy of this big data in smart environments [104]. With the advent of the Cloud Computing technology, it is possible to perform very expensive computational tasks and store tremendous amount of data. Payment is only for the cost of usage, which is better than purchasing expensive resources for time-framed tasks. The Cloud technology removes the overhead of maintenance and resource management, which is usually related to owning resources by small to medium sized companies. Furthermore, the Cloud enables resource and power-limited devices, such as smartphones and IoT devices, to perform heavy computations and store huge amounts of data. Those devices need to only use a lightweight remote interface with the cloud as a solution to mitigate their power and resource limitations. Cloud Computing is the outcome of integrating parallel, distributed, and grid computing [105]. Although it has been proposed in the 60s, the technology has started to be widely used commercially in 2006 by Amazon [106]. Services are usually provided as packages, namely Software-as-a-Service (SaaS), Platform-as-a-Service (PaaS), and Infrastructure-asa-Service (IaaS). Zhou *et al.* [107] added the possibility of having Data-as-a-Service (DaaS), Identity and Policy Management as-a-Service (IPMaaS), Network-as-a-Service (NaaS). X-as-a-Service (XaaS) is a term called for the countless number of services that can be provided by cloud computing [108]. XaaS allows for more service packages, which enable the creation of various applications and systems based on the services provided. Jadeja and Modi [106] categorized the deployment of the cloud infrastructures into public, private, hybrid, and community clouds. They listed easy management, cost reduction, uninterrupted services, disaster management, and green computing as the main advantages of Cloud Computing. A tremendous number of systems and applications were built based on Cloud services since 2010 [107]. Ma and Zhang [108] studied a provider for cloud services called Google App Engine (GAE). They explored three of its services, i.e. Google File System (GFS), MapReduce, and Bigtable. They showed how these services opened the doors for Big Data Analysis in IoT environments. Before discussing how Cloud integrates into Blockchainbased IoT solutions, we first show how Blockchain helped the Cloud technology itself. Blockchain has been used for Cloud Exchange [21], which allows for provisioning and management of multiple Cloud providers. It can lower the price and can provide flexible options for Cloud users. Xie *et al.* [21] proposed using Blockchain to decentralize Cloud exchange services. It mitigates malicious attacks and the cheating behaviors of third-party auctioneers. Furthermore, Blockchain form new models for security-aware Cloud schedulers, like the lightweight Proof–of–Schedule (PoSch) consensus algorithm [109]. In addition, integrating the Cloud with Blockchain-IoT solutions enables for seamless authen tication, data privacy, security, easy deployment, robustness against attacks, and self-maintenance [7]. Because of limited power, storage, and computational resources, IoT devices heavily depend on Cloud resources. However, relying on the cloud can create unacceptable delays, specially when a feedback is required. Fog and Edge technologies can be used to reduce these delays; in addition, they provide better privacy by processing IoT data in proximity to IoT devices. Therefore, they are more convenient than the Cloud, specially for Blockchain-IoT applications. Samaniego and Deters [91] evaluated both edge-based and cloud-based Blockchain implementations for IoT networks. They did show, via simulations, that edge-based Blockchain implementations outperform cloud-based implementations. Transferring massive amounts of data, that is produced by IoT devices, to the Cloud consumes considerable amount of network bandwidth. Flooding the core network with massive traffic, like in streaming IoT applications, will create network bottlenecks and single points of failures. There are also problems with privacy exposure, and context and geographical location unawareness. Edge Computing solves these problems by forming a distributed and collaborative computing resources. It reduces power consumption, provides real-time service, and improves scalability for many industrial IoT solutions [110]. In addition, Unmanned Aerial Vehicles (UAVs) can act as limited-resource edge servers in environments with limited or no infrastructures [111]. Blockchain can be used to provide mutual-confidence between UAVs of different providers [112], and to preserve privacy and security for their data [113]. Fog Computing has been also proposed to pre-process and trim IoT data before sending it to the cloud for computationally expensive analysis [114]. Fog servers are usually deployed in smart gateways, which are equipped with decent computational resources. They eliminate unnecessary communication to the cloud to save the network bandwidth and reduce the load in its data centers. Fog Computing should not be mixed with Edge Computing, as Edge Computing brings the computations very close to the end devices. Edge servers are usually deployed in Radio Access Networks (RANs) or mobile Base Stations (BSs). On the other hand, Fog Computing provides distributed mini-cloud resources between the end devices and the cloud (see Fig. 6). Both technologies reduce the network delay, and provide better Quality-ofService and Experience (QoS & QoE). These technologies are essential parts in real-time/streaming IoT applications, which might also require location/context-aware information processing. Mobile Edge Computing, also called Multiaccess Edge 9 ----- Cloud Layer → Fog Layer → Edge Layer → IoT Layer → **FIGURE 6.** Cloud, Fog, and Edge Computing for IoT networks. Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey devices provides real-time services for the users. To mitigate malicious attacks on fog nodes, Wu and Ansari [121] proposed partitioning the fog nodes into different clusters. The nodes in each cluster have their own access control list, which is protected and managed by Blockchain. They showed, using simulations, the effectiveness of their approach in reducing the computational and storage requirements of Blockchain. They included a heuristic algorithm to reduce the time needed to solve the consensus puzzle by having all fog nodes perform the computations cooperatively. Cloud, Fog, and Edge computing technologies provided IoT devices with more capabilities, which increased their adoption in new applications and domains. The continuous expansion of IoT networks and their dynamic nature ask for intelligent and dynamic management of such networks. Adaptive control of the network does not only save in terms of hardware cost, but also dynamically optimizes the operations in the network. Network optimization can be done using Software-Defined Networking, which dynamically changes data flow in the network based on its state. SDN can incorporate Artificial Intelligence algorithms to optimally choose between using Cloud, Fog, or Edge resources, or even a combination of them, to process IoT data. Computing (MEC), is a specific type of Edge Computing that leverages mobile Base Stations. It complements cloud computing by offloading computations closer to mobile and IoT devices. MEC supports ultralow latency and delay-sensitive IoT applications in 5G networks [115]. Xiong *et al.* [116] proposed a prototype for MEC-enabled Blockchain for mobile IoT applications. However, the limited MEC resources makes it critical to optimally offload heavy computations to different Edge, Fog, or Cloud resources. Liu *et al.* [117], for example, optimized the joint computation offloading and content caching problems to tackle the intensive computations in PoW consensus algorithm. Xiong *et al.* [118] studied the relationship between cloud or fog providers and PoW-based Blockchain miners with limited computation resources. They chose to offload the computational intensive part of PoW to the cloud and/or fog nodes. The computing nodes offer services to the miners for a given price using a game theoretic approach. The miners can then decide on the amount of service to purchase from the computing nodes. Tuli *et al.* [119] also used Blockchain to provide authentication and encryption services to secure IoT sensitive data and operations. They proposed FogBus, a lightweight end-to-end platform-independent framework for IoT applications. It enables easy deployment, scalability, and cost efficiency by integrating IoT into cloud, fog, and edge computing with the help of Blockchain. Blockchain has been used to provide distributed access control for IoT devices [28]. Almadhoun *et al.* [120] proposed a user authentication system for IoT devices using fog computing. In their proposed system, fog nodes utilize an Ethereum smart contract to authenticate the users and man age access permissions. The proximity of fog nodes to IoT 10 **VII. BLOCKCHAIN & SOFTWARE-DEFINED** **NETWORKING (SDN)** SDN Technology demonstrated its importance in managing routing decisions in smart environment IoT networks since they are usually vulnerable to node/link failures [122]. SDN has been used to mitigate latency issues, like congestion and transmission delays, in time-sensitive smart industrial IoT environments [123]. SDN also balances the load between Fog nodes, which can be vehicles in IoV environments, and the Cloud to allow time-sensitive tasks meet their deadlines [124]. When a Fog is overloaded, SDN dynamically makes offloading decisions to select the best offloading Fog node based on computational and network resource information [125]. Blockchain augments SDN benefits for IoT networks by providing security, privacy, flexibility, scalability, and confidentiality to increase energy utilization and throughput while reducing end-to-end delay [126]. SDN enables network management and protocols to be adaptable and programmable. It was proposed in 2006, in the OpenFlow whitepaper, to test experimental protocols in university campus networks [127]. SDN fits to the rapid changes and demands in network applications, and eliminates the need for pre-programmed, vendor-specific, and expensive network devices. SDN achieves this flexibility by separating the control and data planes in the network. The control plane makes decisions on the traffic flow in the network, whereas the data plane is responsible for forwarding that traffic. Indeed, the control plane is the network brain, and is usually a centralized software entity called SDN controller. The SDN architecture uses the concept of Application Programming Interface (API) in a three-layer structure. Fig. 7 shows two common interfaces between these layers, i.e. ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey **Defines Network Policies** ### Control Plane: Data Plane: **Defines Forwarding Rules** **Performs Packet Forwarding** **FIGURE 7.** The three-layers SDN architecture. Northbound and Southbound APIs. Two additional interfaces are sometimes considered to allow the communication be tween multiple controllers in the control layer, i.e. Eastbound and Westbound APIs [27]. A master controller is usually needed to coordinate the decisions of multiple controllers. A controller is the interface between network elements and applications like firewalls and load balancing applications [128]. It provides agility to network infrastructures, like routers and switches, by dynamically optimizing the network resources. Like other emerging technologies, SDN introduces new security challenges to network infrastructures. At the same time Blockchain can mitigate those challenges by providing confidentiality, integrity, and availability to the network devices [18]. Denial of Service Attacks (DoS) and Distributed Denial of Service Attacks (DDoS) target centralized network architectures. These attacks stop network services from serving legitimate users, devices, and applications. Such attacks can target SDN controllers to malfunction and paralyze the whole network. Blockchain can mitigate these attacks and help avoiding single-points-of-failures in centralized network architectures. Blockchain can provide decentralized trust in physically distributed, logically centralized, SDN controller architectures [18]. Using SDN and Blockchain, network administrators can easily program and configure network components using smart contracts. Those components can securely perform their software updates by accessing policies and configurations from Blockchain-based SDN controllers. Blockchain integration with SDN did not attract the required level of attention yet by the research community [18]. Alharbi [18] explained the role of Blockchain in providing and improving security features in SDN architectures. The dynamism, adaptability, and remote configuration of SDN networks provide a lot of support for IoT networks. Bera *et al.* [27] showed how these features can provide efficient, scalable, seamless, and cost-effective management of IoT devices. SDN also realizes the real-time demands of IoT applications by its ability to optimize traffic flow and load balancing. Such optimization improves the bandwidth utilization in the network and mitigates bottlenecks. Jararweh *et al.* [129] proposed a comprehensive SDNbased IoT framework to simplify IoT management and mitigate several problems in traditional IoT architectures. They integrated software-defined networks, storage, and security into a single software-defined control model for IoT devices. The software-defined storage manages big IoT data by separating the data control layer, which controls storage resources, from the underlying infrastructure of storage assets. Finally, the software-defined security separates the data forwarding plane from the security control plane. They included a proof-of-concept to show the performance of their framework in handling huge amounts of IoT data. The global network view in SDN controllers addresses the heterogeneity, scalability, optimal routing, and bottleneck issues in IoT networks. Kalkan and Zeadally [130] discussed the benefits and drawbacks of SDNs in IoT networks and focused on single-point-of-failure issues in traditional centralized SDN controllers. They separated the roles of single controllers to multiple hosts using a distribution-of-risks scheme. The bandwidth utilization was improved by distributing the communication traffic among three different controllers, i.e. Intrusion, Key, and Crypto Controllers. The intrusion controller mitigates possible intrusions besides managing and securing the routes. The key controller controls symmetric and asymmetric key distribution in the whole ecosystem. The crypto controller provides cryptographic services for authentication, integrity, confidentiality, privacy, and identity management. LI *et al.* [19] focused on the security challenges and solutions for Blockchain-based SDN systems. They focused on DoS and DDoS attacks on centralized SDN controllers, and insider attacks on distributed SDN controllers. They highlighted Blockchain ability to secure distributed SDN controllers and data plane forwarding devices. They also listed scanning, spoofing, hijacking, DoS, and Man-in-themiddle attacks as SDN vulnerabilities. At the end, they listed traffic-flow control, policy enforcement, and DoS defence mechanisms as possible solutions for those vulnerabilities. Blockchain was also used to ensure the security and consistency of the statistics in SDN-based IoT networks [131]. Medhane *et al.* [132] proposed a security framework for next generation IoT by integrating Blockchain with SDN, edge, and cloud computing technologies. The framework features security attack mitigation, continuous confidentiality, authentication, and robustness. The attacks in IoT networks are detected in the cloud and reduced at the edge nodes. SDN controllers examine and manage the traffic flow to actively mitigate doubtful traffic. Similarly, Sharma *et al.* [133] de 11 ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey scribed a distributed cloud architecture at the edge of the network; i.e. Fog nodes, which is secured using Blockchain and SDN. They securely aggregate IoT data using fog nodes before it is sent to the cloud for heavier analysis and/or longterm storage. Their architecture supports large IoT data using low-cost, high-performance, and on-demand secure services. They significantly reduced traffic load and delay compared to traditional IoT architectures. Sharma *et al.* [89] proposed DistBlockNet, a decentralized and secure Blockchain-based SDN architecture that updates flow-rule tables in large-scale IoT networks. Blockchain was used to verify flow-rule tables’ versions, and securely download them to IoT/forwarding devices. Sharma *et al.* [134] also proposed SoftEdgeNet to extend their previous works [89], [133]. SoftEdgeNet improved their previous designs by pushing the storage and computations to the extreme edge to manage real-time traffic and avoid resource starvation. It is a distributed network management architecture for edge computing networks. It mitigates flooding attacks and provides real-time network analytics using Blockchain, SDN, fog, and edge nodes. SoftEdgeNet has an efficient flow-rule allocation and partitioning algorithm at the edge of the network that minimizes traffic redirection, and creates a sustainable network. Blockchain has been also used to secure the configuration, management, and migration of Virtual Network Functions (VNFs) [135]. VNF, also called Network Function Virtualization (NFV), allows devices with adequate resources to perform multiple tasks simultaneously, or at least in realtime. VNF achieves multitasking by separating the control plane from the physical devices [136]. Alvarenga *et al.* [135] implemented a prototype that makes VNF configuration immutable, auditable, nonrepudiable, consistent, and anonymous. Their design eliminates single-points-of-failures and provides high availability of the network’s configuration information with a delay of about two seconds. The architecture is resilient to Blockchain collusion attacks, and the configuration information cannot be compromised even with a successful collusion attack. Such resiliency is achieved by using a variant of the Byzantine Fault Tolerant (BFT) consensus protocol called Ripple protocol [137]. Blockchain was also used in wireless network virtual ization ecosystems to prevent double spending of wireless resources at a given time and location [138], [139]. Wireless network virtualization enables sharing physical wireless infrastructures and radio frequency slices to improve coverage, capacity, and security. Proof-of-Wireless-Resources (PoWR) [139] has been proposed to mitigate double spending of the same wireless resources. Rawat [138] used SDN to provide dynamic and efficient network configuration, and used Edge Computing to decrease delays by avoiding the use of highspeed backhauls. Such fusion of Blockchain with SDN and edge computing guarantees QoS for end users, and provides trust, transparency, and seamless subleasing of resources in trustless wireless networks. The use of Blockchain makes it practically infeasible to create malicious attempts to sublease 12 other’s wireless resources. The dynamic capabilities of SDN-based networks open the doors for many applications, specially in IoT ecosystems. These capabilities can be further enhanced by adding intelligent decision making into their controllers. Such intelligent decisions are now possible to be inferred using AI and machine learning algorithms. With the advent of Deep Neural Networks (DNN), these algorithms can learn on highly complex environments. Recent DNN-based algorithms achieved accuracies that exceed human abilities in different domains. SDN controllers can deploy DNNbased Reinforcement Learning algorithms to create intelligent agents that dynamically adapt to network changes. These capabilities are a must for IoT networks in future smart environments. **VIII. BLOCKCHAIN & ARTIFICIAL INTELLIGENCE (AI)** AI and Machine Learning algorithms play an essential role in adding automation and intelligence into different smart environment applications, including smart cities applications [140]. Besides supporting smart applications, AI also augments various underlying technologies, like optimizing SDN monitoring to minimize network latency [141]. Furthermore, Blockchain empowers AI decision-making by making it more secure and efficient [142]. For example, integrating Blockchain and AI provides decentralized authentication for smart cities [143], where user identities are kept secret while attackers are automatically identified. In smart health systems, Blockchain integration with AI helped secure medical data sharing [144] and protect personal healthcare records [145]. For smart energy trading, Blockchain-enforced Machine Learning predictive analysis models provide real-time support and monitoring as well as immutable transaction logs for decentralized trading [146]. Likewise, Blockchain integration with Machine Learning in smart factories can secure system transactions and deliver smarter quality control schemes [147]. Akter *et al.* [31] defined the ABCD of digital business as AI, Blockchain, Cloud, and Data analytic. They considered these emerging technologies as transformation factors for future digital business models. For successful digital business, Garcia [29] proposed a complete legal doctrine for smart digital economy. Technologies like Blockchain, AI, IoT, and big data guide governments’ annual budget plans to simplify and maximize the application of taxes for digital businesses. Garcia also showed that Blockchain and cryptocurrencies augments AI and IoT to create smart economy in smart digital world. Ekramifard *et al.* [30] studied how AI algorithms improve Blockchain designs and operations. They discussed the effect of this integration in the medical field, like the ability to gather, analyse, and make decisions on medical datasets. AI and Blockchain helped in different medical applications, including systems for the ongoing COVID-19 pandemic [148], [149]. Mashamba-Thompson and Crayton [149] integrated Blockchain and AI to create a low cost self testing and tracking system for COVID-19. Their system ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey is ideal in environments with poor access to laboratory infrastructures. Similarly, Nguyen *et al.* [148] discussed how Blockchain and AI were used in the literature to compact the COVID-19 pandemic. Al-Garadi *et al.* [26] discussed the role of machine learning and deep learning in IoT security. They suggested integrating Edge computing and Blockchain into machine learning and deep learning to provide reliable and effective IoT security methods. In their work, they investigated the use of Neural Networks (NN) and other machine learning algorithms (see Fig. 8) to detect attacks in IoT networks. Based on IoT and network data, IoT network state is classified into normal (secure), early warning, or attacked state. Fig. 8 shows a brief taxonomy of different algorithms in the field of AI that can be adopted in IoT ecosystems. Those algorithms can be classified into classical machine learning algorithms and algorithms based on DNN. Each of those classes can be further categorized into supervised, unsupervised, and semisupervised methods. The supervision in learning algorithms means introducing training labels with training examples that were prepared by professional or computer software. It is sometimes impractical or hard to create labeled datasets; in this case, we can use unsupervised algorithms to categorize and cluster the data into different groups. Semi-supervised algorithms lay between those two classes, where only a small portion of the data is labeled, whereas there are no labels for the rest of the data. AI integration is essential to provide smart decisionmaking capabilities into different technologies in a smart environment ecosystem. The breakthroughs in machine learning and Deep Learning algorithms make them suitable for solving complex problems in rapidly changing environments, such as IoT networks. AI is important to enhance the performance of technologies, like Blockchain, IoT, SDN, Cloud, Fog, and Edge Computing. Self-driving vehicles, smart transportation, automatic delivery robots are some examples of smart environment applications that need AI integration into Blockchain-IoT solutions. AI can also be used to optimize the global energy consumption in a smarter and greener world to decrease the effect of climate change and local air pollution. Kumari *et al.* [150], for example, studied the advantages and challenges of integrating Blockchain with AI in Energy Cloud Management (ECM) systems. Using IoT and Smart Grids (SG), this integration allows for sustainable energy management and efficient load prediction in a trustless environment (see Fig. 9). They also proposed a decentralized Blockchain-based AI-powered ECM framework for energy management to mitigate security and privacy issues in traditional implementations. AI can provide best pricing for IoT data to be sold and/or computations to be performed. AI can also empower SDN controllers to choose optimal network routes to forward data traffic. AI-based SDN traffic control can minimize network delay and bandwidth consumption. It is challenging to optimally offload computations between end-devices and the Cloud, Fog, and/or Edge servers jointly considering **FIGURE 8.** A taxonomy of Artificial Intelligence algorithms for IoT ecosystems. delay, computations, and power resources. AI algorithms, like Deep Reinforcement Learning (DeepRL), have been recently considered for task offloading and orchestration in edge computing applications [151], [152]. Dai *et al.* [151] used DeepRL to dynamically orchestrate edge computing and caching resources in complex vehicular networks. The complexity of such networks comes from vehicles mobility and content popularity/localization, e.g. a car accident at a given location. Furthermore, Dai *et al.* [152] proposed to integrate AI and Blockchain to provide intelligent architectures for flexible and secure resource sharing and content caching in 5G networks. 13 ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey selection algorithm to increase trust, provide protection, and save time and cost in automotive supply chain ecosystems that are owned by different organizations. **FIGURE 9.** Smart energy management for sustainable energy and green smart environments. Qiu *et al.* [153] provided trust in SDN Industrial IoT networks using Blockchain consensus protocols. In their design, Blockchain collects, synchronizes, and distributes network views between distributed SDN controllers. To im prove the throughput, they used a Dueling Deep Q-Learning (DDQL) approach to jointly optimize view changes, access selection, and computational resource allocation. Deep QLearning (DQL) was also used in Distributed SoftwareDefined Vehicular Network (DSDVN) to adapt to the variety of data, network-flow, and vehicle types [154]. To increase the throughput of a permissioned Blockchain, consensus schemes were used to reach consensus efficiently and securely in DSDVN using DQL. It optimizes compute and network resources by jointly considering the trust features of Blockchain nodes to improve the throughput. Blockchain integration with all those technologies was a major factor in the deployment of smart environment applications. Sharma *et al.* [155], for example, used Blockchain to allow for decentralized coordination and control for vehicular networks in smart cities. Moreover, Sharma and Park [156] integrated Blockchain and SDN to provide a two-layer network architecture that was specially designed for smart cities. It is composed of core and edge networks, and leverages the benefits of both centralized and distributed architectures. Their design supports IoT heterogeneity and provides a scalable and secure architecture using edge computing. They used a memory-hardened PoW scheme to enforce distributed privacy and security, and to avoid tampering of information by attackers. Sharma *et al.* [157] used a private version of Ethereum to simulate a distributed framework for automotive industries in smart cities. They proposed a novel miner node 14 **IX. OPEN RESEARCH PROBLEMS** A lot of work is still needed to allow for smoother integration between Blockchain and IoT technologies to create smarter things. Law and regulation issues are some of the main problems that are discussed in the literature for using Blockchain for basic monetary transactions. Hence, these issues will directly impact machine-to-machine monetary transactions using Blockchain. Akins *et al.* [67] discussed income taxation of cryptocurrency transactions, such as Bitcoins, when used for purchases with monetary value. Sapovadia [158] discussed the legal issues in cryptocurrencies that are similar to those of foreign currencies. Emelianova and Dementyev [159] argued for a unified supranational legal act for cryptocurrencies, similar to the European Union (EU) directives. They discussed the provisions in many European and Asian countries for the use and taxation of cryptocurrencies, which differs from one government to another. Omololu [160] emphasized the need for full law enforcement of Blockchain applications as they still do not conform to current legal structures. This requires countries to supervise Blockchain integration in different applications and domains, including IoT, to ensure that they comply with the law. To create an IoT-oriented Blockchain platform, both hardware and software should be highly optimized to perform Blockchain operations. IBM has taken the lead in this path by creating a 10-cents tiny edge CPU architecture that can efficiently run Blockchain operations [161]. It can be embedded into IoT devices to support Blockchain operations. IBM called this project "Crypto Anchors", as they want to anchor physical objects into Blockchain-IoT applications. However, the difference between such CPU architectures and standard Computer CPUs requires specific Blockchain implementations for Crypto Anchor CPUs. Hence, we believe that this is a good research direction to create an operating system for those CPUs that is capable of bridging this gap, such as the work by Wright and Savanah [162]. Building Blockchain-oriented CPUs, firmware, and operating systems is a great step towards more robust Blockchain-IoT integration. However, we believe that the research should also continue on improving Blockchain architectures, including consensus algorithms and communication protocols. In terms of consensus algorithms, there has been an effort to replace the most secure consensus algorithm in practice, i.e. PoW. The main reason is the power consumption of PoW, which does not fit resource-limited IoT devices. PoS [66] and DPoS [54] are two promising alternatives, but they are still criticized for not being as secure as PoW [163]. Ethereum plans to migrate from using PoW to PoS [164], because it is currently the best alternative for PoW [165]. DPoS is currently adopted in EOS [166] and few other Blockchain implementations. There is a strong debate around the level of decentralization in DPoS and PoW [167]. To show the ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey difference, Li and Palanisamy [167] studied a DPoS-based cryptocurrency for social media, called Steem [168], and the PoW-based Bitcoin Blockchain. They showed that Bitcoin is more decentralized compared to Steem among top miners, but less decentralized in general due to Bitcoin mining pools. Consensus algorithms are a major factor in determining Blockchain performance, and that is why researchers try to increase the security and lower power consumption of these algorithms [169], [170]. Even lightweight Blockchain platforms that are built specifically for resource-limited IoT applications have some drawbacks. DAG-based platforms for example, like IoTA, suffer from double-spending attacks. Hence, there is always a security vs. performance tradeoff when choosing between different Blockchain platforms or consensus algorithms. Such trade-off should be carefully selected to meet the different requirements of different IoT applications. Pongnumkul *et al.* [53], for example, studied the trade-off between choosing the most secure PoW vs. the fastest PBFT consensus algorithms in Ethereum and Hyperledger Fabric, respectively. However, it is sometimes necessary to compare different Blockchain platforms using factors other than consensus algorithms. Developers can falsify Blockchain performance, and attract investors only based on the consensus performance. However, choosing a Blockchain platform based solely on the consensus performance can badly affect its performance in large scale IoT networks or in smart city applications. Hence, Zheng *et al.* [10] argued for Blockchain testing schemes to help practitioners select Blockchain platforms that fit the requirements of different IoT applications. They also discussed the drawbacks of mining pools in public Blockchain implementations, which can cause loss of decentralization. Another Blockchain related problem in IoT applications are Smart Contract bugs. These bugs shorten the life-cycle of that code and loosen its agility in the ever changing IoT networks. Furthermore, using Oracles [171], which are trusted third party information sources, exposes Blockchain to lose its inherited security. Oracles provide external data and information to Blockchain Smart Contracts to enrich the capabilities of Blockchain applications, including IoT applications. Oracles query, verify, and authenticate external data sources to provide trust for such sources. Blockchain should trust oracles, since they make decisions based on the data they provide. However, the trust issues of Oracles will directly impact Blockchain security that was meant to work in trustless environments. Oracles might still suffer from centralization, collusion, Sybil, and Man-in-the-middle attacks [172]. They are also possibly exposed to physical attacks on IoT devices that are usually the source for external Blockchain data. An example of physical attacks in IoT food or drug supply chain systems is the displacement of temperature or GPS sensors that are usually attached on supply chain shipping trucks. The displacement of IoT devices feeds Blockchain ecosystems with falsified information, which causes those systems to loose their security features, and hence to fail. 51% attacks, cost, regulations, confirmation time, forks, and scalability are common technical Blockchain-related problems [173]. Scalability, for example, has been discussed a lot in literature, and different solutions have been proposed [64]. In addition, there are also domain-specific challenges, such as the challenges for using Blockchain for AI [68], Security [68], Healthcare [174], Education [175], Product Traceability [176], E-Voting [73], [74], [177], [178], and IoT applications. There is a need to reduce Blockchain energy consumption and operation cost to make it feasible to integrate with various technologies, including IoT. However, security, privacy, scalability, and Oracle inherited Blockchain issues need to be addressed before looking for integration issues [20]. Finally, real deployment of Blockchain-IoT solutions in smart city prototypes might reveal new issues compared to what simulation results currently demonstrate. To focus more on Blockchain-IoT integration challenges, Makhdoom *et al.* [14] used a test case for a supply chain monitoring system. The challenges include the lack of IoTcentric consensus protocols, IoT-based transaction validation rules, IoT-oriented Blockchain interfaces, and storage capacity for IoT data. They added other Blockchain-related challenges, like consensus finality, resistance to DoS attacks, fault tolerance, scalability, and transaction volume. The test case required a secure and synchronized software upgrade scheme for IoT devices and the underlying Blockchain platform. The upgrade scheme can fix bugs and protect the system against new vulnerabilities. In addition, we need to be careful when integrating SDN and Blockchain technologies to fully benefit from SDN features for IoT applications. Blockchain is decentralized by nature, whereas SDN controllers are supposed to be centralized. Moreover, SDN controllers need to update the flow-rule tables to the forwarding devices in real-time, while Blockchain consensus works periodically on a longer time frame. The challenges brought by integrating Blockchain with Cloud, Fog, and Edge Computing directly relate to IoTBlockchain challenges. That is because these technologies are mainly meant to support IoT networks. The lack for perfectly implemented Blockchain-specific IoT infrastructures, and the absence of energy-efficient mining are some of those challenges [13]. Authentication, adaptability, network security, data integrity, verifiable computation, and low latency are requirements for integrating Blockchain with Cloud, Fog, and specially Edge Computing [16]. Yang *et al.* [16] identified load balancing, task offloading, resource management and function integration on heterogeneous platforms as challenges to be addressed for successful integration. Addressing all these issues is essential to support next generation applications in fully automated futuristic smart environments. In such environments, all these technologies should be smoothly integrated, flawlessly functioning, and securely handling IoT data. 15 ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey **X. DISCUSSION AND CONCLUSION** In this paper, we present the strengths of Blockchain beyond its traditional use for monetary and digital asset trading. We focus on Blockchain integration with IoT to create a futuristic view of smart environments. Implementing such autonomous smart environment architectures will simplify human lives and increase their effectiveness. We discuss the role of AI, SDN, Cloud, Fog, and Edge Computing in enhancing the capabilities of Blockchain-IoT applications, and providing such automation in smart environments. Blockchain augments IoT applications with automation, security, privacy and many other features that are essential for smart environments. We showed in this work how Blockchain was able to address a number of issues, limitations, and challenges in all those technologies. Table 2 shows the level of technological integration of some Blockchain-IoT applications and prototypes in the literature. The table also shows the recent research interest to integrate more technologies into such systems. Powered by Blockchain-IoT integration, these prototypes served different needs and solved different problems in different smart applications. To build those systems, some of the authors had to create their own Blockchain implementation and/or its consensus algorithm. Specific application requirements usually ask for different Blockchain characteristics that are not usually available in traditional implementations. Hence, there is a need for a Blockchain implementation with the possibility to plug in new features when needed by developers and practitioners. This will allow the industry and the research community to focus more on developing more smart applications, and mitigating different technological limitations. Figure 10 shows a simplified architecture for smart environments using the technologies discussed in this paper. It shows the use of Blockchain to securely share and store IoT data in trustless environments. In this architecture, Blockchain can be deployed using Cloud, Fog, or Edge resources, or even a combination of them. The AI-powered SDN traffic control can manage the traffic flow of IoT data in a dynamic smart way. AI can also be used to provide best pricing for IoT data that needs to be sold in such a system, like sensor data. It can be also used to provide the best target to process this data using the Cloud, Fog, or Edge computing. In addition, private data can be securely stored in Blockchain, and can be obtained by authentic users using different encryption and security measures. Blockchain will work as an access control mechanism for data access, including IoT data, for IoT devices and their users. There are a number of challenges and open research problems that still need to be tackled to provide a smoother technological integration. A major problem of such integration is the difficulty of testing using physical deployment in real smart environments in order to reveal hidden issues. Physical deployment is needed because simulation results alone are insufficient to demonstrate the performance and issues in such systems. In addition, creating a general purpose 16 **AI** -based data pricing **FIGURE 10.** Architecture of Smart Envrinoments. Blockchain platform that can be easily adapted to solve different problems is of a great importance for the research community. Creating such a general purpose platform will remove the burden that is imposed by modifying Blockchain architectures to meet certain features and requirements. **REFERENCES** [1] U. Ghosh, P. Chatterjee, S. Shetty, and R. Datta, “An sdn-iot-based framework for future smart cities: Addressing perspective,” 2020. [2] M. Conoscenti, A. Vetrò, and J. C. De Martin, “Blockchain for the internet of things: A systematic literature review,” in 2016 IEEE/ACS 13th International Conference of Computer Systems and Applications (AICCSA), Nov 2016, pp. 1–6. [3] K. Christidis and M. Devetsikiotis, “Blockchains and smart contracts for the internet of things,” IEEE Access, vol. 4, pp. 2292–2303, 2016. [4] A. Reyna, C. Martín, J. Chen, E. Soler, and M. Díaz, “On blockchain and its integration with IoT. challenges and opportunities,” Future Generation Computer Systems, vol. 88, pp. 173 – 190, 2018. [Online]. Available: [http://www.sciencedirect.com/science/article/pii/S0167739X17329205](http://www.sciencedirect.com/science/article/pii/S0167739X17329205) [5] G. S. Ramachandran and B. Krishnamachari, “Blockchain for the IoT: Opportunities and challenges,” CoRR, vol. abs/1805.02818, 2018. [[Online]. Available: http://arxiv.org/abs/1805.02818](http://arxiv.org/abs/1805.02818) [6] A. Panarello, N. Tapas, G. Merlino, F. Longo, and A. Puliafito, “Blockchain and IoT integration: A systematic survey,” Sensors, vol. 18, [no. 8, 2018. [Online]. Available: https://www.mdpi.com/1424-8220/18/](https://www.mdpi.com/1424-8220/18/8/2575) [8/2575](https://www.mdpi.com/1424-8220/18/8/2575) [7] T. M. Fernández-Caramés and P. Fraga-Lamas, “A review on the use of blockchain for the internet of things,” IEEE Access, vol. 6, pp. 32 979– 33 001, 2018. [8] M. Banerjee, J. Lee, and K.-K. R. Choo, “A blockchain future for internet of things security: a position paper,” Digital Communications and Networks, vol. 4, no. 3, pp. 149 – 160, 2018. [Online]. Available: [http://www.sciencedirect.com/science/article/pii/S2352864817302900](http://www.sciencedirect.com/science/article/pii/S2352864817302900) [9] H. F. Atlam, A. Alenezi, M. O. Alassafi, and G. Wills, “Blockchain with internet of things: Benefits, challenges, and future directions,” International Journal of Intelligent Systems and Applications, vol. 10, [no. 6, pp. 40–48, June 2018. [Online]. Available: https://eprints.soton.ac.](https://eprints.soton.ac.uk/421529/) [uk/421529/](https://eprints.soton.ac.uk/421529/) [10] Z. Zheng, S. Xie, H.-N. Dai, X. Chen, and H. Wang, “Blockchain challenges and opportunities: A survey,” International Journal of Web and Grid Services, vol. 14, no. 4, pp. 352–375, 2018. ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey **TABLE 2.** Technological integration of Cloud, Edge, Fog, SDN, and AI technologies into Blockchain-IoT prototypes. |Paper|Year|Integration with Edge/Fog Cloud SDN AI|Platform|Functionality|Mining|Type|Transactions| |---|---|---|---|---|---|---|---| |[97]|2014||Bitcoin|Data Market|PoW|Public|UTXOs| |[82]+|2015||Bitcoin|Data Market|PoW|Public|UTXOs & Scripts| |[179]|2015||Bitcoin6|Decentralized Data Management|PoW|Public|Data & Queries| |[87]|2016||Ethereum|D2D Communication & Marketplace|PoW|Private|Smart Contracts| |[92]+|2016||Bitcoin6|Smart Homes||Public|IoT Data| |[98]|2017||NameCoin|Privacy-Aware PKI|PoW|Public|Digital Signatures| |[94]+|2017||Bitcoin6|Smart Homes||Private|IoT Data| |[88]|2017||Ethereum|D2D Communication|PoW|Public|Smart Contracts| |[112]|2017||Ethereum|D2D Communication|PoW|Public|Smart Contracts| |[89]+|2017|||Decentralized SDN (DSDN)|PoW||| |[96]|2017||MultiChain|D2D Marketplace|Round Robin|Private|Digital Assets| |[133]+|2018|  ||Distributed Cloud|PoSer3|Private|Smart Contracts1| |[120]|2018| |Ethereum|Authentication|PoW|Public|Smart Contracts| |[117]|2018||Wireless BC7|Offloading PoW Computations|PoW||| |[139]|2018||Bitcoin6|Wireless Network Virtualization|PoWR5|Public|Network Data| |[154]|2018|  |Fabric|Vehicular Networks|PBFT|Private|Smart Contracts1| |[134]+|2018|  ||DSDN for Edge Computing|PoW||| |[156]+|2018| |Ethereum|Smart Cities|PoW|Private|IoT Data Hashes| |[116]|2018| |Ethereum|Edge Computing Mining|PoW|Private|IoT Data| |[152]+|2019|  ||Intelligent Wireless Networks|PBFT|Consortium|Network Data| |[153]|2019|  ||DSDN Consensus|PBFT|Private|Traffic/Control Data| |[138]|2019| |Bitcoin6|Wireless Network Virtualization||Public|Network Data| |[157]|2019||Ethereum|Supply Chain Management|PoW|Private|Smart Contracts1| |[118]|2019| |Ethereum|Offloading PoW Computations|PoW|Private|Random Data| |[119]|2019| |Bitcoin6|Data Privacy and Integrity|PoW||IoT Data| |[131]|2020|  ||Securing Traffic Measurements|||Network Statistics| |[132]+|2020|  ||Distributed Network Security|||IoT & Network Data| |[109]|2020||Their Own|Secure Cloud Scheduler|PoSch2|Public|Task Scheduling Info| |[121]|2020|  |Their Own|Secure Fog Computing|Cooperative||Access Control Lists| |[180]|2020| |Bitcoin6|DSDN Consensus|BFT4|Private|Traffic/Control Data| 1 Contract execution beside monetary and data transactions. 2 Proof-of-Schedule. 3 Proof-of-Service. 4 Aardvark, RBFT, & PBFT. 5 Proof-of-Wireless-Resources. 6 A different Bitcoin implementation based on the same concepts and architecture. 7 A simulation for a wireless Blockchain implementation. + The design or prototype has a hierarchical or multilayer architecture. **Blank Cells:** Information was not provided in the corresponding paper. [11] M. S. Ali, M. Vecchio, M. Pincheira, K. Dolui, F. Antonelli, and M. H. Rehmani, “Applications of blockchains in the internet of things: A comprehensive survey,” IEEE Communications Surveys Tutorials, vol. 21, no. 2, pp. 1676–1717, Secondquarter 2019. [12] H. Dai, Z. Zheng, and Y. Zhang, “Blockchain for internet of things: A survey,” CoRR, vol. abs/1906.00245, 2019. [Online]. Available: [http://arxiv.org/abs/1906.00245](http://arxiv.org/abs/1906.00245) [13] M. A. Ferrag, M. Derdour, M. Mukherjee, A. Derhab, L. Maglaras, and H. Janicke, “Blockchain technologies for the internet of things: Research issues and challenges,” IEEE Internet of Things Journal, vol. 6, no. 2, pp. 2188–2204, 2019. [14] I. Makhdoom, M. Abolhasan, H. Abbas, and W. Ni, “Blockchain’s adoption in iot: The challenges, and a way forward,” Journal of Network and Computer Applications, vol. 125, pp. 251 – 279, [2019. [Online]. Available: http://www.sciencedirect.com/science/article/](http://www.sciencedirect.com/science/article/pii/S1084804518303473) [pii/S1084804518303473](http://www.sciencedirect.com/science/article/pii/S1084804518303473) [15] K. Salah, M. H. U. Rehman, N. Nizamuddin, and A. Al-Fuqaha, “Blockchain for AI: Review and open research challenges,” IEEE Access, vol. 7, pp. 10 127–10 149, 2019. [16] R. Yang, F. R. Yu, P. Si, Z. Yang, and Y. Zhang, “Integrated blockchain and edge computing systems: A survey, some research issues and challenges,” IEEE Communications Surveys Tutorials, vol. 21, no. 2, pp. 1508–1532, 2019. [17] L. Lao, Z. Li, S. Hou, B. Xiao, S. Guo, and Y. Yang, “A survey of IoT applications in blockchain systems: Architecture, consensus, and traffic modeling,” ACM Comput. Surv., vol. 53, no. 1, Feb. 2020. [Online]. [Available: https://doi.org/10.1145/3372136](https://doi.org/10.1145/3372136) [18] T. Alharbi, “Deployment of blockchain technology in software defined networks: A survey,” IEEE Access, vol. 8, pp. 9146–9156, 2020. [19] W. LI, W. MENG, Z. LIU, and M.-H. AU, “Towards blockchain-based software-defined networking: Security challenges and solutions,” IEICE Transactions on Information and Systems, vol. E103.D, no. 2, pp. 196– 203, 2020. [20] C. Luo, L. Xu, D. Li, and W. Wu, “Edge computing integrated with blockchain technologies,” in Complexity and Approximation: In Memory of Ker-I Ko, D.-Z. Du and J. Wang, Eds. Cham: Springer International Publishing, 2020, pp. 268–288. [Online]. Available: [https://doi.org/10.1007/978-3-030-41672-0_17](https://doi.org/10.1007/978-3-030-41672-0_17) [21] S. Xie, Z. Zheng, W. Chen, J. Wu, H.-N. Dai, and M. Imran, “Blockchain for cloud exchange: A survey,” Computers and Electrical [Engineering, vol. 81, p. 106526, 2020. [Online]. Available: http:](http://www.sciencedirect.com/science/article/pii/S0045790618332750) [//www.sciencedirect.com/science/article/pii/S0045790618332750](http://www.sciencedirect.com/science/article/pii/S0045790618332750) [22] B. K. Mohanta, D. Jena, U. Satapathy, and S. Patnaik, “Survey on IoT security: Challenges and solution using machine learning, artificial intelligence and blockchain technology,” Internet of Things, vol. 11, [p. 100227, 2020. [Online]. Available: http://www.sciencedirect.com/](http://www.sciencedirect.com/science/article/pii/S2542660520300603) [science/article/pii/S2542660520300603](http://www.sciencedirect.com/science/article/pii/S2542660520300603) [23] V. Chamola, V. Hassija, V. Gupta, and M. Guizani, “A comprehensive review of the covid-19 pandemic and the role of iot, drones, ai, blockchain, and 5g in managing its impact,” IEEE Access, vol. 8, pp. 90 225–90 265, 2020. [24] B. L. R. Stojkoska and K. V. Trivodaliev, “A review of internet of things for smart home: Challenges and solutions,” Journal of Cleaner Production, vol. 140, pp. 1454 – 1464, 2017. [Online]. Available: [http://www.sciencedirect.com/science/article/pii/S095965261631589X](http://www.sciencedirect.com/science/article/pii/S095965261631589X) [25] B. Bhushan, A. Khamparia, K. M. Sagayam, S. K. Sharma, 17 ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey M. A. Ahad, and N. C. Debnath, “Blockchain for smart cities: A review of architectures, integration trends and future research directions,” Sustainable Cities and Society, vol. 61, p. 102360, [2020. [Online]. Available: http://www.sciencedirect.com/science/article/](http://www.sciencedirect.com/science/article/pii/S2210670720305813) [pii/S2210670720305813](http://www.sciencedirect.com/science/article/pii/S2210670720305813) [26] M. A. Al-Garadi, A. Mohamed, A. Al-Ali, X. Du, I. Ali, and M. Guizani, “A survey of machine and deep learning methods for internet of things (iot) security,” IEEE Communications Surveys Tutorials, pp. 1–1, 2020. [27] S. Bera, S. Misra, and A. V. Vasilakos, “Software-defined networking for internet of things: A survey,” IEEE Internet of Things Journal, vol. 4, no. 6, pp. 1994–2008, Dec 2017. [28] I. Riabi, H. K. B. Ayed, and L. A. Saidane, “A survey on blockchain based access control for internet of things,” in 2019 15th International Wireless Communications Mobile Computing Conference (IWCMC), 2019, pp. 502–507. [29] A. R. Garcia, AI, IoT, Big Data, and Technologies in Digital Economy with Blockchain at Sustainable Work Satisfaction to Smart Mankind: Ac cess to 6th Dimension of Human Rights. Cham: Springer International Publishing, 2020, pp. 83–131. [30] A. Ekramifard, H. Amintoosi, A. H. Seno, A. Dehghantanha, and R. M. Parizi, A Systematic Literature Review of Integration of Blockchain and Artificial Intelligence. Cham: Springer International Publishing, 2020, pp. 147–160. [31] S. Akter, K. Michael, M. R. Uddin, G. McCarthy, and M. Rahman, “Transforming business using digital innovations: the application of ai, blockchain, cloud and data analytics,” Annals of Operations [Research, May 2020. [Online]. Available: https://doi.org/10.1007/](https://doi.org/10.1007/s10479-020-03620-w) [s10479-020-03620-w](https://doi.org/10.1007/s10479-020-03620-w) [32] H. Baniata and A. Kertesz, “A survey on blockchain-fog integration approaches,” IEEE Access, vol. 8, pp. 102 657–102 668, 2020. [33] Y. Hajjaji, W. Boulila, I. R. Farah, I. Romdhani, and A. Hussain, “Big data and iot-based applications in smart environments: A systematic review,” Computer Science Review, vol. 39, p. 100318, 2021. [Online]. Available: [https://www.sciencedirect.com/science/article/pii/S1574013720304184](https://www.sciencedirect.com/science/article/pii/S1574013720304184) [34] D. M. El-Din, A. E. Hassanein, and E. E. Hassanien, Smart Environments Concepts, Applications, and Challenges. Cham: Springer International Publishing, 2021, pp. 493–519. [Online]. Available: [https://doi.org/10.1007/978-3-030-59338-4_24](https://doi.org/10.1007/978-3-030-59338-4_24) [35] H. S. Woods, “Smart homes: domestic futurity as infrastructure,” Cultural Studies, vol. 0, no. 0, pp. 1–24, 2021. [Online]. Available: [https://doi.org/10.1080/09502386.2021.1895254](https://doi.org/10.1080/09502386.2021.1895254) [36] A. Molnar, “Smart cities education: An insight into existing drawbacks,” Telematics and Informatics, vol. 57, p. 101509, 2021. [Online]. Available: [https://www.sciencedirect.com/science/article/pii/S0736585320301684](https://www.sciencedirect.com/science/article/pii/S0736585320301684) [37] U. Majeed, L. U. Khan, I. Yaqoob, S. A. Kazmi, K. Salah, and C. S. Hong, “Blockchain for iot-based smart cities: Recent advances, requirements, and future challenges,” Journal of Network and Computer Applications, vol. 181, p. 103007, 2021. [Online]. Available: [https://www.sciencedirect.com/science/article/pii/S1084804521000345](https://www.sciencedirect.com/science/article/pii/S1084804521000345) [38] T. Roy, A. Tariq, and S. Dey, “A socio-technical approach for resilient connected transportation systems in smart cities,” IEEE Transactions on Intelligent Transportation Systems, pp. 1–10, 2021. [39] B. Xu and G. Thakur, “Introduction to the special issue on smart transportation,” GeoInformatica, Mar 2021. [Online]. Available: [https://doi.org/10.1007/s10707-021-00432-3](https://doi.org/10.1007/s10707-021-00432-3) [40] X.-F. Shao, W. Liu, Y. Li, H. R. Chaudhry, and X.-G. Yue, “Multistage implementation framework for smart supply chain management under industry 4.0,” Technological Forecasting and Social Change, vol. 162, [p. 120354, 2021. [Online]. Available: https://www.sciencedirect.com/](https://www.sciencedirect.com/science/article/pii/S004016252031180X) [science/article/pii/S004016252031180X](https://www.sciencedirect.com/science/article/pii/S004016252031180X) [41] J. Lee, Y. C. Lee, and J. T. Kim, “Migration from the traditional to the smart factory in the die-casting industry: Novel process data acquisition and fault detection based on artificial neural network,” Journal of Materials Processing Technology, vol. 290, p. 116972, 2021. [[Online]. Available: https://www.sciencedirect.com/science/article/pii/](https://www.sciencedirect.com/science/article/pii/S0924013620303939) [S0924013620303939](https://www.sciencedirect.com/science/article/pii/S0924013620303939) [42] M. Pech, J. Vrchota, and J. Bednáˇr, “Predictive maintenance and intelligent sensors in smart factory: Review,” Sensors, vol. 21, no. 4, [2021. [Online]. Available: https://www.mdpi.com/1424-8220/21/4/1470](https://www.mdpi.com/1424-8220/21/4/1470) [43] Z. Guan, X. Lu, W. Yang, L. Wu, N. Wang, and Z. Zhang, “Achieving efficient and privacy-preserving energy trading based on blockchain and abe in smart grid,” Journal of Parallel and Distributed Computing, vol. 147, pp. 34–45, 2021. [Online]. Available: [https://www.sciencedirect.com/science/article/pii/S0743731520303609](https://www.sciencedirect.com/science/article/pii/S0743731520303609) 18 [44] J. Sahoo and K. Barrett, “Internet of things (iot) application model for smart farming,” 2021. [45] E. M. Ouafiq, A. Elrharras, A. Mehdary, A. Chehri, R. Saadane, and M. Wahbi, “Iot in smart farming analytics, big data based architecture,” in Human Centred Intelligent Systems, A. Zimmermann, R. J. Howlett, and L. C. Jain, Eds. Singapore: Springer Singapore, 2021, pp. 269–279. [46] N. G. Rezk, E. E.-D. Hemdan, A.-F. Attia, A. El-Sayed, and M. A. El-Rashidy, “An efficient iot based smart farming system using machine learning algorithms,” Multimedia Tools and Applications, [vol. 80, no. 1, pp. 773–797, Jan 2021. [Online]. Available: https:](https://doi.org/10.1007/s11042-020-09740-6) [//doi.org/10.1007/s11042-020-09740-6](https://doi.org/10.1007/s11042-020-09740-6) [47] G. Madaan, B. Bhushan, and R. Kumar, Blockchain-Based Cyberthreat Mitigation Systems for Smart Vehicles and Industrial Automation. Singapore: Springer Singapore, 2021, pp. 13–32. [Online]. Available: [https://doi.org/10.1007/978-981-15-7965-3_2](https://doi.org/10.1007/978-981-15-7965-3_2) [48] S. Nakamoto et al., “Bitcoin: A peer-to-peer electronic cash system,” [2008. [Online]. Available: https://bitcoin.org/bitcoin.pdf](https://bitcoin.org/bitcoin.pdf) [49] C. Esposito, M. Ficco, and B. B. Gupta, “Blockchain-based authentication and authorization for smart city applications,” Information Processing & Management, vol. 58, no. 2, p. 102468, 2021. [[Online]. Available: https://www.sciencedirect.com/science/article/pii/](https://www.sciencedirect.com/science/article/pii/S0306457320309584) [S0306457320309584](https://www.sciencedirect.com/science/article/pii/S0306457320309584) [50] C. Oham, R. A. Michelin, R. Jurdak, S. S. Kanhere, and S. Jha, “B-ferl: Blockchain based framework for securing smart vehicles,” Information Processing & Management, vol. 58, no. 1, p. 102426, 2021. [[Online]. Available: https://www.sciencedirect.com/science/article/pii/](https://www.sciencedirect.com/science/article/pii/S0306457320309183) [S0306457320309183](https://www.sciencedirect.com/science/article/pii/S0306457320309183) [51] A. Hasankhani, S. Mehdi Hakimi, M. Bisheh-Niasar, M. Shafie-khah, and H. Asadolahi, “Blockchain technology in the future smart grids: A comprehensive review and frameworks,” International Journal of Electrical Power & Energy Systems, vol. 129, p. 106811, 2021. [[Online]. Available: https://www.sciencedirect.com/science/article/pii/](https://www.sciencedirect.com/science/article/pii/S014206152100051X) [S014206152100051X](https://www.sciencedirect.com/science/article/pii/S014206152100051X) [52] C. Cachin, “Architecture of the hyperledger blockchain fabric,” in Workshop on Distributed Cryptocurrencies and Consensus Ledgers, vol. 310. IBM Research, Zurich, 2016. [53] S. Pongnumkul, C. Siripanpornchana, and S. Thajchayapong, “Performance analysis of private blockchain platforms in varying workloads,” in 2017 26th International Conference on Computer Communication and Networks (ICCCN), July 2017, pp. 1–6. [54] D. Larimer, “Delegated proof-of-stake (dpos),” Bitshare whitepaper, 2014. [55] M. Castro, B. Liskov et al., “Practical byzantine fault tolerance,” in OSDI, vol. 99, 1999, pp. 173–186. [56] D. Kempe and J. Kleinberg, “Protocols and impossibility results for gossip-based communication mechanisms,” in The 43rd Annual IEEE Symposium on Foundations of Computer Science, 2002. Proceedings., Nov 2002, pp. 471–480. [57] P. Maymounkov and D. Mazières, “Kademlia: A peer-to-peer information system based on the xor metric,” in Peer-to-Peer Systems, P. Druschel, F. Kaashoek, and A. Rowstron, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2002, pp. 53–65. [58] E. Ben-Sasson, A. Chiesa, D. Genkin, E. Tromer, and M. Virza, “Snarks for c: Verifying program executions succinctly and in zero knowledge,” in Advances in Cryptology – CRYPTO 2013, R. Canetti and J. A. Garay, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2013, pp. 90–108. [59] W. Ou, M. Deng, and E. Luo, “A decentralized and anonymous data transaction scheme based on blockchain and zero-knowledge proof in vehicle networking (workshop paper),” in Collaborative Computing: Networking, Applications and Worksharing, X. Wang, H. Gao, M. Iqbal, and G. Min, Eds. Cham: Springer International Publishing, 2019, pp. 712–726. [60] Z. Peng, H. Wu, B. Xiao, and S. Guo, “Vql: Providing query efficiency and data authenticity in blockchain systems,” in 2019 IEEE 35th International Conference on Data Engineering Workshops (ICDEW), April 2019, pp. 1–6. [61] V. Buterin et al., “A next-generation smart contract and decentralized application platform,” white paper, vol. 3, no. 37, 2014. [[62] N. Szabo, “Smart contracts,” 1994. [Online]. Available: http://www.](http://www.fon.hum.uva.nl/rob/Courses/InformationInSpeech/CDROM/Literature/LOTwinterschool2006/szabo.best.vwh.net/smart.contracts.html) [fon.hum.uva.nl/rob/Courses/InformationInSpeech/CDROM/Literature/](http://www.fon.hum.uva.nl/rob/Courses/InformationInSpeech/CDROM/Literature/LOTwinterschool2006/szabo.best.vwh.net/smart.contracts.html) [LOTwinterschool2006/szabo.best.vwh.net/smart.contracts.html](http://www.fon.hum.uva.nl/rob/Courses/InformationInSpeech/CDROM/Literature/LOTwinterschool2006/szabo.best.vwh.net/smart.contracts.html) [63] G. Wood et al., “Ethereum: A secure decentralised generalised transaction ledger,” Ethereum project yellow paper, vol. 151, no. 2014, pp. 1–32, 2014. ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey [64] A. Hafid, A. S. Hafid, and M. Samih, “Scaling blockchains: A comprehensive survey,” IEEE Access, vol. 8, pp. 125 244–125 262, 2020. [65] A. Biryukov, D. Khovratovich, and I. Pustogarov, “Deanonymisation of clients in bitcoin p2p network,” in Proceedings of the 2014 ACM SIGSAC Conference on Computer and Communications Security, ser. CCS ’14. New York, NY, USA: Association for Computing Machinery, [2014, p. 15–29. [Online]. Available: https://doi.org/10.1145/2660267.](https://doi.org/10.1145/2660267.2660379) [2660379](https://doi.org/10.1145/2660267.2660379) [66] S. King and S. Nadal, “Ppcoin: Peer-to-peer crypto-currency with proofof-stake,” self-published paper, vol. 19, August 2012. [67] B. W. Akins, J. L. Chapman, and J. M. Gordon, “A whole new world: Income tax considerations of the bitcoin economy,” Pittsburgh Tax Review, vol. 12, p. 25, 2014-2015. [68] W. Meng, E. W. Tischhauser, Q. Wang, Y. Wang, and J. Han, “When intrusion detection meets blockchain technology: A review,” IEEE Access, vol. 6, pp. 10 179–10 188, 2018. [69] C. Pirtle and J. Ehrenfeld, “Blockchain for healthcare: The next generation of medical records?” Journal of Medical Systems, vol. 42, [no. 9, p. 172, 2018. [Online]. Available: https://doi.org/10.1007/](https://doi.org/10.1007/s10916-018-1025-3) [s10916-018-1025-3](https://doi.org/10.1007/s10916-018-1025-3) [70] A. Ekblaw, A. Azaria, J. D. Halamka, and A. Lippman, “A case study for blockchain in healthcare: “MedRec” prototype for electronic health records and medical research data,” in Proceedings of IEEE Open and Big Data Conference, vol. 13. IEEE, 2016, p. 13. [71] C. Noyes, “Bitav: Fast anti-malware by distributed blockchain consensus and feedforward scanning,” CoRR, vol. abs/1601.01405, 2016. [Online]. [Available: http://arxiv.org/abs/1601.01405](http://arxiv.org/abs/1601.01405) [72] Q. Lu and X. Xu, “Adaptable blockchain-based systems: A case study for product traceability,” IEEE Software, vol. 34, no. 6, pp. 21–27, November 2017. [73] M. Seifelnasr, H. S. Galal, and A. M. Youssef, “Scalable open-vote network on ethereum,” Cryptology ePrint Archive, Report 2020/033, [2020. [Online]. Available: https://eprint.iacr.org/2020/033](https://eprint.iacr.org/2020/033) [74] X. Yang, X. Yi, S. Nepal, and F. Han, “Decentralized voting: A self-tallying voting system using a smart contract on the ethereum blockchain,” in Web Information Systems Engineering – WISE 2018, H. Hacid, W. Cellary, H. Wang, H.-Y. Paik, and R. Zhou, Eds. Cham: Springer International Publishing, 2018, pp. 18–35. [75] D. Khoury, E. F. Kfoury, A. Kassem, and H. Harb, “Decentralized voting platform based on ethereum blockchain,” in 2018 IEEE International Multidisciplinary Conference on Engineering Technology (IMCET), Nov 2018, pp. 1–6. [76] F. Þ. Hjálmarsson, G. K. Hreiðarsson, M. Hamdaqa, and G. Hjálmtýsson, “Blockchain-based e-voting system,” in 2018 IEEE 11th International Conference on Cloud Computing (CLOUD), July 2018, pp. 983–986. [77] M. Sharples and J. Domingue, “The blockchain and kudos: A distributed system for educational record, reputation and reward,” in Adaptive and Adaptable Learning, K. Verbert, M. Sharples, and T. Klobuˇcar, Eds. Cham: Springer International Publishing, 2016, pp. 490–496. [78] O. Novo, “Blockchain meets iot: An architecture for scalable access management in iot,” IEEE Internet of Things Journal, vol. 5, no. 2, pp. 1184–1195, 2018. [79] A. Whitmore, A. Agarwal, and L. Da Xu, “The internet of things– a survey of topics and trends,” Information Systems Frontiers, [vol. 17, no. 2, pp. 261–274, 2015. [Online]. Available: https:](https://doi.org/10.1007/s10796-014-9489-2) [//doi.org/10.1007/s10796-014-9489-2](https://doi.org/10.1007/s10796-014-9489-2) [80] Zhihong Yang, Yingzhao Yue, Yu Yang, Yufeng Peng, Xiaobo Wang, and Wenji Liu, “Study and application on the architecture and key technologies for IOT,” in 2011 International Conference on Multimedia Technology, July 2011, pp. 747–751. [81] J. Xie and S. M. Shugan, “Electronic tickets, smart cards, and online prepayments: When and how to advance sell,” Marketing Science, vol. 20, no. 3, pp. 219–243, 2001. [82] Y. Zhang and J. Wen, “An IoT electric business model based on the protocol of bitcoin,” in 2015 18th International Conference on Intelligence in Next Generation Networks, Feb 2015, pp. 184–191. [83] Y. Zhang and J. Wen, “The IoT electric business model: Using blockchain technology for the internet of things,” Peer-to-Peer Networking and Applications, vol. 10, no. 4, pp. 983–994, 2017. [Online]. Available: [https://doi.org/10.1007/s12083-016-0456-1](https://doi.org/10.1007/s12083-016-0456-1) [84] D. Sonstebo, S. Ivancheglo, D. Schiener, and D. S. Popov, “IOTA,” 2015. [[Online]. Available: https://www.iota.org/get-started/what-is-iota](https://www.iota.org/get-started/what-is-iota) [85] B. Shabandri and P. Maheshwari, “Enhancing iot security and privacy using distributed ledgers with iota and the tangle,” in 2019 6th International Conference on Signal Processing and Integrated Networks (SPIN), 2019, pp. 1069–1075. [86] S. Panikkar, S. Nair, P. Brody, and V. Pureswaran, “ADEPT: An IoT Prac[titioner Perspective,” 2015. [Online]. Available: https://www.coindesk.](https://www.coindesk.com/ibm-reveals-proof-concept-blockchain-powered-internet-things) [com/ibm-reveals-proof-concept-blockchain-powered-internet-things](https://www.coindesk.com/ibm-reveals-proof-concept-blockchain-powered-internet-things) [87] A. Bahga and V. K. Madisetti, “Blockchain platform for industrial internet of things,” Journal of Software Engineering and Applications, vol. 9, no. 10, pp. 533–546, 2016. [88] S. Huh, S. Cho, and S. Kim, “Managing iot devices using blockchain platform,” in 2017 19th International Conference on Advanced Communication Technology (ICACT), Feb 2017, pp. 464–467. [89] P. K. Sharma, S. Singh, Y. Jeong, and J. H. Park, “Distblocknet: A distributed blockchains-based secure SDN architecture for IoT networks,” IEEE Communications Magazine, vol. 55, no. 9, pp. 78–85, Sep. 2017. [90] G. Ateniese, M. T. Chiaramonte, D. Treat, B. Magri, and D. Venturi, “Hybrid blockchain,” May 28 2019, uS Patent 10,305,875. [91] M. Samaniego and R. Deters, “Blockchain as a service for iot,” in 2016 IEEE International Conference on Internet of Things (iThings) and IEEE Green Computing and Communications (GreenCom) and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE Smart Data (SmartData), Dec 2016, pp. 433–436. [92] A. Dorri, S. S. Kanhere, and R. Jurdak, “Blockchain in internet of things: Challenges and solutions,” 2016. [93] L. D. Xu, W. He, and S. Li, “Internet of things in industries: A survey,” IEEE Transactions on Industrial Informatics, vol. 10, no. 4, pp. 2233– 2243, Nov 2014. [94] A. Dorri, S. S. Kanhere, R. Jurdak, and P. Gauravaram, “Blockchain for iot security and privacy: The case study of a smart home,” in 2017 IEEE International Conference on Pervasive Computing and Communications Workshops (PerCom Workshops), March 2017, pp. 618–623. [95] A. Dorri, S. S. Kanhere, and R. Jurdak, “Towards an optimized blockchain for iot,” in 2017 IEEE/ACM Second International Conference on Internet-of-Things Design and Implementation (IoTDI), 2017, pp. 173–178. [96] J. J. Sikorski, J. Haughton, and M. Kraft, “Blockchain technology in the chemical industry: Machine-to-machine electricity market,” Applied Energy, vol. 195, pp. 234 – 246, 2017. [Online]. Available: [http://www.sciencedirect.com/science/article/pii/S0306261917302672](http://www.sciencedirect.com/science/article/pii/S0306261917302672) [97] D. Wörner and T. von Bomhard, “When your sensor earns money: Exchanging data for cash with bitcoin,” in Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing: Adjunct Publication, ser. UbiComp ’14 Adjunct. New York, NY, USA: Association for Computing Machinery, 2014, p. 295–298. [Online]. [Available: https://doi.org/10.1145/2638728.2638786](https://doi.org/10.1145/2638728.2638786) [98] L. Axon. and M. Goldsmith., “Pb-pki: A privacy-aware blockchainbased pki,” in Proceedings of the 14th International Joint Conference on e-Business and Telecommunications - Volume 4: SECRYPT, (ICETE 2017), INSTICC. SciTePress, 2017, pp. 311–318. [99] C. Fromknecht, D. Velicanu, and S. Yakoubov, “A decentralized public key infrastructure with identity retention.” IACR Cryptol. [ePrint Arch., vol. 2014, p. 803, 2014. [Online]. Available: https:](https://allquantor.at/blockchainbib/pdf/fromknecht2014decentralized.pdf) [//allquantor.at/blockchainbib/pdf/fromknecht2014decentralized.pdf](https://allquantor.at/blockchainbib/pdf/fromknecht2014decentralized.pdf) [100] H. Shrobe, D. L. Shrier, and A. Pentland, CHAPTER 15 Enigma: Decentralized Computation Platform with Guaranteed Privacy. MITP, [2018, pp. 425–454. [Online]. Available: https://ieeexplore.ieee.org/](https://ieeexplore.ieee.org/document/8333139) [document/8333139](https://ieeexplore.ieee.org/document/8333139) [101] Y. Hajjaji, W. Boulila, I. R. Farah, I. Romdhani, and A. Hussain, “Big data and iot-based applications in smart environments: A systematic review,” Computer Science Review, vol. 39, p. 100318, 2021. [Online]. Available: [https://www.sciencedirect.com/science/article/pii/S1574013720304184](https://www.sciencedirect.com/science/article/pii/S1574013720304184) [102] M. Kumar, K. Dubey, and R. Pandey, “Evolution of emerging computing paradigm cloud to fog: Applications, limitations and research challenges,” in 2021 11th International Conference on Cloud Computing, Data Science Engineering (Confluence), 2021, pp. 257–261. [103] J. Vijaykumar, P. Rajkumar, and S. Rakoth Kandan, “Fog computing based secured mobile cloud for cumulative integrity in smart environment and internet of things,” Materials Today: Proceedings, 2021. [[Online]. Available: https://www.sciencedirect.com/science/article/pii/](https://www.sciencedirect.com/science/article/pii/S2214785320408521) [S2214785320408521](https://www.sciencedirect.com/science/article/pii/S2214785320408521) [104] Z. Lv, L. Qiao, M. S. Hossain, and B. J. Choi, “Analysis of using blockchain to protect the privacy of drone big data,” IEEE Network, vol. 35, no. 1, pp. 44–49, 2021. 19 ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey [105] S. Zhang, S. Zhang, X. Chen, and X. Huo, “Cloud computing research and development trend,” in 2010 Second International Conference on Future Networks, Jan 2010, pp. 93–97. [106] Y. Jadeja and K. Modi, “Cloud computing - concepts, architecture and challenges,” in 2012 International Conference on Computing, Electronics and Electrical Technologies (ICCEET), March 2012, pp. 877–880. [107] M. Zhou, R. Zhang, D. Zeng, and W. Qian, “Services in the cloud computing era: A survey,” in 2010 4th International Universal Communication Symposium, Oct 2010, pp. 40–46. [108] W. Ma and J. Zhang, “The survey and research on application of cloud computing,” in 2012 7th International Conference on Computer Science Education (ICCSE), July 2012, pp. 203–206. [109] A. Wilczy´nski and J. Kołodziej, “Modelling and simulation of security-aware task scheduling in cloud computing based on blockchain technology,” Simulation Modelling Practice and Theory, vol. 99, [p. 102038, 2020. [Online]. Available: http://www.sciencedirect.com/](http://www.sciencedirect.com/science/article/pii/S1569190X19301698) [science/article/pii/S1569190X19301698](http://www.sciencedirect.com/science/article/pii/S1569190X19301698) [110] C. Chen, M. Lin, and C. Liu, “Edge computing gateway of the industrial internet of things using multiple collaborative microcontrollers,” IEEE Network, vol. 32, no. 1, pp. 24–32, Jan 2018. [111] S. Jeong, O. Simeone, and J. Kang, “Mobile edge computing via a UAVmounted cloudlet: Optimization of bit allocation and path planning,” IEEE Transactions on Vehicular Technology, vol. 67, no. 3, pp. 2049– 2063, March 2018. [112] A. Kapitonov, S. Lonshakov, A. Krupenkin, and I. Berman, “Blockchainbased protocol of autonomous business activity for multi-agent systems consisting of UAVs,” in 2017 Workshop on Research, Education and Development of Unmanned Aerial Systems (RED-UAS), Oct 2017, pp. 84–89. [113] A. Kumar, A. Kundu, C. A. Pickover, and K. Weldemariam, “Unmanned aerial vehicle data management,” Sep. 20 2018, uS Patent App. 15/463,147. [114] M. Aazam and E. Huh, “Fog computing and smart gateway based communication for cloud of things,” in 2014 International Conference on Future Internet of Things and Cloud, Aug 2014, pp. 464–470. [115] N. Abbas, Y. Zhang, A. Taherkordi, and T. Skeie, “Mobile edge computing: A survey,” IEEE Internet of Things Journal, vol. 5, no. 1, pp. 450–465, Feb 2018. [116] Z. Xiong, Y. Zhang, D. Niyato, P. Wang, and Z. Han, “When mobile blockchain meets edge computing,” IEEE Communications Magazine, vol. 56, no. 8, pp. 33–39, August 2018. [117] M. Liu, F. R. Yu, Y. Teng, V. C. M. Leung, and M. Song, “Computation offloading and content caching in wireless blockchain networks with mobile edge computing,” IEEE Transactions on Vehicular Technology, vol. 67, no. 11, pp. 11 008–11 021, Nov 2018. [118] Z. Xiong, S. Feng, W. Wang, D. Niyato, P. Wang, and Z. Han, “Cloud/fog computing resource management and pricing for blockchain networks,” IEEE Internet of Things Journal, vol. 6, no. 3, pp. 4585–4600, June 2019. [119] S. Tuli, R. Mahmud, S. Tuli, and R. Buyya, “Fogbus: A blockchain-based lightweight framework for edge and fog computing,” Journal of Systems and Software, vol. 154, pp. 22 – 36, 2019. [Online]. Available: [http://www.sciencedirect.com/science/article/pii/S0164121219300822](http://www.sciencedirect.com/science/article/pii/S0164121219300822) [120] R. Almadhoun, M. Kadadha, M. Alhemeiri, M. Alshehhi, and K. Salah, “A user authentication scheme of iot devices using blockchain-enabled fog nodes,” in 2018 IEEE/ACS 15th International Conference on Computer Systems and Applications (AICCSA), Oct 2018, pp. 1–8. [121] D. Wu and N. Ansari, “A cooperative computing strategy for blockchainsecured fog computing,” IEEE Internet of Things Journal, pp. 1–1, 2020. [122] S. L. Aljohani and M. J. F. Alenazi, “Mpresisdn: Multipath resilient routing scheme for sdn-enabled smart cities networks,” Applied [Sciences, vol. 11, no. 4, 2021. [Online]. Available: https://www.mdpi.](https://www.mdpi.com/2076-3417/11/4/1900) [com/2076-3417/11/4/1900](https://www.mdpi.com/2076-3417/11/4/1900) [123] V. Balasubramanian, M. Aloqaily, and M. Reisslein, “An sdn architecture for time sensitive industrial iot,” Computer Networks, vol. 186, [p. 107739, 2021. [Online]. Available: https://www.sciencedirect.com/](https://www.sciencedirect.com/science/article/pii/S1389128620313256) [science/article/pii/S1389128620313256](https://www.sciencedirect.com/science/article/pii/S1389128620313256) [124] A. J. Kadhim and J. I. Naser, “Proactive load balancing mechanism for fog computing supported by parked vehicles in iov-sdn,” China Communications, vol. 18, no. 2, pp. 271–289, 2021. [125] L.-A. Phan, D.-T. Nguyen, M. Lee, D.-H. Park, and T. Kim, “Dynamic fog-to-fog offloading in sdn-based fog computing systems,” Future Generation Computer Systems, vol. 117, pp. 486–497, 2021. [[Online]. Available: https://www.sciencedirect.com/science/article/pii/](https://www.sciencedirect.com/science/article/pii/S0167739X20330831) [S0167739X20330831](https://www.sciencedirect.com/science/article/pii/S0167739X20330831) 20 [126] A. Rahman, M. J. Islam, A. Montieri, M. K. Nasir, M. M. Reza, S. S. Band, A. Pescape, M. Hasan, M. Sookhak, and A. Mosavi, “Smartblocksdn: An optimized blockchain-sdn framework for resource management in iot,” IEEE Access, vol. 9, pp. 28 361–28 376, 2021. [127] N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson, J. Rexford, S. Shenker, and J. Turner, “Openflow: Enabling innovation in campus networks,” SIGCOMM Comput. Commun. Rev., vol. 38, [no. 2, p. 69–74, Mar. 2008. [Online]. Available: https://doi.org/10.1145/](https://doi.org/10.1145/1355734.1355746) [1355734.1355746](https://doi.org/10.1145/1355734.1355746) [128] S. Badotra and S. N. Panda, “Software-defined networking: A novel approach to networks,” in Handbook of Computer Networks and Cyber Security: Principles and Paradigms, B. B. Gupta, G. M. Perez, D. P. Agrawal, and D. Gupta, Eds. Cham: Springer International Publishing, 2020, pp. 313–339. [Online]. Available: [https://doi.org/10.1007/978-3-030-22277-2_13](https://doi.org/10.1007/978-3-030-22277-2_13) [129] Y. Jararweh, M. Al-Ayyoub, A. Darabseh, E. Benkhelifa, M. Vouk, and A. Rindos, “SDIoT: a software defined based internet of things framework,” Journal of Ambient Intelligence and Humanized Computing, vol. 6, no. 4, pp. 453–461, 2015. [Online]. Available: [https://doi.org/10.1007/s12652-015-0290-y](https://doi.org/10.1007/s12652-015-0290-y) [130] K. Kalkan and S. Zeadally, “Securing internet of things with software defined networking,” IEEE Communications Magazine, vol. 56, no. 9, pp. 186–192, Sep. 2018. [131] L. Huo, D. Jiang, S. Qi, and L. Miao, “A blockchain-based security traffic measurement approach to software defined networking,” [Mobile Networks and Applications, 2020. [Online]. Available: https:](https://doi.org/10.1007/s11036-019-01420-6) [//doi.org/10.1007/s11036-019-01420-6](https://doi.org/10.1007/s11036-019-01420-6) [132] D. V. Medhane, A. K. Sangaiah, M. S. Hossain, G. Muhammad, and J. Wang, “Blockchain-enabled distributed security framework for next generation iot: An edge-cloud and software defined network integrated approach,” IEEE Internet of Things Journal, pp. 1–1, 2020. [133] P. K. Sharma, M. Chen, and J. H. Park, “A software defined fog node based distributed blockchain cloud architecture for iot,” IEEE Access, vol. 6, pp. 115–124, 2018. [134] P. K. Sharma, S. Rathore, Y. Jeong, and J. H. Park, “SoftEdgeNet: SDN based energy-efficient distributed network architecture for edge computing,” IEEE Communications Magazine, vol. 56, no. 12, pp. 104– 111, December 2018. [135] I. D. Alvarenga, G. A. F. Rebello, and O. C. M. B. Duarte, “Securing configuration management and migration of virtual network functions using blockchain,” in NOMS 2018 - 2018 IEEE/IFIP Network Operations and Management Symposium, April 2018, pp. 1–9. [136] T. Wood, K. K. Ramakrishnan, J. Hwang, G. Liu, and W. Zhang, “Toward a software-based network: integrating software defined networking and network function virtualization,” IEEE Network, vol. 29, no. 3, pp. 36– 41, May 2015. [137] D. Schwartz, N. Youngs, A. Britto et al., “The ripple protocol consensus algorithm,” Ripple Labs Inc White Paper, vol. 5, no. 8, 2014. [138] D. B. Rawat, “Fusion of software defined networking, edge computing, and blockchain technology for wireless network virtualization,” IEEE Communications Magazine, vol. 57, no. 10, pp. 50–55, 2019. [139] D. B. Rawat and A. Alshaikhi, “Leveraging distributed blockchain-based scheme for wireless network virtualization with security and qos constraints,” in 2018 International Conference on Computing, Networking and Communications (ICNC), 2018, pp. 332–336. [140] R. P. França, A. C. B. Monteiro, R. Arthur, and Y. Iano, An Overview of the Machine Learning Applied in Smart Cities. Cham: Springer International Publishing, 2021, pp. 91–111. [Online]. Available: [https://doi.org/10.1007/978-3-030-60922-1_5](https://doi.org/10.1007/978-3-030-60922-1_5) [141] K. Rajkumar, M. Ramachandran, F. Al-Turjman, and R. Patan, “A reinforcement learning optimization for future smart cities using software defined networking,” International Journal of Machine Learning and [Cybernetics, Jan 2021. [Online]. Available: https://doi.org/10.1007/](https://doi.org/10.1007/s13042-020-01245-w) [s13042-020-01245-w](https://doi.org/10.1007/s13042-020-01245-w) [142] S. S. Panda and D. Jena, “Decentralizing ai using blockchain technology for secure decision making,” in Advances in Machine Learning and Computational Intelligence, S. Patnaik, X.-S. Yang, and I. K. Sethi, Eds. Singapore: Springer Singapore, 2021, pp. 687–694. [143] W. Serrano, “The blockchain random neural network for cybersecure iot and 5g infrastructure in smart cities,” Journal of Network and Computer Applications, vol. 175, p. 102909, 2021. [Online]. Available: [https://www.sciencedirect.com/science/article/pii/S1084804520303696](https://www.sciencedirect.com/science/article/pii/S1084804520303696) [144] R. Kumar, W. Wang, J. Kumar, T. Yang, A. Khan, W. Ali, and I. Ali, “An integration of blockchain and ai for secure data sharing ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey and detection of ct images for the hospitals,” Computerized Medical Imaging and Graphics, vol. 87, p. 101812, 2021. [Online]. Available: [https://www.sciencedirect.com/science/article/pii/S0895611120301075](https://www.sciencedirect.com/science/article/pii/S0895611120301075) [145] S. Aich, N. K. Sinai, S. Kumar, M. Ali, Y. R. Choi, M. I. Joo, and H. C. Kim, “Protecting personal healthcare record using blockchain federated learning technologies,” in 2021 23rd International Conference on Advanced Communication Technology (ICACT), 2021, pp. 109–112. [146] F. Jamil, N. Iqbal, Imran, S. Ahmad, and D. Kim, “Peer-to-peer energy trading mechanism based on blockchain and machine learning for sustainable electrical power supply in smart grid,” IEEE Access, vol. 9, pp. 39 193–39 217, 2021. [147] Z. Shahbazi and Y.-C. Byun, “Integration of blockchain, iot and machine learning for multistage quality control and enhancing security in smart manufacturing,” Sensors, vol. 21, no. 4, 2021. [Online]. Available: [https://www.mdpi.com/1424-8220/21/4/1467](https://www.mdpi.com/1424-8220/21/4/1467) [148] D. Nguyen, M. Ding, P. N. Pathirana, and A. Seneviratne, “Blockchain and ai-based solutions to combat coronavirus (covid-19)-like epidemics: A survey,” Apr 2020. [Online]. Available: [https://www.techrxiv.](https://www.techrxiv.org/articles/preprint/Blockchain_and_AI-based_Solutions_to_Combat_Coronavirus_COVID-19_-like_Epidemics_A_Survey/12121962/1) [org/articles/preprint/Blockchain_and_AI-based_Solutions_to_Combat_](https://www.techrxiv.org/articles/preprint/Blockchain_and_AI-based_Solutions_to_Combat_Coronavirus_COVID-19_-like_Epidemics_A_Survey/12121962/1) [Coronavirus_COVID-19_-like_Epidemics_A_Survey/12121962/1](https://www.techrxiv.org/articles/preprint/Blockchain_and_AI-based_Solutions_to_Combat_Coronavirus_COVID-19_-like_Epidemics_A_Survey/12121962/1) [149] T. P. Mashamba-Thompson and E. D. Crayton, “Blockchain and artificial intelligence technology for novel coronavirus disease-19 self-testing,” Diagnostics, vol. 10, no. 4, p. 198, Apr 2020. [Online]. Available: [http://dx.doi.org/10.3390/diagnostics10040198](http://dx.doi.org/10.3390/diagnostics10040198) [150] A. Kumari, R. Gupta, S. Tanwar, and N. Kumar, “Blockchain and AI amalgamation for energy cloud management: Challenges, solutions, and future directions,” Journal of Parallel and Distributed Computing, vol. 143, pp. 148 – 166, 2020. [Online]. Available: [http://www.sciencedirect.com/science/article/pii/S074373152030277X](http://www.sciencedirect.com/science/article/pii/S074373152030277X) [151] Y. Dai, D. Xu, S. Maharjan, G. Qiao, and Y. Zhang, “Artificial intelligence empowered edge computing and caching for internet of vehicles,” IEEE Wireless Communications, vol. 26, no. 3, pp. 12–18, June 2019. [152] Y. Dai, D. Xu, S. Maharjan, Z. Chen, Q. He, and Y. Zhang, “Blockchain and deep reinforcement learning empowered intelligent 5g beyond,” IEEE Network, vol. 33, no. 3, pp. 10–17, May 2019. [153] C. Qiu, F. R. Yu, H. Yao, C. Jiang, F. Xu, and C. Zhao, “Blockchain-based software-defined industrial internet of things: A dueling deep *Q* -learning approach,” IEEE Internet of Things Journal, vol. 6, no. 3, pp. 4627–4639, June 2019. [154] C. Qiu, F. R. Yu, F. Xu, H. Yao, and C. Zhao, “Blockchain-based distributed software-defined vehicular networks via deep q-learning,” in Proceedings of the 8th ACM Symposium on Design and Analysis of Intelligent Vehicular Networks and Applications, ser. DIVANet’18. New York, NY, USA: Association for Computing Machinery, 2018, p. [8–14. [Online]. Available: https://doi.org/10.1145/3272036.3272040](https://doi.org/10.1145/3272036.3272040) [155] P. K. Sharma, S. Y. Moon, and J. H. Park, “Block-vn: A distributed blockchain based vehicular network architecture in smart city.” Journal of information processing systems, vol. 13, no. 1, pp. 184–195, 2017. [[Online]. Available: http://jips-k.org/q.jips?cp=pp&pn=440](http://jips-k.org/q.jips?cp=pp&pn=440) [156] P. K. Sharma and J. H. Park, “Blockchain based hybrid network architecture for the smart city,” Future Generation Computer Systems, [vol. 86, pp. 650 – 655, 2018. [Online]. Available: http://www.](http://www.sciencedirect.com/science/article/pii/S0167739X1830431X) [sciencedirect.com/science/article/pii/S0167739X1830431X](http://www.sciencedirect.com/science/article/pii/S0167739X1830431X) [157] P. K. Sharma, N. Kumar, and J. H. Park, “Blockchain-based distributed framework for automotive industry in a smart city,” IEEE Transactions on Industrial Informatics, vol. 15, no. 7, pp. 4197–4205, July 2019. [158] V. Sapovadia, “Chapter 13 - legal issues in cryptocurrency,” in Handbook of Digital Currency, D. L. K. Chuen, Ed. San Diego: [Academic Press, 2015, pp. 253 – 266. [Online]. Available: http:](http://www.sciencedirect.com/science/article/pii/B9780128021170000138) [//www.sciencedirect.com/science/article/pii/B9780128021170000138](http://www.sciencedirect.com/science/article/pii/B9780128021170000138) [159] N. N. Emelianova and A. A. Dementyev, “Cryptocurrency, taxation and international law: Contemporary aspects,” in Artificial Intelligence: Anthropogenic Nature vs. Social Origin, E. G. Popkova and B. S. Sergi, Eds. Cham: Springer International Publishing, 2020, pp. 725–731. [160] A. A. Omololu, “Legal ramifications of blockchain technology,” in Decentralised Internet of Things: A Blockchain Perspective, M. A. Khan, M. T. Quasim, F. Algarni, and A. Alharthi, Eds. Cham: Springer International Publishing, 2020, pp. 217–230. [Online]. Available: [https://doi.org/10.1007/978-3-030-38677-1_10](https://doi.org/10.1007/978-3-030-38677-1_10) [161] V. S. K. Balagurusamy, C. Cabral, S. Coomaraswamy, E. Delamarche, D. N. Dillenberger, G. Dittmann, D. Friedman, O. Gökçe, N. Hinds, J. Jelitto, A. Kind, A. D. Kumar, F. Libsch, J. W. Ligman, S. Munetoh, C. Narayanaswami, A. Narendra, A. Paidimarri, M. A. P. Delgado, J. Rayfield, C. Subramanian, and R. Vaculin, “Crypto anchors,” IBM Journal of Research and Development, vol. 63, no. 2/3, pp. 4:1–4:12, 2019. [162] C. S. Wright and S. Savanah, “Operating system for blockchain iot devices,” May 23 2019, uS Patent App. 16/097,497. [163] A. Poelstra et al., “Distributed consensus from proof of stake is impossible,” Self-published Paper, 2014. [164] P. B, M. I, and M. M, “Migration from pow to pos for ethereum,” Oct 2019. [Online]. Available: engrxiv.org/ad8en [165] F. Saleh, “Blockchain without waste: Proof-of-stake,” Available at SSRN [3183935, 2020. [Online]. Available: https://ssrn.com/abstract=3183935](https://ssrn.com/abstract=3183935) [166] D. Larimer, “EOS.IO Technical White Paper v2,” 2018. [Online]. Available: [https://github.com/EOSIO/Documentation/blob/](https://github.com/EOSIO/Documentation/blob/master/TechnicalWhitePaper.md) [master/TechnicalWhitePaper.md](https://github.com/EOSIO/Documentation/blob/master/TechnicalWhitePaper.md) [167] C. Li and B. Palanisamy, “Comparison of decentralization in DPoS and PoW blockchains,” 2020. [168] ——, “Incentivized blockchain-based social media platforms: A case study of steemit,” in Proceedings of the 10th ACM Conference on Web Science, ser. WebSci ’19. New York, NY, USA: Association for Computing Machinery, 2019, p. 145–154. [Online]. Available: [https://doi.org/10.1145/3292522.3326041](https://doi.org/10.1145/3292522.3326041) [169] M. Vukoli´c, “The quest for scalable blockchain fabric: Proof-of-work vs. bft replication,” in Open Problems in Network Security, J. Camenisch and D. Kesdo˘gan, Eds. Cham: Springer International Publishing, 2016, pp. 112–125. [170] J. Garay, A. Kiayias, and N. Leonardos, “The bitcoin backbone protocol: Analysis and applications,” in Advances in Cryptology - EUROCRYPT 2015, E. Oswald and M. Fischlin, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2015, pp. 281–310. [171] A. Beniiche, “A study of blockchain oracles,” 2020. [172] H. Al-Breiki, M. H. U. Rehman, K. Salah, and D. Svetinovic, “Trustworthy blockchain oracles: Review, comparison, and open research challenges,” IEEE Access, vol. 8, pp. 85 675–85 685, 2020. [173] I.-C. Lin and T.-C. Liao, “A survey of blockchain security issues and challenges.” IJ Network Security, vol. 19, no. 5, pp. 653–659, 2017. [174] A. A. Siyal, A. Z. Junejo, M. Zawish, K. Ahmed, A. Khalil, and G. Soursou, “Applications of blockchain technology in medicine and healthcare: Challenges and future perspectives,” Cryptography, vol. 3, [no. 1, p. 3, Jan 2019. [Online]. Available: http://dx.doi.org/10.3390/](http://dx.doi.org/10.3390/cryptography3010003) [cryptography3010003](http://dx.doi.org/10.3390/cryptography3010003) [175] V. Garcia-Font, “Blockchain: Opportunities and challenges in the educational context,” in Engineering Data-Driven Adaptive Trustbased e-Assessment Systems: Challenges and Infrastructure Solutions, D. Baneres, M. E. Rodríguez, and A. E. Guerrero-Roldán, Eds. Cham: Springer International Publishing, 2020, pp. 133–157. [Online]. [Available: https://doi.org/10.1007/978-3-030-29326-0_7](https://doi.org/10.1007/978-3-030-29326-0_7) [176] S. S. Kamble, A. Gunasekaran, and R. Sharma, “Modeling the blockchain enabled traceability in agriculture supply chain,” International Journal of Information Management, vol. 52, p. 101967, 2020. [Online]. Available: [http://www.sciencedirect.com/science/article/pii/S0268401218312118](http://www.sciencedirect.com/science/article/pii/S0268401218312118) [177] K. M. Khan, J. Arshad, and M. M. Khan, “Investigating performance constraints for blockchain based secure e-voting system,” Future Generation Computer Systems, vol. 105, pp. 13 – 26, 2020. [Online]. Available: [http://www.sciencedirect.com/science/article/pii/S0167739X19310805](http://www.sciencedirect.com/science/article/pii/S0167739X19310805) [178] U. C. Çabuk, E. Adıgüzel, and E. Karaarslan, “A survey on feasibility and suitability of blockchain techniques for the e-voting systems,” IJARCCE, vol. 7, no. 3, p. 124–134, Mar 2018. [Online]. Available: [http://dx.doi.org/10.17148/IJARCCE.2018.7324](http://dx.doi.org/10.17148/IJARCCE.2018.7324) [179] G. Zyskind, O. Nathan, and A. . Pentland, “Decentralizing privacy: Using blockchain to protect personal data,” in 2015 IEEE Security and Privacy Workshops, 2015, pp. 180–184. [Online]. Available: [https://ieeexplore.ieee.org/document/7163223/keywords#keywords](https://ieeexplore.ieee.org/document/7163223/keywords#keywords) [180] J. Luo, Q. Chen, F. R. Yu, and L. Tang, “Blockchain-enabled softwaredefined industrial internet of things with deep reinforcement learning,” IEEE Internet of Things Journal, pp. 1–1, 2020. 21 ----- Ebrahim *et al.* : Blockchain as privacy and security solution for smart environments: A Survey MAAD EBRAHIM is currently a Ph.D. student at the Department of Computer Science and Operations Research (DIRO), University of Montreal, Canada. He received his M.Sc. degree in 2019 from the Computer Science Department, Faculty of Computer and Information Technology, Jordan University of Science and Technology, Jordan. His B.Sc. degree in Computer Science and Engineering has been received from the University of Aden, Yemen, in 2013. His research experience includes Computer Vision, Artificial Intelligence, Machine learning, Deep Learning, Data Mining, and Data Analysis. His current research interests include Fog and Edge Computing technologies, Blockchains, and Reinforcement Learning. ABDELHAKIM HAFID spent several years as the Senior Research Scientist with Bell Commu nications Research (Bellcore), NJ, USA, working in the context of major research projects on the management of next generation networks. He was also an Assistant Professor with Western University (WU), Canada, the Research Director of Advance Communication Engineering Center (venture established by WU, Bell Canada, and Bay Networks), Canada, a Researcher with CRIM, Canada, the Visiting Scientist with GMD-Fokus, Germany, and a Visiting Professor with the University of Evry, France. He is currently a Full Professor with the University of Montreal. He is also the Founding Director of the Network Research Laboratory and Montreal Blockchain Laboratory. He is a Research Fellow with CIRRELT, Montreal, Canada. He has extensive academic and industrial research experience in the area of the management and design of next generation networks. His current research interests include the IoT, fog/edge computing, blockchain, and intelligent transport systems. ETIENNE ELIE is Solutions and Systems Architect and Engineering Lead at Intel Corporation, California, USA. Prior to joining Intel Corporation, Dr. Elie was the technology and engineering manager for CARTaGENE, a public research platform and biobank of the Sainte-Justine Learning Hospital. He also served as ASIC Architecture Engineer at Nortel Networks and Advanced Micro Devices (AMD). Before moving to the US, Elie spent a short period of time with PSP Investments, one of Canada’s largest pension investment managers. Beside his role at Intel Corporation, Dr. Elie is a key contributor for the development of a largescale general-purpose neuromorphic Community Infrastructure (CI). Dr. Elie holds a Ph.D. in Computer Architecture from Université de Montréal, with focus on optimization of data movements in computer systems. He also holds a master’s degree, and Bachelor of Science in Engineering with great distinction. 22 -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2203.08901, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "http://arxiv.org/pdf/2203.08901" }
2,022
[ "JournalArticle", "Review" ]
true
2022-03-16T00:00:00
[ { "paperId": "ffffb2a637606031e130307997218949fd5776ef", "title": "A Socio-Technical Approach for Resilient Connected Transportation Systems in Smart Cities" }, { "paperId": "7a46f1b84bb84f1ef33fc2d074eb56a51461b5c8", "title": "Blockchain technology in the future smart grids: A comprehensive review and frameworks" }, { "paperId": "8e8f3511597b2ac55458f702d0a17c2eb4b730f3", "title": "Dynamic fog-to-fog offloading in SDN-based fog computing systems" }, { "paperId": "afd020100ef1e125463ac10bb1770996e49e3fdb", "title": "Migration from the traditional to the smart factory in the die-casting industry: Novel process data acquisition and fault detection based on artificial neural network" }, { "paperId": "2a62e54d86f6fe5bd55c37e93ed778295700a98a", "title": "Smart homes: domestic futurity as Infrastructure" }, { "paperId": "dbc0af9f52857be49c75502ff08def22a96fc137", "title": "Introduction to the special issue on smart transportation" }, { "paperId": "a39c07b6394a6289c3c84c91fd433ea73a3eef81", "title": "MPResiSDN: Multipath Resilient Routing Scheme for SDN-Enabled Smart Cities Networks" }, { "paperId": "33f6f9987b07cb69c774b35c99e3aa58f2f5ab3e", "title": "Blockchain for IoT-based smart cities: Recent advances, requirements, and future challenges" }, { "paperId": "5d5633c4df9109419bc7c204005c16074c1af15e", "title": "Fog computing based secured mobile cloud for cumulative integrity in smart environment and Internet of Things" }, { "paperId": "9ea707a2bb054ab2fd08c237aafcefd6d40cce80", "title": "Protecting Personal Healthcare Record Using Blockchain & Federated Learning Technologies" }, { "paperId": "ee5c614a08ea83dd62c1c9f01d26cdd94cc5e24c", "title": "Predictive Maintenance and Intelligent Sensors in Smart Factory: Review" }, { "paperId": "4af84b71e614671902137699069310d769543836", "title": "The Blockchain Random Neural Network for cybersecure IoT and 5G infrastructure in Smart Cities" }, { "paperId": "5e984d5f260355dcd93d2045aa86974a73ae3d1b", "title": "Integration of Blockchain, IoT and Machine Learning for Multistage Quality Control and Enhancing Security in Smart Manufacturing" }, { "paperId": "b4b74ee11a19fef85edb72c1e93ac2336c335644", "title": "Proactive load balancing mechanism for fog computing supported by parked vehicles in IoV-SDN" }, { "paperId": "3b529988da89bf94c4ebdf183831d07d4f31978e", "title": "Big data and IoT-based applications in smart environments: A systematic review" }, { "paperId": "56d5f9c1a25a47ff372529f9320defd218c72d52", "title": "Evolution of Emerging Computing paradigm Cloud to Fog: Applications, Limitations and Research Challenges" }, { "paperId": "3b208ae43be2fb69527669633c2dba67994925da", "title": "Internet of Things (IoT) Application Model for Smart Farming" }, { "paperId": "40f83ee10dd331f3408d939c98d11cd8a3281051", "title": "A reinforcement learning optimization for future smart cities using software defined networking" }, { "paperId": "48c706defd38fd100c81b62b224fc0e694a32ed8", "title": "Achieving efficient and Privacy-preserving energy trading based on blockchain and ABE in smart grid" }, { "paperId": "b554078f4ae1bc7c74cf9739cda77997fd8e36a5", "title": "Analysis of Using Blockchain to Protect the Privacy of Drone Big Data" }, { "paperId": "76bfb2316f4fb7b0122f727ad6b9db2459647396", "title": "An SDN architecture for time sensitive industrial IoT" }, { "paperId": "59ff2b08ad109a3330280183ec4c11a82af4da52", "title": "Smart Environments Concepts, Applications, and Challenges" }, { "paperId": "97f7540d7a270468b9d63be0d3b0bc3650a24de3", "title": "An Overview of the Machine Learning Applied in Smart Cities" }, { "paperId": "15188a99033c1beddeb14f19a4c6f9558fcb9568", "title": "An Integration of blockchain and AI for secure data sharing and detection of CT images for the hospitals" }, { "paperId": "bc93af4375996f95dc6017740d19dd779c734856", "title": "Multistage implementation framework for smart supply chain management under industry 4.0" }, { "paperId": "1a415ec978c083c81148ad9f29a7a807c1acd9f4", "title": "Blockchain for smart cities: A review of architectures, integration trends and future research directions" }, { "paperId": "66836b52b0b658a0d5f448579d0e87906a505d11", "title": "Smart cities education: An insight into existing drawbacks" }, { "paperId": "9aa7afe4664347d65fdaf40a4b3191f808b22f39", "title": "An efficient IoT based smart farming system using machine learning algorithms" }, { "paperId": "b2e1390f9e1d26bf0df91b9d992774a18fa0ecb7", "title": "Blockchain and AI amalgamation for energy cloud management: Challenges, solutions, and future directions" }, { "paperId": "4939f2ab9400c2c8ebc1b0a5fbe515ebdf495bcf", "title": "Survey on IoT security: Challenges and solution using machine learning, artificial intelligence and blockchain technology" }, { "paperId": "67b8efd1d1d67b266447d495d8ad72795ee4ffd3", "title": "An SDN-IoT-based Framework for Future Smart Cities: Addressing Perspective" }, { "paperId": "7e3acf8a52e19479bf759959be67e814368fe4ce", "title": "B-FERL: Blockchain based Framework for Securing Smart Vehicles" }, { "paperId": "03d1b883e9d8474212094e5764646bc6450cf565", "title": "Blockchain Without Waste: Proof-of-Stake" }, { "paperId": "bb7ec6cfa97bbb13e7bcd2d34e40f6288d8768dc", "title": "Scaling Blockchains: A Comprehensive Survey" }, { "paperId": "21b2c213b10b88bdd9976c64bf44bafa49b1c849", "title": "Modeling the blockchain enabled traceability in agriculture supply chain" }, { "paperId": "33469f86bc7ef904c1f4664d1cadf22227e3462a", "title": "IoT in Smart Farming Analytics, Big Data Based Architecture" }, { "paperId": "9fbbd520f24a810a4278e1054f04f481d417c8bd", "title": "A Comprehensive Review of the COVID-19 Pandemic and the Role of IoT, Drones, AI, Blockchain, and 5G in Managing its Impact" }, { "paperId": "a44dff5c3376b5029c8dff928d5c0c5272537e4b", "title": "Transforming business using digital innovations: the application of AI, blockchain, cloud and data analytics" }, { "paperId": "1df83e4dc0ab34d77179962e156b7bd1a65f0e0d", "title": "Blockchain and AI-Based Solutions to Combat Coronavirus (COVID-19)-Like Epidemics: A Survey" }, { "paperId": "a740dcc3da0e3086db21aedb196e5e7ba5b094e1", "title": "Investigating performance constraints for blockchain based secure e-voting system" }, { "paperId": "cd4a6840c9facd0855c882bbd5c1447a4615358a", "title": "Blockchain and Artificial Intelligence Technology for Novel Coronavirus Disease 2019 Self-Testing" }, { "paperId": "55ec2762fa782ac2c580f057e076f33b98e6ddd3", "title": "A Study of Blockchain Oracles" }, { "paperId": "888f1bb4bd8a61f2ba4b8dc4d5991876a107a8f0", "title": "Cryptocurrency, Taxation and International Law: Contemporary Aspects" }, { "paperId": "caaa2906ea4c47516063bb37d86d2e4960e7ea67", "title": "Blockchain-Enabled Distributed Security Framework for Next-Generation IoT: An Edge Cloud and Software-Defined Network-Integrated Approach" }, { "paperId": "63d145309d8c3449a608664f8c63a82b683020f4", "title": "Edge Computing Integrated with Blockchain Technologies" }, { "paperId": "377eaad43415f1b1ddbb94888ac6fec88e13e65e", "title": "A Cooperative Computing Strategy for Blockchain-Secured Fog Computing" }, { "paperId": "28046f46268df13b4d40a5863501d8d6f3e43ae3", "title": "Scalable Open-Vote Network on Ethereum" }, { "paperId": "b1d63649db5db22d418946bcef74a23346332f52", "title": "A Survey on Feasibility and Suitability of Blockchain Techniques for the E-Voting Systems" }, { "paperId": "0b03db96ca7128efd8ee011d3531fcfa0a09b7ac", "title": "Comparison of Decentralization in DPoS and PoW Blockchains" }, { "paperId": "d8e49199939494b41ac30fd2672c05cd8cf3546b", "title": "A Survey of IoT Applications in Blockchain Systems" }, { "paperId": "26b3d3d6a81d536923372f46aea665838a379fe6", "title": "Towards Blockchain-Based Software-Defined Networking: Security Challenges and Solutions" }, { "paperId": "4c83e05e10a47121c2ed8c6868bdbfd3c08f83fe", "title": "Modelling and simulation of security-aware task scheduling in cloud computing based on Blockchain technology" }, { "paperId": "50f7f133fe4f72df1ce89a2e88e9dca4e0184f4f", "title": "A Blockchain-Based Security Traffic Measurement Approach to Software Defined Networking" }, { "paperId": "566dee985717d2dabe7a341185ee3b1139bd0dea", "title": "Deployment of Blockchain Technology in Software Defined Networks: A Survey" }, { "paperId": "550bba213ecf92f6b1fd129ad42ec22c1f8aaecd", "title": "Blockchain for cloud exchange: A survey" }, { "paperId": "46e6fc16fee32c96a5c77c3255fa63765c42e60f", "title": "Fusion of Software Defined Networking, Edge Computing, and Blockchain Technology for Wireless Network Virtualization" }, { "paperId": "7c6bc10f47edffd53ff29d38aa1e072090436dbc", "title": "Migration from POW to POS for Ethereum" }, { "paperId": "694e0e8b4c53df21e6cd4083d36ba05629f648cd", "title": "A Decentralized and Anonymous Data Transaction Scheme Based on Blockchain and Zero-Knowledge Proof in Vehicle Networking (Workshop Paper)" }, { "paperId": "6291cf5f4d01de67deb1c4e39e10ffbcb664cdee", "title": "Blockchain-Based Distributed Framework for Automotive Industry in a Smart City" }, { "paperId": "acca913accc1680d9945a70df29ce4eafb830581", "title": "A survey on Blockchain based access control for Internet of Things" }, { "paperId": "7fe4bbce603abe533f688b888a51d597db600609", "title": "Blockchain for Internet of Things: A Survey" }, { "paperId": "bfe4586b650f39b4db7890a827fe35c108aa1f97", "title": "Artificial Intelligence Empowered Edge Computing and Caching for Internet of Vehicles" }, { "paperId": "53d634e2ec960b4c5356fbef9fa020cbe74eac8d", "title": "Blockchain-Based Software-Defined Industrial Internet of Things: A Dueling Deep ${Q}$ -Learning Approach" }, { "paperId": "cae22ac04006c776a9110802f4b14567e9794d54", "title": "Blockchain and Deep Reinforcement Learning Empowered Intelligent 5G Beyond" }, { "paperId": "eab3cc42abcc604eb2b552076093c18847858f35", "title": "Incentivized Blockchain-based Social Media Platforms: A Case Study of Steemit" }, { "paperId": "3765a0c09976c0e2d0e8825f47388d5c80812e2b", "title": "VQL: Providing Query Efficiency and Data Authenticity in Blockchain Systems" }, { "paperId": "e2666d69d9e354b9184e40ebaffd53f0322f14e6", "title": "Crypto anchors" }, { "paperId": "dcc2e13387dc9d70e576ddf87b1998d84fe37011", "title": "Enhancing IoT Security and Privacy Using Distributed Ledgers with IOTA and the Tangle" }, { "paperId": "0b7a10138d8037d28dff301e207ff92f61554c2a", "title": "Integrated Blockchain and Edge Computing Systems: A Survey, Some Research Issues and Challenges" }, { "paperId": "57004d74645ba5cb9e7732ca733ec910eb27b0be", "title": "Applications of Blockchain Technology in Medicine and Healthcare: Challenges and Future Perspectives" }, { "paperId": "1cccd591e7a72501ff4e418e148d3e0848479a35", "title": "Blockchain's adoption in IoT: The challenges, and a way forward" }, { "paperId": "7a916f5e028e00a482486ffc2e2d2a82516ead8c", "title": "Smart contracts" }, { "paperId": "164e7be27409eff36ad4a8f2ebb355fa3bde3540", "title": "FogBus: A Blockchain-based Lightweight Framework for Edge and Fog Computing" }, { "paperId": "657f57075dab025cc79e6f5b06d8a98176cc6c1f", "title": "Decentralized Voting: A Self-tallying Voting System Using a Smart Contract on the Ethereum Blockchain" }, { "paperId": "b3b42a5963440e844cf692236899ae6dad13c87f", "title": "Decentralized Voting Platform Based on Ethereum Blockchain" }, { "paperId": "03f5de0e96b2cf6c9fe2a8602621809902874e01", "title": "Blockchain-Based Distributed Software-Defined Vehicular Networks via Deep Q-Learning" }, { "paperId": "305edd92f237f8e0c583a809504dcec7e204d632", "title": "Blockchain challenges and opportunities: a survey" }, { "paperId": "29b0df183dbba0b3351e1b10b6c92a14977923fc", "title": "A User Authentication Scheme of IoT Devices using Blockchain-Enabled Fog Nodes" }, { "paperId": "868de1bd341d1bcbb9ddf9c5d05ff2af6c824c49", "title": "SoftEdgeNet: SDN Based Energy-Efficient Distributed Network Architecture for Edge Computing" }, { "paperId": "0919faff858302c01be75739f4dd703fea74a82d", "title": "Blockchain based hybrid network architecture for the smart city" }, { "paperId": "c972800be0371d6931030777df3d9548d6906485", "title": "Securing Internet of Things with Software Defined Networking" }, { "paperId": "9233ab0eaa5fc22a00f2fdb6d7d2d0017637b50b", "title": "Computation Offloading and Content Caching in Wireless Blockchain Networks With Mobile Edge Computing" }, { "paperId": "d00afc726b3130909351f3bb2159f12878c303a5", "title": "Blockchain for Healthcare: The Next Generation of Medical Records?" }, { "paperId": "383057f972b11b99cbc8c0d3e6c47170e9d95c1c", "title": "Blockchain and IoT Integration: A Systematic Survey" }, { "paperId": "da8a949f9c9f1df3a38f12c2cac97b789c705465", "title": "A Survey of Machine and Deep Learning Methods for Internet of Things (IoT) Security" }, { "paperId": "54d50269928dafc6a0744e46044c17d973fdb01c", "title": "Blockchain-Based E-Voting System" }, { "paperId": "16de1f889aa6437c0ca7985e45e8c1bf4707f291", "title": "Blockchain Technologies for the Internet of Things: Research Issues and Challenges" }, { "paperId": "01157f7c700e92323a5933e00c71cf001a8bac88", "title": "Blockchain with Internet of Things: Benefits, Challenges, and Future Directions" }, { "paperId": "02458904f9bd718bd8c6a1a36e9847ad83b0410b", "title": "A Review on the Use of Blockchain for the Internet of Things" }, { "paperId": "de8cec33c0d284f9e1ab657193dde70ff44633ba", "title": "Blockchain for the IoT: Opportunities and Challenges" }, { "paperId": "0c5653210f0a6d9a3680521a45b9ee01e9ae8342", "title": "Securing configuration management and migration of virtual network functions using blockchain" }, { "paperId": "b48a3381491aca0844b47237e9045ea95fc45ef5", "title": "Leveraging Distributed Blockchain-based Scheme for Wireless Network Virtualization with Security and QoS Constraints" }, { "paperId": "9ad22da8353ca1799c7026115c8987e3172ae6dd", "title": "Blockchain Meets IoT: An Architecture for Scalable Access Management in IoT" }, { "paperId": "e7f84b1d7f8378ffaadbf85c33bacc8bcd9e28dd", "title": "Mobile Edge Computing: A Survey" }, { "paperId": "051a8fae323f26a9bd2ca551940b4ba52b99c1be", "title": "A Software Defined Fog Node Based Distributed Blockchain Cloud Architecture for IoT" }, { "paperId": "a30b4b52b1e7b0aff4a5085cdc43ace30ca66f5e", "title": "When Intrusion Detection Meets Blockchain Technology: A Review" }, { "paperId": "ff6ae91ae0e500a1e91f368fca33d2b59aedbe55", "title": "Edge Computing Gateway of the Industrial Internet of Things Using Multiple Collaborative Microcontrollers" }, { "paperId": "461d523f9ba942c7474aef332412fe7b53c731be", "title": "When Mobile Blockchain Meets Edge Computing" }, { "paperId": "7d49d03e62907c45ea3f84cdc626dcbd75dc03f0", "title": "Adaptable Blockchain-Based Systems: A Case Study for Product Traceability" }, { "paperId": "b3f538f13b441c969a998350e63efaded86223e6", "title": "Cloud/Fog Computing Resource Management and Pricing for Blockchain Networks" }, { "paperId": "fa33053af899e6ca0da59ce04802c5437e43ee5a", "title": "A blockchain future for internet of things security: a position paper" }, { "paperId": "e64590b78434a38b931cf86d915335850fb67f2a", "title": "Blockchain-based protocol of autonomous business activity for multi-agent systems consisting of UAVs" }, { "paperId": "76bd712e4908a42c5514c50427a168d6d7952c70", "title": "DistBlockNet: A Distributed Blockchains-Based Secure SDN Architecture for IoT Networks" }, { "paperId": "5bbd5bfa99b4554729ae7bbf5de0388f28ed1151", "title": "Software-Defined Networking for Internet of Things: A Survey" }, { "paperId": "a22f52c7555f955cfc6720f7c1b89eb4af613a4d", "title": "Performance Analysis of Private Blockchain Platforms in Varying Workloads" }, { "paperId": "eabb07994b757d329a434a082e72b81fca9f6237", "title": "Blockchain technology in the chemical industry: Machine-to-machine electricity market" }, { "paperId": "61d69925287bd0e3adea7d6fe9ccaffaf29207cd", "title": "Towards an Optimized BlockChain for IoT" }, { "paperId": "28fe6a3fab2f2097a6f9aac5ae9799577badf883", "title": "Blockchain for IoT security and privacy: The case study of a smart home" }, { "paperId": "cbe2f7dd07d869dd1d0b2ba1cbc03b84dd695bb1", "title": "Block-VN: A Distributed Blockchain Based Vehicular Network Architecture in Smart City" }, { "paperId": "631cc57858eb1a94522e0090c6640f6f39ab7e18", "title": "Blockchain as a Service for IoT" }, { "paperId": "628c2bcfbd6b604e2d154c7756840d3a5907470f", "title": "Blockchain Platform for Industrial Internet of Things" }, { "paperId": "0200d453f5c995c87761e50976ed07692e257a30", "title": "The Blockchain and Kudos: A Distributed System for Educational Record, Reputation and Reward" }, { "paperId": "451729b3faedea24771ac4aadbd267146688db9b", "title": "Blockchain in internet of things: Challenges and Solutions" }, { "paperId": "c998aeb12b78122ec4143b608b517aef0aa2c821", "title": "Blockchains and Smart Contracts for the Internet of Things" }, { "paperId": "f572bcaa97e36d79e0cd01fb18dadb2f58eebebd", "title": "The IoT electric business model: Using blockchain technology for the internet of things" }, { "paperId": "96e1bff12b5acf42aae162b6aa339c0b4db49740", "title": "BitAV: Fast Anti-Malware by Distributed Blockchain Consensus and Feedforward Scanning" }, { "paperId": "efb1a85cf540fd4f901a78100a2e450d484aebac", "title": "The Quest for Scalable Blockchain Fabric: Proof-of-Work vs. BFT Replication" }, { "paperId": "658bcb7729770487216843abd422c5088a96b843", "title": "SDIoT: a software defined based internet of things framework" }, { "paperId": "63572fc8fa50afae559feda91b1867c8e336e454", "title": "Toward a software-based network: integrating software defined networking and network function virtualization" }, { "paperId": "4b9184937da308914b9e13c43bfd75845eaf910b", "title": "Decentralizing Privacy: Using Blockchain to Protect Personal Data" }, { "paperId": "9b5618e4d3295ac642e03681e3b9c7bf6db265e3", "title": "The Bitcoin Backbone Protocol: Analysis and Applications" }, { "paperId": "e46b1aca220cc11583b99ca1e612a85d55f7195e", "title": "An IoT electric business model based on the protocol of bitcoin" }, { "paperId": "26157d17ef7f4f42c1b18bb352eaf593059f7999", "title": "When your sensor earns money: exchanging data for cash with Bitcoin" }, { "paperId": "e4b92eccc5bc7ededff579232f5bed5186bf8302", "title": "Deanonymisation of Clients in Bitcoin P2P Network" }, { "paperId": "aa8a6894735dc32aa7212ab7606ccd6d6e7338b9", "title": "Fog Computing and Smart Gateway Based Communication for Cloud of Things" }, { "paperId": "9ace4e2eff677024c730a61146ab2efa06a6bf6a", "title": "The Internet of Things—A survey of topics and trends" }, { "paperId": "7cd16cdb3abb5e08ba253f1c403598c055c69793", "title": "Internet of Things in Industries: A Survey" }, { "paperId": "5fb646dcdbaad786437f37a1b8d54fdcf3729163", "title": "A Whole New World: Income Tax Considerations of the Bitcoin Economy" }, { "paperId": "b179cf70666c9f67fc8c330be3a55a5dd2266c2e", "title": "SNARKs for C: Verifying Program Executions Succinctly and in Zero Knowledge" }, { "paperId": "a8e0c1e13f9fc1798ac4e213ce56dcbb5ca3a43a", "title": "The survey and research on application of cloud computing" }, { "paperId": "5aa5e0dd8f623f6429650ebf95d5e32e8567fc4a", "title": "Cloud computing - concepts, architecture and challenges" }, { "paperId": "a66d64b90b4c4c0f07000d90146906c90efcb6f9", "title": "Study and application on the architecture and key technologies for IOT" }, { "paperId": "8f2397105f49d3b8b241603e2a319abc0983b81d", "title": "Services in the Cloud Computing era: A survey" }, { "paperId": "60eb6ca2993be1330cb845e1ac60d0188f61f08f", "title": "Cloud Computing Research and Development Trend" }, { "paperId": "0742fa40bf9be455fc6338e3a40ed6f0113d4a61", "title": "OpenFlow: enabling innovation in campus networks" }, { "paperId": "e236eb170d5bf37f8a24622715bcb73e7e8d4ede", "title": "Protocols and impossibility results for gossip-based communication mechanisms" }, { "paperId": "eb51cb223fb17995085af86ac70f765077720504", "title": "Kademlia: A Peer-to-Peer Information System Based on the XOR Metric" }, { "paperId": "b2b23720778727e2e4d9ffba751a6ffafb8c7b35", "title": "Electronic Tickets, Smart Cards, and Online Prepayments: When and How to Advance Sell" }, { "paperId": "8132164f0fad260a12733b9b09cacc5fff970530", "title": "Practical Byzantine fault tolerance" }, { "paperId": "f7ddf5129b392cc23411c2ca28ece3bc0ad682d6", "title": "A Survey Of Blockchain Security Issues And Challenges" }, { "paperId": "603b4f0fa052fdea4e1d96eb368c9d29adca562a", "title": "Peer-to-Peer Energy Trading Mechanism Based on Blockchain and Machine Learning for Sustainable Electrical Power Supply in Smart Grid" }, { "paperId": "63009b9c9392e044d30e0877899d43706b186298", "title": "Blockchain-based authentication and authorization for smart city applications" }, { "paperId": "6d789bc632a1d80a9b8333bcfef7821cde10811b", "title": "SmartBlock-SDN: An Optimized Blockchain-SDN Framework for Resource Management in IoT" }, { "paperId": "7f6d0564e546599143cd028c729c1ade5089552e", "title": "A Survey on Blockchain-Fog Integration Approaches" }, { "paperId": "2933e981ce051e9893dbb1059f36bfd38085ef56", "title": "Correction to: Legal Ramifications of Blockchain Technology" }, { "paperId": "44a5532e4a80d1e4f95e070ca4639811f82f8b95", "title": "Blockchain-Based Cyberthreat Mitigation Systems for Smart Vehicles and Industrial Automation" }, { "paperId": "931bc061fef71c1485946aca1bc3dc391bc44e59", "title": "Software-Defined Networking: A Novel Approach to Networks" }, { "paperId": "f77e99d5d15f32cbdcb464c46849e4d47a477887", "title": "Hybrid Blockchain" }, { "paperId": "bc2207c8f7a3c68a334b6b525b8dbe0866edcf41", "title": "A Systematic Literature Review of Integration of Blockchain and Artificial Intelligence" }, { "paperId": "68a8470244c39ca74b3f4fa1e527ebdb93eb6ae5", "title": "Decentralizing AI Using Blockchain Technology for Secure Decision Making" }, { "paperId": "19fae8ff8eb043efada14903025b3754f4b3d788", "title": "Trustworthy Blockchain Oracles: Review, Comparison, and Open Research Challenges" }, { "paperId": "34162f7c8f69e3e892bb22672042c9d4041317d7", "title": "AI, IoT, Big Data, and Technologies in Digital Economy with Blockchain at Sustainable Work Satisfaction to Smart Mankind: Access to 6th Dimension of Human Rights" }, { "paperId": "2614a0d0213e8cc457ea62b435a8f43dea54245a", "title": "Blockchain for AI: Review and Open Research Challenges" }, { "paperId": "3d267bbcce5a599ac9cc42964fefb40e7b49cbb1", "title": "Applications of Blockchains in the Internet of Things: A Comprehensive Survey" }, { "paperId": "64cce7c078749876b4327cea08df6fc7dd95f0f4", "title": "Blockchain: Opportunities and Challenges in the Educational Context" }, { "paperId": "7247762784cc43596f91adde7a094f85edc30bf6", "title": "CHAPTER 15 Enigma: Decentralized Computation Platform with Guaranteed Privacy" }, { "paperId": null, "title": "Unmanned aerial vehicle data management" }, { "paperId": null, "title": "EOS.IO Technical White Paper v2" }, { "paperId": null, "title": "Mobile edge computing via a UAVmounted cloudlet: Optimization of bit allocation and path planning" }, { "paperId": "bc05bfaae0483caead1660d946c4b557a78bd4f1", "title": "PB-PKI: A Privacy-aware Blockchain-based PKI" }, { "paperId": "24711b2a7a4dc4d0dad74bbbfeea9140abab047b", "title": "Managing IoT devices using blockchain platform" }, { "paperId": "490369507b3a1e425ff4b7150fc9d083783bb908", "title": "A review of Internet of Things for smart home: Challenges and solutions" }, { "paperId": "f852c5f3fe649f8a17ded391df0796677a59927f", "title": "Architecture of the Hyperledger Blockchain Fabric" }, { "paperId": "3ed0db58a7aec7bafc2aa14ca550031b9f7021d5", "title": "A Case Study for Blockchain in Healthcare : “ MedRec ” prototype for electronic health records and medical research data" }, { "paperId": "9d1cb62f7fde9b03882567ee01adba31b7108276", "title": "Distributed Consensus from Proof of Stake is Impossible" }, { "paperId": "0dbb8a54ca5066b82fa086bbf5db4c54b947719a", "title": "A NEXT GENERATION SMART CONTRACT & DECENTRALIZED APPLICATION PLATFORM" }, { "paperId": null, "title": "ADEPT: An IoT Practitioner Perspective" }, { "paperId": null, "title": "Chapter 13 - legal issues in cryptocurrency" }, { "paperId": "3c50bb6cc3f5417c3325a36ee190e24f0dc87257", "title": "ETHEREUM: A SECURE DECENTRALISED GENERALISED TRANSACTION LEDGER" }, { "paperId": "57855fea0eea38a503ae58cbb024a2606002f677", "title": "A Decentralized Public Key Infrastructure with Identity Retention" }, { "paperId": "bff4ecdd2c40bb67abab8d49e99c81287a7b2810", "title": "The Ripple Protocol Consensus Algorithm" }, { "paperId": null, "title": "Delegated proof-of-stake (dpos)" }, { "paperId": "0db38d32069f3341d34c35085dc009a85ba13c13", "title": "PPCoin: Peer-to-Peer Crypto-Currency with Proof-of-Stake" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": "249377e09f6da6eda933ed4f39b4dbe6aa74b592", "title": "the Internet of Things: a Systematic Literature Review" } ]
37,289
en
[ { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/010def07d358187bc10b482d05b77f0e27f833dc
[]
0.929743
Governance of Blockchain and Distributed Ledger Technology Projects
010def07d358187bc10b482d05b77f0e27f833dc
Social Science Research Network
[ { "authorId": "123890248", "name": "Bronwyn E. Howell" }, { "authorId": "1828146", "name": "P. Potgieter" }, { "authorId": "2971618", "name": "B. Sadowski" } ]
{ "alternate_issns": null, "alternate_names": [ "SSRN, Social Science Research Network (SSRN) home page", "SSRN Electronic Journal", "Soc Sci Res Netw", "SSRN", "SSRN Home Page", "SSRN Electron J", "Social Science Electronic Publishing presents Social Science Research Network" ], "alternate_urls": [ "www.ssrn.com/", "https://fatcat.wiki/container/tol7woxlqjeg5bmzadeg6qrg3e", "https://www.wikidata.org/wiki/Q53949192", "www.ssrn.com/en", "http://www.ssrn.com/en/", "http://umlib.nl/ssrn", "umlib.nl/ssrn" ], "id": "75d7a8c1-d871-42db-a8e4-7cf5146fdb62", "issn": "1556-5068", "name": "Social Science Research Network", "type": "journal", "url": "http://www.ssrn.com/" }
Blockchains are the most well-known example of a distributed ledger technology (DLT). Unlike classic databases, the ledger is not maintained by any central authority. The integrity of the ledger is maintained automatically by an algorithmic consensus process whereby nodes vote and agree upon the authoritative version. In effect, the consensus algorithm operates in the manner of a decision-making process within a governance system.<br><br>The technological characteristics of blockchain systems are well documented (Narayanan, Bonneau, Felton and Miller, 2016). We propose that one of the reasons why it has so far proved very difficult to seed large-scale commercial DLT (blockchain) projects lies in the arena of project ownership and governance. Unlike classic centralised database systems, DLTs have no one central point of “ownership” of any of the system’s infrastructure or data. <br><br>In this piece of exploratory research, we propose applying theories of club governance to both the technical design and operational development of a range of DLT (blockchain) systems, including (but not necessarily limited to) cryptocurrencies and enterprise applications to explore how they can explain the development of (or lack of development of) sustainable solutions to real business problems. There are many parallels to the governance arrangements observed historically in the origins of complex distributed telecommunications networks.
Howell, Bronwyn E.; Potgieter, Petrus H.; Sadowski, Bert M. **Conference Paper** ## Governance of Blockchain and Distributed Ledger Technology Projects 2nd Europe - Middle East - North African Regional Conference of the International Telecommunications Society (ITS): "Leveraging Technologies For Growth", Aswan, Egypt, 18th-21st February, 2019 **Provided in Cooperation with:** International Telecommunications Society (ITS) _Suggested Citation: Howell, Bronwyn E.; Potgieter, Petrus H.; Sadowski, Bert M. (2019) :_ Governance of Blockchain and Distributed Ledger Technology Projects, 2nd Europe - Middle East - North African Regional Conference of the International Telecommunications Society (ITS): "Leveraging Technologies For Growth", Aswan, Egypt, 18th-21st February, 2019, International Telecommunications Society (ITS), Calgary This Version is available at: [https://hdl.handle.net/10419/201737](https://hdl.handle.net/10419/201737) **Standard-Nutzungsbedingungen:** Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. **Terms of use:** _Documents in EconStor may be saved and copied for your personal_ _and scholarly purposes._ _You are not to copy documents for public or commercial purposes, to_ _exhibit the documents publicly, to make them publicly available on the_ _internet, or to distribute or otherwise use the documents in public._ _If the documents have been made available under an Open Content_ _Licence (especially Creative Commons Licences), you may exercise_ _further usage rights as specified in the indicated licence._ ----- # Governance of Blockchain and Distributed Ledger Technology Projects ### Bronwyn E. Howell[*], Petrus H. Potgieter[†], Bert M. Sadowski[‡] ## Abstract Blockchains are the most well-known example of a distributed ledger technology (DLT). Unlike classic databases, the ledger is not maintained by any central authority. The integrity of the ledger is maintained automatically by an algorithmic consensus process whereby nodes vote and agree upon the authoritative version. In effect, the consensus algorithm operates in the manner of a decision-making process within a governance system. The technological characteristics of blockchain systems are well documented (Narayanan, Bonneau, Felton and Miller, 2016). We propose that one of the reasons why it has so far proved very difficult to seed large-scale commercial DLT (blockchain) projects lies in the arena of project ownership and governance. Unlike classic centralised database systems, DLTs have no one central point of “ownership” of any of the system’s infrastructure or data. In this piece of exploratory research, we propose applying theories of club governance to both the technical design and operational development of a range of DLT (blockchain) systems, including (but not necessarily limited to) cryptocurrencies and enterprise applications to explore how they can explain the development of (or lack of development of) sustainable solutions to real business problems. There are many parallels to the governance arrangements observed historically in the origins of complex distributed telecommunications networks. _Keywords: blockchain, distributed ledger, governance, club governance, distributed consensus_ ## 1 Introduction “Reform is a profoundly political process, not a technical one.” Fukuyama (2014, 161) Blockchains are the first, and most well-known example of a distributed ledger technology (DLT). A distributed ledger (DL) is a database (or file) spread across several nodes or computing devices. Each node in a network has access to (and probably saves) an identical copy of the ledger. Unlike *School of Management, Victoria University of Wellington, bronwyn.howell@vuw.ac.nz †Department of Decision Sciences, University of South Africa, potgiph@unisa.ac.za / php@grensnut.com ‡Department of Industrial Engineering & Innovation Sciences, Eindhoven University of Technology, b.m.sadowski@tue.nl 1 ----- classic databases, the ledger is not maintained by any central authority. The integrity of the ledger is maintained automatically by an algorithmic consensus process whereby nodes vote and/or agree upon the authoritative version, which is then updated and saved independently on each node. In effect, the consensus algorithm operates in the manner of a decision-making process within a governance system. Blockchain DLs use a chain of blocks linked to one another and secured using public-key cryptography to provide a secure and valid distributed consensus. A blockchain is usually distributed across and managed by peer-to-peer networks. Its append-only structure only allows data to be added to the database: altering or deleting previously entered data on earlier blocks is impossible. Blockchain technology is therefore well-suited for recording events, managing records, processing transactions, tracing assets, and voting. The technological characteristics of blockchain systems are well documented (Narayanan, Bonneau, Felten, Miller and Goldfeder, 2016). Considerable faith has been placed in the technology as a means of revolutionising digital transacting (Mulligan, Scott, Warren and Rangaswami, 2018; Crosby, Nachiappan, Pattanayak, Verma and Kalyanaraman, 2015; Czepluch, Lollike and Malone, 2015; Swan, 2015). However to date, outside of the arena of highly-publicised cryptocurrencies such as Bitcoin and Ethereum, few examples exist of the use of the technology to support significant economic activities. Nonetheless, plans for many other blockchain systems have been announced – for example Sovrin for identity management, and Halo for supply chain management. Comparatively little has however been documented so far about the governance of blockchain systems and the commercial activities they support, beyond the algorithmic voting processes via which the nodes agree on the authoritative version of the DL. This paper represents an exploratory endeavour to address this gap. We begin by reviewing current interpretations of blockchain “ownership” and “governance”. Neither the data in the ledgers nor the software governing blockchain operations is claimed to be owned by anyone in particular. Nonetheless ongoing responsibility for the rules governing their ongoing operation must be assumed by someone if they are to be created and operated successfully for commercial endeavours. We propose that the governance of DL systems can be analogised to that of clubs. While some of the governance rules are embedded in system software, and may be costly and difficult to change, leading to stable ledger content, other rules are embedded in the institutional arrangements linking system participants (club members) outside of the software, and may not be so costly to change, depending upon how key decision-making rights are distributed across them. The stability of a given DL system will depend upon the interaction of the decision rights allocated and exercised within the software with rights allocated and exercised outside of it. From the theoretical discussion, we develop a framework for examining a given blockchain DL system to identify and evaluate the effectiveness of its governance arrangements given its specific commercial application. We apply the framework in two case studies: the cryptocurrency Bitcoin and the identity management system Sovrin. While both claim to be public blockchains using very similar “proof of work” algorithms to agree ledger content, Sovrin’s structure as a “permissioned” blockchain (it poses barriers to entry for node operators (they are required to become Stewards) and requires end-users to have a relationship with node operators separately verifiable from their blockchain relationship) differentiates it from Bitcoin. We suggest that the costs of changing Sovrin’s governance arrangements as circumstances change are much lower 2 ----- than those of Bitcoin. Thus, Bitcoin is less susceptible to successful forking than Sovrin. Sovrin’s stability relies more strongly upon the alignment of the interests of its node operators (stewards) than does Bitcoin (miners are node operators), because Sovrin’s governance arrangements allow them both greater control over changes to software content and lower costs of co-ordinating successful forking than their Bitcoin comparators. ## 2 Ownership and governance of distributed ledgers Who owns and governs a blockchain system? One view (prevailing with the development of cryptocurrencies) is that a specific DL application is “governed by no-one”, because it is “owned by no-one” (Sovrin, 2018). Anyone who wishes to use the blockchain application may do so (it is “public”). Unlike web-based applications, any user operating as a node has a copy of both the ledger and the software required to participate in the system (although some classes of user may interact via web pages managed locally by a node operator). Neither the data nor the software are proprietary to a single controlling entity. As the software code is open source, any node operator is free at any time to make changes to the code (which they themselves are using) and institute a new blockchain operating independently of the first (termed “forking”), without facing the disadvantage of a centralised system of not having access to the accumulated historic data. All blocks up to the point of forking are identical in both the original and the new chain. Although the components of a blockchain are “owned by no-one”, it cannot be said that a blockchain system is “governed by no-one”. All systems operate within a framework of rules, either derived implicitly from the norms and cultures of the participants or explicitly articulated in formal agreements (such as constitutions and contracts) (Williamson, 1999; 2000). Collectively, these rules comprise the governance arrangements under which systemic interaction takes place. Within them, in order to co-ordinate participants and streamline decision-making, selected groups of individuals are granted superior decision-making rights by assuming those ceded to them by specific subsets of system users. Governance arrangements can emerge endogenously over time (bottom-up) (for example, as has occurred with the constitutional arrangements of nation states), or be imposed from the outset (top-down)(for example, following military conquest, or in the Constitutions and Articles of Incorporation of firms, marketplaces (e.g. stock exchanges, clubs and trusts) (Ostrom, 1990; 2005). Efficient and effective governance arrangements will specify both the set of rules prevailing for normal transacting, and provisions whereby those rules can be changed in response to changing circumstances. When the provisions for rule changes are explicit, and users have clear means of observing those charged with the responsibility for managing the rules process and holding them to account for their actions (or inactions), then the systems will tend to be more efficient than if the rules and/or the identity of the decision-makers are unclear, those with decision-making rights can exercise them covertly, and there are no clear means of users holding the decision-makers to account or the costs of doing so are so high as to render the probability of occurrence remote (Hansmann, 1996; Cordery & Howell, 2017). The classic shareholder-owned firm provides an example of one such explicit system, where shareholders give up their rights to make decisions about the day-to-day use of the firm’s assets to boards and management to facilitate more efficient firm functioning than if the shareholders themselves were required to undertake co-ordination (Berle and Means, 1937; Williamson, 1985). Other examples include the arrangements pertaining to the management of non-owned 3 ----- non-rival and non-excludable “public” goods where governments as trustees exercise decisionmaking rights on behalf of all citizens, and those pertaining to the management of non-rival but excludable “club” goods enjoyed by an identified population (as a subset of general citizenry). ### 2.1 Distributed ledger systems as clubs Club theory, proposed initially by Buchanan (1965) in respect of clubs dealing in rival, excludable goods provided and consumed by volunteer-members has been expanded subsequently to take account of the separate dimensions of non-rivalry and non-excludability of the goods provided by the clubs, and the exclusivity of club membership (e.g. Olson, 1989; Comes and Sandler, 1996). Important work by E Ostrom and V Ostrom (Ostrom, 1990; 2005; 2010; Ostrom, 2014) melded the concept of the club with theories of self-organising governance systems, federalism and polycentrism in government, demonstrating that common resources could be managed successfully without government regulation or privatisation, by way of decentralised entities operating as polities using representation relationships (i.e. “membership”) rather than contractual assignment of rights proportional to asset ownership to allocate decision-making control. In this view, a DL system (DLS) is a club with (arguably) open membership. It is a ‘public system’ as any member of the public agreeing to abide by the rules can join in order to use it. At any point in time, the club membership is defined by those participating in the DLS. The DLS is governed by rules covering both membership and operation. Various classes of membership are usually determined by the nature and form of interactions the members have with it. Operational rules cover how routine operations will occur and how, in the event that conflicts arise that cannot satisfactorily be resolved within the existing rules, the rules can be changed to maintain the integrity of the system and thereby ensure ongoing use by the members (Cordery and Howell, 2017). The consensus arrangements by which the DLS resolves conflicts about the content of the DL constitute but one part of the system’s operational rules. Importantly, decision-making powers attach to and differ by membership status. Axiomatically, users operating nodes that agree the ledger content participate differently in decision-making from end-users participating only by using the DLS to transact with other end-users. The over-arching institutional arrangements under which the DLS operates – including the allocation to membership status and associated decision-making rights – must be decided first by founding members in order for the system to be created. The founding members will determine the original governance rules – both those coded into the software and other non-coded arrangements. They exercise considerable design control. It matters, therefore, whether these individuals exert decision-making influence in either or both of operational and other representative roles. The rules must address the potential for conflicts in these decision-making responsibilities to be resolved. The founding members assume fiduciary duties to both future members and the DLS club as a whole. To the extent that the initial rules establish hierarchies of membership and allocation of decision-making responsibilities, these can be thought of as defining the club committees and sub-committees, with the founding members being allocated to different roles. Once the DLS becomes operational, the roles may transfer to new members as they join and as the new membership begins exercising its rights. This may include the replication of “branches” or “sub-branches” of the club with their own committees and sub-committees as the number of node operators expands. The DLT rules specifying how these distributed entities are federated 4 ----- into the overall club governance functions and how decision-making rights and responsibilities are distributed across them are effectively constitutions. However, there is also a threat that new members are offered discriminatory treatment by incumbent members or are excluded altogether. Rey and Tirole (2007) have shown that incumbent members have an incentive to exploit their monopoly power or restrict entry by new players. Within the blockchain context, that is a massive centralization problem due to the concentration of mining power within a small group of initial members. However, there is a movement from the concept of Proof of Work (PoW) towards the Proof of Stake (PoS) that tries to address this problem by giving greater voting power to those who have a stake in the venture. For a discussion of blockchains as constitutional entities, where club members are equated to citizens, see Berg, Berg and Novak (2018) and Berg, Novak, Potts and Thomas (2018). Our analysis differs from theirs in that they focus only on the ongoing operation and management of a DLS once it has been created, whereas we examine both instantiation and ongoing operation. We also extend our analysis to include relationships between different classes of member outside of the operation of the ledger – that is, we see the constitutional rules encompassing both software and non-software elements. Their analysis focuses predominantly on the software-mediated elements. ### 2.2 DLS governing rules and club stability The initial DLS is offered as a “take it or leave it” package to the first public users – that is, as a ‘top-down’ imposition as per Ostrom (2009) (e.g. implementation of a new stock exchange). This differs from the voluntary agreement of federated arrangements when extant groups negotiate the rules under which their activities become linked (e.g. when two stock exchanges merge) – that is, bottom-up arrangements enabling large-scale co-operation (Bednar, 2009). In either case, once the rules are agreed, they can change via either gradual reinterpretation of the rights and obligations defined by apparently stable governance rules, or substantive episodic change in the formal structure of the governance rules (analogous respectively to Buchanan’s distinction between regular political activity and constitutional moments – Buchanan and Tullock, 1962; Buchanan and Brennan 1985; Congleton, 2014). Both clubs and political entities will function effectively so long as there are credible commitments by members/citizens to monitoring, enforcement/sanction and conflict resolution within the existing rules (North, 1993; Ostrom, 2005), and the ability to bring about changes to those rules to meet the demands of changing circumstances (Tarko, Schlager and Lutter, 2018). The former are enhanced by rule stability and certainty; yet too much stability can lead to fragility if the governance rules are not well-adapted to new challenges. But allowing for ready change may also lead to instability as competing centres of authority may attempt to devise and impose rules to benefit themselves at the expense of others. The challenge is to find a workable balance between stability and flexibility in the governing arrangements. As long the original (or extant) arrangements suit the (ever-changing) membership, a DLS will survive – albeit in a dynamically-adapting form as constitutionally defined. However, when an issue arises which cannot be agreed using the current rules, a fork may occur (a ‘break-away club’ forms). The ongoing success of both clubs/DSLs now depends on the proportion of the members who move to the new DLS or remain with the original one. 5 ----- On the one hand, the ease with which disaffected node operators can break away provides significant pressure on the existing DLS to be designed for and operate consistently in node operator interests, which may not necessarily be in the interests of end-users. On the other hand, if end-users’ interests are compromised, they will not participate in the DLS in the first place. Careful balancing of interests of both node operators and end users is necessary to both attract a critical mass of node operators and user-members and maintain DLS stability. If these are not well-balanced, the DLS will be unstable – that is, prone to either failure (no members) or forking. ### 2.3 Trading flexibility and stability in a dynamic context However, to the extent that many of the governance arrangements must be coded into the software in advance of the system beginning operation, DLS design is subject to the bounded rationality of human designers. Arrangements that are satisfactorily balanced at one point in time within one set of wider environmental circumstances may not be optimal if those circumstances change (e.g. substantial changes in electricity prices for node operators). Typically, the more flexible the governance rules are, the more easily they can be altered to take account of the changes and the DLS as an institutional entity will be more stable. However, the co-ordination required to institute the changes can be costly. A strength of DLS consensus algorithms is that the costs of co-ordinating to change the softwaregoverned outcomes are very high. This leads to high confidence in the integrity of the data held in the ledgers. However, the costs of changing the non-software-governed elements can, but need not, be high. The lower are the costs of co-ordinating the activities of the human entities using the system, the easier it is to institute changes – either to enhance the operation of the existing DLS by a general agreement of all with sufficient powers to amend the software to maintain its stability, or to facilitate a successful fork by persuading a critical mass of users of the existing system to support the break-away rules and system, instead of the original. If all end users are definitively linked via mechanisms that enable them to be easily identified, then it is cheaper to communicate with them to co-ordinate any action than if they cannot be easily identified. Change of either type (forking; mutually-agreed rule and software changes) is more likely as co-ordination costs are lower. Furthermore, the greater is the extent to which the member activities are linked outside the day-to-day operation of the DL, the cheaper is the cost of co-ordinating activities for (for example), end users to follow the decisions of their relevant node operator when it becomes necessary to decide whether to support a software change for the existing system or to support a fork. Assume, for example, that the node operators are required by the DLS rules to be identified and known to each other in order to be authorised to act in this capacity. That is, the DLS is a permissioned system. The cost of co-ordinating activity amongst a known number of identified individuals is less than where neither the number nor the identity of the members is known. Both agreed changes maintaining existing DLS integrity and forking will be less costly, suggesting that permissioned systems may be less stable. In the long run, this could have the effect of making it harder for the DLS to attract and retain new node operators, and thus build membership scale quickly. This effect could be overcome in part by adopting rules that make forking more expensive – for example, requiring the payment of a substantial bond (membership fee) that is forfeited if the member initiates or joins a fork. 6 ----- The requirement for a node operators to develop their own software to interface with the DLS offers only a weak form of a bond against forking once the desired system has been selected from the range available. If a successful fork occurs, then the software will be equally useful on either variant (at least initially). Thus expected DLS scale rather than stability likely has a greater influence on the selection of the system by a given node operator. However, the advantages of a flexible, permissioned system become more evident when certainty about the future environment in which the DLS will operate is lower. Assume now that a DLS is operating in a volatile environment, where changes in these external circumstances may alter the returns to a subset of node operators such that they may find it desirable to change the existing rules or create/support a fork with rules more conducive to their interests. Such actions will be frustrated by the high costs of identifying the likely disaffected operators and co-ordinating the requisite change. If no one operator once having joined a system can co-ordinate a defection at low cost, the incentives to join in the first instance, and to invest in developing the requisite code for a fork are likely low. In these circumstances, it may be feasible to instigate a new club only if, in the first instance, the rules serve to lower the costs of changing them when circumstances indicate. This could be the case in the early days of developing the business case for a DLS to be used for a specific application or in a specific industry (e.g. identity verification or supply chain management). However, it is not axiomatic that this state will prevail indefinitely. Generally, the more mature is the DL application, the more widely used it is, and the more diverse are the interests of ts user groups, the more costly it will be to gain a consensus on changes to the non-software-mediated governance rules, and the more stable will be the DLS. We note that, by analogy, the internet did not originate as the public, open entity governed by the cultures, norms and formal arrangements currently prevailing. Rather, it began as a closed, permissioned entity with substantial restrictions placed by its governance arrangements in its users. It was initially a network of peers in government and academia with very narrow, homogeneous research interests in network technology development, beginning in 1969. The governance arrangements expanded gradually to include users with broader, more general research-oriented interests with their own network resources that specifically excluded commercial network operators. While users may have utilised PSTN connections to make their connections, the public telephone companies were unable to participate meaningfully in the internet ecosystem until changes in the governance arrangements in 1995 enabled wide private sector participation (Leiner, Cerf, Clark, Kahn, Kleinrock, Lynch, Postel, Roberts and Wolff, 2009). Over time, as more and more user groups were added, and users became more and more heterogeneous, the changes in governance arrangements became less and less frequent as the costs of co-ordinating their interests in order to institute changes became greater. Substantive changes now occur very rarely indeed, via costly consensus-seeking international processes organised by entities such as the International Telecommunications Union and the Internet Corporation for Assigned Names and Numbers. Furthermore, as DLs exist to serve applications for communities of specific interest, the extent to which the environments in which they operate are stable or volatile, and hence the costs of coordination will vary depending on a wide range of context-specific factors. In part, this explains why, despite the existence of several thousand cryptocurrencies, only a handful are operating at a meaningful scale. Unless nearly all operators are equally affected by the exogenous change in circumstances or an internally-agreed rule change, then even if the high co-ordination costs could be overcome, the proportion of defections to a forked variant will be small and it will 7 ----- be unlikely to appeal to a significant number of end-users. The greater is the choice of nodes available to an end-user to interact with the original DLS, the lower are the incentives for a single node to defect, unless the choice of end-users to patronise other nodes is also restricted. It now matters how end-users’ interaction with the DL is mediated. If their choice of mediating node is restricted to a limited number of operators, then the costs of co-ordinating the migration of end-users’ future use from the original DLS to the forked version are substantially reduced compared to the alternative of multiple (or ‘free’) node choices. ## 3 Developing an inquiry framework To analyse a DLS as a club, it is first necessary to identify its membership and the rules governing its operation and governance. As with any club, there may be many different states of membership, defined in the rules. The rules will specify their relationships with each other, along with the various powers each class of member may exercise in both regular operation and in club governance. ### 3.1 Membership The fundamental technical entities in any DLS are the nodes on which the DL copies are stored. Each node is managed by human actor, which may be either a real (unique human) or a legal (corporate) person. The human actor makes the decision to join the club, and in doing so agrees to abide by the club rules. Human node operators will subsequently be termed node-members of the DLS club. Upon joining, node operators must acquire the current version of the DLS operating software from the club’s software bank or find equivalent and indistinguishable (from the point of the network) software elsewhere. The software bank is managed by a sub-committee of members, who may or may not fulfil other roles within the club at the point of time the analysis is being undertaken. However, as the origin of any DLS relies upon the development of the relevant software, and no nodes can join the until the software has been ‘released’ for production, most DLS clubs will begin with a small number of members all of whom have strong stakes in the software development. Members who participate in the development and maintenance of software alone will be termed software-members. At the origin of a DLS, the club is most likely comprised almost exclusively of software-members. Over time, however, as more nodemembers join, the proportions will change. Software-members play a vital role as they have the knowledge and skills necessary to evaluate the effectiveness of the existing software and implement changes to it – such as those necessary to generate a fork. Members of the software subcommittee therefore exert significant power (control) over the software content and hence the likelihood of forking or changes to the existing software occurring. Node-members are typically remunerated for holding ledger copies and processing transactions via a combination of payments from the DLS (in redeemable tokens - system currency) upon becoming the ‘winner’ who first posts the ultimately-agreed block, and from the entities who requested the transaction in the first place. They have strong vested interests in both the systemgenerated rules for token payments and any other agreements about the setting and collection of transaction fees paid by those using the system. 8 ----- The development of the software for a DLS and its promotion, are not costless. Softwaremembers may contribute to software developments without being paid, but nonetheless, they face an opportunity cost for the time invested. In effect, they donate that time (albeit that they may expect to be rewarded subsequently via returns from DLS operation as node operators, transaction generators or end-users). However, the DLS may have members who support the club’s activities with explicit financial contributions. Members acting in this capacity will subsequently be termed donor-members. The club nature of a DLS is distinguished from a proprietary firm by the fact that these donor-members are not shareholders. They have no defined claim on either the club’s assets or any profits made from operating (though of course, like software-members they may obtain benefits in other capacities of interaction in the system). For a DLS to be operational, it requires two other classes of member. These are - end user-members, who wish to use the system to undertake transactions with other end users or request information held in the DL, and - transaction-members, who manage the interfaces via which end-users participate. Transaction members may hold copies of the DL content, to facilitate raising transactions and answering queries. However, they do not participate in processing the transactions, which is undertaken by the node-members. In some cases, the transaction-members may also be end-users, interacting with the DL for their own purposes. These users are typically dependent upon using software and/or code provided by the DLS in order to generate transactions to it or queries on it. However, they do not exercise any rights in the development of that software/code, unless they also participate separately as software-members. They must take the code ‘as given’. An example is a cryptocurrency wallet, managed by an individual end-user. In other cases, transaction-members may undertake a vast range of activities separate and distinct from the DL, with a vast range of end-users, as well as posting transactions for node-members to process and queries to be responded to. These transaction-members may generate their own bespoke code and applications (e.g. web pages) that build on code supplied by the DLS, but once again, unless they engage separately as software-members, they exert no influence on the DLS code per se. An example is a currency exchange, which may interact with many different cryptocurrency DLs in addition to banks handing traditional fiat currencies and payment mechanisms. However, to the extent that transaction-members have access to the DLS code, and have the capacity to understand and modify it, they provide an important discipline on the DLS because of their potential to create a fork. Transaction-members are typically remunerated in fees paid to them by end-users. These may (but do not need to be) determined by club rule processes. However, transaction-members must pay fees to node operators when transactions are successfully completed. These may be encoded within the DLS and paid using system tokens, or ‘off-system’ via club rules or other ‘private’ contracts agreed between members. Club rules can be used to outlaw the latter agreements, but enforcement is contingent upon the ability to detect their existence and use. Their primary governance concerns will pertain to the level of, and rules setting, these fees, and rules concerning how they must relate with equivalent and adjacent club members - that is, other transaction members, node-members and end user-members. End user-members are those who participate in the DLS only to the extent that they are originators or beneficiaries of transactions, or they make inquiries on the ledger. They are the equivalent of 9 ----- customers of a shareholder-owned firm. They will pay fees to transaction-members for services requested. They may pay these in stocks of system tokens via processes encoded in the DLS, but equally, in other currencies, via arrangements that need not be agreed by or encoded in the DLS. End user-member interests in DLS governance will pertain largely to the rules via which these fees are determined. As for transaction-members, their primary governance concerns will pertain to how they must relate with equivalent and adjacent club members - that is, other user-members and transaction-members. ### 3.2 Governance As identified above, a single club member may interact with the DL in many different capacities. The potential overlaps are illustrated in Figure 1. In the early stages of the DLS life cycle, and especially during its development, all roles (except perhaps donor-membership) may overlap, as in the hatched portion of Figure 1. However, as the DLS matures, and especially as it increase its scale of operations, the roles would be expected to gradually separate out (i.e. specialisation, as per Williamson, 1986 emerges). The nature of the separation will now be governed by the rules implicitly embedded in the DLS software and explicitly articulated in its “offline” rules. Figure 1: Membership Status **3.2.1** **Control** Broadly speaking, Figure 1 identifies a hierarchy in membership status for mature systems. The higher-up in the hierarchy a member sits, the more power potentially is conferred in decisionmaking in the governance arrangements. End-user members and transaction-members can exert very little formal power via the governance and decision-making processes, as they must ‘take as given’ the package offered by node members. Their power is confined to ‘voting with their 10 ----- feet’ and either choosing not to patronise the DLS, or (to the extent possible given the costs), co-ordinating a successful fork. The costs of organising a successful fork depend on the extent to which the disaffected transaction-members can ensure that if they leave, end-users will follow them and not defect to transaction-members remaining on the original DLS. This is largely a matter of the design of the contractual relationships between transaction members and end-users. If these allow a transaction-member to limit the extent to which an end-user can patronise other transactionmembers, then the costs of co-ordinating to organise a fork will be lower. On the one hand, DLS designers may not want to place many restrictions on these relationships, as reducing the likelihood of forking reinforces system stability. While the power of members higher up the membership hierarchy in Figure 1 is reinforced, it will rarely need to be exercised to change the software and/or other governance rules. On the other hand, as discussed in the preceding theory, if change is anticipated, then it may be necessary to co-ordinate the actions of all members in order to change key elements of the DLS rules (software and other rules) without exposing the DLS to undue risks of forking. Node-members are pivotal, as the DLS cannot operate without them, but equally, they too may have little choice but to accept a ‘take it or leave it’ package offered by the founder-members. Once again, they can opt not to join in the first place, or like transaction members, co-ordinate to instigate a successful fork. However, to the extent that they are formally engaged in the process of DLS governance outside of the software channels, they can work constructively with the software and donor members to change the rules in a manner that ensures their ongoing patronage. By either custom or explicit design, therefore, DLS governance is effectively controlled by a small coalition of software-members, who may also participate as node-members or be closely affiliated with influential node-members (i.e. they form the club committee). In order to motivate their participation, it would be expected they anticipate remuneration from either their node operation activities, or some other arrangement such as an honorarium paid from DLS funds held off the ledger – for example, financial or in-kind contributions (e.g. time, computing resources) made by donor-members. Donor-members without other membership stakes are unlikely to make substantial contributions of this kind unless they too exercise some influence over DLS governance and management – for example, having some powers to appoint or veto candidates to the club committee, or specifying in advance how their donations are to be managed and/or applied – in the same manner as expected by donors to clubs. Thus, it is more likely that formal articulation of DLS governance arrangements outside of the software itself (e.g. formal agreement of rules, club or trust agreements, etc.) will be necessary the more donor-members there are, and the greater is their contribution of resources towards DLS operation. Sponsorship of these formal arrangements may arise in the event that a group of donor-members form a club to establish a new DLS for a specific purpose (e.g. to serve a trade organisation or similar). Formal governance arrangements may also be necessary for a group of informally-organised software-members wishing to make use of existing entities (e.g. firms, trade organisations) to take an embryonic DLS from test-state to production. 11 ----- **3.2.2** **Rule and relationship formalisation** When a DLS is new and small, all members have homogeneous interests, and all are known to each other (i.e. all participate in the club in the same manner, as illustrated by the memberships intersecting in the hatched area of Figure 1), then the need for formal rules articulating the relationships between member groups and how conflicts will be resolved are less necessary. However, as it grows, role specialisation increases and member interests begin to diverge, then rule formalisation becomes more likely to be important. In particular, the allocation of important decision-making powers, processes of appointments to decision-making bodies, the relationships between different member categories and expectations and obligations of the relevant members should be made explicit in order to allow members to make appropriate decisions and enabling them to expect consistent predictable outcomes when interacting with the club. Figure 2: Transacting Relationships Figure 2 illustrates some potential patterns in relationships between different classes of members in a hypothetical DLS. This can be used to illustrate how different restrictions placed upon the interrelationships of club members affect costs of co-ordination. For example, node N3 operates in a closed environment with a limited number of transactionmembers (TP4 and TP5) who interact with no other node operator. Furthermore, the transactionmembers interact with a limited number of end-users (EU7, EU8 and EU9), who do not interact with any other transaction-members who do not operate through node N3. This arrangement could be achieved by having rules restricting interactions to a closed subset of members. That is, 12 ----- N3 will only accept transactions from transaction members known to or recognised by it, and these members are precluded by software-mediated rules from interacting via any other nodemember. Similar obligations can attend the interactions of end user-members with transaction members. In this example, EU8 can interact with any transaction member affiliated with N3 (TP4 or TP5), but EU7 and EU9 may be limited to interacting with TP4 and TP5 respectively. The N3 limb of Figure 2 is an example of a “permissioned” DLS - each member needs the ‘permission’ of one higher up the membership tree to interact with the DLS. As these arrangements prevent ‘client’ transaction- and end user-members from interacting via any other node, considerable power is vested in N3. If it instigates a fork, then it can be sure of maintaining its existing transaction volume at negligible cost. The higher are a node-member’s investments in the DLS and its operation, and the greater the share of its remuneration it gets from fees paid by transaction members, as opposed to the DLS, the more likely it is that a node-member will prefer to use the governance rules to restrict transaction-member and end user member choices. If all node operators are comparatively homogeneous in their identities and operations, and the proportion of their remuneration received from payments agreed ‘off system’ with downstream affiliates (rather than the internal DLS payments), then the more likely it is that a strictly hierarchical system will emerge. Even though the ledger and software are decentralised, each node will operate as the principal of its own federated ‘sub-branch’ of the DLS club. A commercial analogy is franchisees operating with exclusive territories. As with the franchise system, inducing participation by the node operators is contingent upon these protections. However, unlike franchise systems, the node operator can relatively costlessly exit, taking existing systems and customers along. If the club is to attract node operators in the first place, and remain stable into the future, those members must be protected from the effects of competition emerging from forking ex-members. The governance rules must contain provisions that make forking costly (e.g. very large membership fees, forfeited on forking). By contrast, nodes N1 and N2 in Figure 2 can interact with any of TP1, TP2 and TP3. It is an example of a “permissionless’ or fully public system. If N2 forks, TP1 and TP3 can shift their interaction to N1. Fewer (or no) interaction restrictions lower the likelihood of forking and therefore the risks of joining for a new node. There is less need for governance arrangements to constrain member defection by forking. Indeed, such a system may be able to operate without any special rules governing interaction. Competition within and between members may be satisfactory. However, in both cases, the more heterogeneous are the node members, the more likely it is that a ‘one size fits all’ set of rules (especially for remuneration) will be optimal for all member types. Tensions between members are more likely to arise in these circumstances. However, unless members are identifiable and known to each other, and formal channels established for resolving disputes, the costs of achieving a satisfactory resolution are likely so large as to be prohibitive. Change is unlikely to occur, either to the existing rules or via forking. To the extent that these problems can be anticipated, cost-reducing dispute resolution mechanisms may be contained within the DLS governance provisions. If they are not, then arrangements external to it may also facilitate co-ordination - for example, if specific member groups are affiliated in other ways, such as by being members of an industry association. Knowledge of such potentials may alter the strategies by which a specific DLS may seek to expand - for example, by engaging with the aggregating entry directly, or seeking to include it as a member, and thereby relying on its resources to assist in dispute resolution. In this case, it may not be necessary for the DLS to have direct knowledge of the identity of, or direct communication with, individual club 13 ----- members. Nonetheless, it is noted that the outcomes of such co-ordination may not be aligned with preserving the viability of the DLS unless its governance rules contain means of ensuring the aggregate members are required to prioritise this outcome. ## 4 Case study: Bitcoin The Bitcoin blockchain is a distributed ledger (DL) which is used to record transactions in the Bitcoin cryptocurrency. In this paper, however, we consider only the mechanism by which new blocks are added to the ledger rather than the operation of the cryptocurrency which is well described elsewhere, for examply by Böhme, Christin, Edelman and Moore (2015). ### 4.1 Decription of the Bitcoin protocol On the most basic level, the Bitcoin “network” consists of a large number of entirely independent computers that exchange messages conforming to certain specifications using the same protocol and each with a copy of the Bitcoin DL. The Internet protocol (IP) addresses of some key servers are published on authoritative websites and it is free to join. These servers can store and distribute the addresses of other servers on the network. Each server (or, node) can check whether its version of the DL corresponds to those stored by others (up to a certain number of blocks) on the peer-to-peer network but this is really most easily done buy consulting some authoritative website. The basic function of a node is to copy the DL but it can also submit transactions for possible inclusion in the chain using an identity based on a randomly generated public-private key pair. Before considering which transactions (which are actually simply messages in a specific format) are included in the DL, we have to consider the integrity of the system as a whole. Suppose, as a thought experiment, the Internet were suddenly split into two fully functioning parts. For example, by a single large country detaching itself from the global network. Bitcoin nodes on the detached part of the network might be unable to find some of the servers with IP addresses published on the authoritative websites (if these were available) but as long as some of the authoritative servers are based in the detached portion of the Internet, Bitcoin nodes in the rump would continue to functions as normal. The same would be true for the other portion of the Internet and the two versions of the Bitcoin DL would simply grow differently. For the paranoid, in short, there is no way of knowing that they are operating on the “true” DL. Nevertheless, Bitcoin has proven quite successful as a payment system and has maintained its integrity and support remarkably well. The main reason for this is the ingenuous design of its proof-of-work system for generating new blocks for inclusion in the ledger. ### 4.2 Adding new blocks The work is done by “miners” which are nodes on the network that generate candidates for new blocks. Each of these candidates must contain valid transactions (moderately easily checked by other nodes) and the solution (very easily checked) to a mathematical problems that necessarily involves generating a lot of random candidate solutions, on average. The solution is included in 14 ----- the candidate block broadcast to the network, as is a transaction that includes awarding a bounty to the miner. As nodes receive valid candidate blocks from miners, they accept them, add them to their copy of the ledger and rebroadcast them. Subsets of the nodes can at this stage receive and accept different new blocks. This is a dilemma but one that is completely resolved (usually within about an hour) by the nature of the Bitcoin protocol. Whichever new block is accepted by more nodes will tend to be the block that is accepted by miners and that they use to build subsequent blocks. This is entirely consistent with miners’ self-interest – they would have no incentive to mine on chains that are likely to be abandoned. With the operation of this majoritarian mechanism, nodes that have accepted a less favoured block will eventually find that the chain is a dead-end and will revert to the surviving chains. This mechanism delivers a consensus that is driven entirely by the self-interest of all the parties and is subject to no prior explicit arrangement. This majoritarian system of vetting new blocks[1] is the source of fear of the so-called “51% attack” which would consists of putative malicious control of a majority of the mining capacity and the possible introduction of improper blocks (for example, containing invalid transaction messages) that are then included in the DL. Since all nodes are able to relative inexpensively check the validity of blocks, this is extremely unlikely to go unnoticed for very long but it is likely to cause a great deal of confusion and distrust. Nevertheless, since the miners are awarded in Bitcoin, they probably have very little incentive to engage in behaviour that reduces trust in the underlying cryptocurrency. The governance of the Blockchain ledger is therefore mechanically implied by the protocol, which is the genial invention that engenders great robustness and stability. Nevertheless, this does not exclude the use of explicit agreements among participating nodes (or, indeed natural or juristic persons). Should, for example, a 51% attack introduce and invalid block, there is nothing preventing a large number of stakeholders to agree to make a certain change to the ledger and to restart from that point onward and in a specific way. This could however be costly and disruptive because of the ongoing demand for the DL to record the processing of payments. ### 4.3 Forking the chain It has happened on several occasions that a sufficiently large section of Bitcoin users managed to agree to change the protocol that they at a certain point started following new rules (“forking” the blockchain at that point) and that this change has been sufficiently sustainable. Bitcash is one example. At the point of creating Bitcash, anyone with (say) 2.3 Bitcoin would retain that amount but would also have 2.3 units of Bitcash as well, attributed to the same public key identity. Such a fork does not require more than for a viable number of participants to (agree to) do it. In the thought experiment above the splitting of the Internet in two, the fork would have been involuntary. In August 2010, a notable fork to correct a technical error, took place. A block had been mined that created 184 467 440 737.09551616 units of Bitcoin (van Wirdum, 2016) and sent them to two addresses. The number is remarkable since the Bitcoin protocol only allows for a total of 21 million Bitcoin to ever exist. A bulletin board message on 15 August from “Satoshi Nakamoto” warned [1New blocks and who mines them can be observed directly at https://www.blockchain.com/explorer.](https://www.blockchain.com/explorer) 15 ----- *** WARNING *** We are investigating a problem. DO NOT TRUST ANY TRANSACTIONS THAT HAPPENED AFTER 15.08.2010 17:05 UTC (block 74638) until the issue is resolved. and the error was reversed by a software update within a few hours. This was the most serious protocol or software error in the history of Bitcoin and it happened when the DL was only two years old. A similar breakdown today would hardly be tolerable in view of the number of transactions per day. ### 4.4 Informal and formal governance In addition to the large miners, “a small core of highly skilled developers” (De Filippi and Loveluck, 2016) for Bitcoin Core, the most widely used Bitcoin client, have an outsize influence in practice on the arrangements for this DL. The software was initially published by Satoshi Nakamoto (pseudonym) who also released the Bitcoin founding whitepaper. Development of Bitcoin Core has been funded by MIT Media Lab and others (van Wirdum, 2016). Given the practical dominance of Bitcoin Core software, it would not be entirely out of place to view governance of Bitcoin as identical to the governance of the software project, as a first-order approximation. This is nevertheless very informal and unstructured, even anarchic, since there is still absolute freedom to fork the open source software project (at the same time as the chain). It would not be incorrect to say that there is no formal governance arrangement for Bitcoin. ### 4.5 Applying the analytical framework to Bitcoin Applying Figure 1 to Bitcoin, the following membership stakes are identified: - software-members are an unidentified person/group of people acting under the soubriquet “Satoshi Nakamoto”. It would appear that this group exercises ultimate control over Bitcoin governance; - MIT Media Lab was an original donor-member, but it is not known whether it or any other unidentified funders continue to contribute actively to Bitcoin governance; - node members are the miners, who can freely enter and exit of their own accord. There is no explicit relationship between them and any other members; - transaction members and end user members can freely enter and exit of their own accord. There are no explicit requirements governing their interactions. One entity may participate as all of end user, transaction- and node-members. Overlaps almost certainly occur, given the majority of issued bitcoin are held by small number of system participants Applying Figure 2, we conclude that in the absence of any apparent rules to the contrary, Bitcoin is a fully-public DLS with no explicit or externally-articulated governance arrangements. All applicable rules are embedded in and executed by the software. Changing the rules is extremely costly - as evidenced by the perpetuation and growth of the DLS despite the absence of substantive changes to the software-based rules since the August 2010 fork. Given the large number of nodes and the comparative inability of member interests to organise successful rule changes or forks where large numbers of members defect, it would seem that Bitcoin members 16 ----- are heterogeneous, and lack ‘off-system’ means of cost-reducing co-ordination. The inherent anonymity of Bitcoin members also militates against such actions. That said, we note that the Bitcoin DLS is underpinned by a relatively simple and reasonably well-understood financial transacting business case. While the distributed ledger component of the system is is novel, the payments processing function is not. It is much easier at the outset to identify the governance requirements for a stable, well-understood system where change in the business model is unlikely to occur. Arguably, the Bitcoin DLS has been remarkably stable because these characteristics have meant that the circumstances of tensions arising between different members or class of members have simply not come about, since 2010 at least. And that the changes made in 2010 were successful was likely in large part attributable to the fact that at that stage, membership was small, much more homogeneous and more likely to be comprised of people known to each other (or at least, a sufficiently large-enough coalition was well-enough known to each other to co-ordinate the forking at a comparatively lower price. To the extent that forking has occurred to start new currencies, it s likely that this has been steered by software=-skilled members, iof not software-members per se.That none of the forked currencies has grown to rival Bitcoin simply serves to reinforce the dominance of the existing arrangements. ## 5 Case study: Sovrin Sovrin is a “global public, permissioned identity utility for exchanging identity more securely” (Patel, 2018) based on a distributed ledger overseen by the Sovrin Foundation. It is based on opensource blockchain software and trusted participants that issue and verify identities and other identifying pertinent information about natural and juristic persons. The main aim of the project is to facilitate the reuse of verified information (with the permission of the data subject), to incentivise the release of information and to record the withdrawal of the right to use such information (Ldapwiki, 2018). ### 5.1 Description of the Sovrin network Figure 3 shows the Sovrin Goverance Network in which the Sovrin Governance Framework Master Document defines the “constitution” of the Sovrin Network laying down the purpose, core principles and links to other main documents. The Sovrin organisation is formally constituted as a nonprofit organisation incorporated in Utah, USA on February 2 2018. Its purposes include, but are not limited to (a) To develop, govern and promote an international nonprofit private sector self-sovereign digital identity system based on the Sovrin distributed ledger; (b) To own, lease, sell, exchange or otherwise deal with all property, real and personal, tangible or intangible, to be used in furtherance of these purposes; and (c) To engage in any and all lawful activities incidental, useful or necessary to the accomplishment of the above-referenced purposes.[2] While it has no legally-defined members, the term “members” is used “to refer to donors, technology contributors, ledger stewards, 2Sovrin Articles of Incorporation, February 2, 2018. [https://drive.google.com/file/d/](https://drive.google.com/file/d/1QC7Ma9DZUiOjY3G4S1URLXD2CBJzvxqw/view) [1QC7Ma9DZUiOjY3G4S1URLXD2CBJzvxqw/view](https://drive.google.com/file/d/1QC7Ma9DZUiOjY3G4S1URLXD2CBJzvxqw/view) 17 ----- Figure 3: The Sovrin Governance Network members of Corporation committees or work groups, and other participants in the Sovrin community whose roles may be further defined in the Bylaws, agreements, or other governing documents”.[3] Its principal office and mailing address is identified as 151 S 1150 E, Lindon, UT 84042. The Sovrin Foundation is overseen by a Board of Trustees with no less than three and no more than twenty one members. The original Trustees are identified as Phillip J Windley and Jason A Law, both of Utah, and Drummond S Reed of Washington State. In mid January 2019, it comprised 12 members. A nominations committee selected by the Board identifies eligible nominees, who are elected annually at the Annual Meeting. Trustees are not remunerated, but are (subject to approval by the Board), reimbursed for expenses incurred on behalf of the organisation. The Board appoints an Executive Director to supervise operations. The Chief Executive Officer Heather Dahl is the Executive Director. Along with the CEO, the Chief Financial Officer Roy Avondet, Chief Technology Officer Nathan George and Director of Marketing Helen Garneau comprise the Executive Leadership. Seventeen other staff are identified on the website. The Board also has the power to create Advisory Councils and other committees as required, in addition to the Executive Committee and Finance Committee identified in the Bylaws. In mid-January 2019, a fifteen-member Technical Governance Board, including CTO Nathan George and founder-trustee Jason Law is identified as having been appointed to govern the “technical design, architecture, implementation and operation of the Sovrin Network as a global public utility for self-sovereign identity”. Day-to-day activities are managed by an executive team comprising CEO Heather Dahl, Chief Financial Officer Roy Avondet, Chief Technology Officer Nathan George and Director of Marketing Helen Garneau. CEO Heather Dahl is one of the twelve Board Members. Seventeen staff are identified on the website. The Board 3Sovrin Bylaws, Jan 31 2018 https://drive.google.com/file/d/1kkuiEp0vA620ydcAND9pIY_hLHVjHKsG/view 18 ----- Central to the mode of governance within the Sovrin network is the Sovrin Steward Agreement document specifying the legal obligations, liabilities, etc. for Stewards and the Sovrin Foundation[4]. The agreement, governed by the law of the State of Delaware,contains explicit provisions to be followed in the event of a dispute between the Stewards and Sovrin. The trusted participants acting as node members in the project are called Stewards. In mid January 2019, forty-eight are identified on the Sovrin website.[5] They include banks, telecommunications companies and universities as well as IT companies such as Cisco and IBM. On the Sovrin network, it is these Stewards that approve transactions for inclusion in the ledger and that, in fact, submit transactions for inclusion (Tobin, 2018). There is a strong emphasis on anonymous identity in ledger records (as actual users can create unique reference numbers for each relationship) as well as zero-knowledge proofs where a user can, for example, use the system to prove that they are over 18 without revealing their actual age – based on the trusted information embedded in the system. This feature is unique to the Sovrin Identity Network (SIDN) which allows to create self-sovereign identity (SSI) for end user members (Muhle et. al 2018). The software used has been part of The Linux Foundation’s Hyperledger project since 2017 under the name Hyperledger Indy.[6] Like the Bitcoin software, it is opensource. Sovrin facilitates engagement of individuals in the Hyperledger Indy project, including direct links on its website to the weekly Indy Group calls, Chat room and Mailing List (https://sovrin.org/developers/). No sensitive data is stored in the DL at all – only the Stewards’ identifying information as well as pointers to the end-users’ data are included in the ledger. Stewards act as validators (similar to Bitcoin miners) as well as clients (who submit transactions). ### 5.2 Adding to the ledger The validation of information submitted to the Sovrin network is entirely the work of the Stewards. This is strikingly dissimilar to the case of Bitcoin where anyone may start mining. One cannot be a Steward without formally entering into an agreement with the Foundation. The validity of transaction is only vouched for by other Stewards and cannot necessarily (as in the case of Bitcoin) be easily checked by anyone with access to the ledger. ### 5.3 Forking the chain It is definitely technically possible for a subset of Stewards to agree to defect with a current copy of the ledger but this act itself would deprive them of the governance arrangements embodied by the Sovrin Foundation. The consensus algorithm in Sovrin is called plenum, an enhancement of the redundant byzantine fault tolerance algorithm.[7] In most general terms, this is a voting algorithm that executes very quickly and resolves faults possibly introduced by errant nodes. 4https://sovrin.org/library/steward-agreement/ 5https://sovrin.org/stewards/ [6https://github.com/hyperledger/indy-sdk/blob/rc/doc/getting-started/getting-started.md](https://github.com/hyperledger/indy-sdk/blob/rc/doc/getting-started/getting-started.md) [7https://github.com/hyperledger/indy-plenum/wiki](https://github.com/hyperledger/indy-plenum/wiki) 19 ----- ### 5.4 Informal and formal governance Essentially all governance in the case of Sovrin is formal. There is however a certain devolution of power as regards the Stewards’ dealing with individuals. ### 5.5 Applying the analytical framework to Bitcoin Applying Figures 1 and 2 to Sovrin, the following membership stakes and interactions are identified: - The three founder-trustees appear to have acted as initial donor-members. At least one of these has a technical background, so appears to have been an original software-member. As Sovrin has expanded, the Technical Governance Board appears to have assumed responsibility for software oversight. This includes “reviewing the technical qualifications of the Steward candidates to ensure they meet the requirements and principles of the Sovrin Trust Framework” - suggesting a degree of central control over not just the blockchain application but also the applications used by node-member stewards in their engagement with transaction-members and end user-members (the Sovrin Steward Agreement 3b obliges the Stewards to “only run software code that has been approved by the Sovrin Foundation as referenced in the Sovrin Governance Framework”). - Stewards as node-members are central. Entry is strictly controlled by Sovrin, and steward identities are known to all other members. Only stewards can raise transactions on the ledger (although other entities may read data from it), so they fulfil the role of transactionmembers as well. The reliability of the identification system relies upon the fact that the stewards themselves can be trusted by other stewards and users of identity information. The steward agreement does not mention any membership payments, but there is likely some considerable expense involved in satisfying the Foundation that admission as a steward will not bring undue risk to the DLS. The more arduous are these processes, the more costly they are for the firms concerned. Only firms who are serious about belonging will put in the effort; if they walk away from the arrangements (i.e. instigate a fork) this cost is truly sunk. This means the number of nodes is likely to be much smaller than a public system like Bitcoin, but it also means that the stewards are both known and can be pursued for any costs caused by seceding. - End user-members may have many different identities on the ledger, and be effectively anonymous on it, but each has to raise identity transactions via a Steward who is prepared to vouch for their identity when posting to the ledger. Hence end user-members cannot be anonymous in respect of at least that Steward. The knowledge comes from interactions the stewards have with individuals in other ways, for example as customers of a financial institution or telecommunications provider. Stewards thus provide the bedrock of trust in the identities lodged on the DLS. User-member participation will not be possible without an of-system interaction with at least one steward. The Sovrin system is in a very early stage of development. It is not clear yet how payments for services undertaken will be made. However, Sovrin does embody a token system to reward end-user and steward/node participation. In contrast to Bitcoin, Sovrin is a new application with a business case that is yet to be fully 20 ----- proven in any context - either physical or virtual. It is clear that the current Sovrin Framework is a ‘work in progress’, though one which has moved from the truly novel innovation to small-scale testing phase. Arguably, the stewards who have joined so far are participating as a means of learning how they can make use of the system as it scales up rather than to achieve an already clearly-understood outcome. It is not clear at this point what the future might bring for this venture. However, from our analysis, the much tighter control exerted from the centre, with clearly-articulated responsibilities, rules and disputes resolution processes binding all parties is consistent with a system where future developments are uncertain. The strict governance rules allow for greater flexibility in the direction that can be taken in incorporating rules into the software. While current abilities for Stewards to formally influence directions are less than for the software and donor members of the system, they exert significant commercial power as they mediate the relationships with end users. This close economic co-dependence between them and the Sovrin Foundation affords them a degree of influence in the governance of their DLS far greater than that of the Bitcoin miners over their DLS. ## 6 Discussion of the case studies Ultimately, in the case of Bitcoin, the validity of the entire blockchain can be checked by any interested party. This includes that veracity of the solutions to the mining problems. It also includes checking that the cryptographic signatures of the individual transactions are valid. The only thing that is not possible to check is whether all other nodes might not have conspired to pick certain valid mine blocks rather than others but the Bitcoin protocol (involving free entry) makes it rather intuitively unlikely that this would take place. It has been suggested by Vitalik Buterin, creator of Ethereum, that the degree of (de)centralisation of a network can be examined along three axes (Siriwardena, 2017). 1. Architecture (de)centralisation – what is the physical nature of the system and how robust is it? 2. Political (de)centralisation – how is membership of and participation in the system governed? 3. Logical (de)centralisation – how flexible are interfaces and data structures in the system? Considering Bitcoin and Sovrin, we suggest that the decentralisation of the two systems can be categorised as follows. Degree of centralisation Bitcoin Sovrin Architectural Low Medium Political Low High Logical High High Centralisation is linked to governance, a topic to be explored further in future work. 21 ----- ## 7 Conclusion The authors have investigated distributed ledger (DL) governance in the context of the theory of clubs. The incentives to passively join and to take part in operating the consensus mechanism of the DL can be understood using this theory. The examples of Bitcoin and Sovrin illustrate how formal and informal arrangements operate – either through formal agreement or through technical arrangements embedded in the software. We note that DL systems tend to be effectively controlled by a small coalition of softwaremembers, who may also participate as node-members or be closely affiliated with influential node-members. In order to motivate their participation, it would be expected they anticipate remuneration from either their node operation activities, or some other arrangement such as an honorarium paid from contributions made by donor-members. The more donor members there are, the more likely it is that formal articulation of DLS governance arrangements outside of the software itself would be required. ## References Berg, Alastair, Berg, Chris, & Novak, Mikayla. 2018a. Blockchains and Constitutional Catallaxy. Available at SSRN 3295477. Berg, Chris, Novak, Mikayla, Potts, Jason, & Thomas, Stuart J. 2018b. From Industry Associations to Ecosystem Associations: Blockchain, Interest Groups and Public Choice. Interest Groups and Public Choice (November 16, 2018). Berle, Adolph, & Means, Gardiner. 1932. Private property and the modern corporation. New York: Mac-millan. Buchanan, James M. 1962. Predictability: The criterion of monetary constitutions. In search of a monetary constitution, 155–83. Buchanan, James M. 1965. An economic theory of clubs. Economica, 32(125), 1–14. Buchanan, James M. 1987. The constitution of economic policy. The American economic review, 77(3), 243–250. Böhme, Rainer, Christin, Nicolas, Edelman, Benjamin, & Moore, Tyler. 2015. Bitcoin: Economics, Technology, and Governance. Journal of Economic Perspectives, 29(2), 213–38. http://www.aeaweb.org/articles?id=10.1257/jep.29.2.213 Cordery, Carolyn, & Howell, Bronwyn. 2017. Ownership, Control, Agency And Residual Claims In Healthcare: Insights On Cooperatives And Non-profit Organizations. Annals of Public and Cooperative Economics, 88(3), 403–424. Cornes, Richard, & Sandler, Todd. 1996. The theory of externalities, public goods, and club goods. Cambridge University Press. Crosby, Michael, Nachiappan, Pattanayak, Pradhan, Verma, Sanjeev, & Kalyanaraman, Vignesh. 2015. Blockchain technology: beyond Bitcoin. Sutardja Center for Entrepreneurship & Technology. http://scet.berkeley.edu/wp-content/uploads/BlockchainPaper.pdf 22 ----- Czepluch, Jacob Stenum, Lollike, Nikolaj Zangenberg, & Malone, Simon Oliver. 2015. The use of block chain technology in different application domains. The IT University of Copenhagen, Copenhagen. De Filippi, Primavera, & Loveluck, Benjamin. 2016. The invisible politics of Bitcoin: governance crisis of a decentralised infrastructure. Internet Policy Review, 5(3). Elinor, Ostrom. 1990. Governing the commons: the evolution of institutions for collective action. Foroglou, George, & Tsilidou, Anna-Lali. 2015. Further applications of the blockchain. Columbia University PhD in Sustainable Development, 10. Fukuyama, Francis. 2014. Political Order and Political Decay: From the Industrial Revolution to the Globalization of Democracy. New York: Farrar, Straus and Giroux, 455–466. Hansmann, Henry. 1996. The changing roles of public, private, and nonprofit enterprise in education, health care, and other human services. Pages 245–276 of: Individual and social responsibility: Child care, education, medical care, and long-term care in America. University of Chicago Press. Krecké, Elisabeth. 2004. 14 The emergence of private lawmaking on the Internet. Markets, Information and Communication: Austrian Perspectives on the Internet Economy, 289. Ldapwiki. 2018. Sovrin. Retrieved on 2019-01-13. https://ldapwiki.com/wiki/Sovrin Leiner, Barry M, Cerf, Vinton G, Clark, David D, Kahn, Robert E, Kleinrock, Leonard, Lynch, Daniel C, Postel, Jon, Roberts, Larry G, & Wolff, Stephen. 2009. A brief history of the Internet. ACM SIGCOMM Computer Communication Review, 39(5), 22–31. Mattila, Juri. 2016. The blockchain phenomenon. Berkeley Roundtable of the International Economy. Mazieres, David. 2015. The stellar consensus protocol: A federated model for internet-level consensus. Stellar Development Foundation. Mulligan, CJ, Scott, Z, Warren, S, & Rangaswami, JP. 2018. Blockchain Beyond the Hype. In: World Economic Forum. http://www3.weforum.org/docs/48423_Whether_Blockchain_WP.pdf. Accessed, vol. 2. Narayanan, Arvind, & Clark, Jeremy. 2017. Bitcoin’s academic pedigree. Communications of the ACM, 60(12), 36–45. Narayanan, Arvind, Bonneau, Joseph, Felten, Edward, Miller, Andrew, & Goldfeder, Steven. 2016. Bitcoin and cryptocurrency technologies: a comprehensive introduction. Princeton University Press. Olson, Mancur. 1989. Collective Action. London: Palgrave Macmillan UK. Pages 61–69. https://doi.org/10.1007/978-1-349-20313-0_5 Ostrom, Elinor. 2005. Understanding institutional diversity. Princeton University Press. Ostrom, Elinor. 2010. Beyond markets and states: polycentric governance of complex economic systems. American economic review, 100(3), 641–72. 23 ----- Ostrom, Vincent. 2014. Polycentrictiy: The Structural Basis of Self-Governing Systems. Choice, Rules and Collective Action: The Ostrom’s on the Study of Institutions and Governance, 45. Patel, Milan. 2018. IBM Blockchain Trusted Identity: Sovrin Steward closed beta offering. Retrieved on 2019-01-13. https://www.ibm.com/blogs/blockchain/2018/08/ibm-blockchaintrusted-identity-sovrin-steward-closed-beta-offering/ Reijers, Wessel, O’Brolcháin, Fiachra, & Haynes, Paul. 2016. Governance in blockchain technologies & social contract theories. Ledger, 1, 134–151. Swan, Melanie. 2015. Blockchain: Blueprint for a new economy. O’Reilly Media, Inc. Szabo, Nick. 1997. Formalizing and securing relationships on public networks. First Monday, 2(9). Tarko, Vlad, Schlager, Edella, & Lutter, Mark. 2018. The Faustian Bargain: Power-Sharing, Constitutions, and the Practice of Polycentricity in Governance. Governing Complexity: Analyzing and Applying Polycentricity, eds. William A. Blomquist, Dustin Garrick and Andreas Thiel (Cambridge University Press, Forthcoming). Tobin, Andrew. 2018. Sovrin: What Goes on the Ledger? Retrieved on 2019-01-11. https://sovrin.org/wp-content/uploads/2018/10/What-Goes-On-The-Ledger.pdf van Wirdum, Aaron. 2016 (Apr). Who Funds Bitcoin Core Development? How the Industry Supports Bitcoin’s ‘Reference Client’. https://bitcoinmagazine.com/articles/who-funds-bitcoincore-development-how-the-industry-supports-bitcoin-s-reference-client-1459967859/ Williamson, OE. 1985. 1985: The economic institutions of capitalism. Firms, markets, relational contracting. New York: Free Press. Williamson, Oliver E. 1999. Strategy research: governance and competence perspectives. Strategic management journal, 20(12), 1087–1108. Williamson, Oliver E. 2000. The new institutional economics: taking stock, looking ahead. Journal of economic literature, 38(3), 595–613. Windley, Phillip J. 2016. How Sovrin Works. Retrieved on 2019-01-11. https://sovrin.org/wpcontent/uploads/2018/03/How-Sovrin-Works.pdf 24 -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.2139/ssrn.3365519?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.2139/ssrn.3365519, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBYNCND", "status": "GREEN", "url": "https://www.econstor.eu/bitstream/10419/201737/1/ITS2019-Aswan-paper-24.pdf" }
2,019
[]
true
2019-02-01T00:00:00
[ { "paperId": "50f1ae445429b401d3fbce145266b05f37f9af62", "title": "Blockchains and constitutional catallaxy" }, { "paperId": "2ac3eb4a601b831c2c88d4960f5cc57cddb10d68", "title": "The Faustian Bargain: Power-Sharing, Constitutions, and the Practice of Polycentricity in Governance" }, { "paperId": "50e03e0797144e6d69aaeb89a024dd6124fccab7", "title": "From Industry Associations to Ecosystem Associations: Blockchain, Interest Groups and Public Choice" }, { "paperId": "1b5dd2f547344ec010a79c2d228ca2470921dab4", "title": "Ownership, Control, Agency and Residual Claims in Healthcare: Insights on Cooperatives and Non‐Profit Organizations" }, { "paperId": "ce64501d692952193be72d450cabe285f7637b9f", "title": "Bitcoin's academic pedigree" }, { "paperId": "01f08210264ee1f338a86e5c95a00d7531c83fe2", "title": "Governance in Blockchain Technologies & Social Contract Theories" }, { "paperId": "2d04c4388dd2b5929cd273b8d1cfb197304cf52b", "title": "The Invisible Politics of Bitcoin: Governance Crisis of a Decentralized Infrastructure" }, { "paperId": "c2de5385c197aab309abd859b36bee1362147688", "title": "Bitcoin and Cryptocurrency Technologies - A Comprehensive Introduction" }, { "paperId": "97fddbbfd681bce9eeb8e0a013353b4d5b2ba0db", "title": "Blockchain: Blueprint for a New Economy" }, { "paperId": "f02d4551fb7832a5209b7c05c2f46a335217bd92", "title": "Political order and political decay: From the industrial revolution to the globalization of democracy" }, { "paperId": "7b8de1a148d61005e92d2cc0de745674fb6faf2f", "title": "Bitcoin: Economics, Technology, and Governance" }, { "paperId": "3a5c55353bb7d4b29bc6ed45d062e78bd8291a66", "title": "Beyond Markets and States: Polycentric Governance of Complex Economic Systems" }, { "paperId": "78be4997d0e889fe88ad2fd3d5dc830219d76103", "title": "The emergence of private lawmaking on the Internet" }, { "paperId": "11040e39cb84397ac26f063058fefcb394da9400", "title": "The New Institutional Economics: Taking Stock, Looking Ahead" }, { "paperId": "1c49353abaea57d33f3297c2ab0919aa52dea5be", "title": "STRATEGY RESEARCH: GOVERNANCE AND COMPETENCE PERSPECTIVES" }, { "paperId": "44773b8f9248b861e80e76166f42e47db556f88d", "title": "A brief history of the internet" }, { "paperId": "5b4cf1e37954ccd1ca6b315986d45904f9d2f636", "title": "Formalizing and Securing Relationships on Public Networks" }, { "paperId": "f1f7964241c12d4c55e2c46195791c115f754080", "title": "Governing the Commons: The Evolution of Institutions for Collective Action" }, { "paperId": "d8c08a4abbd5a601df2c7e11d9b754b1d9af9f34", "title": "The constitution of economic policy." }, { "paperId": "7254b5d903ebe9ec164706304080458e7d821b03", "title": "The Economic Institutions of Capitalism: Firms, Markets, Relational Contracting." }, { "paperId": "81c7ca626f1587b342ccc79eb37156fc5c2ed2f7", "title": "The Theory of Externalities, Public Goods, and Club Goods" }, { "paperId": "3dfb55a740236ab335674b3697c7268d5f75d01e", "title": "VI PREDICTABILITY: THE CRITERION OF MONETARY CONSTITUTIONS" }, { "paperId": "a13b9f1931a71c3f958bec5f62b88cbed1041e93", "title": "In search of a monetary constitution" }, { "paperId": null, "title": ": What Goes on the Ledger?" }, { "paperId": null, "title": "Sovrin: What Goes on the Ledger? Retrieved on 2019-01-11" }, { "paperId": null, "title": "IBM Blockchain Trusted Identity: Sovrin Steward closed beta offering" }, { "paperId": null, "title": "Blockchain Beyond the Hype" }, { "paperId": null, "title": "Who Funds Bitcoin Core Development? How the Industry Supports Bitcoin’s ‘Reference Client’" }, { "paperId": null, "title": "How Sovrin Works" }, { "paperId": null, "title": "The blockchain phenomenon. Berkeley Roundtable of the International Economy" }, { "paperId": "3babb89369eed603ce3c702f447ee6274429cda1", "title": "The Stellar Consensus Protocol: A Federated Model for Internet-level Consensus" }, { "paperId": null, "title": "Further applications of the blockchain" }, { "paperId": null, "title": "Blockchain technology: beyond Bitcoin" }, { "paperId": null, "title": "The use of block chain technology in different application domains. The IT University of Copenhagen, Copenhagen" }, { "paperId": "fde4ebe0f9747f285df5cfaabd95b54e59ecd3e4", "title": "Collective Action" }, { "paperId": null, "title": "Polycentrictiy: The Structural Basis of Self-Governing Systems. Choice, Rules and Collective Action: The Ostrom's on the Study of Institutions and Governance" }, { "paperId": "46881a8dfd3655945909ded6a9c2ff17c2703191", "title": "Understanding Institutional Diversity" }, { "paperId": "588d7b2ce8cbdf172f3321594aff458282170e65", "title": "The Economic Theory of Clubs" }, { "paperId": "5652ad2015c852f02df63c5cc5c9a38e0920305c", "title": "The Changing Roles of Public, Private, and Nonprofit Enterprise in Education, Health Care, and Other Human Services" }, { "paperId": null, "title": "1985: The economic institutions of capitalism" }, { "paperId": null, "title": "Private property and the modern corporation" }, { "paperId": null, "title": "b) To own, lease, sell, exchange or otherwise deal with all property, real and personal, tangible or intangible, to be used in furtherance of these purposes" } ]
17,699
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/010f3407c141dfafbcddf8db50f4b8914f9a3d0d
[ "Computer Science" ]
0.87953
Scheduling for Responsive Grids
010f3407c141dfafbcddf8db50f4b8914f9a3d0d
Journal of Grid Computing
[ { "authorId": "3086404", "name": "C. Germain" }, { "authorId": "143932705", "name": "C. Loomis" }, { "authorId": "29120186", "name": "J. Moscicki" }, { "authorId": "48840171", "name": "R. Texier" } ]
{ "alternate_issns": null, "alternate_names": [ "J Grid Comput" ], "alternate_urls": [ "https://link.springer.com/journal/10723", "http://www.springer.com/journal/10723" ], "id": "993eb4fa-4cb7-4aed-980e-20e5298efad0", "issn": "1570-7873", "name": "Journal of Grid Computing", "type": "journal", "url": "https://www.springer.com/journal/10723" }
null
#### EGEE-PUB-2008-002 # Scheduling for Responsive Grids ## Germain-Renaud, C LRI - CNRS and Universit Paris-Sud et al ### 20 December 2008 Journal of Grid Computing #### EGEE-II is a project funded by the European Commission Contract number INFSO-RI-031688 The electronic version of this EGEE Publication is available on the CERN Document Server at the following URL: ``` <http://cdsweb.cern.ch/search.py?p=EGEE-PUB-2008-002> #### EGEE-PUB-2008-002 ``` ----- ### Scheduling for Responsive Grids C´ecile Germain-Renaud LRI and LAL Charles Loomis LAL Jakub T. Mo´scicki CERN Romain Texier LRI June 2006 Abstract. Grids are facing the challenge of seamless integration of the grid power into everyday use. One critical component for this integration is responsiveness, the capacity to support on-demand computing and interactivity. Grid scheduling is involved at two levels in order to provide responsiveness: the policy level and the implementation level. The main contributions of this paper are as follows. First, we present a detailed analysis of the performance of the EGEE grid with respect to responsiveness. Second, we examine two user-level schedulers located between the general scheduling layer and the application layer. These are the DIANE (DIstributed ANalysis Environment) framework, a general-purpose overlay system, and a specialized, embedded scheduler for gPTM3D, an interactive medical image analysis application. Finally, we define and demonstrate a virtualization scheme, which achieves guaranteed turnaround time, schedulability analysis, and provides the basis for differentiated services. Both methods target a brokering-based system organized as a federation of batch-scheduled clusters, and an EGEE implementation is described. Keywords: Responsiveness, Interactive Grids, Meta-scheduler, User-level Scheduling 1. Introduction The exponential increases in network performance and storage capacity [41], together with ambitious national and international efforts, have already enabled the virtualization and pooling of processors and storage in advanced and relatively stable systems such as the EGEE grid. However, it is more and more evident that the exploitation model for these grids is somehow lagging behind. At a time where industry acknowledges interactivity as a critical requirement for enlarging the scope of high performance computing [35, 43, 6], grids cannot anymore be envisioned only as very large computing centres providing batch-oriented ⃝c 2006 Kluwer Academic Publishers. Printed in the Netherlands. GCJv14.tex; 17/12/2006; 0:08; p.1 ----- 2 Grid Scheduling for Interactive Analysis access to complex scientific applications with high job throughput as the primary performance metric. A much larger range of grid usage scenarios is possible. Seamless integration of the grid power into everyday use calls for unplanned and interactive access to grid resources. We define responsive grids as grid infrastructures that support on-demand computing and interaction. This paper describes a set of scheduling methods providing different levels and types of Quality of Service (QoS) required by responsiveness. Compared to many recent proposals in this area, our methods target production grids. They have been implemented within EGEE, on top of the gLite middleware. EGEE is the largest production grid worldwide, comprising more than 20000 CPUs, 200 sites and 20000 jobs per day and requiring the strongest constraints on dependability. In this framework, responsiveness must be built on top of the traditional grid scheduling tools, which are batch-oriented and dominated by fair-share policies at institutional time-scales. The associated constraints are: − delays incurred by non-interactive jobs are bounded, − resource utilization is not degraded (e.g. by idling processors), and − the local policies governing resource sharing (Virtual Organizations, advance reservation, etc. ) are not impacted. This rest of this paper is organized as follows. Section 2 describes use-cases for grid responsiveness. Section 3 presents the scheduling architecture of the EGEE grid and an experimental study of the EGEE profiles of execution time and overhead. Section 4 presents two examples of user-level scheduling deployed on top of gLite, which is the EGEE middleware. The first one exemplifies a generic overlay system. The second one is an application-dedicated environment, which exemplifies grid-enabled computational steering in medical image analysis. We show that user-level scheduling does improve the quality of service, by eliminating the middleware overhead, providing a sustained job output rate, and optimizing the failure recovery. On the other hand, user-level scheduling is limited to best-effort. Section 5 presents the Virtual Reservation scheme, which provides guarantees on the overall turnaround time, and its implementation inside gLite. Section 5 discusses related work, and Section 6 presents the conclusions. GCJv14.tex; 17/12/2006; 0:08; p.2 ----- Scheduling for Responsive Grids 3 2. Motivation Responsiveness is a key component for real-world grid usage; this section presents a few examples. The first one is grid-enabling medical image analysis [11, 45, 23]. In a clinical context, medical image analysis (segmentation, registration) and exploitation (augmented reality for intervention planning or intra-operative support) require full interaction because computer programs are not yet competitive with the human visual system for mining these structured and noisy data. Analyzing large images at a sufficient speed to support smooth visualization requires not only substantial computing power, which can be provided by the grid, but also unplanned access and sophisticated interaction protocols. The second use case is digital libraries. Most of the resource consumption in digital libraries management is related to bulk, off-line tasks such as indexing. When humans query this massive amount of data, various actions are triggered such as feature extraction in a queryby-example scheme, which must take place before the actual search can be carried out, or content protection (e.g. watermarking). User satisfaction requires nearly instantaneous response. Finally, in the larger perspective of ubiquitous computing and ambient intelligence, multi-modal interfaces that are capable of natural and seamless interaction with and among individual human users are mandatory. Responsiveness is a key component for grid-enabling the methods and technologies that form the back-end of these interfaces, such as pattern analysis, statistical modelling and computational learning. Interactive grid applications require a specific grid guarantee, namely a bound on the overall turnaround time of the grid jobs contributing to the application. Because such jobs have typically a short execution time and require completion by a deadline, we call them Short Deadline Jobs (SDJ) in the remainder of this paper. 3. A case for responsiveness 3.1. EGEE Scheduling EGEE combines globally-distributed computational and storage resources into a single production infrastructure available to EGEE users. Each participating site configures, runs, and maintains a batch system containing its computational resources and makes those resources available to the grid via a gatekeeper. The scheduling policy for each GCJv14.tex; 17/12/2006; 0:08; p.3 ----- 4 Grid Scheduling for Interactive Analysis site is defined by the site administrator. Common scheduling policies use either FIFO (often with per-user or per-group limits) or fair-share algorithms. Consequently the overall EGEE scheduling policy is not centrally defined, but the effect of the interaction of largely autonomous policies. The gLite middleware deployed on the EGEE infrastructure integrates the sites’ computing resources through the Workload Management System (WMS) [3]. The WMS is a set of middleware-level services responsible for the distribution and management of jobs. The site computational resources present a common interface to the WMS, the Computing Element (CE) service. The CE specification is one of the core parts of the Glue information model [4], which is the current basis for interoperability between EGEE and other grids. From the middleware point of view, a CE has multiple functions: running jobs, staging the files required by the job, providing information about resource availability, and notifying the WMS of the job-related events. In the framework of this paper, a CE can be simply considered as a batch queue, subject to the above-mentioned policies. The core of the WMS is the Workload Manager which accepts jobs from users and dispatches them to computational resources based on the users requirements on one hand, and the characteristics (e.g. hardware, software, localization) and state of the resources on the other hand. The WM is implemented as a distributed set of resource brokers, with some tens of them currently installed; all the brokers get an approximatively consistent view of the resource availability through the grid information system. Each broker reaches a decision of which resource should be used by a matchmaking process between submission requests and available resources. Job requirements are exposed to the various services of the WMS via the Job Description Language (JDL) [38], derived from the Condor ClassAd language [39]. The users can rank acceptable resources (in JDL language) by using an arbitrary expression which uses state information published by the resources. Once a job is dispatched, the broker only reschedules it if it failed; it does not reschedule jobs based on the changing state of the resources. 3.2. EGEE usage The relevant quantities for measuring the responsiveness of the grid are the running time t, the on-site queuing delay q, and the middleware overhead s, which includes the various delays experienced by the job in the WMS. The turnaround time m = s + q + t is the total time from submission to notification that the job has completed. For the study presented here, these quantities were derived from information in the GCJv14.tex; 17/12/2006; 0:08; p.4 ----- Scheduling for Responsive Grids 5 1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55100 101 102 103 104 105 106 t(seconds) Figure 1. Cumulative distribution of execution times. Logging and Bookkeeping service (LB). This is a companion service to the resource broker which maintains the state of all jobs managed by the resource broker. Because the detailed LB data were not available for all jobs, the analysis below is limited to a particular broker (grid09.lal.in2p3.fr). These data cover one year (October 2004 to October 2005) and include more than 50000 successful production jobs from 66 distinct users. Fig. 1 shows the distribution of execution time from this trace. The striking feature is the importance of short jobs: the 80% quantile is 20 s. The second important point is the dispersion of t; the mean is 2 s, but the standard deviation is of the order of 10[4] s. The very large fraction of extremely short jobs is partially due to the high usage of this particular broker by the EGEE Biomed Virtual Organization. However, for more than 50% of the overall EGEE jobs at the same period, the execution time is less than 3 minutes. Fig. 2 shows the distribution of the dimensionless overhead factor or = (m − t)/t, which is the overhead normalized by the execution time. The leftmost histogram shows the distribution of the full sample: only 26% of the 53000 jobs are in the first bin, meaning than 74% of the jobs suffer an overhead factor larger than 25. A closer look at the small overheads (rightmost histogram) shows that only 13% of the jobs experience an overhead factor lower than 2. Clearly, the EGEE infrastructure can make no claims for responsiveness using only the base middleware services. The next question is the respective impacts of the middleware and the queuing time on the global overhead. Fig. 3 plots the distribution GCJv14.tex; 17/12/2006; 0:08; p.5 ----- 6 Grid Scheduling for Interactive Analysis 7E3 1.4E4 6E3 1.2E4 1.0E4 5E3 8.0E3 4E3 6.0E3 3E3 4.0E3 2E3 2.0E3 1E3 0 0 100 200 300 400 500 0 0 10 20 30 40 50 (m−t)/t (m−t)/t Figure 2. Distribution of the overhead factor.The left histogram is the distribution of the full sample, the right histogram is the distribution of the small overheads. 8000 7000 6000 5000 4000 3000 2000 1000 0 0 0.5 1 1.5 2 2.5 3 q/s Figure 3. Impact of the queuing time and the middleware on the overhead. of q/s, and shows that the queuing time is a significant component of the overhead. This behaviour was exhibited at an early stage of EGEE usage, where the pressure on the resource was only starting to increase. Finally, the median queuing time is 91 seconds, and the median middleware overhead is 221 seconds. 4. User-level scheduling Submitting, scheduling and mapping of jobs on a grid take at least one order of magnitude more time than the execution time for SDJ even in absence of competition for resources. For instance, with the most recent and tuned EGEE middleware, gLite 3.0, the middleware latency remains on the order of minutes. User-level scheduling is the GCJv14.tex; 17/12/2006; 0:08; p.6 ----- Scheduling for Responsive Grids 7 most promising way to address the difference of scale between short execution times and the large grid middleware latencies. User-level (or application-level) scheduling is a virtualization layer on the application side. Instead of being executed directly, the application is executed via an overlay scheduling layer (user-level scheduler). The overlay scheduling layer runs as a set of regular user jobs and therefore it operates entirely inside user space. Because user-level scheduling does not require any modification to the Grid middleware and infrastructure nor the deployment of special services in the Grid sites, it provide immediate exploitation of the full range of a Grid sites which are available for a given user. The user-level scheduling approach has the following constraints: − user jobs must be instrumented with the scheduling functionality, and − jobs run under user-level scheduling must compete on the same basis with all other jobs on the grid, and their resource usage be fully reported to the corresponding user. A user-level scheduler may be embedded into the application or external to it. A scheduler embedded into the application is developed and optimized specifically for a given application, typically by re-factoring and instrumenting the original application code. It allows fine tuning and customizing the scheduling according to the specific execution patterns of the application. Such a scheduler is intrusive at the application source code level which means that the code reuse of the scheduler is reduced and the development effort is high for each application. A scheduler external to the application relies on the general properties of the application such as a particular parallel decomposition pattern (e.g. iterative decomposition, geometric decomposition or divide-andconquer). An application adapter connects the external scheduler to the application at runtime. Depending on the decomposition pattern, the application re-factoring at the source code level may or may not be required. The disadvantage of external schedulers is that it may be very hard to generalize execution patterns for irregular or speculative parallelism. In this case, which occurs in various situations ranging from medical image processing to portfolio optimization [50], a development of a specialized embedded scheduler may be necessary. In the next sections we examine two user-level schedulers: an external scheduler for generic master-worker applications (DIANE) and an embedded scheduler for medical image processing (gPTM3D). GCJv14.tex; 17/12/2006; 0:08; p.7 ----- 8 Grid Scheduling for Interactive Analysis 4.1. DIANE: a generic, external scheduler 4.1.1. Overview DIANE (DIstributed ANalysis Environment) is a R&D project developed in the Information Technology Department at CERN in Geneva Switzerland. [36]. It is a generic user-level scheduler based on the extended task farm (master/slave) processing . The runtime behaviour of the framework, such as failure recovery or task dispatching, may be customized with a set of hot-pluggable policy functions. This enables fine-tuning of the scheduler according to the needs of particular application and provides support for other parallel decomposition patterns (e.g. divide-and-conquer). 4.1.2. Applications DIANE provides a python-based framework and enables a rapid integration with existing applications. Both transparent and intrusive application integrations have been demonstrated. Data analysis in Athena framework for Atlas experiment [1], is an example of transparent application integration; the application adapters in the form of python packages have been developed without modifying the original application code. The examples of intrusive integrations include the particle simulation in medical physics using Geant 4 toolkit [22]. The parallelization of these applications has been based on the iterative decomposition and master/worker processing model with fully independent tasks. Other applications using DIANE include, among others, the Geant 4 statistical regression testing application [34], Autodock [10] tools for bioinformatics and telecommunication applications [32]. 4.1.3. Execution model In the DIANE execution model, a temporary virtual master/worker overlay network is created for each user job and is destroyed when the job terminates. The job is split into a number of tasks which are executed by a number of lightweight worker agents in the Grid. The worker agents run as regular Grid jobs submitted with credentials and authentity of a single user. Therefore the full user-based accounting from the system administration point of view is possible. The agents are time-limited and the computing resources are released when the processing terminates (all tasks processed) or if they exceed the time limit on the batch queue, whatever occurs first. The number of resouces acquired by a user is limited by standard mechanisms i.e. the fair-share policies in the Grid and in the local sites. Each task is defined by a set of application-specific parameters. The dispatching of tasks is the process of allocating the tasks to workers by GCJv14.tex; 17/12/2006; 0:08; p.8 ----- Scheduling for Responsive Grids 9 sending appropriate parameters to the worker agents. The communication overhead is typically much smaller than in the systems based on checkpointing and task migration and it allows scheduling with a high rate of incoming and outgoing tasks. For example the DIANE Master routinely achieves peaks of 110-120 Hz without observable degradation in the performance. This means that scheduling overhead is negligible for N ∗ 120 worker agents if average task duration is N seconds. The default scheduling algorithm used in DIANE is based on dynamic pull approach also known as self-load-balancing. DIANE makes it easy to plug-in alternative algorithms, however the results described in this paper use the default one. DIANE allows the standard GSI-based authentication and authorization of the worker agents. The Grid proxy certificate is shipped via standard Grid submission mechanisms to the worker node, while the master retains the original. The secure mode prevents the accidental mixing of user credentials in a single overlay. The results described in this paper refer to the default, non-authenticated mode. The following sections present three examples of improved QoS characteristics with DIANE User Level Scheduling: the job turnaround time, job completion rate, and error recovery. 4.1.4. Job turnaround time with high-granularity splitting DIANE supports high-granularity job splitting, i.e. partitioning a job into a large number of short or very short tasks. For example, the radio-frequency compatibility analysis jobs for ITU RRC06 conference [32], have been split into approximately 40 000 tasks performed simultaneously by around 200 worker agents at 6 EGEE Grid sites across Europe. Task duration was highly variable (Fig. 4) lasting from few seconds (majority of the tasks) to 20 minutes (few individual tasks). The exact distribution of the task duration was not known until the job was fully executed. Consequently, it was not possible to a priori aggregate short tasks and isolate long tasks. The efficiency of user-level scheduling was high with the number of tasks executing in parallel very close to the size of the worker pool (Fig. 5). As shown in previous sections (Fig. 2) the job turnaround time is orders of magnitude higher in a plain grid environment. 4.1.5. Job completion rate User-level scheduling provides a more sustained job completion rate. Fig. 6 shows the job completion rate for a Geant 4 release statistical regression testing application [34]. The job has been split in 207 tasks and average task duration was around 400 seconds. In the Grid, the GCJv14.tex; 17/12/2006; 0:08; p.9 ----- 10 Grid Scheduling for Interactive Analysis Task duration histogram 100000 10000 1000 100 10 1 1 10 100 1000 10000 Time [seconds] Figure 4. High-granularity splitting with exponential distribution of the task execution time. Most of 40000 tasks execute in less then 10 seconds, with individual tasks executing in 1000 seconds. Efficiency of resource utilization 200 150 100 50 0 0 1000 2000 3000 4000 5000 Time [seconds] Figure 5. Comparison of the number of concurrently processed tasks (the number of busy workers) and the number of available workers (the worker pool size). The difference represents the scheduling overhead, including the network communication cost. Currently, the scheduler does not remove excessive workers from the pool, hence the number of idle workers increases at 4000s due to few long-lasting tasks. GCJv14.tex; 17/12/2006; 0:08; p.10 ----- Scheduling for Responsive Grids 11 Job completion rate 100 90 80 DIANE scheduling 70 62 workers 60 35 workers 50 21 workers 40 30 plain Grid scheduling 20 10 0 100 1000 10000 100000 Time [seconds] Figure 6. Comparison of job completion rate between user-level scheduling based on DIANE (A) and plain Grid scheduling(B). Geant 4 regression testing jobs were run simultaneously in both scheduling modes. Equal number of available computing resources (85 worker nodes) within EGEE Grid in each mode was guaranteed. The figure shows three selected jobs with typical behaviour. This figure has been taken from [34]. load on the Computing Elements (queuing time) and the load on the Resource Broker (efficiency of matchmaking) may change dynamically in short periods of time resulting in a job completion curve which is less predictable (B1 and B3) or jobs being stuck in the Grid for a very long time and appear as incomplete (B2). The user-level scheduler assures that, even if the number of effectively available resources is low and varying, the job output throughput is stable if splitting granularity is correctly chosen. 4.1.6. Error recovery Efficient and accurate failure recovery is an important factor for Quality of Service. Large distributed systems such as the grid are prone to diverse configuration and system errors. A generic strategy of handling errors does not exist and the specific strategies depend on the application as well as the environment. An application-oriented scheduler such as DIANE is capable of distinguishing application and system errors and reacting appropriately via customizable error recovery methods. Crashing worker agents are automatically taken out of the worker pool. Transient connectivity problems in the WAN are detected; the GCJv14.tex; 17/12/2006; 0:08; p.11 ----- 12 Grid Scheduling for Interactive Analysis failed tasks are automatically re-dispatched to another worker agents. The mechanism uses a direct, highly efficient communication links in the virtual master/worker network and is much more efficient than a standard metascheduling techniques implemented in the middleware (JDL RetryCount parameter) which involve the full submission cycle. A part of recent Avian Flu Drug Search [29] have been performed using DIANE scheduler. A master agent spanning several weeks was taking care of efficient error recovery so the system could be operated by a single person. Because of the long duration of the job, the worker agents were often aborted because they exceeded the time limits in the queues at the Computing Elements. The operator was adding new worker agents to the system so that at least 200 were available at any time. DIANE was able to dynamically reconfigure the virtual master/worker network to accommodate the new worker agents. The overall efficiency of DIANE user-level scheduling was 84%, compared to 38.4% efficiency of pure grid scheduling. 4.2. gPTM3D PTM3D [42] is a fully-featured DICOM image analyzer developed at LIMSI. PTM3D transfers, archives and visualizes DICOM-encoded data. Besides moving independently along the usual three axes, the user is able to view the cross-section of the DICOM image along an arbitrary plane and to move it. PTM3D provides computer-aided generation of three-dimensional (3D) representations from CT, MRI, PET-scan, or echography 3D data. A reconstructed volume (organ, tumour) is displayed inside the 3D view. The reconstruction also provides the volume measurement required for therapeutic decisions. The system currently runs on standard PC computers and it is used online in radiology centres. Clinical motivation for grid-enabled volume reconstruction is described in [21]. The first step in grid-enabling PTM3D (gPTM3D) is to speedup compute-intensive tasks such as the volume reconstruction of the whole body used in percutaneous nephrolithotomy planning [37]. The volume reconstruction algorithm includes a semi-automatic segmentation component based on an active contours method where the user initiates the segmentation, and can correct it at anytime. It also includes a tessellation component which is the compute-intensive part of the algorithm. The gPTM3D application requires fine-grained parallelism. The parallel tasks are the reconstruction of one slice; in the examples presented Fig. 7, the execution time of the majority of the tasks is in the order of a few hundreds of milliseconds but with high variability. GCJv14.tex; 17/12/2006; 0:08; p.12 ----- Scheduling for Responsive Grids 13 Figure 7. gPTM3D performance When the geometry of the volume becomes complex, the reconstruction of the critical slices can last for 20 seconds or more. The architecture has two components: scheduler/worker agents at the user-level and the Interaction Bridge (IB) as an external service. The IB acts as a proxy between the PTM3D workstation, which is not EGEE-enabled and the EGEE world. When opening an interactive session, the PTM3D workstation connects to the IB. In turn, the IB launches a scheduler and a set of workers on an EGEE site, through fully standard requests to an EGEE User Interface. A stream is established between the scheduler and the PTM3D front-end through the IB. When the actual volume reconstruction is required, the scheduler receives contours. The scheduler/worker agents follow a pull model with each worker computing one slice of the reconstructed volume at a time, and sending it back to the scheduler, which forwards them to IB from where they finally reach the front-end. The overall response time is compatible with user requirements (less than 2 minutes), while the sequential time on a 3GHz PC with 2GB of memory can reach 20 minutes and more than 30 minutes on less powerful front-ends. So far, the only bottleneck is the rate at which the front-end is able to generate contours. Fig. 7 presents the speedup achieved on EGEE, with one scheduler and up to 14 workers in the largest case. For small reconstructions, the grid is obviously not necessary; we have included them to prove that there is no penalty (in fact a small advantage) in this case. Thus there is no need to switch from a local mode to a grid one in an interactive session. For the largest reconstruction, the speedup is nearly optimal. Lowering the execution time to this point has strictly no impact on the local interaction scheme, which includes stopping, restarting and improving locally the segmentation. GCJv14.tex; 17/12/2006; 0:08; p.13 ----- 14 Grid Scheduling for Interactive Analysis 5. Grid differentiated services 5.1. Virtual Reservations As a shared resource, a grid supports a broad spectrum of workloads ranging from long-running batch workloads executed under best-effort policy to workflows [28, 20] or parallel applications for which specific scheduling strategies have been proposed. Examples of these strategies include static [18] or dynamic [47] gang-scheduling using advance reservation [38] and middleware mechanisms favouring simultaneous allocation such as the EGEE DAG job type. Grid advance reservation suffer from two drawbacks: first, planning is not consistent with the goal of seamless integration with everyday computing practice, for instance the use cases described in Section 2; second, reservation is inherently not work-conserving, meaning that processors might idle while eligible jobs are queued [46]. Providing differentiated QoS either at the processor or network level usually relies on some implementation of Generalized Processor Sharing (GPS). However, the fundamental concept required for schedulability analysis and schedule construction in these frameworks is that the allocation of resources may be broken along quanta of time. The problem for grid scheduling is that such quanta do not exist. Jobs are not partitionable. Except for checkpointable jobs, a job that has started running cannot be suspended and restarted later. Moreover, as shown before, the execution times exhibit an extremely high variance. We have defined and implemented the concept of a Virtual Reservation (VRes), which addresses both issues of advance reservation and scheduling quanta by allowing controlled time-sharing. VRes permits the definition of time quanta and their exposure at the grid level. At the site level, each of the p physical processors is virtualized into k virtual processors, providing pk slots to the site scheduler. When a virtual slot is unused, the computing bandwidth is transparently returned to the other classes sharing the same physical processor. Thus, a fraction of these slots can then be permanently reserved for some class of applications without jeopardizing utilization. The mapping of classes first to the virtual processors, then onto the physical ones is obviously the key for full processor utilization. This mapping must be controlled so that each class maps to the full range of physical processors, as shown in Fig. 8. Provided that the mapping is controlled, the reservation ensures both application isolation with respect to computational bandwidth and full processor utilization. GCJv14.tex; 17/12/2006; 0:08; p.14 ----- Scheduling for Responsive Grids 15 3 3 3 3 2 2 2 2 12 virtual processors 1 1 1 3 4 physical processors Figure 8. Example of VRes: class 1, 2 and 3 are allocated respectively 1/4, 1/3 and 5/12 of the computational bandwidth 5.2. EGEE Implementation An implementation of VRes has been developed for the MAUI scheduler and the gLite middleware. It can be downloaded from the EGEE SDJ Working Group site http://egee-na4.ct.infn.it/wiki/index.php/ShortJobs. The Job Description language (JDL) has been modified to include a Boolean attribute SDJ. Sites willing to accept SDJ jobs set up a CE which permits running one job per SDJ slot. Jobs submitted to this CE either are immediately scheduled or rejected. The broker is notified in case of rejection and can either reschedule the job on another resource or notify the user. These sites also configure their scheduler with parameters controlling the computational bandwidth dedicated to SDJ. In particular, the wall-clock time and CPU time of SDJ jobs are limited. While these parameters are lower for SDJ jobs than for the usual batch jobs, all EGEE jobs are subject to the same kind of limitations, and all are aborted if they exceed these. This work has exposed a problem with scheduling in the EGEE middleware. The system does not permit a CE to provide access control based on job type, which is required for application isolation in general and for QoS in our case. As a temporary solution, a namebased dispatch has been set up in gLite 3.2. The SDJ-dedicated CEs GCJv14.tex; 17/12/2006; 0:08; p.15 |3|3|3|3| |---|---|---|---| |2|2|2|2| |1|1|1|3| ----- 16 Grid Scheduling for Interactive Analysis SDJ 5 dteam 4 3 2 1 0 58000 58200 58400 58600 58800 59000 59200 Time (seconds) Figure 9. Number of concurrent jobs on a single dual-processor node as a function of time. are named such that they have a trailing “sdj”. The submission system introduces an appropriate regular expression in job requirements so as the WMS will select select SDJ CEs for short deadline jobs and prevents batch jobs from being scheduled on SDJ CEs. It is worth mentioning that this method can be adopted for early experiments of other classes, because it requires only minor modifications of the gLite code. A more elegant and general solution is being investigated. However, the Glue schema must be modified and such modifications are a long process. Tests that have been conducted at LAL to ensure the correct behaviour of the SDJ configuration. Fig. 9 shows a breakdown of the occupation of one dual-processor node. On a background of batch jobs, which never exceed 2 (one per processor), SDJ can run within the same limit, and also concurrently with a third class (dteam) required by EGEE operational monitoring. Hence there are five slots per dualprocessor node. Fig. 10 exemplifies control of the global computational bandwidth at the site level dedicated to SDJ. In this configuration, a maximum of ten concurrent SDJ were permitted. The virtual reservation mechanism and the SDJ CE have been put in production at LAL since May 2006. The LAL site is equipped with a mixture of IBM and HP 1U rack-mounted dual-processor (AMD Opteron, 2.2 GHz) machines with 1 GB of RAM per CPU (2 GB total) and 80 GB of disk. The SDJ slots are routinely used in production by several biomedical applications and also for EGEE demonstrations (one cannot wait in queues when the audience is waiting for a live demonstration), and run concurrently with the usual batch jobs. The site utilization is extremely high, approaching a steady 100%. This experimental result provides an empirical answer to the often raised GCJv14.tex; 17/12/2006; 0:08; p.16 ----- Scheduling for Responsive Grids 17 40 35 30 25 20 15 10 5 0 58000 58200 58400 58600 58800 59000 59200 Time (seconds) Figure 10. Number of concurrent jobs on the site as a function of time. issue of the negative impact of concurrency (from cache to IO) on real-world workloads running on high-end processors. 6. Related work Existing approaches to grid scheduling for QoS follow three distinct paths: Virtual Machines (VM) encapsulation, statistical prediction, and service level agreements. Virtual machines provide a powerful new layer of abstraction in centralized computing environments in order to ensure fault isolation. Distributed scheduling based on VM encapsulation has been explored as a general tool in the PlanetLab project [7]. The Virtuoso project has more specifically explored virtualization for differentiated services [30, 31], and the Virtual Workspaces project [27] investigates the large-scale deployment of VM inside the Globus middleware. Virtual machines provide complete freedom of scheduling and even migrating an entire OS and associated computations which considerably eases time-sharing between deadline-bound short jobs and long running batch jobs. On the other hand, the virtual machines strategy is extremely invasive. All of, or a significant fraction of, the computations must be run inside virtual machines to provide scheduling opportunities—something for which traditional batch users have little incentive. Another issue is that VM interactivity follows the remote desktop model. In this model, which has been often been adopted for grid-enabling computational steering [44, 24, 26, 40], the user front-end is a passive terminal. With Grid Differentiated Services and user-level scheduling, we provide a much more modular environment that can support any combination of local and remote computations. GCJv14.tex; 17/12/2006; 0:08; p.17 ----- 18 Grid Scheduling for Interactive Analysis Accurate statistical prediction of the workloads is possible in large range of situations including shared clusters [16] and batch-scheduled parallel machines [13]. In particular, [48] shows that statistical prediction allows efficient support of interactive computations in unreserved cluster environments. At the grid scale, in the current status where time-sharing is possible only through control mechanisms such as VRes, predictive methods would apply for instance to the availability of SDJ slots provided by VRes. Service level agreements (SLAs) are the standard to represent the agreed constraints between service consumers and service providers on a grid [2]. SLAs by themselves do not provide scheduling solutions, but allow expressing flexible requirements and incorporating multicriteria approaches. SLAs could be applied to differentiated services in our context. For instance proposing a choice between a quick and reliable turnaround time, with strong completion constraints, and a more unreliable turnaround time without constraints. SLAs also offer the perspective of a general framework for renegotiation of resources [33] by running jobs. In our context this could be used to switch from the first mode to the second one, for instance when a SDJ approaches the end of its allocated time and must be prorogated. User-level scheduling has been proposed in many other contexts, and a case for it has been made in the AppLeS [14, 8] project. In a production grid framework, the DIRAC [51] project has proposed a permanent grid overlay where scheduling agents pull work from a central dispatching component. Our work differs from DIRAC on a major point: both for DIANE and gPTM3D, the execution agents are regular gLite jobs, and are thus subject to all grid policies and accounting. The abuse of glideIn techniques, which would permanently launch execution agents, would be counter-productive. The local EGEE schedulers (typically MAUI or PBS) do enforce fair share across VO and users. Thus, running a useless execution agent will prevent it to be run on the same site at the next scheduler decision. Obviously, if the site allows infinite execution, there will never be a scheduler decision, but the resource usage of this agent would be charged to the appropriate user or VO. 7. Conclusion We have presented complementary strategies to address the QoS requirements of a responsive grid: Grid Differentiated Services and userlevel schedulers. Grid Differentiated Services provide a general framework for the isolation of classes of applications and the realization GCJv14.tex; 17/12/2006; 0:08; p.18 ----- Scheduling for Responsive Grids 19 at the grid level of the concepts required for hard or soft real-time scheduling. User-level schedulers cope with high latencies associated with grid middleware. Equally important is a clean separation between two optimization problems: at the grid level, the optimization is related to fair-share and load balancing, while at the user-level, the optimization is for a specific application workload. Depending on the application requirements, Grid Differentiated Services and user-level schedulers can be used separately or combined. In the example of gPTM3D, combining Grid Differentiated Services and an embedded user-level scheduler provides a fully transparent coupling of the grid resources with an augmented reality desktop software. The scope of applications deployed on top of the DIANE generic scheduler exemplify the impact of user-level scheduling on a number of QoS characteristics. Both strategies have been deployed on the EGEE grid, as autonomous site decisions (for the Grid Differentiated Services) or as regular user jobs (for the user-level schedulers). They are fully compatible with gLite, the existing EGEE middleware. Their architecture and to a large extent their implementation depend only on generic grid concepts. We are convinced that this non-intrusiveness is a key to a progressive convergence of QoS and grid technology. Acknowledgements This work was partially funded by the EGEE EU project (INFSO-RI508833 Grant). gPTM3D is part of the AGIR project funded by the ACI Masses de Donn´ees program of the French ministry of research . References 1. Atlas Computing - Technical Design Report CERN-LHCC-2005-022. http://doc.cern.ch//archive/electronic/cern/preprints/lhcc/public/lhcc-2005022.pdf 2. R. AlAli, K. Amin, G. von Laszewski, O. Rana, D. Walker, M. Hategan and N. Zaluzec. Analysis and Provision of QoS for Distributed Grid Applications. Journal of Grid Computing, 2(2):163-182. June 2004 3. P. Andreetto et al. Practical approaches to Grid workload and resource management in the EGEE grid. In Procs. CHEP’04 4. S. Andreozzi et al. GLUE Schema Specification version 1.2. http://infnforge.cnaf.infn.it/glueinfomodel/ 5. S. Baruah, N. Cohen, C. Plaxton and D. Varvel. Proportionate Progress: A Notion of Fairness in Resource Allocation. Algorithmica 15(6), pp. 600-625, 1996. 6. S. Basu, V. Talwar, B. Agarwalla and R. Kuma. Interactive Grid Architecture for Application Service Providers. HP Technical Report HPL-2003-84R1. 2003. GCJv14.tex; 17/12/2006; 0:08; p.19 ----- 20 Grid Scheduling for Interactive Analysis 7. A. Bavier, M. Bowman, B. Chun, D. Culler, S. Karlin, S. Muir, L. Peterson, T. Roscoe, T. Spalink, and M. Wawrzoniak. Operating System Support for Planetary-Scale Network Services. In Procs. of the First Symposium on Networked Systems Design and Implementation (NSDI, pp 253–266. 2004. 8. F. Berman, R. Wolski and H. Casanova. Adaptive Computing on the Grid Using AppLeS. IEEE Trans. Parallel Distrib. Syst., 14(4): 369–382. 2003. 9. G. Shao and R. Wolski and F. Berman: Master/Slave Computing on the Grid; in Proceedings of the 9th Heterogeneous Computing Workshop, Cancun, Mexico, May 2000, pp. 3-16. http://citeseer.ist.psu.edu/wolski00masterslave.html 10. Morris, G. M. et al.(1998), Automated Docking Using a Lamarckian Genetic Algorithm and and Empirical Binding Free Energy Function J. Computational Chemistry, 19: 1639-1662 11. D. Berry, C. Germain-Renaud, D. Hill, S. Pieper and J. Saltz. Report on the Workshop IMAGE’03: Images, medical analysis and grid environments. TR UKeS-2004-02, UK National e-Science Centre, Feb. 2004. 12. Christophe Blanchet et al. GPSA: Bioinformatics grid portal for protein sequence analysis on EGEE grid. In Procs of Healthgrid 2006, Studies in Health Technology and Informatics 120, IOS Press. May 2006 13. J. Brevik, D. Nurmi and R. Wolski. Predicting bounds on queuing delay for batch-scheduled parallel machines. In Procs. of the eleventh ACM SIGPLAN symposium on Principles and practice of parallel programming (PPoPP ’06), pp 110–118. 2006. 14. H. Casanova, G. Obertelli, F. Berman and R. Wolski. The AppLeS parameter sweep template: user-level middleware for the grid. In Procs 2000 ACM/IEEE conference on Supercomputing SC’00. 2000. 15. A. Chandra, M. Adler, and P. Shenoy. Deadline fair scheduling: Bridging the theory and practice of proportionate-fair scheduling in multiprocessor servers. In Procs. of the 7th IEEE Real-Time Technology and Applications Symposium. June 2001. 16. P. Dinda. Design, Implementation, and Performance of an Extensible Toolkit for Resource Prediction In Distributed Systems. IEEE Transactions on Parallel and Distributed Systems, 17(2). 2006. 17. D.H.J. Epema, M. Livny, R. van Dantzig, X. Evers and J. Pruyne. A worldwide flock of Condors: Load sharing among workstation clusters. Future Generation Computer Systems, 12:53–65. 1996. 18. D. G. Feitelson and M. Jette. Improved Utilization and Responsiveness with Gang Scheduling. In Procs of the IPPS ’97 Workshop on Job Scheduling Strategies for Parallel Processing, 1997 19. I. Foster, M. Fidler, A. Roy, V, Sander and L. Winkler. End-to-End Quality of Service for High-end Applications. Computer Communications, 27(14):13751388. 2004. 20. T. Glatard, J. Montagnat and X. Pennec. Efficient services composition for grid-enabled data-intensive applications. In Procs of the IEEE Int. Symp. on High Performance Distributed Computing (HPDC’06), June 2006. 21. C. Germain-Renaud, R. Texier and A. Osorio. Interactive Reconstruction and Measurement on the Grid. Methods of Information in Medecine, 44(2):227- 232. 2005. 22. S. Guatelli et al.: Geant4 Simulation in A Distributed Computing Environment; submitted to IEEE Trans. Nucl. Sci. 2006 23. S. Hastings, M. Kurc, S. Langella, U. V. Catalyurek, T. C Pan and J. H. Saltz, Image Processing for the Grid: A Toolkit for Building Grid-enabled Image GCJv14.tex; 17/12/2006; 0:08; p.20 ----- Scheduling for Responsive Grids 21 Processing Applications. In Proc. 3rd Int. Symp. on Cluster Computing and the Grid, pp 36–43. 2003 24. P. Heinzlreiter and D. Kranzlm??ller. Visualization Services on the Grid: The Grid Visualization Kernel. Parallel Processing Letters 13(2):135–148. 2003. 25. E. Heymann, M. A. Senar and E. Luque and M. Livny”. Adaptive Scheduling for Master-Worker Applications on the Computational Grid. In Procs 1st IEEE/ACM Intl Workshop on Grid Computing (GRID 2000). 2000 26. R. Kumar, V. Talwar and S. Basu. A resource management framework for interactive Grids. Concurrency and Computation: Practice and Experience, 16(5):489–50. April 2004. 27. K. Keahey, I. Foster, T. Freeman, and X. Zhang. Virtual Workspaces: Achieving Quality of Service and Quality of Life in the Grid. To appear in The Scientific Programming Journal. 2006. 28. G. von Laszewski, M. Hategan and D. Kodeboyina. Work Coordination for Grid Computing. To appear 2006. http://www.mcs.anl.gov/ gregor/papers/vonLaszewski- workcoordination.pdf 29. H.C.Lee et al.: Grid-enabled High Throughput in-silico Screening Against Influenza A Neuraminidase; NETTAB 2006, Santa Margherita di Pula, July 1013, 2006; submitted to a special issue of IEEE Transactions on Nanobioscience 30. B. Lin, P. A. Dinda. VSched: Mixing Batch And Interactive Virtual Machines Using Periodic Real-time Scheduling. In Procs. ACM/IEEE conference on Supercomputing 2005 (SC’05). 2005 31. B. Lin, P. A. Dinda and D. Lu. User-driven Scheduling of Interactive Virtual Machines. In Procs. of the Fifth IEEE/ACM Int.Workshop on Grid Computing (GRID’04), pp 380–387. 2004 32. A.Manara et al.: Integration of new communities in the Grid for mission critical applications: distributed radio-requency compatibility analysis for the ITU RRC06 conferece; submitted to EGEE’06 Conference, 25-29 September 2006, Geneva, Switzerland 33. J. MacLaren, R. Sakellariou, K. T. Krishnakumar, J. Garibaldi, and D. Ouelhadj. Towards Service Level Agreement Based Scheduling on the Grid. In Procs 14th Int. Conf. on Automated Planning and Scheduling (ICAPS 04). 2004 34. P. Mendez-Lorenzo et al.: Distributed Release Validation of the Geant4 Toolkit in the LCG/EGEE Environment; submitted to IEEE Nuclear Science Symposium 2006 35. I. Mirman. Going Parallel the New Way. Desktop Computing, 10(11), June 2006 36. J.T. Mo´scicki: Distributed analysis environment for HEP and interdisciplinary applications; Nuclear Instruments and Methods in Physics Research A 502 (2003) 426-429 37. A. Osorio, O. Traxer, S. Merran, F. Dargent, X. Ripoche, J. Ati. Real time fusion of 2D fluoroscopic and 3D segmented CT images integrated into an Augmented Reality system for percutaneous nephrolithotomies (PCNL). In InfoRAD 2004, RSNA’04. 2004. 38. F. Pacini and P. Kunzt. Job Description Language Attribute Specification. EGEE Tech. Rep. 555796. 2005 http://glite.web.cern.ch/glite/documentation/ 39. R. Raman, M. Livny, and M. Solomon. Matchmaking: Distributed resource management for high throughput computing. In Procs. of the Seventh IEEE International Symposium on High Performance Distributed Computing. July 1998. GCJv14.tex; 17/12/2006; 0:08; p.21 ----- 22 Grid Scheduling for Interactive Analysis 40. H. Rosmanith and D. Kranzlmuller. glogin - A Multifunctional, Interactive Tunnel into the Grid. In Procs 5th IEEE/ACM Int. Workshop on Grid Computing (GRID’04), 2004. 41. L. G. Roberts. Beyond Moore’s law: Internet growth trends. IEEE Computer, 33(1),pp 117-119, Jan. 2000 42. V. Servois, A. Osorio et al. A new PC based software for prostatic 3D segmentation and measurement : application to permanent prostate brachytherapy (PPB) evalution using CT and MR images fusion. In InfoRAD RSNA 2002. 88th Annual Meeting of the Radiological Society of NorthAmerica. Dec. 2002. 43. SGI. High-performance computers become interactive. www.sgi.com/company info/newsroom/3rd party/111505 starp.pdf. Jan. 2006 44. J. Shalf and E. W. Bethel Cactus and Visapult: An Ultra-High Performance Grid- Distributed Visualization Architecture Using Connectionless Protocols. IEEE Computer Graphics and Applications, 23(2):51–59. 2003. 45. S. Smallen, H. Casanova and F. Berman. Applying scheduling and tuning to on-line parallel tomography. In Procs 2001 ACM/IEEE conference on Supercomputing (SC’01). 2001. 46. Q. Snell, M. J. Clement, D. B. Jackson and C. Gregory. The Performance Impact of Advance Reservation Meta- scheduling. In Procs of the Workshop on Job Scheduling Strategies for Parallel Processing at IPDPS ’00/JSSPP ’00, pp 137–153. Springer-Verlag. 2000. 47. A. C. Sodan, S. Doshi, L. Barsanti and Darren Taylor. Gang Scheduling and Adaptive Resource Allocation to Mitigate Advance Reservation Impact. InProcs Sixth IEEE Int. Symp. on Cluster Computing and the Grid, pp 649-653, IEEE Comp. Soc. Press, 2006. 48. A. Sundararaj, A. Gupta, and P. Dinda. Increasing Application Performance In Virtual Environments Through Run-time Inference and Adaptation. In Procs of the 14th IEEE Int. Symp. on High Performance Distributed Computing (HPDC 2005). 2005. 49. M. Thomas, J Burruss, L Cinquini, G Fox, D. Gannon, I. Glilbert, G. von Laszewski, K. Jackson, D. Middleton, R. Moore, M. Pierce, B. Plale, A. Rajasekar, R. Regno, E. Roberts, D. Schissel, A. Seth, and W. Schroeder. Grid Portal Architectures for Scientific Applications. Journal of Physics, 16, pp 596-600. 2005. 50. V. Tola, F. Lillo, M. Gallegati and R. N. Mantegna. Cluster analysis for portfolio optimization. Preprint. 2005. 51. A. Tsaregorodtsev, V. Garonne, and I. Stokes-Rees. DIRAC: A Scalable Lightweight Architecture for High Throughput Computing. In Procs 5th IEEE/ACM Int. Workshop on Grid Computing (GRID’04), 2004. GCJv14.tex; 17/12/2006; 0:08; p.22 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/s10723-007-9086-4?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/s10723-007-9086-4, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "http://cds.cern.ch/record/1152703/files/EGEE-PUB-2008-002.pdf" }
2,008
[ "JournalArticle" ]
true
2008-03-01T00:00:00
[]
12,706
en
[ { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0112123cfb29e45be6d0a80751184ab7f20888df
[]
0.825609
An Approach to Develop a Transactional Calculus for Semi-Structured Database System
0112123cfb29e45be6d0a80751184ab7f20888df
International Journal of Computer Network and Information Security
[ { "authorId": "8419191", "name": "R. Ganguly" }, { "authorId": "144341858", "name": "A. Sarkar" } ]
{ "alternate_issns": null, "alternate_names": [ "Int J Comput Netw Inf Secur" ], "alternate_urls": null, "id": "75298332-e3d9-44ab-af5e-03b23d7bd4e6", "issn": "2074-9090", "name": "International Journal of Computer Network and Information Security", "type": "journal", "url": "http://www.mecs-press.org/ijcnis/" }
— Traditional database system forces all data to adhere to an explicitly specified, rigid schema and most of the limitations of traditional database may be overcome by semi-structured database. Whereas a traditional transaction system guarantee that either all modifications are done or none of these i.e. the database must be atomic (either occurs all or occurs nothing) in nature. In this paper transaction is treating as a mapping from its environment to compensable programs and provides a transaction refinement calculus. The motivation of the Transactional Calculus for Semi Structured Database System (TCSS) is-finally, on a highly distributed network, it is desirable to provide some amount of fault tolerance. The paper proposes a mathematical framework for transactions where a transaction is treated as a mapping from its environment to compensable programs and also provides a transaction refinement calculus. It proposes to show that most of the semi structured transaction can be converted to a calculus based model which is simply consists of a forward activity and a compensation module of CAP (consistency, availability, and partition tolerance) [12] and BASE (basic availability, soft state and eventually consistent) [45] theorem. It proposes to show that most of the semi-structured transaction can be converted to a calculus based model which is simply consists of a forward activity and a compensation module of CAP and BASE theorem
Published Online September 2019 in MECS (http://www.mecs-press.org/) DOI: 10.5815/ijcnis.2019.09.04 # An Approach to Develop a Transactional Calculus for Semi-Structured Database System ## Rita Ganguly Department of Computer Applications; Dr.B.C.Roy Engineering College; Durgapur: 713206; West Bengal, India E-mail: ganguly.rita@gmail.com ## Anirban Sarkar Department of Computer Science; National Institute of Technology; Durgapur;713209; West Bengal, India E-mail: sarkar.anirban@gmail.com Received: 06 August 2019; Accepted: 25 August 2019; Published: 08 September 2019 **_Abstract—Traditional database system forces all data to_** adhere to an explicitly specified, rigid schema and most of the limitations of traditional database may be overcome by semi-structured database. Whereas a traditional transaction system guarantee that either all modifications are done or none of these i.e. the database must be atomic (either occurs all or occurs nothing) in nature. In this paper transaction is treating as a mapping from its environment to compensable programs and provides a transaction refinement calculus. The motivation of the Transactional Calculus for Semi Structured Database System (TCSS) is-finally, on a highly distributed network, it is desirable to provide some amount of fault tolerance. The paper proposes a mathematical framework for transactions where a transaction is treated as a mapping from its environment to compensable programs and also provides a transaction refinement calculus. It proposes to show that most of the semi structured transaction can be converted to a calculus based model which is simply consists of a forward activity and a compensation module of CAP (consistency, availability, and partition tolerance) [12] and BASE (basic availability, soft state and eventually consistent) [45] theorem. It proposes to show that most of the semistructured transaction can be converted to a calculus based model which is simply consists of a forward activity and a compensation module of CAP and BASE theorem. It is important that the service still perform as expected if some nodes crash or communication links fail, Verification of several useful properties of the proposed TCSS includes in this article. Moreover, a detailed comparative analysis has been providing towards evaluation of the proposed TCSS. **_Index Terms—Semi-structured, transactional calculus,_** X-Query, GOOSSDM, CAP, BASE, GQL-SS. I. INTRODUCTION In recent years, researches have produced several proposals [2, 3, 4, 5, 7, 8, and 9] towards conceptual modelling of semi-structured database system compare to the proposals of conceptual modelling. To overcome traditional transactional problems, extending the transactional processing system in semi-structured database by addition of compensation and coordination of consistency, availability, and partition tolerance (CAP)[12] and basic availability, soft state and eventually consistent (BASE) [45] theorem and enrich a standard design model with new healthiness conditions. There is no specific transactional calculus for semistructured data. The proposed Transactional Calculus for Semi-structured database (TCSS) puts forward a mathematical framework for transactions where a transaction is treated as a mapping from its environment to compensable program. Further, the transactional calculus is derive from an algebra based query language GQL-SS [11] and illustrated using examples of real life. The motivation of the Transactional Calculus for Semistructured System, it is desirable to provide some amount of fault tolerance, on a highly distributed network. It is important that the service still perform as expected, when some nodes crash or communication links fail. The ACID (Atomicity, Consistency, Isolation and Durability) acronym says that database transactions should be firstly, seem indispensable, and yet they are incompatible with availability and performance in very large systems. The semi-structured database violates the ACID properties. According to ACID properties in Atomic the entire transaction will fail if one node element of a transaction fails, but in semi-structured database, it is not possible. In semi-structured database, if one node is damaged the entire network should not be affected. **Secondly, no** ----- transaction has access to any other transaction in Isolation that is in an intermediate or unfinished state. Thus, each transaction is independent unto itself. This is required for both performance and consistency of transactions within a database. The semi-structured database violates this property because it works in path basis and every node is inter linked to each other. The benefits of the transactional calculus for Semi-structured databases are manifold. It provides supports towards (1) structural and functional design concerns with enriched semantics and syntaxes for transactional calculus of semi-structured database represented by precise knowledge of domain independent conceptualization;(2) a systematic methodology which used to transforming calculus for functional design; (3)Transactional Calculus to Semi-structured database query system provides guidelines for the purpose of mapping .The proposed Transactional system for semi-structured is based on path expression. The path expressions may also contain label variables to preserve labels or tags. Three types of algorithms are using to evaluate the path in Graph Object Oriented Semi-Structured Data Model (GOOSSDM)[2, 19, 20, and 21] schema and Graphical Query Language for Semi-structured (GQL-SS) [11] schema, one for searching return node, second for searching the path from root of GOOSSDM schema to the desired node and the third one is for the searching and listing of the tail nodes.. Here trying to use the CAP theorem in the broader context of distributed computing theory. An important contribution of this paper is to discuss some of the practical implication of CAP Theorem of a transactional calculus for Semi-structured database. There are some proposal; they are only using CAP [12] or BASE [25] theorem or without these. To introduce the transactional calculus for Semi-structured database, with the help of CAP theorem, the CAP theorem was introducing as a trade-off between consistency, availability and partition tolerance. **Consistency: A read sees all previously** completed writes i.e. all nodes see the same data at the same time. **Availability: A guarantee that every request** receives a response about whether it succeeded or failed i.e. read and write always succeed. This means that in GOOSSDM schema there should be a searching path and its return some value. The path value should not be null. **Partition** **Tolerance:** Guaranteed properties are maintained even when network failures prevent some machines from communicating with others. The system continues to operate despite arbitrary partitioning due to network failures. However, developers face some challenges despite of several advantages of existing Semi-structured databases, when they apply the transaction processing system. Such challenges are as follows _Ch1:_ Lack of transactional methodology that blends semi-structured databases specification with syntaxes of transactional calculus for semi-structured database system. _Ch.2: Majority of existing transactional procedure are not_ usable for large semi-structured database queries. _Ch.3:_ Few transactional calculus for semi-structured database approaches are present in literatures that may represent evolving knowledge of transaction in semistructured databases but not in precise. _Ch.4: Appropriate guidelines and tools are absent which_ may help designers for specification. _Ch.5: XML-based semi-structured database systems_ characterized by an expressive global schema. The main issue here concerns the presence of a significant set of integrity constraints expressed over the schema and the concept of node identity, which requires particular attention when data come from autonomous data sources. This paper fulfils the deficiency of systematic methodology in transactional calculus of GOOSSDM model[44]. The paper is structuring as follows. Several related works in this field specified in Section 2 briefly. Section 3 is about the GOOSSDM modelling framework and this portion is subdividing into two parts components of GOOSSDM and Illustration of GOOSSDM. The proposed Transaction calculus for semi-structured database system (TCSS) has been describing and formalised in Section 4. Next, guidelines about the way in which the validation of TCSS can be applied databases by using CAP and BASE theorem and application specific conceptualisations have been suggesting in Section 5. Further, the proposed TCSS have been implementing and visualised using different operators and practically illustrates the proposed work using suitable example in Section 6. Following this, Section 7 practically illustrates the proposed work using a suitable programming code. Finally, the paper is concluding in Section 8.Aiming to overcome issues explained in above mentioned challenges this paper proposes several objectives. First, the proposed framework of Transactional system for semi-structured is based on path expression. They may also contain path variables, which, are evaluating to the empty path or to a path having a length of n edges. The path expressions may also contain label variables to preserve labels or tags. At second, the path operator is using to set the root node in GOOSSDM [2, 19, 20, and 21] schema and useful to find the path from the root node to desired node for any transaction. At Third, the propose work facilitate the early verification of the semi-structured data schema structure in correspondence with the desired transactional calculus. Finally, the transactional calculus is introducing to Semistructured database, with the help of CAP and BASE theorem. This objective addresses the issues described in Ch.2, Ch.3, Ch.4 and Ch.5.The benefits of the Transactional Calculus for Semi-structured system will represents a framework for specifying the semantics of a transactional facility integrated within a Semi-structured database system. The motivation of the Transactional Calculus for Semi-structured System is-finally, on a highly distributed network, is that when some nodes crash or communication links fail, it is important that the service still perform as expected. This paper fulfils the deficiency of systematic methodology in transactional calculus of GOOSSDM model. In addition, this paper proposes a formal transactional calculus called ----- Transactional Calculus for Semi-structured database (TCSS) in terms of concepts, relations and axioms for domain independent systems. It provides syntaxes and semantics for TCSS. Further, the transactional calculus is derived from a algebra based query language GQL-SS [11] and illustrated using examples of real life. Moreover, TCSS are proved by CAP and BASE theorems properties to show the expressiveness of the propose calculus. II. RELATED WORK In previous work [11], focused on path expression in semi-structured database system. More precisely (i) described GOOSSDM [2,19,20 and 21] schema and GQL-SS [11] data are amalgamate to leaves so the path expression may carry data variables as abstractions of the content of leaves. They may also carry path variables those are evaluating to the void path or to a path having a length of n edges. The path expressions may also contain label variables to preserve labels or tags. (ii) Develop three types of algorithms. Three types of algorithms use to evaluate the path in GOOSSDM schema, one for searching return node, second for searching the path from root of GOOSSDM schema to the desired node and the third one is for the searching and listing of the tail nodes. (iii) Define the GQL-SS algebra for GOOSSDM model that operate on semi-structured schema concept and / or several constructs described in the model. The algebra consists of a set of operators and few of them can be using with the constructs like ESG, CSG separately. As a result, point out that have to develop a transactional calculus related to this GQL-SS model. To the best of knowledge, there are no other global solutions addressing the transactional calculus for semi-structured database system. A small number of research works exist in the literatures those are in general semi-structured and used query language. However, still there is no specific transactional calculus, which is devoted enough to conceal the five challenges specified in the introduction section. The work in Supporting Multi Data Stores Applications in Cloud Environments [23] has given some idea about the semi-structured query but no proposed calculus. The amalgamation of transactions with programming control structures has provenance in systems such as Argus [28, 29].There is a composition of work that enquire into the formal specification of various zest of transactions [35, 36, 37]. However, these act of striving do not explore the semantics of transactions when integrated into a high-level programming language. Most closely related to goal is the work of Black et. al. [38], Choithia, and Duggan [39]. The former presents a theory of transactions that specify atomicity, isolation and durability properties in the form of an equivalence relation on processes. Beyond significant technical differences in the specification of the semantics, results differ most significantly from theirs insofar as [6] present a stratified semantics for a realistic kernel language intended to express different concurrency control models within the same framework. Choithia and Duggan present the pik-calculus and pike-calculus, extension of the pi calculus that supports various abstractions for distributed transactions and optimistic concurrency. Their work is relating to other efforts [40, 41] that encode transaction-style semantics into the pi-calculus and its variants. Haines et.al. [31] describes a compassable transaction facility in ML that supports persistence; undo ability, locking and threads. Their abstractions are modular and first class, although their implementation does not rely on optimistic concurrency mechanisms to handle commits. Consequently, none of the existing approaches is appropriate enough to cover the 5 challenges specified in the introduction section. In this regard, devising a new proposal, which is essential to resolve the issues, addressed in the 5 challenges. In this case, since dealing with the combination of CAP and BASE theorem, this proposal for expressing and executing queries and real time applications shown by using the calculus. Introducing an approach for a mapping language to map attributes of the data sources to the global schema and bridge query language to write the calculus. III. GOOSSDM: THE BASIC Extending the object-oriented paradigm to semi structured data model, the GOOSSDM introduced. It’s specifying the irregular and heterogeneous structure, hierarchical and non-hierarchical relations, n – array relationships, cardinality and participation constraint of instances with all details that are required for semistructured data model. The entire semi-structured database to be viewing as a Graph (V, E) in layered organization that is allowed by the proposed data model (GOOSSDM).At the lowest layer, each vertex represents an occurrence of an attribute or a data item. Let consider an example of Project Management System (PMS),[11], associated with Project. Project has attributes like members, department and publications. Several members are associated with project and each member can participated in any project. Department contains member, and each individual members may have or not have publication. The PMS is semi-structured is in nature. The GOOSSDM schema for PMS has been shown in fig. 1. The sample data is showing in Table 1. Fig.1. GOOSSDM Schema for PMS ----- Table 1. Sample Data Set for PMS |Project 1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10| |---|---|---|---|---|---|---|---|---|---| |Pname|PID|Topics|Member|||Department||Publication|| ||||MID|MName|Maddress|DID|DeptName|PuID|Ptopics| |ABC|P1001|AAAA|M01|Bipin|XX|D01|CSE|P001|RRR| |XYZ|P1003|CCCC|M03|Ashu|PP|D02|CA|P003|SSS| |DEF|P1004|DDDD|M04|Rashi|YY|D03|EE|P004|TTT| |XYZ|P1005|QQQQ|M06|Sashi|RR|D03|EE|P005|VVV| |ABC|P1001|BBBB|M07|Priya|CC|D01|CSE|P006|MMM| |Project 2|||||||||| |Pname|PID|Topics|Member|||Department||Publication|| ||||MID|MName|Maddress|DID|DeptName|PuID|Ptopics| |PQR|P1006|YYYY|M07|Priya|CC|D02|CA|P007|NNN| IV. CALCULUS FOR SEMI-STRUCTURED DATABASE SYSTEM In previous work, defining the GQL-SS algebra for GOOSSDM model that employ on semi-structured schema impression and / or various form reportein the model. Using GOOSSDM schema the semi-structured data seen as single rooted or multi rooted graph. In every case, while initiating any query, one needs to set an immediate root for the desired CSG and then need to find the tail nodes in respect to the desired CSG. In all the algorithms, the searching node and return node must be a type of CSG in GOOSSDM semantics. The GOOSSDM schema will use as input for the algorithms. The algorithms will invoke when the path operator (ρ) will execute. In case of proposed calculus whenever any operator will invoke, internally it will also invoke the path operator (ρ) to set the path from root node to the desired node in GOOSSDM schema by invoking algorithm 1 and algorithm 2. Moreover, the tail node list will create by invoking the algorithm 3 on next. If algorithm 1 and / or algorithm 2 return null value, then the actual operator need not to execute as there is no root available for the transactional calculus. This will facilitate the early verification of the semi-structured data schema structure in correspondence with the desired transactional calculus. The running example of Project Management System (PMS) used to illustrate the functionalities of operators. As specified earlier path operator (ρ) is also inclusive part of the algebra and invoked every time it is required to invoke any other operator specifically defined for management of semistructured data. In the example, if Project is set as root then the path from Project to Department can be established and expressed as Project (Root)MemberDepartment [11]. _Algorithm 1: Searching of Node in GOOSSDM Schema_ Step 1: Start Step 2: Input a node C= (CSG). Step 3: let op: = search node C And return node C Step 4: let P1:=layer 0 P2:= Immediate to layer 0 P3:= Next to immediate layer P4:= Next to Next Immediate layer Step 5: for i = 1 to 4 If (op Pi (C) ≠ᴓ) then Goto for next layer Else (Op Pi (C))= (Root) Step 6: stop Fig.2. Searching tail node Fig.3. Searching path from root to desired node Searching tail node from the desired node layer by layer, when the return operator of path is equal to the preceding one, then it is the last node i.e. tail node. ----- _Algorithm 2: Searching path from Root to Desired Node_ Step 1: Start Step 2: Input C=CSG // CSG for searching path. Step 3: If (IS Root(C) =false) then N1:=op Pi (CSG, P, ϴ) Step 4: If (N1==ᴓ) then Path =N1 Else Goto step 3. Step 5: Exit Root is Project, and then it searches the desired node layer by layer. Let N1is a path operator ρ with arguments layer no, CSG, and N1 value should not be Root. If N1 value is null, then the path value will be N1and if not then it will be check again from Root node. _Algorithm 3: Searching Tail Nodes from the Desired_ _Node_ Step 1: Start Step 2: Input G=GOOSSDM schema Step 3: let the path structured σ= (r,(E)), where E is a binary relation of(CSG,P,ϴ) Step 4: For i = 1 to n // n= No. of iterations If( IS Root(CSG))=True then Op<get the i[th] node>(CSG,P,ϴ)={CSGR,PR,ϴR} Step 5: for i = 1 to n Op< get the i[th]node>( CSG,P,ϴ)={CSG i[,P ]i[,ϴ ]i[} ] If( (op(CSGi-1,Pi-1,ϴi-1 ))== (op(CSG i,P i[,ϴ ]i[))) then // Finding the Tail node ] Tail =(CSG i,P i,ϴ i) Else goto step 4. Step 6 : The destination will be denoted as path ρ(CSG i,P i,ϴ i)= ρ{(CSG R,P R,ϴ R ), (CSGR-1,PR-1,ϴR-1),( CSGR-2,PR-2,ϴR 2[),.........,(CSG]i[, P]i[, ϴ]i[)} ] Else goto step 3. Step 7: Stop. Fig.4. Searching tail node from the desired node Searching tail node from the desired node layer by layer, when the return operator of path is equal to the preceding one, then it is the last node i.e. tail node. _A. Propose Operator_ In this section, the propose operator of Transactional Calculus for Semi-structured (TCSS) of GOOSSDM model is defined. It consists of a set of operators that take one or two CSG as input and produce a new list of CSG. The fundamental operators of TCSS consist of a set of operators and few of them also can be used with constructs like CSG, ESG separately. - _Select (σ) Operator_ The select operator will select CSG and returns CSG that satisfy a given predicate of a given list of ESGs or CSGs from the GOOSSDM schema. Thus to select those CSG from GOOSSDM schema, the tuple relational calculus (TRC) notation may be write as, { | (1) Its denote that tuple C is in CSG. { | (2) Its mean, it is the set of all tuples C such that predicate list is true for C. [List (CSG) =OUTPUT CSG where list= {list of ESG}] (3) If the set of all CSG for which the _List(C) evaluates_ true. And the path expression will be like that [ for all levels, existential CSG set the path and if it does not have any edge then it is set to Root] ( ) ( ) [Searching for desired CSG level by level and get ultimate CSG.] (4) - _Retrieve (π) Operator:_ The retrieve operation allows producing the CSG from GOOSSDM schema that satisfies a given condition. The retrieve operator extracts ESG or CSG from the CSG using some constraints _CON over one or more ESG or_ CSG defined in GOOSSDM schema. { | )}[C1 belongs to some CSG with satisfied condition] (5) It is meaning that the set of all tuples C such that for all tuples C1 is in predicate CSG is true for CON. ----- ( ) (6) [C1 belongs to CSG with specified Condition and that returns the restricted CSG.]It’s mean that for all tuplesC1 there exists predicate CON is true for C1is exists in CSG implies predicate CON is true for specified CSG. Let; Constraints=CON ( ) (7) [The dot operator extracts ESG or CSG from the CSG using some specified constraints CON over one or more ESG or CSG defined in schema.]CON1 contains all tuples of C1 extracts the exists predicate such that C1 is exits CSG and filename (f1) and CON (C1.f1) is true. (8) CON2 contains all tuples of C2 extracts the exists predicate such that C2 is exits CSG and filename (f2) and CON (C2.f2) is true. { | ( ) (9) - _Union, Intersection (ᴗ,ᴖ)operators:_ These operators will have usual meaning. The union of any two sets A and B, denoted by AB, is the set of all elements which belong to A or B or both. Hence, A  B _={_ (( ) ) (10) [C1 or _C2 or specify constraints of dot product of_ _C1_ and C2 that returns CSG or ESG which belongs to C1 or _C2 or both.] For all C1 and C2, C1 is in exist CSG or C2_ is in exist CSG or CON over both CSG implies the C1 union C2. Intersection denoted by AB, is the set of elements which belong to A and B both and can be expressed as _AB= { x: x_ _A AND x_ _B }._ (( ) ) (11) [C1 or C2 or specified constraints of dot product of C1 and C2 that returns CSG or ESG which belongs to _C1_ and C2.] For all C1 and C2, C1 is in exist CSG and C2 is in exist CSG and CON over both CSG implies the C1 union C2. - _Join (|X|) operator:_ The join operator is a special case of Cartesian product operator. It is a binary operator to relate two CSGs where one identical ESG must be common. Let, two CSGs are _CSG1 and CSG2. Also let, a set of ESG E1=(E11, E12,...,_ _E1R)_ and a set of ESG _E2=( E21, E22,..., E2s) is related_ with theCSG1 and _CSG2 respectively. The join operator_ between _CSG1and_ _CSG2 is possible iffE1ɅE2≠. Now_ let E1ɅE2= {Ea, Eb, Ec} then, { | _Ʌ_ _Ʌ_ (12) [SpecifiedCSG in _C1 with Existential CSG in_ _C2 and_ both will satisfy a common ESG field.] (( ) ) (13) [All CSG in _C1 and CSG in_ _C2 and a common ESG_ field is satisfied then this will return the all common ESG.] V. ILLUSTRATION OF TRANSACTIONAL CALCULUS OF SEMI-STRUCTURED (TCSS) DATABASE BY CAP THEOREM AND BASE THEOREM In this section, CAP theorem is as described in propose Semi-structured calculus system is as follows: _In a web concern to transmission collapse, it is difficult_ _for any web service to execute an atomic read/write_ _shared memory that promises a response to every request._ **Proof Sketch: Having stated the CAP theorem, it is** relatively straightforward to prove it correct. Consider an execution in which the nodes (servers) are partitioned into 2 disjoint set :{ N1} and (N2 ...Nn}. Some node (client) sends a read request to server node N2.Since N1 is in a divergent component of the partition from N2, every message from N1to N2 is lost. Thus it is intolerable for N2 to differentiate the following 2 expressions: **_i._** There has been a preceding write of path value p1 requested of node N1, and N1has sent an ok response. **_ii._** There has been a preceding write of path value p2 requested of node N1, and N1has sent an ok response. No matter how long N2 waits, it cannot differentiate these 2 cases, and as a consequence it cannot ascertain whether to return response p1 or p2. Server node N2 eventually must return a response, even if the system is segregated; if the message delay from N1 to N2 is ----- sufficiently large that N2 believes the system to be differentiated, then it may return an erroneous response, despite the scarcity of partitions. The paramount explanation for extending the CAP theorem is to make the point that in the majority of instances, a distributed system can only guarantee two of the features, not all three. To ignore such a decision could have catastrophic results that include the possibility of all three elements falling apart simultaneously. **_Consistency: A read sees all previously completed writes_** i.e. all nodes see the same data at the same time .g: As the above figure I show that, if Project is set as Root then the Path from Project to department can be established and expressed as ; Project (Root)→Member→Department. Let; the path denoted as ρ. Then, it can be expressed as_ρ(R,C)= the path from Root to CSG._ _Root_ denoted as _R, C(CSG) and E is a_ trinary relation of(CSG,P,ϴ) [ ] [ ] (14) [ ] . (15) (16) For all i,ρ satisfies the layer, for all i and existential C if operator _ρ_ with layer and CSG is satisfied Root then Root implies the operator _ρ with layer and CSG. If the_ operator _ρ_ with layer and CSG satisfies the preceding layer and CSG then it implies the tail node. _Therefore, all nodes see the same data at the same time._ _In addition, it also satisfy the Base Theorem Basic_ _Availability that means it response to any request._ **_Availability: Guarantee that every request receives a_** response about whether it succeeded or failed i.e. read and write always succeed. This means that in GOOSSDM schema there should be a searching path and its return some value. The path value should not be null. Here defining a path means it guarantees that every request receives a response about whether it succeeded or failed i.e. read and write always succeed. When it succeeded then it is succeeded path otherwise, it is failed path. Succeeded path =N1 Failed Path or [ ] (17) [ ] (18) (19) ( ) [ ] (20) ( ) ( ( )) (21) For existential C, let succeeded path is not root. Succeeded path implies for existential N1 if N1 value is null then this will be the path value, should not be null, then again for existential N1 returns failed path or succeeded path or succeeded path with not null value. _Therefore, all searching path must return some value._ _Again, it is also satisfying the Base Theorem Soft State_ _that according to the users’ requirement the desired path_ _will change and it must return some value._ **_Partition_** **_Tolerance:_** Guaranteed properties are maintained even when network failures prevent some machines from communicating with others. The system continues to operate despite arbitrary partitioning due to network failures. [ ] [ ] (22) [ ] (23) (24) For all i, the Root implies operator _ρ with layer and_ CSG. If the operator _ρ_ with layer and CSG satisfies the preceding layer and CSG then it implies the tail node. The all-possible paths of OR operation implies the desired node. _Therefore, the every Node will cultivate q to everywhere_ _it should sooner or later, but the path will continue to_ _receive input and is not checking the consistency of every_ _transaction before it moves onto the next node._ _Read-Write Operation Algorithm_ Assuming node R is the Root node. The algorithm behaves as follows and A is desired node. _Algorithm 1: Read at node A_ _Step 1: A sends a request to R for the recent value._ _Step 2: If A receives a response from R that means find a_ _path value, then save the value and send it to \\\the client._ By applying algorithm R is the root node and scanning from R to the desired node, A returns the path value with arguments in operator layer no and CSG and it is the finding of path value. _Algorithm 2: Write at node A_ _Step 1: A sends a message to R with the new path value._ ----- _Step 2: If A receives an ACK from R, then A sends an_ _ACK to the client and stop._ _Step 3: If A has not yet received an ACK from R, then A_ _sends a message to R with the new value._ Fig.5. Example of read at node A Fig.6. Example of write at node A A sends request to R for the new path value and R scans it from right to left, i.e. R→B→A; A have to wait to get the ACK and B will get the ACK prior to A and then A sends a message to R with the new value. _Algorithm 3: New value is receiving at node R_ _Step 1: R increments its sequence no by 1._ _Step 2: R sends out the new value and sequence no to_ _every node._ Fig.7. Example of New value is received at node R. According to previous algorithm Root will increment its layer value by 1 and every node will getting there layer no i.e. sequence no. VI. VALIDATION OF TRANSACTIONAL CALCULUS OF SEMI-STRUCTURED (TCSS) DATABASE BY CAP AND BASE THEOREM Data validation intended to provide certain well defined guarantees for fitness, accuracy, and consistency for any of various kinds of user input into an application or automated system. Data validation rules can be defined and designed using any of various methodologies, and be deployed in any of various contexts. Data validation, as explained above, is making sure that all data (whether user input variables, read from file or read from a database) are valid for their intended data types and stay valid throughout the application that is driving this data. What this means is data validation, in order to be as successful as it can be, must implemented at all parts that get the data, processes it and saves or prints the results. **_Validation_** In evaluating the basics of data validation, generalizations can made regarding the different types of validation, according to the scope, complexity, and purpose of the various validation operations to be carried out. For example: **_Data type validation: Data type validation customarily_** carried out on one or more simple data fields. The simplest kind of data type validation verifies that the individual characters provided through user input are consistent with the expected characters of one or more known primitive data types; as defined in a programming language or data storage and retrieval mechanism. As the above figure I show that, if Project are set as Root then the Path from Project to Department can be established and expressed as ; **_Project_** **_(Root)→Member→Department._** Let ; the path denoted as ρ. Then, it can be expressed as _ρ(R,C)=i.e. the path from Root to CSG._ [ ] _And_ (25) _Then, [ ]_ (26) [ ] (27) (28) This is the simple example of data validation that verifies that the individual characters provided through user input are consistent with the expected characters of ----- one or more known primitive data types; as defined in a programming language or data storage and retrieval mechanism and in previous section it is already proved that it satisfy the CAP and BASE Theorem. **_Constraint_** **_validation:_** Constraint validation may examine user input for consistency with a minimum/maximum range, or consistency with a test for evaluating a sequence of characters, **_Consistency: A read sees all previously completed writes_** i.e. all nodes see the same data at the same time. _E .g: As the above figure I show that, if Project are set as_ Root then the Path from Project to department can be established and expressed as ; _Project_ _(Root)→Member→Department._ Let ; the path denoted as ρ. Then, it can be expressed as_ρ(R,C)=i.e. the path from Root to CSG._ _Root_ denoted as _R, C(CSG) and E_ trinary relation of(CSG,P,ϴ) _ρ(i) [_ _is the layer]_ _And_ (29) _Then, [ ]_ (30) [ ] (31) (32) **_Therefore, all nodes see the same data at the same time._** This is the simple example of constraint validation and in constraint validation examine for consistency. In previous section it is already proved that consistency satisfy the CAP and BASE Theorem. **_Structured validation: Structured validation allows for_** the combination of any of various basic data-type validation steps, along with more complex processing. Such complex processing may include the testing of conditional constraints for an entire complex data object or set of process operations within a system. _Path(Root,E) [ C(CSG) and E_ trinary relation of(CSG,P,ϴ)] [ ] (33) [ ] (34) (35) _Therefore, the every Node will propagate to everywhere_ _it should sooner or later, but the path will continue to_ _receive input._ This is the example of Structured validation it include complex processing such complex processing may include the testing of conditional constraints for an entire complex data object or set of process operations within a system. VII. TCSS OPERATORS WITH EXAMPLE In previous work defining the GQL-SS algebra for GOOSSDM model that operate on semi-structured schema concept and / or several constructs described in the model. The algebra consists of a set of operators and few of them also can be used with the constructs like ESG, CSG separately. The running example of _Project_ _Management System (PMS)_ used to illustrate the functionalities of operators. As specified earlier _path_ _operator (ρ)_ is also inclusive part of the algebra and invoked every time it is required to invoke any other operator specifically defined for management of semistructured data. Let consider an example of Project Management System (PMS) where a project has several members and members are associated with some departments. Individual members either may or may not have publications. Moreover, each member may participate in any number of projects. The database for PMS is purely semi-structured in nature. The sample data has been showing in table I. _A. Operators in GOOSSDM_ Let us note that in GOOSSDM the data are seen as single rooted graphs or multi rooted graph. In every cases have to set an immediate root for the desired CSG and then also find the tail node in respect to the desired CSG. - **_Select (σ) Operator: The select operator will_** select CSG and returns CSG that satisfy a given list of ESGs or CSGs from the GOOSSDM schema. The tuple relational calculus (TRC) notation is, {C|C CSG (36) {C|List C (37) [List(CSG)=OUTPUT CSG where list={list of ESG}] If the set of all CSG for which the List(C) evaluates true. And the path expression will be like that (38) [for all levels, existential CSG set the path and if it does not have any edge then it is set to Root] (39) (40) ( ) (41) ----- [Searching for desired CSG level by level and get ultimate CSG.] - **_Retrieve (π) Operator:_** The retrieve operator extracts ESG or CSG from the CSG using some constraints _CON over one or more ESG or CSG_ defined in GOOSSDM schema. { | )} (42) [C1 belongs to some CSG with satisfied condition] ( ) (43) [C1 belongs to CSG with specified condition and that returns the restricted CSG.]Let; Constraints=CON ( ) (44) [The dot operator extracts ESG or CSG from the CSG using some specified constraints CON over one or more ESG or CSG defined in schema.] (45) { | ( ) (46) - **_Union, Intersection and Difference (ᴗ,ᴖ,and -_** **_)operators:_** These operators will have usual meaning. The union of any two sets _A and_ _B,_ denoted by AB, is the set of all elements which belong to A or B or both. Hence, A  B ={ x: x _A OR x_ _B}._ (( ) ) (47) [C1 or C2 or specified constraints of dot product of C1 and C2 that returns CSG or ESG which belongs to C1 or C2 or both.] Intersection denoted by AB, is the set of elements, which belong to A, and B both, expressed as _AB= { x: x_ _A AND x_ _B }._ (( ) ) (48) [C1 or C2 or specified constraints of dot product of C1 and C2 that returns CSG or ESG which belongs to C1 and C2.] - **_Join (|X|) operator:_** The join operator is a special case of Cartesian Product operator. It is a binary operator to relate two CSGs where one identical ESG must be common. Let, two CSGs are _CSG1_ and CSG2. Also let, a set of ESG _E1= (E11, E12..._ _E1R)_ and a set of ESG _E2= (E21, E22... E2s) is_ related with theCSG1 and CSG2 respectively. The join operator between _CSG1and_ _CSG2 is possible_ iffE1ɅE2≠. Now letE1ɅE2={Ea,Eb, Ec} then, { | (49) [Specified CSG in C1 with Existential CSG in C2 and both will satisfy a common ESG field.] (( ) ) (50) [All CSG in C1 and CSG in C2 and a common ESG field is satisfied then this will return the all common ESG.] _B. Capabilities of the proposed calculus TCSS_ In this section, the expressiveness capabilities of the proposed calculus of TCSS demonstrated by applying the tuple relational calculus to suitable example queries. _a. Find the project name and project id from the CSG_ _project1._ In this query, the _Select operator has been using to_ select list like _Pname and_ _PID from_ _Project1.The_ calculus can be expressed as follows, {P.Pname, P.PID|Project1 (P)}. Result: <Project1> <PName> ABC</PName> <PID>P1001</PID> <PName> XYZ</PName> <PID>P1003</PID> <PName> DEF</PName> <PID>P1004</PID> <PName> XYZ</PName> <PID>P1005</PID> <PName>ABC</PName> <PID>P1001</PID> </Project1> _b. Find the details of publication whose Member Id_ _MID= M03 and Publication Id PuID= P003 ._ In this query, the Retrieve operator has been used with the constraints of _select operation on select list asMID_ _= M03 from_ _Member CSG and also select the list_ asPID= P003 from Publication CSG. The calculus can be expressed as follows, {P.Publication|Project1(p)Ʌ(Ǝ)(Member(M)ɅM.MID=’ M03’)Ʌ(Ǝ)(Publication(B)ɅB.PuID=’P003’)} ----- Result: <Project1> <Publication> <PuID> P003 </PuID> <Ptopics>SSS</Ptopics> </Publication> </Project1> _c. Find the details of member where MName= Bipin_ _from project1 and also find the details of Member where_ _MName= Priya from Project2._ In this query, the Retrieve operator has been used with the constraints of _select operation on the list_ _Mname_ _= Bipin andMname = Priya_ from _Member CSG. The_ calculus can be expressed as follows, {P.Member|Project1(P)Ʌ(Ǝ)(Member(M)ɅM.MName=’ Bipin’)}V{P.Member|Project2(P)Ʌ(Ǝ)(Member(M)ɅM. MName=’Priya’} Result: <Project1> <Member> <MID> M01</MID> <MName>Bipin</MName> <MAddress> XX </MAddress> </Member> </Project1> <Project2> <Member> <MID> M07</MID> <MName>Priya</MName> <MAddress> CC </MAddress> </Member> </Project2> _d. Find the name of all members who have the same_ _department id “DID=D03 and department name “EE ._ In this query, the Retrieve operator has been used with the constraints of select operation as the list DID= D03 from _Department CSG. Also another Retrieve operator_ has been used with constraints on select operation as the list DName= Electrical from Department CSG. Finally the intersection operator has been used. The calculus can be expressed as follows {P.Member|Project1(P)Ʌ (Member(M))Ʌ(Ǝ)(Depart ment(D)ɅD.DID=’D03’ɅD.Dname=’EE’} Result: <Project1> <Member> <MName>Rashi</MName> </Member> <Member> <MName>Sashi</MName> </Member> </Project1> _e. Find the name of the all members who have the_ _department id same._ In this query, required to set the custom root and then required to apply the join operator. For the purpose, theMemberCSG needs to set the root. The calculus can be expressed by semantics and corresponding result are as follows {P.Member|Project1(P)Ʌ( (M)((Member(M))Ʌ(Depart ment(D))→ D.DID=D.DID)} Result: <Member> <MName>Bipin</MName> <MName>Rashi</MName> <MName>Sashi</MName> <MName>Priya</MName> </Member> _f. Find the project name and project id from the CSG_ _project1 and CSG project2._ In this query, the _Select operator has been used to_ select list like _Pname and_ _PID from_ _Project1 and also_ _Project2.The calculus can be expressed as follows:_ {P.Pname,P.PID|Project1(P)}.V{P.Pname,P.PID|Project2 (P)}. Result: <Project1> <PName> ABC</PName> <PID>P1001</PID> <PName> XYZ</PName> <PID>P1003</PID> <PName> DEF</PName> <PID>P1004</PID> <PName> XYZ</PName> <PID>P1005</PID> <PName>ABC</PName> <PID>P1001</PID> </Project1> <Project2> <PName> ABC</PName> <PID>P1001</PID> <Project2> _g._ _Find_ _the_ _details_ _of_ _publications_ _where_ _MName= Bipin from project1 and also find the details_ _of publication where MName= Priya from Project2._ In this query, the Retrieve operator has been used with the constraints of _select operation on the list_ _Mname_ _= Bipin andMname = Priya_ from _Member CSG. The_ calculus can be expressed as follows {P.Publication|Project1(P)Ʌ(Ǝ)(Member(M)ɅM.MName =’Bipin’)}V{P.Publication|Project2(P)Ʌ(Ǝ)(Member(M) ɅM.MName=’Priya’} Result: <publication> <puid> P001</puid> <ptopics> RRR </ptopics> <puid> P007</puid> ----- <ptopics> NNN</ptopics> </publication> VIII. AN IMPLEMENTATION OF PROPOSED TCSS _A. Transaction Execution:_ Fig.8. Example of transaction execution The above figure 8 shows the root node is 1 and then scanning from right, the next node is 2 and the next after next node is 4 after that it scans for the left node 3.Focusing on a simplified variant of TCSS, that is dynamically typed. To introduce the syntaxes and semantics of TCSS, let us starting with a simple example of transactional query by using x-query. In this section, the expressiveness capabilities of the proposed transactional calculus of TCSS demonstrated by applying the calculus to suitable example queries. <project> <project1> <pname>ABC</pname> <pid>P1001</pid> <topics>AAAA</topics> <member> <mid>M01</mid> <mname>BIPIN</mname> <maddress>xx</maddress> <department> <did>D01</did> <dname>CSE</dname> <publication> <puid>P001</puid> <ptopics>RRR</ptopics> </publication> </department> </member> <pname>XYZ</pname> <pid>P1003</pid> <topics>CCCC</topics> <member> <mid>M03</mid> <mname>ASHU</mname> <maddress>PP</maddress> <department> <did>D02</did> <dname>CA</dname> <publication> <puid>P003</puid> <ptopics>SSS</ptopics> </publication> </department> </member> <pname>DEF</pname> <pid>P1004</pid> <topics>DDDD</topics> <member> <mid>M04</mid> <mname>RASHI</mname> <maddress>YY</maddress> <department> <did>D03</did> <dname>EE</dname> <publication> <puid>P004</puid> <ptopics>TTT</ptopics> </publication> </department> </member> <pname>XYZ</pname> <pid>P1005</pid> <topics>QQQQ</topics> <member> <mid>M06</mid> <mname>SASHI</mname> <maddress>RR</maddress> <department> <did>D03</did> <dname>EE</dname> <publication> <puid>P005</puid> <ptopics>VVV</ptopics> </publication> </department> </member> <pname>ABC</pname> <pid>P1001</pid> <topics>BBBB</topics> <member> <mid>M07</mid> <mname>PRIYA</mname> <mid>M07</mid> <mname>PRIYA</mname> <maddress>CC</maddress> <department> <did>D01</did> <dname>CSE</dname> <publication> <puid>P006</puid> <ptopics>MMM</ptopics> </publication> </department> </member> </project1> <project2> <pname>PQR</pname> <pid>P1006</pid> <topics>YYYY</topics> <member> <mid>M07</mid> <mname>PRIYA</mname> <maddress>cc</maddress> <department> <did>D02</did> <dname>CA</dname> -<publication> <puid>P007</puid> <ptopics>NNN</ptopics> </publication> </department> </member> </project2> </project> **_1._** **_Find the project name and project id from the CSG_** **_project1._** _for $p1 in doc("demo1.xml")//project1_ _for $p2 in doc("demo1.xml")//project1_ _where $p1//topics != $p2//topics_ _return<table ID="project">_ _<pname>{data($p1//pname)}</pname>_ _<pid>{data($p1//pid)}</pid>_ _</project>_ _</table>_ ----- <table ID=” project”> <pname> ABC XYZ DEF XYZ ABC </pname> <pid> P1001 P1003 P1004 P1005 P1001 </pid> </project> </table> **_2._** **_Find the details of publication whose Member Id_** **_MID=”M03” and Publication Id PID=”P003”._** _for $p in doc("demo1.xml")//member_ _where $p//mid = "M03"_ _and $p//puid = "P003"_ _return $p//publication_ <publication> <puid> P003 </puid> <ptopics> SSS</ptopics> </publication> **_3._** **_Find the details of member where MName=”Bipin” from_** **_project1 and also find the details of Member where_** **_MName=”Priya” from Project2._** _for $p1 in doc("demo.xml")/project/project1/member_ _for $p2 in doc("demo.xml")/project/project2/member_ _where $p1//mname = "BIPIN"_ _and $p2//mname = "PRIYA"_ _return<table ID= "project">_ _<member>_ _{$p1//(mid,mname,maddress)}_ _{$p2//(mid,mname,maddress)}_ _</member>_ _</table>_ < table ID= ”project”> <member> <mid> M01</mid> <mname> BIPIN </mname> <maddress> XX </maddress> <mid> M07</mid> <mname> PRIYA</mname> <maddress> CC</maddress> </member> </project> </table> **_4. Find the name of all members who have the same department_** **_id “DID=D03” and department name “EE”._** _for $p in doc("demo1.xml")//member_ _where $p//dname = "EE"_ _and $p//did = "D03"_ _return<project1>_ _<member>_ _{$p//(mid,mname,maddress)}_ _</member>_ _</project1>_ <project1> <member> <mname> RASHI </mname> </member> </project1> <project1> <member> <mname> SASHI </mname> </member> </project1> **_5._** **_Find the name of the all members who have the_** **_department id same_** _for $p1 in doc("demo1.xml")/project/project1/member_ _for $p2 in doc("demo1.xml")/project/project1/member_ _where $p1//did = $p2//did_ _and $p1//puid != $p2//puid_ _return<member>_ _<mname>{data($p1//mname)}</mname>_ _</member>_ <member> <mname> BI PIN</mname> </member> <member> <mname> RASHI </mname> </member> <member> <mname> SASHI </mname> </member> <member> <mname> PRIYA </mname> </member> **_6._** **_Find the project name and project id from the CSG_** **_Project1 and Project2_** _for $p1 in doc("demo1.xml")//project1_ _for $p2 in doc("demo1.xml")//project_ _where $p1//topics != $p2//topics_ _return<table ID="project">_ _<pname>{data($p1//pname)}</pname>_ _<pid>{data($p1//pid)}</pid>_ _<pname>{data($p2//pname)}</pname>_ _<pid>{data($p2//pid)}</pid>_ _</table>_ < table ID=”project”> <pname>ABC XYZ DEF XYZ ABC</pname> <pid> P1001 P1003 P1004 P1005 P1001</pid> <pname> PQR</pname> <pid> P1006</pid> </table> **_7._** **_Find the details of publications where MName=”Bipin”_** **_from project1 and also find the details of publication_** **_where MName=”Priya” from Project2._** _for $p1 in doc("demo1.xml")/project/project1/member_ _for $p2 in doc("demo1.xml")/project/project2/member_ _where $p1//mname = "BIPIN"_ _and $p2//mname = "PRIYA"_ _return<table ID= "project">_ _<publication>_ _{$p1//(puid,ptopics)}_ _{$p2//(puid,ptopics)}_ _</publication>_ _</table>_ <table ID=”project”> <publication> <puid> P001</puid> <ptopics> RRR </ptopics> <puid> P007</puid> <ptopics> NNN</ptopics> </publication> </table> _B. Implementation of TCSS X-Query_ To examine the scalability of proposed TCSS X-Query implementation, trying to perform an experimental evaluation using “Project” xml data. Here also trying to perform a comparison of TCSS X-Query with open source xml processors: BASE-X. _Queries_ Here considering 5 basic types of queries: Selection, Retrieve, Union, Intersection and Join. **_Selection:_** Query 1 finds the project name and project id from the CSG project1 _for $p1 in doc("demo1.xml")//project1_ _for $p2 in doc("demo1.xml")//project1_ _where $p1//topics != $p2//topics_ _return<table ID="project">_ _<pname>{data($p1//pname)}</pname>_ _<pid>{data($p1//pid)}</pid>_ _</project>_ _</table>_ _Query 1_ **_Retrieve: Query 2 finds the details of publication whose_** Member Id MID=”M03” and Publication Id PID=”P003”. _for $p in doc("demo1.xml")//member_ _where $p//mid = "M03"_ _and $p//puid = "P003"_ _return $p//publication_ _Query 2_ ----- **_Union:_** Query 3 finds the details of member where MName=”Bipin” from project1 and also find the details of Member where MName=”Priya” from Project2. _for $p1 in_ _doc("demo.xml")/project/project1/member_ _for $p2 in_ _doc("demo.xml")/project/project2/member_ _where $p1//mname = "BIPIN"_ _and $p2//mname = "PRIYA"_ _return<table ID= "project">_ _<member>_ _{$p1//(mid,mname,maddress)}_ _{$p2//(mid,mname,maddress)}_ _</member>_ _</table>_ _Query 3_ **_Intersection:_** Query 4 finds the name of all members who have the same department id “DID=D03” and department name=“EE”. _for $p in doc("demo1.xml")//member_ _where $p//dname = "EE"_ _and $p//did = "D03"_ _return<project1>_ _<member>_ _{$p//(mid,mname,maddress)}_ _</member>_ _</project1>_ _Query 4_ **_Join:_** Query 5finds the name of the all members who have the department id same. _for $p1 in_ _doc("demo1.xml")/project/project1/member_ _for $p2 in_ _doc("demo1.xml")/project/project1/member_ _where $p1//did = $p2//did_ _and $p1//puid != $p2//puid_ _return<member>_ _<mname>{data($p1//mname)}</mname>_ _</member>_ _Query 5_ _C. Experimental Results_ This paper performance study explores TCSS X-Query ability. Here in Fig 9 it shows that in case of TCSS XQuery each query execution time is near about same to each other and its maintain a parity, whereas BASE-X xquery processor takes more time for selection procedure and takes less time for join queries. Whereas TCSS xquery time remains comparable, i.e. the additional data is processing in the same amount of time. Here TCSS XQuery demonstrated using a real 10 KB XML dataset (trying to perform an experimental evaluation using “Project” xml data.’) for various XML selection, retrieve, union, and intersection and join queries. In future, planning to analysing of big xml data and optimization of the query compiler. Fig.9. above TCSS X-Query and below BASE-X X-Query VIII. CONCLUSION The proposed framework blends semantic of transactional calculus specification and abstraction mechanism with syntaxes in specific modelling. Thus, the paper fulfils the deficiency of systematic methodology in transactional calculus of GOOSSDM model. In addition to this paper proposes a formal transactional calculus called Transactional Calculus for Semi-structured database (TCSS) Further, the transactional calculus is derived from a algebra based query language [11] and illustrated using examples of real life. The benefits of the proposed work are manifold. It provides supports towards (1) representation of precise knowledge of domain independent conceptualisation from structural and functional design concerns with enriched semantics and syntaxes for transactional calculus of semi-structured. (2) realisation of proposed TCSS working with CAP and BASE theorem. (3) a systematic methodology that pave the way of transforming domain analysis. (4) providing guidelines for the purpose of mapping of Transactional Calculus for Semi-structured database. (5) the proposed Transactional system for semi-structured is based on path expression. (6) the path operator is used to set the root node in GOOSSDM schema and also useful to find the path from the root node to desired node for any transaction. (7) facilitate the early verification of the semi-structured data schema structure in correspondence with the desired transactional calculus. The perspective is an extension to this calculus allowing to support larger class of complex queries like aggregates, group by operations. REFERENCES [1] Conrad R., Scheffner D., Freytag J. C., "XML conceptual modeling using UML", 19[th]Intl. Conf. on Conceptual Modeling, PP: 558-574, 2000. ----- [2] Anirban Sarkar, “Design of Semi-structured Database System: Conceptual Model to Logical Representation”, Book Titled: Designing, Engineering, and Analyzing Reliable and Efficient Software, Editors: H. Singh and K. Kaur, IGI Global Publications, USA, PP 74 – 95, 2013. [3] McHugh J., Abiteboul S., Goldman R., Quass D., Widom J., "Lore: a database management system for semistructured data", Vol. 26 (3), PP: 54 - 66, 1997. [4] Badia, A., "Conceptual modeling for semistructured data", 3[rd]International Conference on Web Information Systems Engineering, PP: 170 – 177, 2002. [5] Mani M., “EReX: A Conceptual Model for XML”, 2[nd]International XML Database Symposium, PP 128-142, 2004. [6] Suresh Jagannathan, Jan Vitek,Adam Welc, Antony Hosking, A Transactional Object Calculus, Dept of Comp.sc,Purdue University, West Lafayette, IN 47906, United States. [7] Liu H., Lu Y., Yang Q., "XML conceptual modeling with XUML", 28[th]International Conference on Software Engineering, PP: 973–976, 2006. [8] Combi C., Oliboni B., "Conceptual modeling of XMLdata", ACM Symposium on Applied Computing, PP: 467– 473, 2006. [9] Wu X., Ling T. W., Lee M. L., Dobbie G.," Designing semistructured databases using ORA-SSmodel", 2[nd]International Conference on Web Information Systems Engineering, Vol. 1, PP: 171 –180, 2001. [10] Seth Gilbert and Nancy Lynch. Brewer’s conjecture and the feasibility of consistentavailable, partition tolerant web services.SigActNews, June2002. [11] Rita Ganguly, RajibKumarchatterjee,Anirban Sarkar. “Graph Semantic based Approach for Quering Semistructured Database System.”22[nd] International Conference on SEDE-2013, pp: 79-84. [12] Seth Gilbert National University of Singapore and Nancy Lynch. Brewer’sMassachusetts Institute of Technology,”Perspectives on the CAP Theorem. [13] Soichiro Hidaka Zhenjiang Hu Kazuhiro Inaba Hiroyuki Kato, “Bidirectionalizing Structural Recursion on Graphs”,Techical Report, National Institute of Informatics, The University of Tokyo/JSPS Research Fellow, The University of Electro-Communications, August 31, 2009 [14] Data Validation, Data Integrity, Designing Distributed Applications with Visual Studio NET, Arkady Maydanchik (2007), "Data Quality Assessment", Technics Publications, LLC [15] Object Oriented Transaction Processing in the KeyKOS® Microkernel. William S. Frantz,Periwinkle Computer Consulting, 16345 Englewood Ave. Los Gatos, CA USA 95032 rantz@netcom.com Charles R. Landau,Tandem Computers Inc. 19333 [16] Vallco Pkwy, Loc 3-22,Cupertino, CA USA 95014 landau_charles@tandem.com. Introduction to ObjectOriented Databases. Prof. Kazimierz Subieta,subieta@pjwstk.edu.pl,http://www.ipipan.waw.pl /~subieta Ni W., Ling T. W., “GLASS: A Graphical Query Language for Semi-structured Data”, 8th International Conference on Database Systems for Advanced Applications, PP 363 –370, 2003. [17] R. K. Lomotey and R. Deters, “Datamining from document-append NoSQL,” Int. J. Services Comput., vol. 2, no. 2, pp. 17–29, 2014. [18] Braga, D., Campi, A. and Ceri, S., “XQBE (XQuery By Example): A visual interface to the standard XML query language”, ACM Transactions on Database Systems(TODS), Vol.30 (5), pp. 398 – 443, 2003. [19] AnirbanSarkar, "Conceptual Level Design of Semi structured Database System: Graph-semantic Based Approach", International Journal of Advanced Computer Science and Applications, The SAI Pubs., New York, USA, Vol. 2, Issue 10, PP 112 – 121,November, 2011. [ISSN: 2156-5570(Online) &ISSN : 2158-107X(Print)]. [20] T. W. Ling. A normal form for sets of not-necessarily normalized relations. In Proceedings of the 22nd Hawaii International Conference on System Sciences, pp. 578586. United States: IEEE Computer Society Press, 1989. [21] T. W. Ling and L. L. Yan. NF-NR: A Practical Normal Form for Nested Relations. Journal of Systems Integration. Vol4, 1994, pp309-340. [22] Rita Ganguly,Anirban Sarkar “ Evaluations of Conceptual Models for Semi-structured Database system “. International Journal of ComputerApplications.Vol 50, Issue 18, PP 5-12,july,2012.[ISBN:973-93-80869-67-3]. [23] Rami Sellami, Sami Bhiri, and Bruno Defude, “Supporting Multi Data Stores Applications in cloud Environments.” IEEE Transactions on services computing, vol-9, No-1,pp-59-71, January/February2016. [24] O. Cur_e, R. Hecht, C. Le Duc, and M. Lamolle, “Data integration over NoSQL stores using access path based mappings,” inProc. 22nd Int. Conf. Database Expert Syst. Appl., Part I, 2011, pp. 481–495. [25] ACID vs. BASE: The Shifting pH of Database Transaction Processing, By Charles Roe, www.dataversity net [26] Martin Abadi Microsoft Research.university of california santa cruz, Tim Harris, Microsoft Research, Katherine F Moore Microsoft Research, University of Washington, “A Model of Dynamic Seperation for Transactional Memory”. [27] Manfred Schmidt-Schau_, David Sabel Goethe University, Frankfurt, Germany, ICFP '13, Boston, USA, Correctness of an STM Haskell Implementation. [28] B. Liskov and R. Scheifler. Guardians and actions: Linguistic support for robust distributed programs. ACM Transactions on Programming Languages and Systems, 5(3):381–404, July 1983. [29] J. Eliot B. Moss. Nested Transactions: An Approach to Reliable Distributed Computing.MIT Press, Cambridge, Massachusetts, 1985. [30] Jeffrey L. Eppinger, Lily B. Mummert, and Alfred Z. Spector, editors. Camelot and Avalon: A Distributed Transaction Facility. Morgan Kaufmann, 1991. [31] D. D. Detlefs, M. P. Herlihy, and J. M. Wing. Inheritance of synchronization and recovery in Avalon/C++. IEEE Computer, 21(12):57–69, December 1988. [32] Nicholas Haines, Darrell Kindred, J. Gregory Morrisett, Scott M. Nettles, and Jeannette M. Wing. Composing first-class transactions. ACM Transactions on Programming Languages and Systems, 16(6):1719–1736, November 1994. [33] Alex Garthwaite and Scott Nettles. Transactions for Java. In Malcolm P. Atkinson and Mick J. Jordan, editors, Proceedings of theFirst International Workshop on Persistenceand Java, pages 6–14. Sun Microsystems Laboratories Technical Report 96-58, November1996. [34] Richard J. Lipton. Reduction: a new method of proving properties of systems of processes. InProceedings of the 2nd ACM SIGACT-SIGPLAN symposium on Principles of programming languages, pages 78–86. ACM Press, 1975. [35] Shaz Qadeer, Sriram K. Rajamani, and JakobRehof. Summarizing procedures in concurrent programs. In ----- Proceedings of the 31st ACM SIGPLAN-SIGACT symposium on Principlesof programming languages, pages 245–255. ACM Press, 2004. [36] Nancy Lynch, Michael Merritt, William Weihl, and Alan Fekete. Atomic Transactions. Morgan-Kaufmann, 1994. [37] Panos Chrysanthis and Krithi Ramamritham. Synthesis of Extended Transaction Models Using ACTA. ACM Transactions on Database Systems, 19(3):450–491, 1994. [38] Jim Gray and Andreas Reuter. Transaction Processing: Concepts and Techniques. Data Management Systems. Morgan Kaufmann, 1993. [39] Andrew Black, Vincent Cremet, Rachid Guerraoui, and Martin Odersky. An Equational Theory for Transactions. Technical Report CSE 03-007, Department of Computer Science, OGI School of Science and Engineering, 2003. [40] Tom Chothia and Dominic Duggan. Abstractions for Fault-Tolerant Computing. Technical Report 2003-3, Department of Computer Science, Stevens Institute of Technology, 2003. [41] N. Busi, R. Gorrieri, and G. Zavattaro. On the Serializability of Transactions in Java Spaces. InConCoord 2001, International Workshop on Concurrency and Coordination, 2001. [42] R. Bruni, C. Laneve, and U. Montanari. Orchestrating Transactions in the Join Calculus.In 13th International Conference on Concurrency Theory, 2002. [43] E. Preston Carman, Jr.1, Till Westmann2§, Vinayak R. Borkar3*, Michael J. Carey4, Vassilis J. Tsotras11University of California, Riverside 2Couchbase _3X15 Software, Inc. 4University of California, Irvine_ _Email: ecarm002@ucr.edu A Scalable Parallel XQuery_ Processor 2015 IEEE International Conference on Big Data (Big Data)978-1-4799-9926-2/15/$31.00 ©2015 IEEE 164. [44] Shreya Banerjee and Anirban Sarkar “Ontology-driven approach towards domain-specific system design “.International journal semantics and ontologies, vol 11, no 1, pp- 39-60. [45] ACID vs. BASE: The Shifting pH of Database Transaction Processing,By Charles Roe,www.dataversity net. **Authors’ Profiles** **Rita Ganguly, received the M.Tech degree** from the NIT, Durgapur, India and entitled her name as a Research Scholar (Part-time) in Computer Science department(formerly known as Computer Application department.),NIT, Durgapur under the supervision of Dr. Anirban Sarkar. Presently she is working as an Asst. Prof of Computer Application Department, in Dr. B.C.Roy Engineering College, Durgapur, India. **Anirban Sarkar is presently a faculty** member in the Department of Computer Applications, National Institute of Technology, Durgapur, India. He received his PhD degree from National Institute of Technology, Durgapur, India in 2010. His areas of research interests are Database Systems and Software Engineering. His total numbers of publications in various international platforms are above 100. He is actively involved in collaborative research with several Institutes in India and USA and has also served in the committees of several international conferences in the area of software engineering and computer applications. **How to cite this paper: Rita Ganguly, Anirban Sarkar, "An** Approach to Develop a Transactional Calculus for SemiStructured Database System", International Journal of Computer Network and Information Security(IJCNIS), Vol.11, No.9, pp.24-39, 2019.DOI: 10.5815/ijcnis.2019.09.04 -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.5815/ijcnis.2019.09.04?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5815/ijcnis.2019.09.04, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GOLD", "url": "http://www.mecs-press.org/ijcnis/ijcnis-v11-n9/IJCNIS-V11-N9-4.pdf" }
2,019
[]
true
2019-09-08T00:00:00
[ { "paperId": "02bc9dffd65ad41edd6b3a51ebfe747f64dc9d2d", "title": "DATA MINING FROM DOCUMENT-APPEND NOSQL" }, { "paperId": "1c3554d61c41996149e2edde56f8732633061293", "title": "Correctness of an STM Haskell implementation" }, { "paperId": "21595186c0570f654034cc189faa8ec00fa2dc44", "title": "Evaluations of Conceptual Models for Semi-structured Database System" }, { "paperId": "5feaafaa4103a751dc39b1be389a56004536c218", "title": "Conceptual Level Design of Semi-structured Database System: Graph-semantic Based Approach" }, { "paperId": "e7cf963e61e767f480a6e03300b0a4cd74686435", "title": "Perspectives on the CAP Theorem" }, { "paperId": "05352f0809cd42eabd275738af180d9311df7448", "title": "Data Integration over NoSQL Stores Using Access Path Based Mappings" }, { "paperId": "b69295dafcd49459eef48d69cf6a5d06b26ddbf2", "title": "XML conceptual modeling with XUML" }, { "paperId": "7b8d3ff9559708c7bc61ab6b8458eeab452fdaae", "title": "A transactional object calculus" }, { "paperId": "180570e4a015111874cc7bfde1aee923da93338e", "title": "XQBE (XQuery By Example): A visual interface to the standard XML query language" }, { "paperId": "0ad47673386f7630a25b2ab571cbfc17bd6229f8", "title": "EReX: A Conceptual Model for XML" }, { "paperId": "9311704307e73b74e9860b55e12b0c1e3cacd32f", "title": "Summarizing procedures in concurrent programs" }, { "paperId": "24b3d948ad277862d96c0a77579b9649176a9949", "title": "An Equational Theory for Transactions" }, { "paperId": "18c84d6e03edc5b482ea76326a81b8fd89d8890c", "title": "Conceptual modeling for semistructured data" }, { "paperId": "f8525b50a77f8730a661807be7b04f394a6dbd44", "title": "Orchestrating Transactions in Join Calculus" }, { "paperId": "bbe7a291e74d12d0bcad6f623d8a13e084b83255", "title": "Brewer's conjecture and the feasibility of consistent, available, partition-tolerant web services" }, { "paperId": "f192c76b6fcda1b274cd1fd65efd1467f20e0127", "title": "Designing semistructured databases using ORA-SS model" }, { "paperId": "acf312724831f680eeb9198cc7cacbb522f3b03e", "title": "XML Conceptual Modeling Using UML" }, { "paperId": "ff486da7373f75a6f20d3b64f4c07c0067542683", "title": "Transactions for Java" }, { "paperId": "099f9eccf15aed58801051148a5e937d05f4ede3", "title": "Lore: a database management system for semistructured data" }, { "paperId": "34e72fa4beed59a8fa9adbaafe0dcf7923887dc5", "title": "NF-NR: A practical normal form for nested relations" }, { "paperId": "9e6263d1e02acd28601c2e2ccfc0fbed0c36fca9", "title": "Composing first-class transactions" }, { "paperId": "b7d2b569049706ed33b36605adcaefef22da08f8", "title": "Synthesis of extended transaction models using ACTA" }, { "paperId": "66f8fcd2cbbaa169943b6b6b3b7633e4c08f85e7", "title": "Object-Oriented Transaction Processing in the KeyKOS Microkernel" }, { "paperId": "66ec4f8c1e6f471881cae3e2d97735713c5c071d", "title": "Transaction Processing: Concepts and Techniques" }, { "paperId": "9d4fb7fba5b1a3fe006cdc5bc083629826a2df69", "title": "Camelot and Avalon: A Distributed Transaction Facility" }, { "paperId": "38d1c6d72afd3e215d537226236f0e18d16fd29b", "title": "A normal form for sets of not-necessarily normalized relations" }, { "paperId": "b422a8fff5ce4a46eb4c964c88b1003c2a2a710e", "title": "Guardians and actions: linguistic support for robust, distributed programs" }, { "paperId": "6a5632a789cbe7c1a690ca310731c2cde0155c76", "title": "Ontology-driven approach towards domain-specific system design" }, { "paperId": "4a41efb7d12d11eec7e672c836e7a45ed7e3b66a", "title": "Supporting Multi Data Stores Applications in Cloud Environments" }, { "paperId": null, "title": "Tsotras11University of California, Riverside 2Couchbase 3X15 Software, Inc. 4University of California, Irvine Email: ecarm002@ucr.edu A Scalable Parallel XQuery Processor 2015" }, { "paperId": null, "title": "Graph Semantic based Approach for Quering Semistructured Database System.”22" }, { "paperId": "715653150bddd42c017220596bef7e9601991f8c", "title": "Design of Semi-Structured Database System: Conceptual Model to Logical Representation" }, { "paperId": null, "title": "“Bidirectionalizing Structural Recursion on Graphs”" }, { "paperId": null, "title": "Data Validation, Data Integrity, Designing Distributed Applications with Visual Studio NET" }, { "paperId": null, "title": "Abstractions for Fault-Tolerant Computing" }, { "paperId": null, "title": "USA 95014 landau_charles@tandem.com. Introduction to Object-Oriented Databases. Prof. Kazimierz Subieta ,subieta@pjwstk" }, { "paperId": null, "title": "On the Serializability of Transactions in Java Spaces" }, { "paperId": "39667791d5a1e289d47f2fb93b557a503ce6603f", "title": "Atomic Transactions" }, { "paperId": null, "title": "Inheritance of synchronization and recovery in Avalon/C++" }, { "paperId": "cfdc0b56a1789e2e67128a0e48534641f80443c9", "title": "Nested Transactions: An Approach to Reliable Distributed Computing" }, { "paperId": "f4a00bdc978e316a85509d9b872dfc996c3da1cd", "title": "Reduction: a new method of proving properties of systems of processes" }, { "paperId": null, "title": "ACID vs. BASE: The Shifting pH of Database Transaction Processing, By Charles Roe" }, { "paperId": null, "title": "“A Model of Dynamic Seperation for Transactional Memory”" } ]
18,891
en
[ { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0112d65d2ea11c065d7eb5f6fada9287002aa158
[]
0.895491
A Data Management Model for Intelligent Water Project Construction Based on Blockchain
0112d65d2ea11c065d7eb5f6fada9287002aa158
Wireless Communications and Mobile Computing
[ { "authorId": "2005591", "name": "Zhoukai Wang" }, { "authorId": "2124633689", "name": "Kening Wang" }, { "authorId": "2143491502", "name": "Yichuan Wang" }, { "authorId": "2158400663", "name": "Zheng Wen" } ]
{ "alternate_issns": null, "alternate_names": [ "Wirel Commun Mob Comput" ], "alternate_urls": [ "https://onlinelibrary.wiley.com/journal/15308677", "http://www.interscience.wiley.com/jpages/1530-8669/" ], "id": "501c1070-b5d2-4ff0-ad6f-8769a0a1e13f", "issn": "1530-8669", "name": "Wireless Communications and Mobile Computing", "type": "journal", "url": "https://www.hindawi.com/journals/wcmc/" }
The engineering construction-related data is essential for evaluating and tracing project quality in industry 4.0. Specifically, the preservation of the information is of great significance to the safety of intelligent water projects. This paper proposes a blockchain-based data management model for intelligent water projects to achieve standardization management and long-term preservation of archives. Based on studying the concrete production process in water conservancy project construction, we first build a behavioral model and the corresponding role assignment strategy to describe the standardized production process. Then, a distributed blockchain data structure for storing the production-related files is designed according to the model and strategy. In addition, to provide trust repository and transfer on the construction data, an intelligent keyless signature based on edge computing is employed to manage the data’s entry, modification, and approval. Finally, standardized and secure information is uploaded onto the blockchain to supervise intelligent water project construction quality and safety effectively. The experiments showed that the proposed model reduced the time and labor cost when generating the production data and ensured the security and traceability of the electronic archiving of the documents. Blockchain and intelligent keyless signatures jointly provide new data sharing and trading methods in intelligent water systems.
Hindawi Wireless Communications and Mobile Computing Volume 2022, Article ID 8482415, 16 pages [https://doi.org/10.1155/2022/8482415](https://doi.org/10.1155/2022/8482415) # Research Article A Data Management Model for Intelligent Water Project Construction Based on Blockchain ## Zhoukai Wang,[1,2] Kening Wang,[3] Yichuan Wang,[1,2] and Zheng Wen 4 1School of Computer Science and Engineering, Xi’an University of Technology, Xi’an, 710048, China 2Shaanxi Provincial Key Laboratory of Network Computing and Security Technology, Xi’an, 710048, China 3School of Automation and Information Engineering, Xi’an University of Technology, Xi’an, 710048, China 4School of Fundamental Science and Engineering, Waseda University, Tokyo 169-8050, Japan Correspondence should be addressed to Zhoukai Wang; zkwang@xaut.edu.cn Received 7 December 2021; Accepted 16 February 2022; Published 9 March 2022 Academic Editor: Qingqi Pei [Copyright © 2022 Zhoukai Wang et al. This is an open access article distributed under the Creative Commons Attribution](https://creativecommons.org/licenses/by/4.0/) [License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is](https://creativecommons.org/licenses/by/4.0/) properly cited. The engineering construction-related data is essential for evaluating and tracing project quality in industry 4.0. Specifically, the preservation of the information is of great significance to the safety of intelligent water projects. This paper proposes a blockchain-based data management model for intelligent water projects to achieve standardization management and long-term preservation of archives. Based on studying the concrete production process in water conservancy project construction, we first build a behavioral model and the corresponding role assignment strategy to describe the standardized production process. Then, a distributed blockchain data structure for storing the production-related files is designed according to the model and strategy. In addition, to provide trust repository and transfer on the construction data, an intelligent keyless signature based on edge computing is employed to manage the data’s entry, modification, and approval. Finally, standardized and secure information is uploaded onto the blockchain to supervise intelligent water project construction quality and safety effectively. The experiments showed that the proposed model reduced the time and labor cost when generating the production data and ensured the security and traceability of the electronic archiving of the documents. Blockchain and intelligent keyless signatures jointly provide new data sharing and trading methods in intelligent water systems. ## 1. Introduction In the water conservancy project management, archives have the characteristics of large numbers and comprehensive coverage, and they play an essential role in all aspects of engineering construction. With the increasing investment of water conservancy projects, the scale gradually grows, and the project gradually becomes complex. The management of water conservancy project archives also faces more and more problems, which restrict the development of water conservancy projects. On the other side, the traditional file management mode can no longer adapt to the rapidly developing economic needs, so the introduction of digital archives for water conservancy projects has become an inevitable trend [1, 2, 3]. However, because the construction of water conservancy projects requires the global deployment and management of various units and resources, making the digitization process of its archives difficult, the status of library management leading to the water conservation institutions requires acceleration transformation [4]. At present, digital archives of water conservancy projects have less relevant research in foreign countries, and the research in China is also in the initial stage [5, 6]. Although the new “Archives Law of the People’s Republic of China” provides legal and policy guarantees for the informationization of construction files of water conservancy project construction, the relevant research and application are still focused on the initial stage of construction [7]. Other important aspects of water conservancy project construction, such as concrete production and mixing, and metal structure installation, still lack effective information management means [8]. In addition, the current digital file management ----- 2 Wireless Communications and Mobile Computing methods are relatively simple, with the drawbacks of poor antitampering and antirepudiation capabilities, and their application range is also limited. In total, the current digital file management methods cannot undertake the engineering construction works involving significant safety needs [9]. In response to the shortcomings of traditional file management methods, this paper introduces blockchain and keyless signature techniques [10], takes the concrete mixing process as the research object, and conducts research on data management in intelligent water conservancy construction. The main contributions of this paper are demonstrated as follows. (1) By employing the smart keyless signatures, this paper established a paperless concrete production and operation management model to monitor the concrete mixing process and prevent data tampering during the process (2) With the help of the consortium blockchain, this paper built up an intelligent document storage method to effectively supervise the progress, quality, and safety of concrete production and then explore general methods for encryption, storage, and traceability of production files (3) Integrated with the corresponding model and method, this paper proposed a blockchain-based file management system for concrete mixing procedures and then implemented it in the Hanjiang to Weihe River Project to improve the production management capacity markedly ## 2. Motivation Concrete mixing is a vital link in the construction of water conservancy projects, and many engineering archival documents are generated during the mixing process to record the concrete mixing details [11]. These files are crucial basic information for project quality control and problem tracing and are related to the whole life cycle safety of the project. However, the management of concrete production files has problems such as low informationization and insufficient security at present [12, 13]. Firstly, the current management method wastes paper. The volume of files related to concrete mixing production is enormous. The amount of grouting required for reservoir construction is usually more than 100,000 cubic meters, which will generate a massive amount of paper data that is difficult to store and manage. Secondly, the paper-based management method has less credibility. The manually dumped paper files are not standardized, and falsification of the paper files often occurs. Thirdly, the traceability of the paper files is feeble. Currently, the cataloging and archiving of concrete production files have not yet formed a strict and complete discipline and management system. Therefore, it is difficult to achieve practical traceability issues tracing. At last, the current management methods obtain insufficient security since the lack of security and confidentiality control measures for the massive paper files. In response to the above problems, more and more researchers have devoted their efforts to studying the digital file management of water conservancy projects, especially for the informatization of the concrete mixing process, and preliminary research results have been achieved. The representative projects include the Jingtaichuan Dam in Gansu Province, the Daxing Water Conservancy Hub Project in Guizhou Province, and Chushandian Reservoir in Henan Province [14]. However, these research results still fail to completely solve the shortcomings of low antitampering ability and poor antirepudiation ability of digitized archives [15]. Digital archives are still exposed to risks, and they are difficult to effectively manage the concrete mixing procedure and ensure the procedure’s safety. The quality of concrete production and the management of related production documents are closely associated with the safety of people’s lives and property. They have a high level of tamper-proof and repudiation-proof requirements [16]. Although the Chinese government has established the corresponding laws to push forward electronic signatures steadily, digital files are bound to be severe trust and security concerns when transmitted over the Internet and stored in centralized servers for long periods [17]. The higher the sensitivity of the data, the greater the risk of using high-tech means to “blacken” it. Apparently, there are substantial technical difficulties in achieving the highly informative management of concrete mixing files paperless. How to help the concrete mixing files and the corresponding data get rid of the security threat becomes a hot topic in the intelligent water conservancy field. In 2009, blockchain technology was proposed to guarantee the security of data. Recently, applications based on blockchain have increasingly appeared in various fields of daily social life, such as finance, public services, culture and entertainment, data insurance, and general welfare [18–20]. However, the present archival research on the blockchain mainly focuses on the feasibility of document archive management and specific application methods [21]. Many scholars have proposed their application for standard archival management based on blockchains, such as museum archives, student archives, and medical information archives [22, 23]. Other scholars have also discussed the challenges and troubles that blockchain technology may face when applied in archival management [24, 25]. But there are only a few cases of the practical application of blockchain-based archive management in hydraulic engineering fields [26]. In summary, applying blockchain and related technologies in engineering construction archive management, especially to critical aspects such as concrete production, has received less attention from relevant studies domestic and abroad. Based on the current research foundation in related fields, this paper takes the whole concrete production process as the research object, integrates blockchain technology with the specific needs of water conservancy projects, and improves the management quality of the electronic files in the concrete mixing process. Meanwhile, this paper also proposes a highly integrated information management system to guarantee the data security of each step in the concrete mixing procedure. The specific steps are as follows: first, ----- Wireless Communications and Mobile Computing 3 study the relationship between the different concrete production departments and establish a behavioral model describing the concrete production process; second, design and implement a distributed blockchain data structure for concrete production process management; third, use keyless signature technology to manage the type-in, modification, and approval process of the concrete production files; finally, all the files generated in the concrete production process are uploaded to the blockchain to achieve openness and transparency of the entire process, guaranteeing accurate traceability of production files and quick location of quality problems, thus effectively supervising the data quality and safety in the intelligent water conservancy projects construction. ## 3. Behavioral Model for the Concrete Production Process 3.1. Process Sorting and Role Assignment. To establish a behavioral model for the concrete production process, we first need to sort out the production process. As shown in Figure 1, the concrete production process is divided into raw material preparation and concrete production parts. Specifically, the raw material preparation part can be divided into the import and test subpart. In contrast, the concrete production part can be divided into the mix proportion design subpart, the concrete mixture subpart, and the concrete test subpart. The raw materials for concrete production include cement, fly ash, admixtures, coarse aggregate, and fine aggregate. The first three materials are transported and supplied by the corresponding manufacturers, while the other materials can be produced by the mixing plant itself. As Figure 1 illustrates, in the material import stage, the quality and quantity reports are provided along with the entry of the purchased raw materials. When the raw materials are in storage, the laboratory of the mixing plant will sample and measure them and then record the report of the material test results in the ledger by computer. Besides, self-made raw materials like coarse and fine aggregates are also tested in detail and recorded by the laboratory of the mixing plant either. At last, all these raw material inspection reports are submitted to the supervision, and the supervision’s approval allows the materials to participate in concrete production. In the concrete production stage, the construction unit submits an application of concrete to the mixing plant. Moreover, the required concrete grade and performance requirements, the required quantity, and the use purpose are also informed to the mixing plant at the same time. After receiving the application, the laboratory personnel in the mixing plant will inspect the moisture content of sand and stone, check the exceeding and inferior grain in aggregate according to the relevant regulations, and then design the concrete mixing proportion. After the supervisor confirms the mixing ratio, the relevant mixing information is provided to the mixing plant. The mixing plant strictly follows the ratio, sets the raw material feeding value, and operates the mixing plant for concrete production. Besides, the raw material temperature and weighing information are recorded during the concrete mixing process according to the regulations. After the concrete mixture, samples are taken from the outlet of the mixing plant; then, the construction unit tests the samples’ quality and forms the sample record and test report. The role assignment could be set as follows by sorting the concrete production process. The main characters involved in the production process are the mixing plant, the laboratory of the mixing plant, the construction department of China Railway 12th Bureau (CR-12 in short), the laboratory of CR-12, the supervisor, and the third-party testing center. In specific, the mixing plant and its laboratory worked in the raw material preparation stage, while CR-12 and the corresponding laboratory worked in the concrete production stage. At last, the supervisor and the thirdparty testing center took part in every stage of the concrete production process to ensure the safe and reliable quality of the whole concrete production process. 3.2. Classification of Concrete Production Files. The second step of building the concrete production behavioral model is to classify all the files involved in the concrete production process according to their attributes. The files include the raw material performance testing records before concrete mixing, the concrete supply contact sheets, the descriptions on concrete mixing proportion, the records about the mixing process, the result of the concrete performance testing, forms related to each cycle errata, and summaries. The cooperation of these files is demonstrated as follows: The manufacturers supply the raw materials to the mixing plant for concrete production. After production, the mixing plant’s laboratory samples the concrete and conducts a quality inspection. If the concrete meets the quality standards, it would be transported to the construction department of CR-12 by vehicles. After the additional tests conducted by the laboratory of CR-12, the construction department of CR-12 builds the water conservancy facilities with qualified concrete. At last, as a neutral third party, the supervisor keeps on inspecting the concrete by commissioning a third-party laboratory to sample and test the concrete at all stages during the production. In total, after summarizing the files involved in the concrete production process, 50 categories of forms are obtained. There are a total of 29 forms related to raw materials, 1 contact sheet for material supply, 7 forms related to the concrete mixing process, 12 forms related to testing, and 1 form for erratum summary. The details are in Figure 2. ## 4. Distributed Blockchain Data Structure 4.1. General Framework Design. Based on the behavioral model, the distributed blockchain data structure can be constructed, and then, the preservation, categorization, and management of the concrete production-related archives can be achieved. The general framework design is illustrated in Figure 3. In Figure 3, the archives generated in concrete production are divided into temporal and spatial levels in the order of warehouse blocks, procedure blocks, branch ----- 4 Wireless Communications and Mobile Computing Concrete Material test - Material quality - Aggregate moisture mixture - Sample record - Material quantity content - Test report - Sample record - Operating record - … … - Aggregate inferior - … … - Inspection report - Discharge port record grain - … … - … … - … … Material Mix proportion Concrete test import design Figure 1: Schematic diagram of the concrete production process. Feedback 10 forms Concrete mixtureplant Into storage Material Spot check Concrete mixtureplant Supervision Project supervision (laboratory) 3 forms manufacturer 3 forms (field) 14 forms Concrete Set Constructor requirement Concrete mixtureplant proportion Concrete mixtureplant Sampling Concrete mixtureplant (field) 1 form (laboratory) 2 forms (field) 1 form (laboratory) Feedstock 4forms Supervision report 2 forms Commit Constructor Test Constructor Cooperate Project supervision record Tird-party (field) 2 forms (laboratory) 4 forms 2 forms laboratory Figure 2: Classification and statistics of the concrete production files. Warehouse Warehouse x-1 Warehouse x Warehouse x+1 block Procedure Data before mixing Supply list Data in mixing Outlet sampling Errata and summary block @A @B @C @D @E Mixing Mix Mixing Weight Branch Producer Lab Supervisor Detail list Lab Constructor Supervisor Errata Summary plant proportion records and temp block @Ab @Ac @Ad @Ba @Da @Db @Dc @Ea @Eb @Aa @Ca @Cb @Cc @Aa001 @Ab001 @Ba001 @Ca001 @Cb001 @Cc001 @Da001 @Db001 @Dc001 @Ea001 @Eb001 @Aa002 @Ab002 @Ca002 @Cc002 @Da002 @Db002 @Dc002 @Aa003 @Ab003 @Cc003 @Db003 @Cc004 @Db004 Unit block @Ac001 @Ac006 @Ac011 @Ad001 @Ad006 @Ac002 @Ac007 @Ac012 @Ad002 @Ad007 @Ac003 @Ac008 @Ac013 @Ad003 @Ad008 @Ac004 @Ac009 @Ac014 @Ad004 @Ad009 @Ac005 @Ac010 @Ad005 @Ad010 Figure 3: General framework of the distributed blockchain data structure. |Warehouse block|Warehouse x-1 Warehouse x Warehouse x+1| |---|---| |Procedure block Branch block Unit block|Data before mixing Supply list Data in mixing Outlet sampling Errata and summary @A @B @C @D @E Mixing Mix Mixing Weight Producer Lab Supervisor Detail list Lab Constructor Supervisor Errata Summary plant proportion records and temp @Ab @Ac @Ad @Ba @Da @Db @Dc @Ea @Eb @Aa @Ca @Cb @Cc @Aa001 @Ab001 @Ba001 @Ca001 @Cb001 @Cc001 @Da001 @Db001 @Dc001 @Ea001 @Eb001 @Aa002 @Ab002 @Ca002 @Cc002 @Da002 @Db002 @Dc002 @Aa003 @Ab003 @Cc003 @Db003 @Cc004 @Db004 @Ac001 @Ac006 @Ac011 @Ad001 @Ad006 @Ac002 @Ac007 @Ac012 @Ad002 @Ad007 @Ac003 @Ac008 @Ac013 @Ad003 @Ad008 @Ac004 @Ac009 @Ac014 @Ad004 @Ad009 @Ac005 @Ac010 @Ad005 @Ad010| blocks, and cell blocks. The structure in Figure 3 represents a comprehensive mixing procedure for a warehouse of concrete, and the warehouse is the fundamental quantity unit in the concrete production process. Inside the data structure, the subblock is composed of one or several distributed led gers. The functions and properties of each subblock in the distributed blockchain are described below. The top element in the distributed blockchain data structure is the warehouse block. The warehouse block contains all the files during the concrete mixing process. In the actual ----- Wireless Communications and Mobile Computing 5 environment, the whole construction procedure of the water conservancy project is often divided into unit projects, division projects, and cell projects. Furtherly, the cell projects are refined into a series of sequential subtasks, and the data and corresponding files generated in each subtask are formed as a warehouse block. Like the example in Figure 3, during the mixing process, the computers automatically record the production data for each tray of concrete, including the set and actual usage amount of the raw materials, the mixing time, the use of the concrete, and other detailed information. The warehouse blocks are made up of cyclic packets, and they are numbered sequentially from 0001 onwards in chronological order. The procedure blocks are the blocks that indicate the specific flows of the concrete production. Note that the block could not be formed until the previous one is generated, and all blocks in the same layer are chained together in a tandem pattern. As shown in Figure 3, the concrete production process contains five procedure blocks: data before mixing, supply list, data in mixing, outlet sampling, errata, and summary. The blocks in the branch block layer are the distributed ledgers created and categorized by different roles in the concrete mixing procedure. For example, @Aa,@Ab,@Ac,@Ad are the branch blocks under the same procedure block A in Figure 3; they represent the file collections in the mixing plant, the producer, the lab, and the supervisor, respectively. The bottom layer in the distributed blockchain data structure is the unit block layer. In this layer, the unit blocks are the specific files, forms, images, or other media boundaries with archival requirements in branch blocks, marked with 001, 002, and so on. As shown in Figure 3, each unit block refers to one file created by a specific role. Besides, the naming scheme of the proposed distributed blockchain data structure is as follows: Firstly, “@” and “#” in the front of each unit block number indicate if this block is shared or not. Secondly, the warehouse block is often divided into blocks for the raw material test, blocks for the concrete inspection, and the other blocks. Among them, the raw material inspection blocks record the samples and the test results of raw materials, such as 200~400 t a sampling unit of cement, 100~200 t a sampling unit of fly ash, and 50 t a sampling unit of admixture. Thirdly, if the unit block is shared, the provenance of the shared data should be indicated, and the indication method is to add the name of the warehouse block which contains the shared block. For instance, if the cement test report is “×××××Ac013,” suppose that a new cement test report is generated in warehouse block “0020,” then its number is “0020#Ac013.” If the next five bunker blocks “0021,” “0022,” “0023,” “0024,” and “0025” need to quote the previous report rather than generating new cement inspection reports, then the quoted report is named as “0020@Ac013.” But if the warehouse block “0026” generates a new cement inspection report, then the name of the report is “0026#Ac013.” 4.2. Distributed Storage Architecture for Digital Archives. Based on the design of blockchain data structure, this paper classifies the files according to different production roles and then stores them in distribution. Specifically, the main characters participating in the concrete production procedure keep their own files locally. For instance, the construction department that initialized and transmitted the supply list would leave a copy of the list in the server of CR-12. Similarly, if the mixing plant initiates the batching notification form, then the form is stored in the computer of the mixing plant. The rules for the rest of the file storage locations are similar, except that the files that are shared by different branches should be stored by both the sending and the receiving units. Moreover, the data-sharing scheme is another crucial part of distributed storage architecture, and it consists of two parts: the data sharing between files and inside files. The data sharing between files means keeping the same sections’ consistency and accuracy in different files. There is a mapping or logical relationship between information in some files and information in the other files during transmission. Therefore, when creating such files, we first store this information in public memory and then automatically obtain the corresponding data with the same content on different files. For illustration, “construction site, strength grade, collapse level, and planned quantity” in the batching order are derived from the contents “construction site, seepage and frost resistance, collapse level and outlet temperature, and concrete supply order” in the supply list. Similarly, the “oversize content” and “undersize content” on the batching notice come from the same contents on the coarse aggregate test records. However, if the shared data is inconsistent, we will issue warning messages to senders and receivers. Then, the file is rejected by the receiver until the sender makes corrections. During the revision, the character who makes the file first checks if the inconsistency is indeed caused by himself then resolves this dispute by amending the filled-in content. Otherwise, the inconsistency is caused by the incorrect data in the system. Then, the dispute will be temporarily put on hold through the consensus mechanism and resolved through the errata at the end of this warehouse block. The data-sharing scheme inside files means that the files are shared between different warehouse blocks. For illustration, in the raw material test stage, an inspection form for raw materials may cover more than one warehouse block; then, these blocks share the same inspection form. As mentioned above, the specific method to distinguish the shared and the unshared data uses “@” and “#” symbols as indicators. ## 5. Edge Computing Supported Intelligent Keyless Signature During the concrete mixing, every authentic and valid file requires the principal’s signature of every department, and the signature means the approval of the file content. This signature process is represented as the form-filling operation in the proposed model. However, there is a risk of tampering with the file during the filling process. The traditional approach is to introduce asymmetric encryption technology in the file approval process to ensure the security of ----- 6 Wireless Communications and Mobile Computing transmission and the file’s integrity. But this technical approach has certain management risks because it involves the management of an individual’s private key. Therefore, this paper employs a keyless signature technology based on edge computing to standardize the form-filling process and provide security for electronic files. 5.1. Hash Tree Construction Based on Edge Computing. The fundamental method for data security during the file transmission is to use the hash function to make a calculation on the file and then regard the calculation result as a digital fingerprint to prove the file’s authenticity. In detail, the proposed management model uses the SHA-256 hash algorithm to calculate the file, generate a 256-bit hash value, perform a series of operations with the hash value, and build up a hash tree. The process is in Figure 4. In Figure 4, x1 to x8 represent the hash values calculated with the SHA-256 algorithm, and these values are the input of the leaf nodes in the hash tree. h denotes the hash funcðÞ tion, and the vertical line represents the join operation, but _hðx1 ∣_ _x2Þ ≠_ _hðx2 ∣_ _x1Þ. The hash tree introduces the hash_ function to fulfill zero-knowledge proof and ensure that the file is authentic. For example, suppose the initial data _x3 knows the hash values fx4, x12, x58g and their position_ markers f1,0,1g. In that case, the root value can be recreated, thus proving that x3 is involved in calculating the generated root value. In total, based on hash chains, goals including a fast comparison on massive data, locating the modified data, and constructing zero-knowledge proofs, can be easily achieved. The hash chain computing process is shown in Figure 5. Further, to secure data transfer and file integrity from the spatial dimension, a large number of hash trees need to be aggregated into Merkle trees simultaneously, and edge computing is the best way to achieve such goals. A Merkle tree consists of a root node, some intermediate nodes, and a set of leaf nodes. Each leaf node is labeled with the hash value of the digital file, while intermediate nodes other than the leaf nodes are marked with the cryptographic hash of their child node labels. Creating a complete Merkle tree requires recursively hashing a set of nodes and inserting the generated hash nodes into the tree until only one hash node remains, which is also called the Merkle root. The construction process of the Merkle tree is in Figure 6. As shown in Figure 6, Merkle trees are created and destroyed once per second. These trees are composed of a hierarchical network of geographically independent distributed computing nodes. Each operates in an asynchronous aggregation fashion, generating a hash tree by receiving hash values from its subtrees transmitting the hash root values to multiple parents. The aggregation process is theoretically unbounded and runs on top of virtual machines or dedicated hardware. Moreover, in a keyless signature system with a multilayer aggregation hierarchy, the acceptable theoretical limit of the system is 2[64] signatures per second. 5.2. The Intelligent Keyless Signature System. The keyless signature system based on Merkle trees is shown in Figure 7, and the specific tree construction process can be described as follows. Firstly, the department participating in the concrete mixing procedure submits the hash value (the blue dots in Figure 7) of the file to the customized keyless signature gateway. Secondly, the adjacent hash values are connected in series, and then, an additional hash operation on the concatenated values is performed again to calculate the result. Subsequently, the newly calculated hash value is submitted to the upper layer for serial hash operation until the Merkle tree’s root is created. Finally, the keyless signature gateway returns a keyless signature to the department. The keyless signature contains the hash value submitted in the previous step and the sequence to regenerate the hash root value. This keyless signature is a hash chain composed of coordinates like the red dots in Figure 7. With this keyless signature system, the concrete construction department can ensure the spatial integrity of electronic data. Except for guaranteeing the spatial integrity of the electronic files, the intelligent keyless signature system based on Merkle trees can also ensure the temporal reliability of the electronic files. The mechanism is illustrated as follows: First, the keyless signature system stores the hash root values in a shared database called the calendar database while creating and destroying every second. Specifically, since 0 : 00, 0 seconds on January 1, 1970, each second of hash values has been regarded as a leaf node, forming a particular type of permanent hash tree, also known as a Merkle forest. The calendar hashes are periodically aggregated to generate the integrity code’s hash value. In a keyless signature system, the calendar database’s integrity code is regularly issued in electronic and paper form in the world media, as shown in Figure 8 [27]. After the integrity code is released in the electronic or paper-based public media, the authenticity of all signatures can be evaluated by tracing back the integrity code, thus ensuring the temporal integrity of the data. [28]. 5.3. Signing and Verification of Production Files. Signing and verifying the production files based on keyless signature are illustrated in Figure 9. As the description at the top of Figure 9, when a file is created and needs to be signed during the concrete production process, first, the signatories make a hash calculation on the file with the SHA-256 function and then submit the hash value to the distributed keyless signature server. From the one-way nature of the hash function, it is clear that the hash value is only the credential for applying a keyless signature, so the privacy of the original file is still kept. In the second step, the keyless signature server that receives the hash value performs a calculation through the hash chain and returns a keyless signature starting from the root node of the Merkle tree to the signatories as a response. In the third step, the keyless signature server timely releases the integrity code through newspapers or other forms. Note that the integrity code is preserved in the online calendar database after its release. The verification of signed files usually occurs in the file approval stage. As shown at the bottom of Figure 9, when the validator receives a signed file from the previous signatory, in order to verify the authenticity of the data, first and foremost, the received file and its corresponding keyless signature should be aggregated to conduct a hash ----- Wireless Communications and Mobile Computing 7 xroot = (x14|x58) x14 = (x12|x34) x58 = (x56|x78) x12 = (x1|x2) x34 = (x3|x4) x56 = (x5|x6) x78 = (x7|x8) x1 x2 x3 x4 x5 x6 x7 x8 Data item Figure 4: Schematic diagram on hash tree construction. x4 x58 Figure 5: Schematic diagram on hash chain computing. January 1[st], 1970 00:00:00 Tis second Hash …… Time calendar Hash tree |Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16|Col17|Col18|Col19|Col20|Col21|Col22| |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| Hash Keyless value signature Hash Keyless value signature Hash Keyless value signature Figure 6: Parallel construction of the Merkle tree based on edge computing. computation. Next, the integrity code associated with the signature file is figured out from the online database. Then, a comparison of the integrity code with the hash computation result is conducted subsequently. If the comparison result is consistent, it indicates that the signature data is accurate and trustworthy, the file transmission and approval process is in line with the standard requirements, and there is no tampering with the data. If the comparison result is ----- 8 Wireless Communications and Mobile Computing Value Pos 1 Hash calendar 0 1 Calendar 0 Top 1 aggregation 0 Aggregation 1 Aggregation Bottom aggregation Aggregate hash chain Hash Keyless value signature Figure 7: The keyless signature system legend. Figure 8: The Merkel forest structure. inconsistent, it proves that the concrete production file management process is not standardized and data security risks have occurred. In summary, the implementation of the keyless signature system could standardize the approval and writing process of files during the concrete production process, supervise ----- Wireless Communications and Mobile Computing 9 User A Service provider 1 Meta data Publish 3 integrity code Server Keyless signature 2 4 Store online User B Meta data Keyless signature Calculate by public tools 5 Verification code 6 Make a comparison Figure 9: Data signing and verification process based on the keyless signature. Accredited blockchain Block Block Block X-1 ę ę X ę ę X+1 ę ę Current block's hash value Previous block's Root's hash hash value value Chain structure Figure 10: Schematic diagram of the chain structure. |Col1|4 Store o| |---|---| Block Block Block X-1 ę ę X ę ę X+1 3 2 6 every step of data generation, and eliminate irregular data recording and internal tampering, thus protecting the security of the concrete production file to a higher degree and for a long time. ## 6. Production File Management Based on Blockchains 6.1. Chain Structure Design. Based on the data structure and the keyless signature system, the chain structure of the concrete production file and the corresponding data onchaining process are in Figure 10. When all the files involved in each concrete warehouse are collected, each file’s hash 1 5 value, also known as the unit block in the distributed blockchain data structure, is calculated separately. Then, a series of unit blocks aggregate two by two to form a binary tree, the root of which is called a procedure block. Thirdly, many procedure blocks polymerize to a compound as a Merkle tree, and the root is regarded as the warehouse block. Finally, by aggregating the current warehouse block with the previous warehouse block into a Merkle tree and storing the root of the tree on a trusted blockchain, the information security of the adjacent two warehouse blocks can be ensured. As Figure 10 illustrates, compared with the traditional paper form files, the electronic files are more conducive to data search and analysis. Besides, electronic information 4 ----- 10 Wireless Communications and Mobile Computing Warehouse X Warehouse Hash-X block Data before mixing A Hash-A Procedure block Branch block Unit block Figure 11: The organization of the warehouse block. security and traceability can be improved markedly by the blockchain-based information deposition mechanism compared with the conventional centralized storage database. 6.2. Automatic Data On-Chaining Mechanism. Creating one warehouse block and uploading it onto blockchain means that the mixing plant finishes an entire concrete production task from batching, mixing to the end according to the instructions. A warehouse usually produces tens to hundreds of cubic meters of concrete. In the proposed model, the warehouse blocks are at the top level, and adjacent warehouse blocks are linked in tandem with time stamps. The organization of each warehouse is in Figure 11. As shown in Figure 11, the blocks of each warehouse employ the Merkle tree structure to organize data, which is compatible with the signature generation mechanism in the keyless signature system. In Merkle trees, the two leaf nodes on each set of forks represent two files, and the files are paired two-by-two in the order of their generation time. In Figure 11, the file hash is regarded as a unit hash, and the branch hash is generated by two-by-two aggregation of all unit hashes. Furtherly, the procedure hash is composed of a two-by-two accumulation of branch hashes, and the warehouse hash is made up of pair-wise procedure hashes. Finally, the automatic data uploading is finished when all unit hashes are chained to form a warehouse hash. Due to the structural characteristics of the Merkle tree, any changes in the underlying data will lead to changes in its parent nodes and eventually affect the changes in the Merkle root. So the Merkle tree has the advantages of efficient comparison of a large amount of data, fast location of modified data, and fast verification of incorrect data, which are all demonstrated explicitly in the proposed management model. For illustration, when two Merkle tree roots are the same, the data they represent must be the same, which makes data verification between different users possible. Besides, when the underlying data is changed, its location can be quickly detected by inspecting the corresponding branch. With this feature, the proposed model can easily fulfill fast querying of the information about the abnormal data. Last but not least, when it is necessary to prove the originality and authenticity of the data, only the hash summary of the data needs to be validated without knowing the exact content of the data. 6.3. Smart Contracts and Consensus Algorithm. The smart contracts in our blockchain management model are fulfilled by introducing various forms of notification measures such as emails and cell phone applets to inform users of pending matters and remind them of the approval delays during file flow. Beyond that, to ensure data consistency during the automatic data on-chaining process, our model adopts ----- Wireless Communications and Mobile Computing 11 Transaction Broadcasting Broadcasting Validation Consensus Data writing User Node 1 Node 2 Node 3 Node 4 1 2 3 4 5 Figure 12: The implementation model of the consensus mechanism. |Broadcasting|Broadcasting|Validation|Consensus| |---|---|---|---| ||||| ||||| ||||| ||||| ||||| Byzantine Fault Tolerance (BFT) [29] as the consensus algorithm to synchronize the data to be recorded. The automatic concrete production data on-chaining mechanism based on the BFT consensus algorithm is shown in Figure 12, and it can be precisely divided into five steps, which are as follows: (1) When the supervisor starts the approval of the concrete mixing order, the action will be considered a transaction, and the proposed model broadcasts this transaction to all blockchain nodes, including raw material providers and construction units (2) After hash computation, the supervisor broadcasts the hash value of the transaction to all blockchain nodes (3) Each blockchain node (participating construction unit) makes a hash after receiving the transaction and compares it with the supervisor’s hash sequence (4) After all nodes receive the message that more than half of the comparisons are approved, the transaction is deemed to be established (5) The transaction is recorded into the block 6.4. Validation and Abnormal Block Tracking. The validation and tracking of the production files can also be fulfilled by the blockchain. As Figure 13 shows, the process is to verify the file’s integrity and record the location of the files that failed the verification. Specifically, the information of the abnormal file is retrieved from the database; then, the values of the file and the corresponding warehouse are recalculated according to the calculation rules at the time of uploading. Subsequently, the newly calculated warehouse values are compared with the corresponding uploaded warehouse values on the blockchain in sequence according to the warehouse organization order. Suppose the comparison of the hash values is consistent. In that case, all the electronic files in the warehouse are safe and secure. It has not been tampered with, so it is unnecessary to continue comparing the detailed information of this warehouse. But if inconsistency happens, the files contained in that warehouse are lost or tampered with, so it is neces sary to continue to compare the hash value of each file in that warehouse. The processes of file hash matching and warehouse hash matching are the same; the newly calculated file hash is compared with the file hash recorded on the blockchain. The file that contains inconsistent hash comparison results is recorded. Thus, the traceability of the problematic blocks can be achieved. ## 7. Model Application 7.1. Overall Architecture. In this paper, a concrete production management system based on the proposed model has been developed and implemented in the Hanjiang to Weihe River Project in Shaanxi Province to verify the model’s practicality and security. The system adopts a B-S architecture, and all users can log in and use it directly through a browser. Figure 14 shows the overall architecture. The concrete production information management system mainly manages data related to concrete production in the water conservancy project construction, including standardized management of file filling, unified management of data archiving, and automatic uploading of production files. The management system consists of user management, menu management, process management, parameter management, authority management, and log management. Through the network interface provided by the management system, different construction units in the concrete production system automatically import or manually enter various information about concrete production and create electronic files. After that, the file, branch, procedure, and warehouse hash are generated sequentially, and then, they are organized to the tree structure according to the distributed blockchain data structure. The generated hash values are uploaded to a credible blockchain for deposition. The information interaction between the blockchain and the information management platforms is fulfilled through port calls. In our information management system, the blockchain is the consortium blockchain called the Blockchain-based Service Network. This blockchain was jointly initiated by the State Information Center, China Mobile Communications Corporation, China UnionPay Corporation, and Beijing Red Date ----- 12 Wireless Communications and Mobile Computing Start Consistent? No Local database Online Caculate hash value of database Yes each warehouse Obtain the published integrity code Verification success Compare every hash value Aggregate hash values to generate Merkle tree Compare integrity code End Abnormalities found with Merkle root Figure 13: Schematic diagram of the abnormal block tracking flow. Accredited blockchain System Accredited Accredited Block management Data on-chaining Block tracing consensus smart contract management User Merkel forest management Hash aggregation Warehouse block Procedure block Branch block Unit block Menu management Keyless signatures Merkel tree building Hash calculation Signature release ę ę Process management Summary of production data Raw material Form of Errata and Parameter Supply list Form in mixing Other form form the concrete summary form management Intelligent data reporting Authority management Real-time collection Automatic import Manual dumped ę ę Character Basic operating environment Log management Concrete Third-party Network Storage Distributed Supervisor Constructor Database mix plant lab system system platform Figure 14: Architecture of concrete production information management system. |Concrete mix plant|Supervisor|Constructor| |---|---|---| |Storage system|Distributed platform|Database| |---|---|---| |Interface system|Accredited blockchain Accredited Accredited Block Data on-chaining Block tracing consensus smart contract management|System management User management Menu management Process management Parameter management Authority management Log management| |---|---|---| |||| ||Merkel forest Hash aggregation Warehouse block Procedure block Branch block Unit block|| |||| ||Keyless signatures Merkel tree building Hash calculation Signature release ę ę|| |||| ||Summary of production data Raw material Form of Errata and Supply list Form in mixing Other form form the concrete summary form|| |||| ||Intelligent data reporting Real-time collection Automatic import Manual dumped ę ę|| ||Character Basic operating environment Concrete Third-party Network Storage Distributed Supervisor Constructor Database mix plant lab system system platform|| ||Character Concrete Third-party Supervisor Constructor mix plant lab|| Consistent? Obtain the published integrity code Compare integrity code with Merkle root Compare every hash value End Online database Technology Corporation [30]. Besides, this consortium blockchain provides the storage, verification, and traceability of hash values and facilitates historical data security verification. In practice, the system was implemented in the Hanjiang to Weihe River Project to collect and organize the concrete production-related files in 2020. The total concrete produc Caculate hash value of each warehouse Verification success Start tion volume in the project in 2020 was about 170,000 square meters, which generated about 16,000 related paper forms in total. At present, we have entered and uploaded some of the files, including 18,000 square meters of concrete related to more than 3,500 forms, and stored these records on the consortium blockchain. The data server and the application server configurations in our system are the same: both are Aggregate hash values to generate Merkle tree Local database ----- Wireless Communications and Mobile Computing 13 80 70 60 50 40 30 20 10 ⁎ ⁎ 0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 ⁎ Keyless signature Data on-chaining Create Merkle tree ⁎ Signature verification Figure 15: Comparison of system time consumption by phase. dual-core and quad-threaded, with 8 G memory and 500 G storage space, which meet the minimum requirements for civil engineers [31]. 7.2. Experiment and Analysis. The experiment for testing the system performance is designed as follows: one warehouse block is selected randomly as the experiment object from the actual concrete production process. The target warehouse block contains 40 files, including supply contact sheets, production notification sheets, quality inspection sheets, and errata summary sheets, recording a complete concrete production process. The file creation and transmission are standardized with the keyless signature. After the files are filled and verified, they are uploaded on the blockchain for permanent storage. Besides, the time for the Merkle tree construction, the keyless signature creation, the signature verification, and files’ chaining are recorded separately. The specific time spent on the four steps of the 40 concrete production files is shown in Figure 15. As shown in Figure 15, the overall trend of keyless signature generation time per file raises as the production data increases. By fitting the linear regression model, it can be seen that the slope of the total keyless signature time is about 0.109. For each integrated keyless signature registration, when the size of the Merkle tree increases, the keyless signature generation time of the following file will also increase by about 0.109 s. Secondly, the average time consumption for data onchaining is about 3.39 s, with a slope of -0.009. This erratic fluctuation is caused by blockchain instability and network fluctuations. Thirdly, the time for the Merkle tree generation also shows an increasing trend correlation with the keyless signature generation time because the keyless signature is based on the combination of hash values from the root to the leaf sequence of the Merkle tree and its corresponding sequence coordinates. The slopes of Merkle tree creation and keyless signature generation are similar by linear fitting, which indicates that the creation time of the Merkle tree is the main factor that increases the generation time of keyless signature. Fourthly, the average time to verify the on-chain data is about 1.38 s. The slope of the linear fit function is 0.019, indicating that the verification is swift for on-chain data. The verification efficiency is mainly affected by the structure of the warehouse block. Besides, we also conduct the tests on keyless signature sizes. As shown in Figure 16, the keyless signature size of each file is about 157 kb, and its storage cost is less than 1 penny. The keyless signature storage cost of the whole warehouse is less than 0.1 yuan, and this cost is almost negligible compared with the benefits of data security. Finally, we use the number of transactions processed per second as the criterion for system throughput to evaluate the entire performance. The throughput of the relevant smart contracts is calculated for different concurrent requests. The number of concurrent requests is set from 100 to 1000, and 10 experiments are conducted in sequence. At last, the average values are taken as the experimental results. The throughput of the smart contracts is in Figure 17. In Figure 17, the throughput of the write operation (data on-chaining) is overall lower than that of the read operation (signature verification). In other words, the write operation ----- 14 Wireless Communications and Mobile Computing 180 160 140 120 100 80 60 40 20 0 0 100 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 File number Figure 16: Storage space consumed by keyless signatures. 80 60 40 20 0 0 100 200 300 400 500 600 700 800 900 1000 1100 Concurrent requests + Data on-chaining Signature verification Figure 17: Throughput of the smart contract with the concurrent requests. tends to be more time-consuming than the read operation. It is because the write operation needs to hash the data and generate a historical version of the old data to ensure the traceability of the blockchain. On the other hand, read operation only needs to search and validate data based on index positions, thus taking less time. Besides, the system’s throughput increases with the number of concurrent requests. However, when the number of simultaneous requests reaches a certain value, the growth trend slows down slightly. By calculation, the throughput stays 71 for the smart contract with data on-chaining. In contrast, the throughput for a signature verification smart contract is about 62. ## 8. Conclusions The digital archiving of engineering construction files in intelligent water projects is of great significance. The blockchain can provide security verification and integrity check for electronic files, which guarantees the security of archive ----- Wireless Communications and Mobile Computing 15 informatization and contributes to the realization of electronic archiving of files. This paper proposes a comprehensive data management model for smart water system construction based on blockchain and edge intelligence and then implements it in the Hanjiang to Weihe River Project in Shaanxi Province. Firstly, the behavioral model for the concrete production process is summarized, and the corresponding roles that participate in the process are abstracted out simultaneously. Secondly, the intelligent keyless signature based on parallel edge computing is introduced to ensure data security. The proposed model uses the Merkle tree to construct a chained file structure and standardizes the data entering, uploading, and checking procedure by the consensus mechanism. In the case study, we have created a blockchain of 3,500 blocks according to the decentralization requirement. In total, the proposed model and the corresponding system have already taken a big step forward in saving workforce and material resources and improving the security and traceability of construction archives markedly. We believe that through a more extensive scope of application and continuous improvement, the management of archives in civil engineering, especially in smart water projects, will eventually achieve the goal of digitalization. ## Data Availability The experiment data used to support the findings of this study are available from the corresponding author upon request. ## Conflicts of Interest All authors declare no conflict of interest in this paper. ## Acknowledgments This research work is supported by the National Natural Science Funds of China (62072368), Basic Research in Natural Science and Enterprise Joint Fund of Shaanxi (2021JLM-58), and Special Scientific Research Project of Education Department of Shaanxi (21JK0781). ## References [1] N. Nizamuddin, K. Salah, and M. A. Azad, “Decentralized document version control using ethereum blockchain and IPFS,” Computers & Electrical Engineering, vol. 76, pp. 183–197, 2019. [2] C. Feng, B. Liu, Z. Guo, K. Yu, Z. Qin, and K. K. R. Choo, “Blockchain-based cross-domain authentication for intelligent 5G-enabled Internet of drones,” IEEE Internet of Things Journal, 2021. [3] W. Wang, H. Xu, M. Alazab, T. R. Gadekallu, Z. Han, and C. Su, “Blockchain-based reliable and efficient certificateless signature for IIoT devices,” IEEE Transactions on Industrial Informatics, p. 1, 2021. [4] Y. Zhang, W. Luo, and F. Yu, “Construction of chinese smart water conservancy platform based on the blockchain: technol ogy integration and innovation application,” Sustainability, vol. 12, no. 20, p. 8306, 2020. [5] J. Song, Z. Han, W. Wang, J. Chen, and Y. Liu, “A new secure arrangement for privacy-preserving data collection,” Computer Standards & Interfaces, vol. 80, article 103582, 2022. [6] C. Feng, B. Liu, and K. Yu, “Blockchain-empowered decentralized horizontal federated learning for 5G-enabled UAVs,” IEEE Transactions on Industrial Informatics, 2022. [7] T. Bui, D. Cooper, and J. Collomosse, “Tamper-proofing video with hierarchical attention autoencoder hashing on blockchain,” IEEE Transactions on Multimedia, vol. 22, no. 11, pp. 2858–2872, 2020. [8] B. Zhong, H. Wu, and L. Ding, “Hyperledger fabric-based consortium blockchain for construction quality information management,” Frontiers of Engineering Management, vol. 7, no. 4, pp. 512–527, 2020. [9] L. Zhang, M. Peng, and W. Wang, “Secure and efficient data storage and sharing scheme for blockchain-based mobileedge computing,” Transactions on Emerging Telecommunications Technologies, no. article e4315, 2021. [10] G. Nagasubramanian, R. K. Sakthivel, and R. Patan, “Securing e-health records using keyless signature infrastructure blockchain technology in the cloud,” Neural Computing and Applications, vol. 32, no. 3, pp. 639–647, 2020. [11] J. Zhang, Y. Huang, and Y. Wang, “Multi-objective optimization of concrete mixture proportions using machine learning and metaheuristic algorithms,” Construction and Building Materials, vol. 253, article 119208, 2020. [12] Y. Gong, L. Zhang, and R. Liu, “Nonlinear MIMO for industrial Internet of Things in cyber-physical systems,” IEEE Transactions on Industrial Informatics, vol. 17, no. 8, pp. 5533–5541, 2021. [13] C. Feng, K. Yu, M. Aloqaily, M. Alazab, Z. Lv, and S. Mumtaz, “Attribute-based encryption with parallel outsourced decryption for edge intelligent IoV,” IEEE Transactions on Vehicular Technology, vol. 69, no. 11, pp. 13784–13795, 2020. [14] K. A. Nguyen, R. A. Stewart, H. Zhang, O. Sahin, and N. Siriwardene, “Re-engineering traditional urban water management practices with smart metering and informatics,” Environmental Modelling & Software, vol. 101, pp. 256–267, 2018. [15] J. Feng, L. Liu, and Q. Pei, “Min-Max cost optimization for efficient hierarchical federated learning in wireless edge networks,” IEEE Transactions on Parallel and Distributed Systems, 2022. [16] H. Li, K. Yu, B. Liu, C. Feng, Z. Qin, and G. Srivastava, “An efficient ciphertext-policy weighted attribute-based encryption for the Internet of health things,” IEEE Journal of Biomedical and Health Informatics, vol. PP, 2021. [17] K. Yu, M. Arifuzzaman, and Z. Wen, “A key management scheme for secure communications of information centric advanced metering infrastructure in smart grid,” IEEE transactions on instrumentation and measurement, vol. 64, no. 8, pp. 2072–2085, 2015. [18] L. Tan, K. Yu, and N. Shi, “Towards secure and privacypreserving data sharing for covid-19 medical records: a blockchain-empowered approach,” IEEE Transactions on Network Science and Engineering, 2022. [19] L. Zhao, J. Li, and A. Al-Dubai, “Routing schemes in softwaredefined vehicular networks: design, open issues and challenges,” IEEE Intelligent Transportation Systems Magazine, vol. 13, no. 4, pp. 217–226, 2020. ----- 16 Wireless Communications and Mobile Computing [20] J. Feng, F. R. Yu, and Q. Pei, “Cooperative computation offloading and resource allocation for blockchain-enabled mobile-edge computing: a deep reinforcement learning approach,” IEEE Internet of Things Journal, vol. 7, no. 7, pp. 6214–6228, 2019. [21] L. Tan, K. Yu, F. Ming, X. Chen, and G. Srivastava, “Secure and resilient artificial intelligence of things: a HoneyNet approach for threat detection and situational awareness,” IEEE Consumer Electronics Magazine, p. 1, 2021. [22] K. Yu, L. Tan, and L. Lin, “Deep-learning-empowered breast cancer auxiliary diagnosis for 5GB remote E-health,” IEEE Wireless Communications, vol. 28, no. 3, pp. 54–61, 2021. [23] L. Liu, C. Chen, and Q. Pei, “Vehicular edge computing and networking: a survey,” Mobile Networks and Applications, vol. 26, no. 3, pp. 1145–1168, 2021. [24] S. Hakak, W. Z. Khan, and G. A. Gilkar, “Securing smart cities through blockchain technology: architecture, requirements, and challenges,” IEEE Network, vol. 34, no. 1, pp. 8–14, 2020. [25] K. Yu, Z. Guo, and Y. Shen, “Secure artificial intelligence of things for implicit group recommendations,” IEEE Internet of Things Journal, 2022. [26] T. Alladi, V. Chamola, and R. M. Parizi, “Blockchain applications for industry 4.0 and industrial IoT: a review,” IEEE Access, vol. 7, pp. 176935–176951, 2019. [27] H. Huang, J. Lin, and B. Zheng, “When blockchain meets distributed file systems: an overview, challenges, and open issues,” IEEE Access, vol. 8, pp. 50574–50586, 2020. [28] L. Liu, J. Feng, and Q. Pei, “Blockchain-enabled secure data sharing scheme in mobile-edge computing: an asynchronous advantage actor–critic learning approach,” IEEE Internet of Things Journal, vol. 8, no. 4, pp. 2342–2353, 2020. [29] H. Xiong, C. Jin, M. Alazab et al., “On the design of blockchain-based ECDSA with fault-tolerant batch verication protocol for blockchain-enabled IoMT,” IEEE Journal of Biomedical and Health Informatics, vol. PP, p. 1, 2021. [30] J. Ma, S. Zhang, and H. Li, “Sparse Bayesian learning for the time-varying massive MIMO channels: acquisition and tracking,” IEEE Transactions on Communications, vol. 67, no. 3, pp. 1925–1938, 2018. [31] L. Liu, M. Zhao, M. Yu, M. A. Jan, D. Lan, and A. Taherkordi, “Mobility-aware multi-hop task offloading for autonomous driving in vehicular edge computing and networks,” IEEE Transactions on Intelligent Transportation Systems, pp. 1–14, 2022. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1155/2022/8482415?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1155/2022/8482415, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://downloads.hindawi.com/journals/wcmc/2022/8482415.pdf" }
2,022
[]
true
2022-03-09T00:00:00
[ { "paperId": "c795828f5390d4208bc70b8a110082a7fc3884fd", "title": "Mobility-Aware Multi-Hop Task Offloading for Autonomous Driving in Vehicular Edge Computing and Networks" }, { "paperId": "50bf499b417fae819700f1050a9f4c92c21bf818", "title": "Blockchain-Empowered Decentralized Horizontal Federated Learning for 5G-Enabled UAVs" }, { "paperId": "cfea5d552e6ab2dbfe2de80f6c830640dd09f84f", "title": "Secure and Resilient Artificial Intelligence of Things: A HoneyNet Approach for Threat Detection and Situational Awareness" }, { "paperId": "7cdf2f3e5ff6ced1c31dd6f21ded0f9b40451d9a", "title": "Blockchain-Based Cross-Domain Authentication for Intelligent 5G-Enabled Internet of Drones" }, { "paperId": "04bfc754c0e28df90de3b7e0a3194b8df7707cdc", "title": "A new secure arrangement for privacy-preserving data collection" }, { "paperId": "f0ce78953f7a5f68a13c7f5b8614850f3e73fc24", "title": "Towards Secure and Privacy-Preserving Data Sharing for COVID-19 Medical Records: A Blockchain-Empowered Approach" }, { "paperId": "d5aba97f9927528f304be85cd3133d17b462fb36", "title": "On the Design of Blockchain-Based ECDSA With Fault-Tolerant Batch Verification Protocol for Blockchain-Enabled IoMT" }, { "paperId": "7dd7966b17fb3f97cd7de85bfaa6994951312a1d", "title": "Nonlinear MIMO for Industrial Internet of Things in Cyber–Physical Systems" }, { "paperId": "3b230f14c46e7e177e9bebb2ebc9f46b346b646d", "title": "Secure and efficient data storage and sharing scheme for blockchain‐based mobile‐edge computing" }, { "paperId": "bb258e5d79ea55d9275db74268830b525367fafa", "title": "Deep-Learning-Empowered Breast Cancer Auxiliary Diagnosis for 5GB Remote E-Health" }, { "paperId": "3ffc477b8a6bbc33fa62c126bd64bebf5194c2ac", "title": "Blockchain-Based Reliable and Efficient Certificateless Signature for IIoT Devices" }, { "paperId": "84769d6c3c7c387a49f2fa5562900e5236138adb", "title": "An Efficient Ciphertext-Policy Weighted Attribute-Based Encryption for the Internet of Health Things" }, { "paperId": "42272a6c530601c22f782872061809031c2a50c8", "title": "Secure Artificial Intelligence of Things for Implicit Group Recommendations" }, { "paperId": "9dad453f4b0273354074e1735b3c73fb829f25f3", "title": "Blockchain-Enabled Secure Data Sharing Scheme in Mobile-Edge Computing: An Asynchronous Advantage Actor–Critic Learning Approach" }, { "paperId": "0f8d81afc0574e8da70447c0c060e141b8ea9076", "title": "Construction of Chinese Smart Water Conservancy Platform Based on the Blockchain: Technology Integration and Innovation Application" }, { "paperId": "21358e2b08b18dad28716e38841002d1d05abf2b", "title": "Attribute-Based Encryption With Parallel Outsourced Decryption for Edge Intelligent IoV" }, { "paperId": "d37ae1ef49f11980515703e0ebf0bbd7cc75fc61", "title": "Multi-objective optimization of concrete mixture proportions using machine learning and metaheuristic algorithms" }, { "paperId": "405769586840e523e5f56fe38c68c7f45509b618", "title": "Hyperledger fabric-based consortium blockchain for construction quality information management" }, { "paperId": "ad35de1aeb563d5ce74b7554bde592f20ba6e0a6", "title": "Cooperative Computation Offloading and Resource Allocation for Blockchain-Enabled Mobile-Edge Computing: A Deep Reinforcement Learning Approach" }, { "paperId": "f375e69427fc50251d8fa089fbe089ac6ca721a1", "title": "When Blockchain Meets Distributed File Systems: An Overview, Challenges, and Open Issues" }, { "paperId": "5ae45bad7cbd6e3e174d6b60193790e314583899", "title": "Tamper-Proofing Video With Hierarchical Attention Autoencoder Hashing on Blockchain" }, { "paperId": "ee59451deedd0c7532b5b4c34340a615c1d76312", "title": "Securing Smart Cities through Blockchain Technology: Architecture, Requirements, and Challenges" }, { "paperId": "282112cc8470c0bac43e6e0266adeb8b7800fd67", "title": "Blockchain Applications for Industry 4.0 and Industrial IoT: A Review" }, { "paperId": "5973e94435ad78899e8d94c786e02193c539c76c", "title": "Vehicular Edge Computing and Networking: A Survey" }, { "paperId": "aaa545332c9da2f75ed8399f3f26200c92509ff6", "title": "Decentralized document version control using ethereum blockchain and IPFS" }, { "paperId": "450a3df92973f32af833173a941c703fd4e424a7", "title": "Sparse Bayesian Learning for the Time-Varying Massive MIMO Channels: Acquisition and Tracking" }, { "paperId": "bfbdba5420d45da7679811595e9393a8948d8890", "title": "Securing e-health records using keyless signature infrastructure blockchain technology in the cloud" }, { "paperId": "08e6b3959f1d0c62f84820eafe83237025042156", "title": "Re-engineering traditional urban water management practices with smart metering and informatics" }, { "paperId": "520809931bafc3a95d8ea43b88c68d477a00d53e", "title": "A Key Management Scheme for Secure Communications of Information Centric Advanced Metering Infrastructure in Smart Grid" }, { "paperId": "42f98bfbba0c06653f60c8317373c614911029c1", "title": "Min-Max Cost Optimization for Efficient Hierarchical Federated Learning in Wireless Edge Networks" }, { "paperId": "2273984cc4c17d8e96216d7bdaa481ce4bf10546", "title": "Routing Schemes in Software-Defined Vehicular Networks: Design, Open Issues and Challenges" }, { "paperId": null, "title": "production management system based on the proposed model has been developed and implemented in the Hanjiang to Weihe River Project in Shaanxi Province to verify the model ’ s practicality" }, { "paperId": null, "title": "After hash computation, the supervisor broadcasts the hash value of the transaction to all blockchain nodes" }, { "paperId": null, "title": "Each blockchain node (participating construction unit) makes a hash after receiving the transaction and compares it with the supervisor ’ s hash sequence" }, { "paperId": null, "title": "By employing the smart keyless signatures" } ]
14,370
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01155ceeaefeab3737c49cebde0b4b9e01f7d9cd
[]
0.821515
TCP/UDP-Based Exploitation DDoS Attacks Detection Using AI Classification Algorithms with Common Uncorrelated Feature Subset Selected by Pearson, Spearman and Kendall Correlation Methods
01155ceeaefeab3737c49cebde0b4b9e01f7d9cd
Revue d'Intelligence Artificielle
[ { "authorId": "48536036", "name": "Kishore Babu Dasari" }, { "authorId": "9364224", "name": "N. Devarakonda" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
The Distributed Denial of Service (DDoS) attack is a serious cyber security attack that attempts to disrupt the availability security principle of computer networks and information systems. It's critical to detect DDoS attacks quickly and accurately while using as less computing power as possible in order to minimize damage and cost efficient. This research proposes a fast and high-accuracy detection approach by using features selected by proposed method for Exploitation-based DDoS attacks. Experiments are carried out on the CICDDoS2019 datasets Syn flood, UDP flood, and UDP-Lag, as well as customized dataset. In addition, experiments were also conducted on a customized dataset that was constructed by combining three CICDDoS2019 datasets. Pearson, Spearman, and Kendall correlation techniques have been used for datasets to find un-correlated feature subsets. Then, among three un-correlated feature subsets, choose the common un-correlated features. On the datasets, classification techniques are applied to these common un-correlated features. This research used conventional classifiers Logistic regression, Decision tree, KNN, Naive Bayes, bagging classifier Random forest, boosting classifiers Ada boost, Gradient boost, and neural network-based classifier Multilayer perceptron. The performance of these classification algorithms was also evaluated in terms of accuracy, precision, recall, F1-score, specificity, log loss, execution time, and K-fold cross-validation. Finally, classification techniques were tested on a customized dataset with common features that were common in all of the dataset’s common un-correlated feature sets.
Vol. 36, No. 1, February, 2022, pp. 61-71 Journal homepage: http://iieta.org/journals/ria # TCP/UDP-Based Exploitation DDoS Attacks Detection Using AI Classification Algorithms with Common Uncorrelated Feature Subset Selected by Pearson, Spearman and Kendall Correlation Methods Kishore Babu Dasari[1*], Nagaraju Devarakonda[2] 1 Department of CSE, Acharya Nagarjuna University, Guntur 522510, Andhra Pradesh, India 2 School of Computer Science and Engineering, VIT-AP University, Amaravati 522237, India Corresponding Author Email: dasari2kishore@gmail.com https://doi.org/10.18280/ria.360107 **ABSTRACT** |Corresponding Author Email:|dasari2kishore@gmail.com| |---|---| **Received: 13 January 2022** **Accepted: 24 February 2022** **_Keywords:_** _CICDDoS2019, classification algorithms,_ _DDoS attacks, Kendall correlation, Pearson_ _correlation, spearman correlation, syn flood,_ _UDP flood, UDP-Lag_ **1. INTRODUCTION** The Distributed Denial of Service (DDoS) attack is a serious cyber security attack that attempts to disrupt the availability security principle of computer networks and information systems. It's critical to detect DDoS attacks quickly and accurately while using as less computing power as possible in order to minimize damage and cost efficient. This research proposes a fast and high-accuracy detection approach by using features selected by proposed method for Exploitation-based DDoS attacks. Experiments are carried out on the CICDDoS2019 datasets Syn flood, UDP flood, and UDP-Lag, as well as customized dataset. In addition, experiments were also conducted on a customized dataset that was constructed by combining three CICDDoS2019 datasets. Pearson, Spearman, and Kendall correlation techniques have been used for datasets to find un-correlated feature subsets. Then, among three un-correlated feature subsets, choose the common un-correlated features. On the datasets, classification techniques are applied to these common uncorrelated features. This research used conventional classifiers Logistic regression, Decision tree, KNN, Naive Bayes, bagging classifier Random forest, boosting classifiers Ada boost, Gradient boost, and neural network-based classifier Multilayer perceptron. The performance of these classification algorithms was also evaluated in terms of accuracy, precision, recall, F1-score, specificity, log loss, execution time, and K-fold cross-validation. Finally, classification techniques were tested on a customized dataset with common features that were common in all of the dataset’s common un-correlated feature sets. Availability-based attacks are network security attacks carried out by a malicious node with the goal of denying access to resources on computer networks. Denial of service (DoS) is an available-based security attack in which the attacker aims to make network resources unavailable to its intended users by temporarily or indefinitely disrupting the services of a host connected to a network. A DoS attack launched by more than one attacker is called a Distributed Denial of Service (DDoS) attack [1]. DDoS attacks make use of a variety of vulnerabilities in the TCP/UDP-based protocols at the application layer to deny its users are called Exploitation based DDoS attacks. DDoS has become more prevalent among cyberattacks due to the extensive use of TCP protocol and easier to exploit features of the TCP three-way handshake mechanism. Syn flood is a TCPbased exploitation DDoS attack, UDP flood, and UDP-Lag are UDP-based exploitation DDoS attacks. SYN flood [2] is a commonly used exploitation-based DDoS attack that exploits the advantage of a feature of the TCP three-way handshake to overflow the TCP queue of the server and make it consume resources resulting in it being unavailable to legitimate users' requests. A TCP connection is established between a client and a server using the TCP threeway handshake mechanism. A client must send a synchronized flag packet (SYN) to the server to establish a TCP connection. The server sends the client an acknowledgment flag for the synchronized packet (SYN-ACK) after receiving the SYN packet delivered by the client. The client sends an acknowledgment flag to the server after receiving the SYNACK flag from the server. With these three steps, a connection between the client and the server is established, and data transformation can now commence. In order to launch a TCP SYN flood attack on a server, attackers take advantage of the server's half-opened connection state. This is the state in which the server is waiting for the client's ACK flag before attempting to establish a connection. The server would have already allocated Memory resources to the client at this point. To take advantage of this behavior, the attacker sends a large number of SYN flags to the server for a number of spoofed IP addresses. The server treats these requests as legitimate, allocating memory and resources to these IP sources and sending the client a SYN-ACK flag. The server would now wait in a half-open state for the client to respond with an ACK flag which would never receive. The attacker's large number of illegitimate SYN requests leads the TCP backlog queue to overflow, resulting in half-opened connections until all system resources are consumed. The legitimate user's request is not accepted by the server due to an overflow of the TCP queue. The primary objective of the TCP SYN flood attack is to disrupt the system's availability. ----- [ ] yp p attack in which the attacker overflows random ports on the targeted host with IP packets containing UDP datagrams. UDP flood attack’s main objective is to saturate the Internet pipe. A UDP flood operates by taking advantage of the steps taken by a server when responding to UDP packets transmitted to one of its ports. When a server receives a UDP packet at a specific port, it goes through two steps in response to normal circumstances: First, the server looks to determine if any programs are currently listening for requests on the specified port. If no programs are receiving packets on that port, the server sends an ICMP (ping) message to the sender to alert them that the destination is unavailable. When the server receives a new UDP packet, it goes through a series of steps to process the request, consuming server resources in the process. When a huge flood of UDP packets is received from different sources with spoof IP addresses, the target's resources can quickly become exhausted as a result of the targeted server using resources to check and then respond to each single UDP packet. The UDP-Lag attack [4] is an attempt to break the connection between the client and the server. This attack is most commonly used in online gaming to outsmart other players slowing down/interrupting their movement. This attack can be carried out in two ways: using a hardware switch known as a lag switch, or with a software program that runs on the network and consumes other users' bandwidth. According to research findings on DDoS attacks, due of their distributed nature, fast detection, less computation, and accuracy in detection is three key challenges in DDoS attack detection. DDoS attacks have caused significant damage in all aspects of business; hence, early detection is essential. As computation is so expensive these days, reducing the number of features is essential to make the computation process more cost-effective. To avoid inconvenience to legitimate users, accurate detection is essential. This research proposes a method for select the un-correlated feature subset using three correlation techniques. It builds a fast and high-accuracy DDoS attack detection approach with very few features. This section introduces the TCP/UDP based Exploitation DDoS attacks and the research motivation and objective of detecting DDoS attacks. In section II of this paper, the methodology is explained, including proposed framework, algorithm, preprocessing, and machine learning classification algorithms. The results and discussion are explained with experimental results in section III of this paper. The study's conclusion is found in Section IV of this paper. **2.** **METHODOLOGY** Proposed model framework depicted in Figure 1. **Proposed Algorithm** 1. Start. 2. Read DDoS attack dataset. 3. Preprocessing: 3.1. Remove uninfluential socket features 3.2. Removing missing and infinity values 3.3. Encoding Benign and Attack labels 3.4. Removing constant features (Threshold==0) 3.5. Removing quasi-constant features (Threshold==0.01) 4. Split the dataset into the train and test data in 80:20 5. Apply Pearson, Spearman and Kendall correlations to test and train data. 6. Apply threshold >=80 and collect Pearson, Spearman and Kendall un-correlated feature subsets. 7. Apply intersection of Pearson, Spearman and Kendall un-correlated feature subsets and find common un-correlated feature set. 8. Apply classification algorithms to train and test data to classify Benign and Attack labels. 9. Stop. **Figure 1. Proposed model framework** **Data set** This study uses the CICDDoS2019 data set which includes a wide variety of DDoS attacks and fills up the gaps in the previous data sets. Every DDoS attack dataset contains 87 features. **Data preprocessing** Data Preprocessing [5] is the first and most important step in building a classification model. It is a process of clean and formatted data suitable for the classification model. It increases the accuracy and efficiency of classification models. First, remove socket features that vary from network to network. Next to clean the data by removing missing and infinity values. Encoding the label string values for Benign and attack label to the binary value of 0 and 1 respectively. And finally, standardize the independent feature values. Initially each dataset contains 88 features, after removing uninfluential socket features each dataset contain 81 features. Pre-processing results are statistically shown in Figure 2 with a bar chart in order of the number of records processed. **Feature selection** Feature Selection [6] is a very critical component in Machine learning algorithms. Machine learning algorithms typically choke when provided with data with a large dimensionality because the number of features raises the training time exponentially and an increasing amount of ----- , g methods help in the resolution of these issues by reducing the dimensions while preserving the overall information. It also helps in identifying the features and their importance. Variance threshold and correlation feature selection methods are used in this study. A variance threshold is used to remove constant and quasi-constant features. Correlation methods are used to find uncorrelated features. **Figure 2. Pre-processing results bar-chart** **Variance threshold** A simple baseline technique for feature selection is the variance threshold. This method eliminates features that vary below a specific threshold. It removes all zero-variance features by default, that is, features that have the same value throughout all samples. More useful information is contained in features with a higher variance. The variance threshold doesn’t consider the relationship of features with the target variable. **Correlation** Correlation [5] is a bivariate analysis that determines the level of association and the direction of the relationship between two variables. The value of the correlation coefficient varies between +1 and -1 in terms of the strength of the association. A value of ±1 shows that the two variables are perfectly associated. The value of 0 shows that the two variables are weakly associated. The sign of the coefficient specifies the relationship's direction; +sign indicates a positive relationship that means one variable goes up, then the second variable also goes up, while –sign indicates a negative relation that means one variable increase then another variable decrease. We can predict one variable from the other using correlation. When two features are correlated, the model only needs one of them, as the other does not provide any extra information. This study uses three types of correlations: Pearson correlation, Spearman correlation, and Kendall rank correlation. Pearson correlation The Pearson correlation is the most generally used correlation statistic for determining the degree of association between linearly related variables. The Pearson correlation is based on information about the mean and standard deviation. Pearson correlation coefficient calculated by: , r is the correlation coefficient; xi - is the value of the x-feature in a sample; 𝑥̅ – is the mean of the values of the x-feature; yi - is the value of the y-feature in a sample; 𝑦̅ – is the mean of the values of the y-feature. Spearman correlation Spearman rank correlation is a non-parametric measure of correlation used to determine the degree of relationship between two variables. Non-parametric correlations rely solely on ordinal data and pair scores. The Pearson correlation between the rank scores of two variables is equivalent to the Spearman correlation between those two variables. Spearman's correlation evaluates monotonic relationships, whereas Pearson's correlation evaluates linear relationships. The strength of a monotonic relationship between two variables with the same scaling as the Pearson correlation is measured by the Spearman correlation. Spearman correlation coefficient calculated by: 𝜌= 1 − 6 ∑𝑑𝑖2 (2) 𝑛(𝑛[2 ] −1) 𝑟= ∑(𝑥𝑖 −𝑥̅)(𝑦𝑖 −𝑦̅) √∑(𝑥𝑖 −𝑥̅)[2] ∑(𝑦𝑖 −𝑦̅) (1) Here, _ρ is the Spearman’s rank correlation coefficient;_ di is the difference between the two ranks of each observation; n is the number of observations. Kendall rank correlation Kendall rank correlation is a non-parametric test that assesses the degree of association between two variables. Nonparametric correlations rely solely on ordinal data and pair scores. Kendall correlation outperforms Spearman correlation in terms of robustness and efficiency. When there are few samples or some outliers, Kendall correlation is preferred. Kendall correlation coefficient calculated by: 𝜏= [𝑁][𝑐] [−𝑁][𝑑] 𝑛(𝑛−1) (3) 2 Here, 𝜏 is the Kendall rank correlation coefficient; Nc is the number of concordant; Nd is the number of discordant. **Classification algorithms** Machine learning is becoming more widely used to detect and classify DDoS attacks [7]. One of the most important steps in machine learning algorithms is feature selection. Feature Selection is essential for reducing dimensionality and removing redundant and irrelevant features. Logistic regression Logistic regression [8] is a machine learning classification method borrowed from statistics to predict the target variable. It uses the logistic function also called as the Sigmoid function. Sigmoid function is: 1 ∅(𝑧) = (4) 1 + 𝑒[−𝑧] ----- p weights and features. # z = w xT = w0 + w x1 g g learners. Gradient Boosting classification algorithm depends on the loss function. The gradient descent optimization procedure is used to determine the contribution of the weak learner to the ensemble. Multilayer perceptron A multilayer perceptron (MLP) [15] is the most standard form of feed-forward artificial neural network. MLP consists of an input layer to receive input data, output layers that make predictions about the input, and at least one hidden layer is capable of approximating any continuous function. **3.** **RESULTS AND DISCUSSION** The objective of this study has been to reduce data computation and execution time in order to improve the accuracy of TCP/UDP-based exploitation DDoS attack detection. Data processing or computation is accomplished by reducing the number of features in the input data sets. Data computation is proportionate to the model's execution time. It means that as data computation time reduces, execution time significantly reduces as well. So, the main objectives of this paper is to reduce the number of features in data sets without decreasing the accuracy of exploitation-based DDoS attack detection. In this paper, we propose a model for reducing the number of features with improving DDoS attack detection accuracy. The proposed model depicted in Figure 1. TCP/UDP-based exploitation DDoS attack data sets are collected for this study from the CICDDoS2019 data set, which contains various TCP/UDP based DDoS attack data sets. Syn flood is TCP based exploitation DDoS attack data set while UDP flood and UDP-Lag are UDP based exploitation DDoS attack data sets. Experiments have also been conducted on a customized exploitation DDoS attack data set in this research. Concatenated 400000 records from each of the Synflood, UDP flood, and UDP-Lag datasets to create a customized exploitation DDoS attack data set. In this section results are discussed in the order of removing constant and quasi-constant features by using variance threshold, finding un-correlated feature subsets with Pearson, Spearman and Kendall correlation methods, finding common un-correlated features from un-correlated feature subsets of Pearson, Spearman and Kendall correlation methods, discussed performance evaluation metrics of classification algorithms with common uncorrelated feature subsets on TCP/UDP based exploitation DDoS attack datasets of Synflood, UDP-flood, UDP-Lag and customized DDoS attack and finally discussed performance evaluation metrics of classification algorithms on customized dataset with common features that were common in all of the dataset’s common uncorrelated feature sets. After pre-processing, variance threshold filter-based feature selection is being used to remove constant and quasi-constant features from the data sets in order to reduce the number of features. The features that are almost constant are known as quasi-constant features. Constant features have a variance threshold value of 0, whereas quasi-constant features have a variance threshold value of 0.01. Constant features are those that have the same value across the entire dataset's rows. Remove these features because they provide no information to the classification algorithms. Table 1 shows the number of constant and quasi-constant feature counts for the Syn flood, UDP flood, UDP-Lag, and customized exploitation data sets. # +w x2 2 ++ w xn n (5) _ϕ(z) values limits in the range [0,1]. It indicates that if z goes_ to infinity, the function becomes one, and if z goes minus infinity, the function reaches zero. Decision tree Decision Tree [9] is a supervised learning method that can be used to display a model's visual representation. A decision tree employs a hierarchical model resembling a flow chart with multiple connected nodes. These nodes indicate tests on the dataset's features, with a branch that leads to either another node or a classification result. The prediction data is passed through the nodes until it can be classified, with the training data used to form the tree. K-Nearest neighbor One of the most basic machine learning classification models is K-Nearest Neighbor (KNN) [10]. With KNN, there is no training; the training data is used to make predictions in order to classify the data. KNN works on the notion that comparable data points would group together, and it uses the K value, which can be any number, to locate the closest data points. Naive bayes classifier A typical NB classifier [11] also relies on Bayes’ theorem and applies probability density information to the training data. It is used to calculate the chance of an event occurring based on previous occurrences that have occurred. Random forest The random forest [12] is based on the principle of bagging, which is used to train a number of decision trees and enhance them based on their attributes. Random attribute selection is used in the random forest training process to improve the relative independence of the generated decision tree and hence improve performance. Assuming that there are n nodes, the standard decision tree selects the best attribute based on all of the n nodes' characteristics, but each node of the random forest's decision tree is based on k attributes that are randomly selected in advance. The magnitude of the k parameter, which is commonly set to log2 d, determines the degree of randomness. Furthermore, the k value can be 1 or d, which reflects a random selection of an attribute and a selection procedure utilizing a traditional decision tree, respectively. Ada boost AdaBoost, also known as Adaptive Boosting [13], is a Machine Learning ensemble classification model. It is an iterative ensemble classification algorithm that means weak learners grow sequentially and become strong ones. The classifier should be interactively trained using a variety of weighted training instances. It tries to provide an excellent fit for these instances in each iteration by minimizing training errors. Gradient boost Gradient Boost [14] is an ensemble boosting classification ----- q p **Number of Constant Features** **Number of Quasi-constant Features** **Data Set** **(Variance Threshold=0)** **(Variance Threshold=0.01)** Syn Flood attack 12 7 UDP flood attack 12 8 UDP-Lag attack 12 5 Customized Exploitation DDoS attack 12 6 **Table 2. Number of correlated features, which has a** threshold value >= 80 by Pearson, Spearman, and Kendall correlation methods for TCP/UDP Exploitation-based DDoS attack data sets **Correlation Methods** **Data Sets** **Pearson** **Spearman** **Kendall** Syn Flood attack 37 50 48 UDP flood attack 34 46 46 UDP-Lag attack 36 50 46 Customized Exploitation 39 48 47 DDoS attack **Table 3. Number of common un-correlated features count** with a proposed feature selection method on TCP/UDP Exploitation-based DDoS attack data sets **Number of common un-** **Data Set** **correlated features** Syn Flood attack 9 UDP flood attack 11 UDP-Lag attack 10 Customized Exploitation 12 DDoS attack Apply the Pearson, Spearman, and Kendall correlations individually on the exploitation-based DDoS attack data sets after deleting constant and quasi-constant features, then collect un-correlated features sub-sets of each correlation method. Table 2 shows the number of correlated feature counts for the Syn flood, UDP flood, UDP-Lag, and customized exploitation data sets. To find the common un-correlated feature subset, apply intersection to un-correlated feature subsets of the Pearson, Spearman, and Kendall correlation methods. Table 3 shows the number of common un-correlated feature counts for the Syn flood, UDP flood, UDP-Lag, and customized exploitation data sets. Table 4 shows the common uncorrelated feature list for the Syn flood, UDP flood, UDP-Lag, and customized exploitation data sets. Unnamed: 0, Flow Duration, Flow IAT Min, Total Length of Bwd Packets, and Protocol are common in the lists of common un-correlated features of Syn flood, UDP flood, UDP-Lag, and customized exploitation DDoS attack data sets. Classification algorithms are applied to Syn flood, UDP flood, UDP-Lag, and customized exploitation DDoS attack data sets with their common un-correlated feature subsets and results evaluated. On customized exploitation DDoS attack data set with common features in the lists of common un-correlated features of Syn flood, UDP flood, UDP-Lag, and customized exploitation DDoS attack data sets, classification algorithms are applied and the results also evaluated. **Confusion matrix** The actual and predicted values of label classes are displayed in a confusion matrix. It shows the four key values that are True Positive, False Negative, False Positive, and True Negative. These values are used to calculate the evaluation metrics. TRUE POSITIVE (TP): The amount of DDoS attacks properly identified by the classifier. TRUE NEGATIVE (TN): The number of BENIGN class labels accurately detected by the classifier. FALSE POSITIVE (FP): The number of BENIGN class labels, classified as DDoS attacks by the classifier. FALSE NEGATIVE (FN): The number of DDoS attack labels, classified as BENIGN class labels by the classifier. **Accuracy** Accuracy is defined as the proportion of benign and attack data in the right classification to the total data. 𝑇𝑃+ 𝑇𝑁 𝐴𝐶𝐶𝑈𝑅𝐴𝐶𝑌= (6) 𝑇𝑃+ 𝑇𝑁+ 𝐹𝑃+ 𝐹𝑁 **Precision** Precision refers to the ratio of the number of attacks correctly classified into attacks to the entire proportion of attack data, which indicates the model's capability to detect attack data. 𝑇𝑃+ 𝑇𝑁 𝐴𝐶𝐶𝑈𝑅𝐴𝐶𝑌= (7) 𝑇𝑃+ 𝑇𝑁+ 𝐹𝑃+ 𝐹𝑁 **Recall/TPR** The recall or true positive rate (TPR) is the percentage of accurately detected attack data instances among all attack data. 𝑇𝑃+ 𝑇𝑁 𝐴𝐶𝐶𝑈𝑅𝐴𝐶𝑌= (8) 𝑇𝑃+ 𝑇𝑁+ 𝐹𝑃+ 𝐹𝑁 **F1-Score** The F1 score is the weighted average precision and recall. Logistic Regression, Gradient Boost, and Naive Bayes provide the best F-score value. Ada Boost and KNN almost provide the best F1-score value. Decision Tree also provides a poor F1-score value. 𝐹1 𝑆𝑐𝑜𝑟𝑒= [2 ∗𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗𝑅𝑒𝑐𝑎𝑙𝑙] (9) 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+ 𝑅𝑒𝑐𝑎𝑙𝑙 **Specificity** Specificity is the ratio of the truly classified BENIGN class labels out of the total actual BENIGN class labels. 𝑇𝑁 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦= (10) 𝑇𝑁+ 𝐹𝑃 Table 5 shows performance evaluation metrics in terms of accuracy, precision, recall, F-score, and specificity for the different classification algorithms on Syn flood attack with common un-correlated features. The classification methods of multilayer perceptron and Ada boost produce the best ----- y p, all classification methods produce good results in terms of precision, recall, and F-score. Multilayer perceptron produces better precision and F-score values for benign classification, while KNN produces better recall values. For attack classification, logistic regression produces a better specificity value. All classification methods produce better specificity values for benign classification. Table 6 shows performance evaluation metrics in terms of accuracy, precision, recall, F-score, and specificity for the different classification algorithms on UDP flood attacks with common un-correlated features. The classification methods of KNN, Ada boost, and multilayer perceptron produce the best accuracy results compared to others. In terms of precision, recall, and F-score, all classification algorithms produce good results for attack classification. Multilayer perceptron produces the best precision score, random forest produces the best recall score, and KNN produces the best F-score value for benign classification. Except for Naive Bayes, all classification methods produce good specificity scores for benign classification. For attack classification, logistic regression and Naive Bayes produce a high specificity score. Table 7 shows performance evaluation metrics in terms of accuracy, precision, recall, F-score, and specificity for the different classification algorithms on UDP-Lag attacks with common un-correlated features. Random Forest and multilayer perceptron produce the best accuracy results compared to other classifiers. All classification algorithms produce good results for attack classification in terms of precision, recall, and F-score except Naive Bayes classifier. Ada Boost and Multilayer perceptron produce the best p, g g p recall value and KNN, Random forest, and Multilayer perceptron produce the best F-score values for benign classification. Logistic regression produces the best specificity score for attack classification. All classification algorithms produce good results for benign classification in terms of specificity except the Naive Bayes classifier. Table 8 shows performance evaluation metrics in terms of accuracy, precision, recall, F-score, and specificity for the different classification algorithms on Customized Exploitation DDoS attacks with common un-correlated features. Multilayer perceptron produces the best accuracy results compared to other classifiers. All classification algorithms produce good results for attack classification in terms of precision, recall, and F-score. The random forest produces the best precision score, while Logistic regression produces the best recall value and KNN produces the best F score value for benign classification. All classification algorithms produce good results for benign classification in terms of specificity. Logistic regression produces the best specificity value for attack classification. **K-fold cross validation** Cross-fold validation is a statistical method for evaluating machine learning classification models. A test set should still be kept aside for final evaluation when employing Crossvalidation, but the validation set is no longer required. The training set is partitioned into k smaller sets in a k-Cross-fold validation. The training data for a model is taken from k-1 folds. After that, the model is tested against the remaining data. **Table 4. Common un-correlated feature list with a proposed feature selection method on TCP/UDP Exploitation-based DDoS** attack data sets **Customized Exploitation DDoS** **Syn Flood attack** **UDP flood attack** **UDP-Lag attack** **attack** 1 Unnamed: 0 Unnamed: 0 Unnamed: 0 Unnamed: 0 2 Flow Duration Flow Duration Flow Duration Flow Duration 3 Flow IAT Min Flow IAT Min Flow IAT Min Flow IAT Min Total Length of Bwd Total Length of Bwd Total Length of Bwd 4 Total Length of Bwd Packets Packets Packets Packets 5 Protocol Protocol Protocol Protocol 6 min_seg_size_forward min_seg_size_forward Inbound min_seg_size_forward 7 Fwd Packet Length Std Fwd Packet Length Max Fwd Packet Length Std Fwd Packet Length Std 8 Total Backward Packets Bwd Packet Length Min Total Backward Packets Total Backward Packets 9 Total Fwd Packets Active Std Active Std Active Std 10 Fwd Header Length Fwd Header Length Fwd Header Length 11 Active Mean Active Mean 12 Down/Up Ratio **Table 5. Accuracy, Precision, Recall, F-score and Specificity score values of the classification algorithms with common un-** correlated feature subset selected by the proposed model on Syn flood attack dataset **Precision** **Recall** **F-Score** **Specificity** **Classification algorithms** **Accuracy (%)** **Attack** **Benign** **Attack** **Benign** **Attack** **Benign** **Attack** **Benign** Logistic Regression 1.00 0.01 0.97 0.84 0.99 0.02 0.84 0.97 97.06 Decision Tree 1.00 0.21 1.00 0.72 1.00 0.32 0.72 1.00 99.89 KNN 1.00 0.03 0.99 0.82 0.99 0.05 0.82 0.99 98.91 Naive Bayes 1.00 0.33 1.00 0.80 1.00 0.47 0.80 1.00 99.93 Random Forest 1.00 0.24 1.00 0.78 1.00 0.36 0.78 1.00 99.90 Ada Boost 1.00 0.71 1.00 0.50 1.00 0.59 0.50 1.00 99.97 Gradient Boost 1.00 0.22 1.00 0.79 1.00 0.35 0.79 1.00 99.89 Multilayer Perceptron 1.00 1.00 1.00 0.48 1.00 0.64 0.48 1.00 99.98 ----- y,,, p y g correlated feature subset selected by the proposed model for the UDP flood attack **Precision** **Recall** **F-Score** **Specificity** **Classification algorithms** **Accuracy (%)** **Attack** **Benign** **Attack** **Benign** **Attack** **Benign** **Attack** **Benign** Logistic Regression 1.00 0.58 1.00 1.00 1.00 0.73 1.00 1.00 99.92 Decision Tree 1.00 0.46 1.00 0.77 1.00 0.57 0.77 1.00 99.88 KNN 1.00 0.93 1.00 0.83 1.00 0.88 0.83 1.00 99.98 Naive Bayes 1.00 0.00 0.04 1.00 0.07 0.00 1.00 0.04 3.87 Random Forest 1.00 0.64 1.00 0.94 1.00 0.76 0.94 1.00 99.94 Ada Boost 1.00 0.90 1.00 0.79 1.00 0.84 0.79 1.00 99.97 Gradient Boost 1.00 0.70 1.00 0.23 1.00 0.35 0.23 1.00 99.91 Multilayer Perceptron 1.00 0.95 1.00 0.67 1.00 0.78 0.67 1.00 99.96 **Table 7. Accuracy, Precision, Recall, F-score and Specificity score values of the classification algorithms with common un-** correlated feature subset selected by the proposed model for the UDP - Lag attack **Precision** **Recall** **F-Score** **Specificity** **Classification algorithms** **Accuracy (%)** **Attack** **Benign** **Attack** **Benign** **Attack** **Benign** **Attack** **Benign** Logistic Regression 1.00 0.17 0.95 0.93 0.97 0.28 0.93 0.95 94.77 Decision Tree 1.00 0.28 0.97 0.86 0.99 0.42 0.86 0.97 97.37 KNN 1.00 0.93 1.00 0.89 1.00 0.91 0.89 1.0 99.80 Naive Bayes 1.00 0.01 0.01 1.00 0.01 0.02 1.00 0.01 01.63 Random Forest 1.00 0.94 1.00 0.88 1.00 0.91 0.88 1.00 99.81 Ada Boost 1.00 0.98 1.00 0.76 1.00 0.86 0.76 1.00 99.71 Gradient Boost 1.00 0.90 1.00 0.87 1.00 0.89 0.87 1.00 99.75 Multilayer Perceptron 1.00 0.98 1.00 0.85 1.00 0.91 0.85 1.00 99.81 **Table 8. Accuracy, Precision, Recall, F-score and Specificity score values of the classification algorithms with common un-** correlated feature subset selected by the proposed model for the Customized Exploitation DDoS attack **Precision** **Recall** **F-Score** **Specificity** **Classification algorithms** **Accuracy (%)** **Attack** **Benign** **Attack** **Benign** **Attack** **Benign** **Attack** **Benign** Logistic Regression 1.00 0.15 0.98 0.92 0.99 0.25 0.92 0.98 97.55 Decision Tree 1.00 0.07 0.96 0.71 0.98 0.13 0.71 0.96 95.81 KNN 1.00 0.87 1.00 0.90 1.00 0.89 0.90 1.00 98.90 Naive Bayes 1.00 0.35 1.00 0.55 1.00 0.43 0.55 1.00 99.35 Random Forest 1.00 1.00 1.00 0.66 1.00 0.80 0.66 1.00 99.85 Ada Boost 1.00 0.96 1.00 0.06 1.00 0.11 0.06 1.0 99.58 Gradient Boost 1.00 0.86 1.00 0.74 1.00 0.79 0.74 1.00 99.83 Multilayer Perceptron 1.00 0.99 1.00 0.75 1.00 0.85 0.75 1.00 99.88 **Table 9. K-fold cross-validation accuracy scores (with a standard deviation) in % of the different classification algorithms with** common un-correlated feature subset selected by the proposed model **Classification Algorithms** **Syn flood attack** **UDP flood attack** **UDP-Lag attack** **Customized Exploitation DDoS attack** Logistic Regression 92.4385(0.7917) 99.9141(0.0111) 95.4938(0.1625) 97.0752(0.0245) Decision Tree 99.9974 (0.0009) 99.9831(0.0021) 99.9723(0.0064) 99.9558(0.0036) KNN 99.9960 (0.0010) 99.9774(0.0035) 99.8371(0.0079) 99.9342(0.0033) Naive Bayes 99.9138 (0.0032) 99.4898(0.0181) 99.3405(0.0291) 99.3236(0.0594) Random Forest 99.9978 (0.0006) 99.9630(0.0061) 99.9835(0.0060) 99.9316(0.0042) Ada Boost 99.9868 (0.0022) 99.9731(0.0041) 99.7678(0.0190) 99.8365(0.0112) Gradient Boost 99.9925 (0.0024) 99.9176(0.0360) 99.9165(0.0169) 99.8549(0.0309) Multilayer Perceptron 99.9844(0.0130) 99.9609(0.0059) 99.7955(0.0082) 99.9104(0.0138) **Table 10. ROC-AUC Scores of the different classification algorithms with common un-correlated feature subset selected by the** proposed model **Classification Algorithms** **Syn flood attack** **UDP flood attack** **UDP-Lag attack** **Customized Exploitation DDoS attack** Logistic Regression 0.9375 0.9998 0.9907 0.9892 Decision Tree 0.8593 0.8821 0.9167 0.8364 KNN 0.9070 0.9635 0.9529 0.9777 Naive Bayes 0.9154 0.9997 0.8921 0.9369 Random Forest 0.9566 0.9999 0.9950 0.8364 Ada Boost 0.9037 0.9999 0.9949 0.9950 Gradient Boost 0.9381 0.6153 0.9933 0.9204 Multilayer Perceptron 0.9681 0.9999 0.9941 0.9948 ----- y (with a standard deviation) in % of the different classification algorithms with common un-correlated feature subset on Syn flood, UDP flood, UDP-Lag, and Customized Exploitation DDoS attacks. Random forest produces the best K-fold cross validation accuracy score with less standard deviation while logistic regression produces lowest value on Syn flood DDoS attack dataset. On the UDP flood DDoS attack dataset, decision tree produces the best K-fold cross validation accuracy score with less standard deviation, whereas Naive Bayes produces the lowest value. Random forest produces the best K-fold cross validation accuracy score with less standard deviation while logistic regression produces lowest value on UDP-Lag DDoS attack dataset. On the customized exploitation DDoS attack dataset, decision tree produces the best K-fold cross validation accuracy score with less standard deviation, whereas logistic regression produces the lowest value. **ROC-AUC score** The Receiver Operating Characteristic (ROC) curve is used to evaluate the model's accuracy. The ROC curve depicts the relationship between True and False classes. The area underneath the ROC Curve (AUC) measures separability between false positive and true positive rates. A ROC curve is a graph that shows a classification model's performance overall decision threshold. A decision threshold is a value used to translate a probabilistic prediction into a class label. Scores between 0 and 1 on the ROC-AUC. When the ROC-AUC value is 1, the classifier correctly classifies all labels. When the ROC-AUC value is zero, the classifier classifies all labels not accordingly, that is, it classifies TRUE labels as FALSE labels and FALSE labels as TRUE labels. The ratio of benign data misclassification to the proportion of all attack data filled with abnormal data is known as the false-positive rate. Table 10 shows ROC-AUC Scores of the different classification algorithms with common un-correlated feature subset on Syn flood, UDP flood, UDP-Lag, and Customized Exploitation DDoS attacks. On a Syn flood attack, Multilayer Perceptron produces the best ROC-AUC score, while Decision Tree produces a lesser ROC-AUC score. Figure 3 shows the Receiver Operating Curve (ROC) of the classification algorithms with common un-correlated feature subset selected by the proposed model for Syn flood attack. On UDP flood attacks, Random forest, Ada boost, and Multilayer perceptron produce the best ROC-AUC scores, whereas Gradient boost produces the lowest ROC-AUC scores. Figure 4 shows the Receiver Operating Curve (ROC) of the classification algorithms with common un-correlated feature subset selected by the proposed model for the UDP flood attack. Random forest and Ada boost produce the best ROC-AUC scores for UDP-Lag attacks, whereas Naïve Bayes classifier produces the lowest ROC-AUC score. Figure 5 shows the Receiver Operating Curve (ROC) of the classification algorithms with common un-correlated feature subset selected by the proposed model for the UDP-Lag attack. On customized exploitation DDoS attacks, Ada boost and Multilayer perceptron produce the best ROC-AUC scores, while Decision tree and Random forest produce the lowest ROC-AUC scores. Figure 6 shows the Receiver Operating Curve (ROC) of the classification algorithms with common un-correlated feature subset selected by the proposed model for the Customized Exploitation DDoS attack. Even if Ada boost does not produce the best ROC-AUC score on the Syn-flood attack data set, it does so on the UDP g, exploitation DDoS attack dataset. Multilayer perceptron produces the best scores on Syn flood and UDP flood DDoS attack datasets, good scores on UDP-Lag DDoS attack datasets, and better scores on customized exploitation DDoS attack datasets in terms of ROC AUC. **Figure 3. Receiver Operating Curve (ROC) of the** classification algorithms with common un-correlated feature subset selected by the proposed model for Syn flood attack **Figure 4. Receiver Operating Curve (ROC) of the** classification algorithms with common uncorrelated features subset selected by the proposed model on UDP flood attack **Figure 5. Receiver Operating Curve (ROC) of the** classification algorithms with common un-correlated features subset selected by the proposed model on UDP-Lag attack ----- **Figure 6. Receiver Operating Curve (ROC) of the** classification algorithms with common un-correlated feature subset selected by the proposed model for the Customized Exploitation DDoS attack **Log loss** The most important probability-based classification metric is log loss. The lower the log-loss number, the better the predictions; the log loss value is 0 for a perfect model. 𝑁 𝐿𝑜𝑔−𝑙𝑜𝑠𝑠= − [1] 𝑁 [∑[𝑦][𝑖] [ln 𝑝][𝑖] [+ (1 −𝑦][𝑖][)𝑙𝑛(1] (11) 𝑖=1 −𝑝𝑖)] where, N is the number of observations, p is the prediction probability and y is the actual value. Table 11 shows Log-loss values of the different classification algorithms with common un-correlated features subset on Syn flood, UDP flood, UDP-Lag, and Customized Exploitation DDoS attacks. On the Syn flood DDoS attack dataset, the multilayer perceptron classifier produces the best log value, whereas logistic regression produces the poorest log loss value. On a UDP flood DDoS attack dataset, the KNN classifier produces the best log value, whereas the Naive y p p g lag DDoS attack dataset, the multilayer perceptron classifier produces the best log value, whereas the Naive Bayes classifier produces the poorest log loss value. On a customized exploitation DDoS attack dataset, the KNN classifier produces the best log value, whereas the Decision tree classifier produces the poorest log value. On all exploitation-based DDoS attack datasets, boosting type classifiers perform well in terms of log-loss evaluation metrics. **Run time** Run time means the execution time of the model. Table 12 shows Execution times (in seconds) of the different classification algorithms with common un-correlated feature subset on Syn flood, UDP flood, UDP-Lag, and Customized Exploitation DDoS attacks. In terms of execution time, the Naive Bayes classifier takes less time while the Gradient boosting classifier takes more time on the Syn flood DDoS attack dataset. On the UDP flood DDoS attack dataset, the Naive Bayes classifier takes less time to run, whereas the Gradient boosting classifier takes longer. The Naive Bayes classifier takes less time to execute on the UDP-Lag DDoS attack data set, whereas the multilayer perceptron takes longer. The Naive Bayes classifier takes less time to execute on the customized exploitation DDoS attack data set, whereas the random forest classifier takes longer. Ada boost classifier takes less time for execution than Gradient boost classifier, random forest bagging classifier, and multilayer perceptron neural network classifier. **Results of classification algorithms on customized data set** **with common feature subset** Table 4 shows the common un-correlated feature list for the Syn flood, UDP flood, UDP-Lag, and customized exploitation data sets. Unnamed: 0, Flow Duration, Flow IAT Min, Total Length of Bwd Packets, and Protocol are common in the lists of common un-correlated features of Syn flood, UDP flood, UDP-Lag, and customized exploitation DDoS attack data sets. Now classification algorithms applied to a customized DDoS attack dataset with these common feature subsets and results are evaluated. **Table 11. Log-loss values of the different classification algorithms with common un-correlated feature subset selected by the** proposed model **Classification Algorithms** **Syn flood attack** **UDP flood attack** **UDP-Lag attack** **Customized Exploitation DDoS attack** Logistic Regression 1.0144 0.0273 1.8063 0.8448 Decision Tree 0.0375 0.0427 0.9096 1.4486 KNN 0.3775 0.0085 0.0693 0.0359 Naive Bayes 0.0228 33.2010 33.9747 0.2251 Random Forest 0.0342 0.0221 0.0662 0.0523 Ada Boost 0.0088 0.0109 0.0988 0.1466 Gradient Boost 0.0372 0.0326 0.0864 0.0590 Multilayer Perceptron 0.0065 0.0138 0.0642 0.0400 **Table 12. Execution times (in seconds) of the different classification algorithms with common un-correlated feature subset** selected by the proposed model **Classification Algorithms** **Syn flood attack** **UDP flood attack** **UDP-Lag attack** **Customized Exploitation DDoS attack** Logistic Regression 19.9021 15.6982 3.3212 13.9782 Decision Tree 4.9124 1.6632 0.9543 3.7729 KNN 2.3210 2.7505 0.6286 2.3074 Naive Bayes 0.2645 0.2528 0.0674 0.2312 Random Forest 113.9609 59.7390 22.2232 148.4150 Ada Boost 34.0015 40.5446 9.8165 48.5320 Gradient Boost 121.1507 143.1666 31.9532 125.2996 Multilayer Perceptron 82.8440 78.7160 65.5630 138.0118 ----- y,,, p y g subset selected for Customized Exploitation DDoS attack **Precision** **Recall** **F-Score** **Specificity** **Classification algorithms** **Accuracy (%)** **Attack** **Benign** **Attack** **Benign** **Attack** **Benign** **Attack** **Benign** Logistic Regression 1.00 0.09 0.96 0.84 0.98 0.16 0.84 0.96 95.96 Decision Tree 1.00 0.95 1.00 0.64 1.00 0.76 0.64 1.00 99.82 KNN 1.00 0.85 1.00 0.72 1.00 0.78 0.72 1.00 99.82 Naive Bayes 1.00 0.68 1.00 0.26 1.00 0.38 0.26 1.00 99.61 Random Forest 1.00 0.96 1.00 0.66 1.00 0.78 0.66 1.00 99.83 Ada Boost 1.00 0.82 1.00 0.13 1.00 0.23 0.13 1.00 99.60 Gradient Boost 1.00 0.99 1.00 0.50 1.00 0.67 0.50 1.00 99.77 Multilayer Perceptron 1.00 0.95 1.00 0.60 1.00 0.73 0.60 1.00 99.81 **Table 14. K-fold cross-validation accuracy scores (with a standard deviation) in %, ROC-AUC Scores, Log- loss value and** execution times of the different classification algorithms on Customized Exploitation DDoS attack with common feature subset **Classification Algorithms** **K-fold cross-validation accuracy** **AUC Score** **Log-loss value** **Execution-time** Logistic Regression 88.4218(0.1191) 0.9640 1.3963 7.3079 Decision Tree 99.8955(0.0025) 0.8202 0.0612 4.5834 KNN 99.8627(0.0039) 0.9205 0.0633 1.3982 Naive Bayes 99.5840(0.0143) 0.7968 0.1338 0.1801 Random Forest 99.8859(0.0144) 0.9372 0.0571 105.7972 Ada Boost 99.8089(0.0088) 0.9762 0.1389 37.1230 Gradient Boost 99.8200(0.0087) 0.8213 0.0778 117.3310 Multilayer Perceptron 99.8360(0.0040) 0.9768 0.0673 133.2027 **Figure 7. Receiver Operating Curve (ROC) of the** classification algorithms with common feature subset on Customized Exploitation DDoS attack Table 13 shows performance evaluation metrics in terms of accuracy, precision, recall, F-score, and specificity for the different classification algorithms on Customized Exploitation DDoS attacks with common features which are common in four common un-correlated feature subsets. Decision tree, KNN, and Multilayer perceptron provide better accuracy scores. In terms of precision, recall, and F-score, all classification methods produce good results for attack classification. The Gradient boost classifier has a higher accuracy score, while the Logistic regression classifier has a higher benign score, and the KNN and Random forest classifiers have a higher F-score for benign classification. Except for Logistic regression, all classification methods have a higher specificity score benign classification, while Logistic regression has a higher specificity score for attack classification. Table 14 shows K-fold cross-validation accuracy scores (with a standard deviation) in %, ROC-AUC Scores, Log-loss, value and execution times of the different classification algorithms on Customized Exploitation DDoS attack with common feature subset. Multilayer perceptron gives the best ROC-AUC value while Naive Bayes provides the lowest ROC-AUC score values in customized exploitation DDoS attack dataset. Figure 7 shows the ROC curves of the classification algorithms with common feature subset on the Customized Exploitation DDoS attack. On a customized exploitation DDoS attack dataset with common features set, KNN provides the best log loss value, whereas logistic regression provides the lowest log loss values. On a customized exploitation DDoS attack dataset with common features set, Naive Bayes takes less time for execution, whereas multilayer perceptron takes more time for execution. On a customized exploitation DDoS attack dataset with common features set, the Decision tree provides the best Kfold cross-validation accuracy value, whereas logistic regression provides the lowest K-fold cross-validation accuracy score values. **4.** **CONCLUSIONS** This research evaluates the effectiveness of the classification algorithms for detecting exploitation DDoS attacks on three CIC-DDoS2019 datasets and customized exploitation DDoS attack dataset with common un-correlated feature subset selected by Pearson, Spearman and Kendall correlation methods. The classification methods of multilayer perceptron and Ada boost produce the best accuracy results compared to others on Syn flood DDoS attack dataset. Decision tree, KNN, and Multilayer perceptron provide better accuracy scores on UDP-flood attack dataset. Random Forest and multilayer perceptron produce the best accuracy results compared to other classifiers on UDP-lag attacks. Decision tree, KNN, and Multilayer perceptron provide better accuracy scores on customized exploitation DDoS attacks. Multilayer perceptron produces the best accuracy results compared to other classifiers on customized exploitation DDoS attacks dataset with common features which are common features of in un-correlated feature subsets. Overall, multilayer ----- p p p y p DDoS attacks datasets. It also provides good results in remaining evaluation metrics. **REFERENCES** [1] Dasari, K.B., Nagaraju, D. (2018). Distributed denial of service attacks, tools and defense mechanisms. International Journal of Pure and Applied Mathematics, 120(6): 3423-3437. [2] Ramkumar, B.N., Subbulakshmi, T. (2021). Tcp Syn flood attack detection and prevention system using adaptive thresholding method. ITM Web of Conferences, 37: 01016. https://doi.org/10.1051/itmconf/20213701016 [3] Amaizu, G.C., Nwakanma, C.I., Bhardwaj, S., Lee, J.M., Kim, D.S. (2021). Composite and efficient DDoS attack detection framework for B5G networks. Computer Networks, 188: 107871. https://doi.org/10.1016/j.comnet.2021.107871 [4] Moubayed, A., Aqeeli, E., Shami, A. (2020). Ensemble based feature selection and classification model for DNS typo-squatting detection. IEEE Canadian Conference on Electrical and Computer Engineering (CCECE), pp. 1-6. https://doi.org/10.1109/CCECE47787.2020.9255697 [5] Li, Z.L., Hu, G.M., Yang, D. (2008). Global abnormal correlation analysis for DDoS attack detection. 2008 IEEE Symposium on Computers and Communications, pp. 310-315. http://dx.doi.org/10.1109/ISCC.2008.4625614 [6] Mekala, S., Rani, B.P. (2020). Kernel PCA based dimensionality reduction techniques for preprocessing of Telugu text documents for cluster analysis. International Journal of Advanced Research in Engineering and Technology, 8(12): 785-793. [7] Dasari, K.B., Devarakonda, N. (2021). Detection of different DDoS attacks using machine learning classification algorithms. Ingénierie des Systèmes d’Information, 26(5): 461-468. http://dx.doi.org/10.18280/isi.260505 [8] Yan, Y.D., Tang, D., Zhan, S.J., Dai, R., Chen, J.W., Zhu, ( ) improved logistic regression. IEEE 21st International Conference on High-Performance Computing and Communications, pp. 468-476. http://dx.doi.org/10.1109/HPCC/SmartCity/DSS.2019.0 0076 [9] Lakshminarasimman, S., Ruswin, S., Sundarakandam, K. (2017). Detecting DDoS attacks using decision tree algorithm. Fourth International Conference on Signal Processing, Communication and Networking (ICSCN), pp. 1-6. http://dx.doi.org/10.1109/ICSCN.2017.8085703 [10] Dong, S., Sarem, M. (2019). DDoS attack detection method based on improved KNN with the degree of DDoS attack in software-defined networks. IEEE Access, 8: 5039-5048. http://dx.doi.org/10.1109/ACCESS.2019.2963077 [11] Singh, N.A., Singh, K.J., De, T. (2016). Distributed denial of service attack detection using Naive Bayes classifier through info gain feature selection. ICIA-16: Proceedings of the International Conference on Informatics and Analytics, pp. 1-9. https://doi.org/10.1145/2980258.2980379 [12] Chen, Y., Hou, J., Li, Q.M., Long, H.Q. (2020). DDoS attack detection based on random forest. 2020 IEEE International Conference on Progress in Informatics and Computing (PIC), pp. 328-334. https://doi.org/10.1109/PIC50277.2020.9350788 [13] Rachmadi, S., Mandala, S., Oktaria, D. (2021). Detection of DoS attack using AdaBoost algorithm on IoT system. 2021 International Conference on Data Science and Its Applications (ICoDSA), pp. 28-33. http://dx.doi.org/10.1109/ICoDSA53588.2021.9617545 [14] Chen, Z., Jiang, F., Cheng, Y., Gu, X., Liu, W., Peng, J. (2018). XGBoost classifier for DDoS attack detection and analysis in SDN-based cloud. In 2018 IEEE International Conference on Big Data and Smart Computing (BigComp), pp. 251-256. http://dx.doi.org/10.1109/BigComp.2018.00044 [15] Wang, M., Lu, Y., Qin, J. (2020). A dynamic MLP-based DDoS attack detection method using feature selection and feedback. Computers & Security, 88: 101645. http://dx.doi.org/10.1016/j.cose.2019.101645 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.18280/ria.360107?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.18280/ria.360107, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://www.iieta.org/download/file/fid/69435" }
2,022
[]
true
2022-02-28T00:00:00
[]
14,744
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01157f7c700e92323a5933e00c71cf001a8bac88
[ "Computer Science" ]
0.916781
Blockchain with Internet of Things: Benefits, Challenges, and Future Directions
01157f7c700e92323a5933e00c71cf001a8bac88
International Journal of Intelligent Systems and Applications
[ { "authorId": "17746180", "name": "Hany F. Atlam" }, { "authorId": "144901593", "name": "A. Alenezi" }, { "authorId": "145568788", "name": "M. Alassafi" }, { "authorId": "144118261", "name": "G. Wills" } ]
{ "alternate_issns": null, "alternate_names": [ "Int J Intell Syst Appl" ], "alternate_urls": null, "id": "f9b6bddb-d88a-424d-a940-2e092886ca71", "issn": "2074-904X", "name": "International Journal of Intelligent Systems and Applications", "type": "journal", "url": "http://www.mecs-press.org/ijisa/" }
The Internet of Things (IoT) has extended the internet connectivity to reach not just computers and humans, but most of our environment things. The IoT has the potential to connect billions of objects simultaneously which has the impact of improving information sharing needs that result in improving our life. Although the IoT benefits are unlimited, there are many challenges facing adopting the IoT in the real world due to its centralized server/client model. For instance, scalability and security issues that arise due to the excessive numbers of IoT objects in the network. The server/client model requires all devices to be connected and authenticated through the server, which creates a single point of failure. Therefore, moving the IoT system into the decentralized path may be the right decision. One of the popular decentralization systems is blockchain. The Blockchain is a powerful technology that decentralizes computation and management processes which can solve many of IoT issues, especially security. This paper provides an overview of the integration of the blockchain with the IoT with highlighting the integration benefits and challenges. The future research directions of blockchain with IoT are also discussed. We conclude that the combination of blockchain and IoT can provide a powerful approach which can significantly pave the way for new business models and distributed applications.
Published Online June 2018 in MECS (http://www.mecs-press.org/) DOI: 10.5815/ijisa.2018.06.05 # Blockchain with Internet of Things: Benefits, Challenges, and Future Directions ## Hany F. Atlam Electronic and Computer Science Dept., University of Southampton, Southampton, UK Computer Science and Engineering Dept., Faculty of Electronic Engineering, Menoufia University, Menoufia, Egypt E-mail: hfa1g15@soton.ac.uk ## Ahmed Alenezi, Madini O. Alassafi, Gary B. Wills Electronic and Computer Science Dept., University of Southampton, Southampton, UK E-mail: {aa4e15, moa2g15, gbw}@soton.ac.uk Received: 04 November 2017; Accepted: 09 February 2018; Published: 08 June 2018 **_Abstract—The Internet of Things (IoT) has extended the_** internet connectivity to reach not just computers and humans, but most of our environment things. The IoT has the potential to connect billions of objects simultaneously which has the impact of improving information sharing needs that result in improving our life. Although the IoT benefits are unlimited, there are many challenges facing adopting the IoT in the real world due to its centralized server/client model. For instance, scalability and security issues that arise due to the excessive numbers of IoT objects in the network. The server/client model requires all devices to be connected and authenticated through the server, which creates a single point of failure. Therefore, moving the IoT system into the decentralized path may be the right decision. One of the popular decentralization systems is blockchain. The Blockchain is a powerful technology that decentralizes computation and management processes which can solve many of IoT issues, especially security. This paper provides an overview of the integration of the blockchain with the IoT with highlighting the integration benefits and challenges. The future research directions of blockchain with IoT are also discussed. We conclude that the combination of blockchain and IoT can provide a powerful approach which can significantly pave the way for new business models and distributed applications. **_Index Terms—Blockchain, Blockchain with IoT, Internet_** of Things, Centralized, Decentralized. I. INTRODUCTION The Internet of Things (IoT) has the ability to connect and communicate billions of things simultaneously. It provides various benefits to consumers that will change the way that users interact with the technology. Using a collection of cheap sensors and interconnected objects, information can be collected from our environment that will allow improving our way of living [1]. The IoT concept is not new. In 1999, Ashton, who is the founder of MIT auto-identification center has said, “The Internet of Things has the potential to change the world, just as the Internet did. Maybe even more so”[2]. Later in 2005, the ITU officially define the IoT as “a global infrastructure for the information society, enabling advanced services by interconnecting (physical and virtual) things based on, existing and evolving, interoperable information and communication technologies” [3]. Current IoT systems are built on centralized server/client model, which requires all devices to be connected and authenticated through the server. This model would not be able to provide the needs to outspread the IoT system in the future [4]. Therefore, moving the IoT system into the decentralized path may be the right decision. One of the popular decentralization platforms is blockchain. A blockchain is a distributed database of records that contains all transactions that have been executed and shared among participating parties in the network. This distributed databased is called distributed ledger. Each transaction is stored in the distributed ledger and must be verified by consent of the majority of participants in the network. All transactions that have ever made are contained in the blockchain. Bitcoin, the decentralized peer-to-peer digital currency, is the most popular example that uses blockchain technology [5]. Integrating IoT with blockchain will have many benefits. The decentralization model of the blockchain will have the ability to handle processing of billions of transactions between IoT devices, which will significantly reduce the costs associated with installing and maintaining large centralized data centers and will distribute computation and storage needs across the billions of devices that form IoT networks. In addition, working with the blockchain technology will eliminate the single point of failure associated with the centralized IoT architecture [6]. Moreover, integrating blockchain with IoT will allow the peer-to-peer messaging, file distribution and autonomous coordination between IoT devices with no need for the centralized server-client model [4]. ----- This paper provides an overview of integrating blockchain with the IoT system; this involves an examination of the benefits resulting from the integration process and the implementation challenges encountered. The ultimate goal of this work is to provide a detailed description of the benefits and the challenges that result from combing blockchain with IoT so as to make the decision whether to go with the decentralization for the IoT or not. The remainder of this paper is organized as follows: Section II presents related work discussing integration blockchain with IoT applications; Section III discusses centralized IoT architecture; Section IV presents the blockchain technology and its structure; Section V introduces essential characteristics of the blockchain; Section VI discusses how the blockchain works; Section VII presents blockchain with IoT; Section VIII illustrates benefits of integrating blockchain with IoT; Section IX challenges of blockchain with IoT; Section X discusses future research directions, and Section XI is the conclusion. II. RELATED WORK The integration of blockchain with IoT have investigated in a few papers. For instance, the IBM Autonomous Decentralized Peer-to-Peer Telemetry (ADEPT) project [7] leverages the blockchain to build a distributed network of devices. As for the ADEPT project, many other approaches are trying to design a solution that will be able to merge all the different blockchain based applications [8]. Also, Slock.it introduced the first implementation of IoT and Blockchain using the Ethereum platform [9]. It is so called Slocks to reflect real-world physical objects that can be controlled by the Blockchain. They use the Ethereum Computer which is a piece of electronics that brings Blockchain technology to the entire home, making it possible to rent access to any compatible smart object and accept payments without intermediaries. In addition, Dorri et al. [10] have proposed a new secure, private, and lightweight architecture for IoT, based on blockchain technology that eliminates the overhead while maintaining most of its security and privacy benefits was investigated on a smart home application as a representative case study for broader IoT applications. The proposed architecture was hierarchical, and consists of smart homes, an overlay network and cloud storages coordinating data transactions with blockchain to provide privacy and security. Blockchain in healthcare IoT application has introduced as a solution for many challenges facing healthcare sector. For example, Gupta et al. [11] have proposed an approach to explain how Blockchain could enable an interoperable and secure electronic health records exchange in which health consumers are the ultimate owners. They have proposed a scenario to store only the metadata about health and medical events on the Blockchain. Otherwise the Blockchain infrastructure will have to scale massively to support complete health records. So, metadata such as patient identity, visit ID, provider ID, payer ID, etc. can be kept on a Blockchain, but the actual records should be stored in a separate universal health cloud. Another study for Blockchain in Healthcare utilize Ethereum's smart contracts to create representations of existing medical records [12]. These contracts are stored directly within individual nodes on the network. They have proposed a solution called “MedRec” to structure the large amount of data into three types of contracts. The first one is Registrar Contract. It stores the participants’ identity with all the needed details and of course, the public keys. This kind of identity registration can be restricted only to certified institutions. The second contract is the Patient-Provider Relationship Contract. It is issued when one node stores or manages data for another node. The main usage will be when there is a smart contract between the care provider and patient. The last one is Summary Contract which helps the patient to locate her medical history. As a result of this contract, all previous and current engagements with other nodes in the system are listed. III. CENTRALIZED IOT ARCHITECTURE Basically, the IoT is the connection and communication of different devices over the Internet. These devices are composed of networking nodes whether serves or computer which are connected together to share their data. All devices are provided with sensors, which collect data that can be transmitted, stored, analyzed, and presented in a useful way [13]. There are many architectures for IoT, which is approved commonly. Different researchers and organizations proposed different architectures. According to the ITU, the IoT architecture is composed of four layers as shown in Fig.[3]: - Application layer - Service support and application layer - Network Layer - Device layer Fig.1. IoT reference model and architecture [3] Application layer encompasses IoT applications. There are many IoT application such as healthcare, smart cities, ----- connected car, smart energy, smart agriculture and ...etc. Service support and application layer contains common capabilities which can be used by different IoT applications [14]. The Network layer includes devices such as routers, switches, gateways, and firewalls that are used to construct local and wide-area networks to provide Internet connectivity. In addition, it enables devices to communicate with one another and to communicate with application platforms such as computers, remote-control devices, and smartphones [13]. The device layer is similar to the physical layer of the Open System Interconnection (OSI) model of the network architecture. It is composed of physical devices and controllers that control objects. These objects represent things in the IoT that include a wide range of endpoint devices that send and receive a variety of information. For instance, sensors that collect information about the surrounding environment [15]. The current IoT architecture is built as a centralized model which is known as server/client model. In this model, all devices cannot talk to each other but talk to a centralized gateway instead. The centralized model has used to connect a wide range of computing devices for many years and will continue to support small-scale IoT networks, however, it will not be capable of providing the needs to extend the IoT system in the future [4]. The number of IoT devices will increase dramatically such that a network capacity will be at least 1,000 times the level of 2016. Cisco has reported that the number of IoT devices is about to reach 20 billion in 2020 [16]. Therefore, the amount of communication that needs to be handled will definitely increase costs exponentially. Even if costs and communication challenges are managed, the server/client model will still a point of failure that can interrupt the entire network [6]. In addition, the centralized model is vulnerable to data manipulation. Collecting real-time data does not ensure that the information is put to good and appropriate use. For example, if energy companies found that smart meter data analysis will be the evidence that might result in high costs or lawsuits. They will edit or delete these data [17]. A decentralized approach for the IoT would solve many of these issues. One of the popular decentralization techniques is blockchain. The next section discusses the blockchain technology. IV. BLOCKCHAIN TECHNOLOGY Blockchain technology provides an efficient way of recording transactions or any digital interaction in a way that makes it secure, transparent, highly resistant to outages, auditable. This technology is still new and changing very fast; adopting it in the commercial market is still a few years off. However, decision-makers across industries and business functions should pay their attention now and start to investigate applications of this technology to avoid disruptive surprises or missed opportunities [6]. In 2008, Satoshi Nakamoto has introduced the concept of Bitcoin. This was by releasing the popular paper, “Bitcoin: A Peer-to-Peer Electronic Cash System” [18]. The paper presented a proposal for distributing electronic transactions rather than maintaining it dependent on centralized institutions for the exchange [19]. There are many definitions for the blockchain. According to [5], the blockchain is defined as “ a distributed database of records, or public ledger of all transactions or digital events that have been executed and shared among participating parties”. Each transaction in the public ledger is verified by consensus of a majority of the participants in the system. Once entered, information can never be erased. The blockchain contains a certain and verifiable record of every single transaction ever made [5]. A blockchain consists of two main elements [6], as shown in Fig.1: - _Transactions: are the actions generated by the_ participants in the system. - _Blocks: record the transactions and make sure they_ are in the correct sequence and have not been tampered with. Fig.1. Structure of blockchain [13] V. CHARACTERISTICS OF BLOCKCHAIN The blockchain has many features that make it very attractive for the IoT to solve many of its issues. As shown in Fig.2, according to [10] blockchain characteristics include: 1. **Immutability: Building immutable ledgers is one** of the key value of blockchain. All centralized databases can be corrupted and commonly requires trust in a third party to keep the information integrity. Once you have agreed on a transaction and recorded it, it can never be changed. 2. **Decentralization: The lack of centralized control** ensures scalability and robustness by using resources of all participating nodes and eliminating many-to-one traffic flows, which in turn decreases latency and solve the problem of single point of failure that exists in the centralized model. 3. **Anonymity: The anonymity provides an efficient** way of hiding the identity of users and keeps their identities private. 4. **Better Security: Blockchain provides better** security because there is no single point of failure ----- to shut down the entire network. 5. **Increased Capacity: One of the significant things** about blockchain technology is that it can increase the capacity of an entire network. Having thousands of computers working together as a whole can have greater power than a few centralized servers. Fig.2. Characteristics of blockchain VI. HOW BLOCKCHAIN WORKS? Although the blockchain is still new and in experimenting stage, it is being perceived as a revolutionary solution that addresses modern technology issues such as decentralization, identity, trust, data ownership and data-driven decisions [7]. The blockchain is generally a database that stores all the transactions in blocks. When a new transaction is created, the sender broadcasts it to the Peer–to–Peer communication channel to all other nodes in the network. The transaction is still new and not verified. When the nodes receive the transaction, they validate it and keep it in their ledger [20]. Transaction validation is performed by running predefined checks on the structure and the actions of the transaction. Special node types called miners create a new block and include all or some of the available transactions from their transaction pool. Then the block is mined, which is a process of finding the proof of work using variable data from the new block’s header [20]. Finding the proof of work is the continuous calculation of a cryptographic hash that fits the defined difficulty target. Mining requires a lot of processing power and the miners use a dedicated mining hardware. The miner that first finds a solution for its block is the winner. His candidate block becomes the new block in the chain. Because transactions are added in the mining block as they arrive, therefore, the latest block in the blockchain contains the latest transactions [4]. When a new block is created, it is time-stamped and propagated to all network nodes. Every node receives the block, validates it, validates the transactions, and adds the block to his ledger. When the majority of nodes accepted the block, it becomes authorized and non-reversible part of the blockchain. In addition to transactions, every block stores some metadata and the hash value of the previous block. So, every block has a pointer to its parent block. That is how the blocks are linked, creating a chain of blocks called blockchain [4]. The distributed ledger is available for everyone in the network to check the blocks and the transactions within. However, the users stay anonymous, they only identified by their public key as an address. Moreover, the transactions are encrypted. Invalid transactions are rejected and are not included in the blocks. Malicious attempt to make a change in the transactions will require repeated calculation of the proof of work for the attached block and all the blocks afterwards. These calculations are infeasible unless the majority of the nodes in the network are malicious [21]. Fig.3. Simple example of blockchain technology This section will discuss the blockchain with a simple example. Suppose that we have four nodes A, B, C and D who want to use the blockchain to transfer money, as it’s known as Bitcoin. To transfer the money from one node to another node, there will be no intermediate third party to make the transfer process, which is the idea of decentralization. Therefore, if node A wanted to transfer money to node B, it will be transferred directly. As shown in Fig.3, suppose node A wants to send £5 to node B, then a transaction will be created and verified by all other nodes in the network to include it in the ledger. In addition, if node B wants to send £10 to node D, then a transaction will be created and verified by all other nodes in the network to include it in the ledger. This will be the same scenario when node C wants to send £20 to node D. All the transactions are chained together in what is called ledger. This ledger is distributed across all nodes in the network to make sure that all nodes have the same copy or version from the ledger, that is why it’s called distributed ledger. VII. BLOCKCHAIN WITH IOT The IoT is an interesting developing system that ----- provides unlimited benefits, but there are many challenges with the current centralized IoT architecture such that all devices are identified, authenticated and connected through the centralized servers [4]. This model was used to connect a wide range of computing devices for many years and will continue to support small-scale IoT networks, however, it will not be capable of providing the needs to extend the IoT system in the future [22]. Table presents a comparison between blockchain and IoT. There are many advantages of both technologies, which can be combined, and get an improved outcome. The IoT has unlimited benefits and adopting a decentralized approach for the IoT would solve many issues especially security. Adopting a standardized peerto-peer communication model to process the hundreds of billions of transactions between devices will significantly reduce the costs associated with installing and maintaining large centralized data centers and will distribute computation and storage needs across the billions of devices that form IoT networks. This will prevent failure in any single node in a network from bringing the entire network to a halting collapse [17,18]. Table 2. Comparison between blockchain and IoT **Blockchain** **IoT** Decentralized Centralized Resource consuming Resource restricted Block mining is time Demands low latency consuming IoT considered to contains Scale poorly with large large number of devices network IoT devices have limited High bandwidth consumption bandwidth and resources Security is one of the big Has better security challenges of IoT The decentralized, autonomous, and trustless capabilities of the blockchain make it an ideal component to become a foundational element of IoT solutions. It is no surprise that enterprise IoT technologies have quickly become one of the early adopters of blockchain technology. However, establishing peer-to-peer communications will present its own set of challenges especially security. IoT security is much more than just about protecting sensitive data. Therefore, the blockchain solutions will have to maintain privacy and security in IoT networks and use validation and consent of participants for transactions to prevent spoofing and theft [6]. In addition, blockchain technology is considered the key solutions to solve privacy and reliability issues in the IoT. It can be used in tracking billions of connected devices, enabling the processing of transactions and coordination between devices; this allows for significant savings for IoT industry manufacturers[24]. Moreover, this decentralized approach would eliminate single points of failure, creating a more resilient system for devices to run on. The cryptographic algorithms used by blockchains would make consumer data more private [25]. In an IoT network, the blockchain can keep an immutable record of the history of smart devices. This feature enables the autonomous functioning of smart devices without the need for centralized authority [26]. As a result, the blockchain will open a series of IoT scenarios that were difficult, or even impossible to implement without it. For example, by leveraging the blockchain, IoT solutions can enable secure, trustless messaging between devices in an IoT network [27]. In this model, the blockchain will treat message exchanges between devices similar to financial transactions in a bitcoin network. To enable message exchanges, devices will leverage smart contracts which then model the agreement between the two parties [20]. One of the most exciting capabilities of the blockchain is the ability to maintain a duly decentralized, trusted ledger of all transactions occurring in a network. This capability is essential to enable the many compliances and regulatory requirements of industrial IoT (IIoT) applications without the need to rely on a centralized model [6]. Many large organizations have started to adopt blockchain with IoT systems to get all benefits of the blockchain. For instance, IBM in partnership with Samsung has developed a platform ADEPT (Autonomous Decentralized Peer- To- Peer Telemetry) that uses elements of the bitcoin’s underlying design to build a distributed network of devices, a decentralized IoT. ADEPT uses three protocols-BitTorrent ( file sharing), Ethereum ( Smart Contracts) and TeleHash ( Peer-ToPeer Messaging)-in the platform [28]. VIII. BENEFITS OF INTEGRATING BLOCKCHAIN WITH IOT There are many benefits of adopting blockchain with IoT, as shown in Fig.4. These benefits can be summarized as follows: 1. **Publicity: All participants have the ability to see** the all the transactions and all blocks as each participant has its own ledger. The content of the transaction is protected by participant’s private key [19], so even all participants can see them, they are protected. The IoT is a dynamic system in which all connected devices can share information together and at the same time protecting users’ privacy. 2. **Decentralization: The majority of participants** must verify the transactions in order to approve it and add it to the distributed ledger. There is no single authority that can approve the transactions or set specific rules to have transactions accepted. Therefore, there is a massive amount of trust included since the majority of the participants in the network have to reach an agreement to validate transactions [28]. Therefore, the blockchain will provide a secure platform for IoT devices. In addition, eliminating centralized traffic flows and single point of failure of the current centralized IoT architecture. |Blockchain|IoT| |---|---| |Decentralized|Centralized| |Resource consuming|Resource restricted| |Block mining is time- consuming|Demands low latency| |Scale poorly with large network|IoT considered to contains large number of devices| |High bandwidth consumption|IoT devices have limited bandwidth and resources| |Has better security|Security is one of the big challenges of IoT| ----- 3. **Resiliency: Each node has its own copy of the** ledger that contains all transactions that have ever made in the network. So, the blockchain is better able to withstand attack. Even if one node was compromised, the blockchain would be maintained by every other node [29]. Having a copy of data at each node in the IoT will improve information sharing needs. However, it introduces new processing and storage issues. 4. **Security: Blockchain has the ability to provide a** secure network over untrusted parties which is needed in IoT with numerous and heterogeneous devices [10]. In other words, all IoT network nodes must be malicious to perform an attack. 5. **Speed: A blockchain transaction is distributed** across the network in minutes and will be processed at any time throughout the day [16]. 6. **Cost saving: Existing IoT solutions are expensive** because of the high infrastructure and maintenance cost associated with centralized architecture, large server farms, and networking equipment. The total amount of communications that will have to be handled when there are tens of billions of IoT devices will increase those costs substantially [30]. 7. **Immutability: Having an immutable ledger is one** of the main advantages of blockchain technology. Any changes in the distributed ledger must be verified by the majority of the network nodes. Therefore, the transactions cannot be altered or deleted easily [14, 25]. Having an immutable ledger for IoT data will increase security and privacy which are the major challenges in this technology and all new technologies. 8. **Anonymity: To process the transaction, both** buyer and seller use anonymous and unique address numbers which keep their identity private. This feature has been criticised as it increases the use of cryptocurrencies in the illegal online market. However, it could be seen as an advantage if used for other purposes, for example, electoral voting systems [14, 26]. Fig.4. Benefits of integrating blockchain with IoT IX. CHALLENGES OF BLOCKCHAIN WITH IOT There is no doubt that integrating blockchain would have many advantages. However, the blockchain technology is not a perfect model which has its own flaws and challenges, as shown in Fig.5. These challenges can be summarized as follow: 1. **Scalability: Scalability issues in the blockchain** might lead to centralization, which is casting a shadow over the future of the cryptocurrency. The blockchain scales poorly as the number of nodes in the network increases. This issue is serious as IoT networks are expected to contain a large number of nodes [28]. 2. **Processing Power and Time: The processing** power and time needed to achieve encryption for all the objects included in a blockchain system. IoT systems have different types of devices which have very different computing capabilities, and not all of them will be able to run the same encryption algorithms at the required speed [14, 27]. 3. **Storage: One of the main benefits of blockchain is** that it eliminates the need for a central server to store transactions and device IDs, but the ledger has to be stored on the nodes themselves [33]. The distributed ledger will increase in size as time passes and with increasing number of nodes in the network. As said earlier, IoT devices have low computational resources and very low storage capacity [34]. 4. **Lack of skills: The blockchain technology is still** new. Therefore, a few people have large knowledge and skills about the blockchain, especially in banking. In other applications, there is a widespread lack of understanding of how the blockchain works [6]. The IoT devices exist everywhere, so adopting the blockchain with IoT will be very difficult without public awareness about the blockchain. 5. **Legal and Compliance: The blockchain is a new** technology that will have the ability to connect different people from different countries without having any legal or compliance code to follow, which is a serious issue for both manufacturers and service providers. This challenge will be the major barrier for adopting blockchain in many businesses and applications [35]. 6. **Naming** **and** **Discovery:** The blockchain technology has not been designed for the IoT, meaning that nodes were not meant to find each other in the network. An example is the Bitcoin application in which the IP addresses of some “senders” are embedded within the Bitcoin client and used by nodes to build the network topology. This approach will not work for the IoT as IoT devices will keep moving all the time which will change the topology continuously [23]. ----- Fig.5. Blockchain and IoT challenges X. FUTURE RESEARCH DIRECTIONS The blockchain has changed the concept of centralized authorities. The integration of blockchain with IoT will be the starting point for opening new businesses and applications. This section discusses future research directions of blockchain with IoT. This can be summarized as follows: _A. Smart Contracts_ Smart contracts are scripts stored on the blockchain. They are so powerful because of their flexibility. They can encrypt and store data securely, restrict access to data to only the desired parties and then be programmed to utilize the data within a self-executing logical workflow of operations between parties. Smart contracts translate business process into the computational process, greatly improving operational efficiency [5]. Using smart contracts within the IoT systems will provide an efficient way to improve security and Integrity of IoT data. The research questions that need to be addressed regarding conducting smart contracts within IoT systems are: Q1: Are the smart contracts able to execute all event functions of IoT devices, which are in billions? Q2: How the smart contract will respond to changing environmental conditions of the IoT as it is a dynamic system? Q3: What is the appropriate platform to implement smart contracts within IoT systems? _B. Regulatory Laws_ Regulatory Laws are the procedures created by authorities and local administrative agencies to define legal ways of working with a product or technology within a certain country or region. As said earlier, the blockchain is a new technology which has not any legal or compliance code to follow. The research question that needs to be addressed regarding blockchain legal and compliance issues is: Q1: What are regulatory rules that ensure the best practice of blockchain in IoT globally? _C. Security_ For all new technologies, the security is still the most challenging topic that takes the attention of researchers and organizations. Integrating blockchain with IoT can improve security as it uses use consent of the majority of participants to validate transactions to prevent spoofing and theft. However, IoT devices have low computational resources and storage space that cannot be able to process cryptographic algorithms. The research questions that need to be addressed regarding security are: Q1: What is the optimum platform for IoT to integrate with blockchain? Q2: How to overcome low capabilities of IoT devices to provide a secure IoT system? _D. IOTA_ [IOTA is a new generation of public and distributed](https://iota.org/) ledger that uses a concept called “Tangle”. The Tangle is a new data structure that based on a Directed Acyclic Graph (DAG). IOTA provide an efficient, secure, lightweight, and real-time transaction without fees. It is open-source, decentralized cryptocurrency, designed specifically for the IoT [36]. As IOTA is designed specifically for the IoT, it may be more appropriate to different IoT applications. However, it’s still under construction. The research questions that need to be addressed regarding IOTA are: Q1: What the appropriate decentralization technology for the IoT, blockchain or IOTA? Q2: What are major challenges with IOTA? XI. CONCLUSION The IoT technology has extended that reached to every home in the universe. It has the ability to connect everyday objects to the Internet. Through cheap sensors, a lot of information can be collected from the surrounding environment that results in improving our life. However, current IoT architecture that based on server/client model has many issues that need to be addressed especially scalability and security. One of the solutions to address IoT issues is blockchain. Blockchain provides distributed peer-to-peer communication network where non-trusting nodes can interact with each other without a trusted intermediary, in a verifiable manner. In this paper, we provided an overview of integrating blockchain with IoT with highlighting benefits and challenges. The discussion also focused on future research directions. At the end, we can conclude that integrating blockchain with IoT can bring many advantages that improve many of IoT issues but at the same time, it introduces new challenges that should be addressed. There is still need more research to investigate implementing blockchain with IoT in more details. ACKNOWLEDGMENT We acknowledge Egyptian cultural affairs and missions sector and Menoufia University for their ----- scholarship to Hany Atlam that allows the research to be funded and undertaken. REFERENCES [1] H. F. Atlam, A. Alenezi, R. J. Walters, and G. B. Wills, “An Overview of Risk Estimation Techniques in Riskbased Access Control for the Internet of Things,” in _Proceedings of the 2nd International Conference on_ _Internet of Things, Big Data and Security (IoTBDS 2017),_ 2017, pp. 254–260. [2] K. Ashton, “That ‘Internet of Things’ Thing,” RFiD J., p. 4986, 2009. [3] ITU, “Overview of the Internet of things,” Ser. Y Glob. Inf. _infrastructure, internet Protoc. Asp. next-generation_ _networks - Fram. Funct. Archit. Model., p. 22, 2012._ [4] E. Karafiloski, “Blockchain Solutions for Big Data Challenges A Literature Review,” in _IEEE EUROCON_ _2017_ _-17th_ _International_ _Conference_ _on_ _Smart_ _Technologies, 2017, no. July, pp. 6–8._ [5] A. Stanciu, “Blockchain based distributed control system for Edge Computing,” in _21st International Conference_ _on Control Systems and Computer Science Blockchain,_ 2017, pp. 667–671. [6] A. Banafa, “IoT and Blockchain Convergence: Benefits and Challenges,” _IEEE IoT Newsletter, 2017. [Online]._ Available: http://iot.ieee.org/newsletter/january-2017/iotand-blockchain-convergence-benefits-and-challenges.html. [7] IBM, “ADEPT: An IoT Practitioner Perspective,” 2015. [8] J. H. Ziegeldorf, F. Grossmann, M. Henze, N. Inden, and K. Wehrle, “CoinParty: Secure Multi-Party Mixing of Bitcoins,” in Proceedings of the 5th ACM Conference on _Data_ _and_ _Application_ _Security_ _and_ _Privacy_ _-_ _CODASPY ’15, 2015, no. August, pp. 75–86._ [9] C. Jentzsch, “Decentralized Autonomous Organization to Automate Governance,” white Pap., pp. 1–30, 2016. [10] A. Dorri, S. S. Kanhere, and R. Jurdak, “Blockchain in internet of things: Challenges and Solutions,” _arXiv1608.05187 [cs], no. August, 2016._ [11] N. Gupta, A. Jha, and P. Roy, “Adopting Blockchain Technology for Electronic Health Record Interoperability,” 2016. [12] A. Ekblaw, A. Azaria, J. D. Halamka, and A. Lippman, “A Case Study for Blockchain in Healthcare:‘MedRec’ prototype for electronic health records and medical research data,” _Proc. IEEE Open Big Data Conf., pp. 1–_ 13, 2016. [13] W. Stallings, “The Internet of Things: Network and Security Architecture,” Internet Protoc. J., vol. 18, no. 4, pp. 2–24, 2015. [14] A. Torkaman and M. A. Seyyedi, “Analyzing IoT Reference Architecture Models,” _Int. J. Comput. Sci._ _Softw. Eng. ISSN, vol. 5, no. 8, pp. 2409–4285, 2016._ [15] Cisco, “The Internet of Things Reference Model,” _White_ Pap., pp. 1–12, 2014. [16] Nir Kshetri, “Can blockchain Strengthen the Internet of Things?,” IEEE Compu ter So ciet y, no. August, pp. 68– 72, 2017. [17] M. Conoscenti, D. Torino, A. Vetr, D. Torino, and J. C. De Martin, “Peer to Peer for Privacy and Decentralization in the Internet of Things,” in 2017 IEEE/ACM 39th IEEE _International Conference on Software Engineering_ _Companion Peer, 2017, pp. 288–290._ [18] S. Nakamoto, “Bitcoin: A Peer-to-Peer Electronic Cash System,” Www.Bitcoin.Org, p. 9, 2008. [19] T. Ahram, A. Sargolzaei, S. Sargolzaei, J. Daniels, and B. Amaba, “Blockchain technology innovations,” 2017 IEEE _Technol. Eng. Manag. Conf., no. 2016, pp. 137–141, 2017._ [20] M. Conoscenti, A. Vetro, and J. C. De Martin, “Blockchain for the Internet of Things: A systematic literature review,” _2016 IEEE/ACS 13th Int. Conf._ _Comput. Syst. Appl., pp. 1–6, 2016._ [21] A. M. Antonopoulos, _Mastering Bitcoin: Unlocking_ _Digital Cryptocurrencies., M 1st ed. Sebastopol, CA,_ USA: O’Reilly Media, Inc., 2014. [22] A. Dorri, S. S. Kanhere, R. Jurdak, and P. Gauravaram, “Blockchain for IoT security and privacy: The case study of a smart home,” _2017 IEEE Int. Conf. Pervasive_ _Comput. Commun. Work. (PerCom Work., pp. 618–623,_ 2017. [23] V. Daza, R. Di Pietro, I. Klimek, and M. Signorini, “CONNECT: CONtextual NamE disCovery for blockchain-based services in the IoT,” _IEEE Int. Conf._ _Commun., 2017._ [24] H. F. Atlam, A. Alenezi, R. J. Walters, G. B. Wills, and J. Daniel, “Developing an adaptive Risk-based access control model for the Internet of Things,” in _2017 IEEE_ _International Conference on Internet of Things (iThings)_ _and IEEE Green Computing and Communications_ _(GreenCom) and IEEE Cyber, Physical and Social_ _Computing (CPSCom) and IEEE Smart Data (SmartData),_ 2017, no. June, pp. 655–661. [25] A. Boudguiga _et al., “Towards Better Availability and_ _Accountability for IoT Updates by means of a_ Blockchain,” in _2017 IEEE European Symposium on_ _Security and Privacy Workshops (EuroS&PW), 2017, pp._ 50–58. [26] H. F. Atlam, A. Alenezi, A. Alharthi, R. Walters, and G. Wills, “Integration of cloud computing with internet of things: challenges and open issues,” in _2017 IEEE_ _International Conference on Internet of Things (iThings)_ _and IEEE Green Computing and Communications_ _(GreenCom) and IEEE Cyber, Physical and Social_ _Computing (CPSCom) and IEEE Smart Data (SmartData),_ 2017, no. June, pp. 670–675. [27] H. F. Atlam, A. Alenezi, R. K. Hussein, and G. B. Wills, “Validation of an Adaptive Risk-based Access Control Model for the Internet of Things,” I.J. Comput. Netw. Inf. _Secur., no. January, pp. 26–35, 2018._ [28] M. Samaniego and R. Deters, “Blockchain as a Service for IoT,” _2016 IEEE Int. Conf. Internet Things IEEE_ _Green Comput. Commun. IEEE Cyber, Phys. Soc. Comput._ _IEEE Smart Data, pp. 433–436, 2016._ [29] D. Geist, “Using the Bitcoin Blockchain as a Botnet Resilience Mechanism,” 2016. [30] K. Christidis and G. S. Member, “Blockchains and Smart Contracts for the Internet of Things,” IEEE Access, vol. 4, pp. 2292–2303, 2016. [31] S. Huh, S. Cho, and S. Kim, “Managing IoT Devices using Blockchain Platform,” in _The_ _19th_ _IEEE_ _International Conference on Advanced Communications_ _Technology (ICACT 2017), 2017, pp. 464–467._ [32] T. Bocek, B. B. Rodrigues, T. Strasser, and B. Stiller, “Blockchains Everywhere - A Use-case of Blockchains in the Pharma Supply-Chain,” in _2017_ _IFIP/IEEE_ _International_ _Symposium_ _on_ _Integrated_ _Network_ _Management (IM2017):, 2017, pp. 772–777._ [33] A. Alenezi, N. H. N. Zulkipli, H. F. Atlam, R. J. Walters, and G. B. Wills, “The Impact of Cloud Forensic Readiness on Security,” in _Proceedings of the 7th_ _International Conference on Cloud Computing and_ _Services Science (CLOSER 2017), 2017, pp. 511–517._ [34] H. F. Atlam, G. Attiya, and N. El-Fishawy, “Integration of Color and Texture Features in CBIR System,” _Int. J._ ----- _Comput. Appl., vol. 164, no. April, pp. 23–28, 2017._ [35] Diana Asatryan, “4 Challenges to Blockchain Adoption From Fidelity CEO,” 2017. . [36] H. F. Atlam, M. O. Alassafi, A. Alenezi, R. J. Walters, and G. B. Wills, “XACML for Building Access Control Policies in Internet of Things,” in _in Proceedings of the_ _3rd International Conference on Internet of Things, Big_ _Data and Security (IoTBDS 2018), 2018, pp. 1–6._ **Authors’ Profiles** **Hany F. Atlam has born in Menoufia,** Egypt in 1988. He has completed his Bachelor of Engineering and computer science from Faculty of Electronic Engineering, Menoufia University, Egypt in 2011, then completed the master’s degree in computer science from the same university in 2014. He joined the University of Southampton as a Ph.D. student since January 2016. Hany’s now is a lecturer in Faculty of Electronic Engineering, Menoufia University, Egypt and a Ph.D. candidate at the University of Southampton, UK. He has large experiences in networking as he holds international Cisco certifications, Cisco Instructor certifications, and database certifications. He also a member of Institute for Systems and Technologies of Information, Control and Communication (INSTICC), and Institute of Electrical and Electronics Engineers (IEEE). Hany’s research areas include IoT security and privacy, Cloud computing security, Blockchain, Big data, digital forensics, computer networking and image processing. **Ahmed Alenezi a lecturer at Northern** Border University, Saudi Arabia and a Ph.D. candidate at the University of Southampton, UK. Ahmed is interested in multidisciplinary research topics that related to computer science. His research interests include Parallel Computing, Digital forensics, Cloud Forensics, Cloud Security, Internet of Things Forensics and Internet of Things Security. **Madini O. Alassafi born in Saudi Arabia.** Alassafi received Bachler’s degree in Computer Science from King Abdul-Aziz University in Saudi Arabia, 2006, and his master’s degree in Advanced Computer Science from California Lutheran University, Thousand Oaks, USA, 2013. He works as a lecturer at King Abdul-Aziz University, Saudi Arabia. Now he is a Ph.D. candidate at the University of Southampton, UK. His current research interested in multidisciplinary research topics to pertain to computer science, which includes and not limited to Cloud Computing, Security, Risks, Cloud Migration Project Management and Cloud of Things, Security Threats. **Gary B. Wills is an Associate Professor in** Computer Science at the University of Southampton. He graduated from the University of Southampton with an Honours degree in Electromechanical Engineering, and then a PhD in Industrial Hypermedia system. He is a Chartered Engineer, a member of the Institute of Engineering Technology and a Principal Fellow of the Higher Educational Academy. He is also a visiting associate professor at the University of Cape Town and a research professor [at RLabs. Gary’s research projects focus on Secure System](http://www.rlabs.org/) Engineering and applications for industry, medicine and education. **How to cite this paper:** Hany F. Atlam, Ahmed Alenezi, Madini O. Alassafi, Gary B. Wills, "Blockchain with Internet of Things: Benefits, Challenges, and Future Directions", International Journal of Intelligent Systems and Applications(IJISA), Vol.10, No.6, pp.40-48, 2018. DOI: 10.5815/ijisa.2018.06.05 -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.5815/IJISA.2018.06.05?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5815/IJISA.2018.06.05, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GOLD", "url": "http://www.mecs-press.org/ijisa/ijisa-v10-n6/IJISA-V10-N6-5.pdf" }
2,018
[ "Review" ]
true
2018-06-08T00:00:00
[ { "paperId": "d723c3da0c3b55212092afe755a8d2a8bc4a2ba0", "title": "Anonymity" }, { "paperId": "7b76cdb354e120f0473e0aaf61e4d36a932f493f", "title": "XACML for Building Access Control Policies in Internet of Things" }, { "paperId": "4c6f3d83b1e3097c642931a9ea3bd1fedf447e75", "title": "Validation of an adaptive risk-based access control model for the Internet of Things" }, { "paperId": "e8709e2906361ade9064cc605b9c7637bec474a0", "title": "Can Blockchain Strengthen the Internet of Things?" }, { "paperId": "37d9283061bb8057adff53ff4033dd11ccdf2a0c", "title": "Blockchain solutions for big data challenges: A literature review" }, { "paperId": "28924ee0429260d4c90e3c9c61760b974bc4b466", "title": "Integration of Cloud Computing with Internet of Things: Challenges and Open Issues" }, { "paperId": "9092a7802f6e56dd5b6d1be30c8b5588a22e53fe", "title": "Blockchain technology innovations" }, { "paperId": "9ee3c4be30d0e33ed9e43e53ebd7a1ab98e67217", "title": "Developing an Adaptive Risk-Based Access Control Model for the Internet of Things" }, { "paperId": "45fd39f70614062cdfc59309b64e9d48892d993d", "title": "Blockchain Based Distributed Control System for Edge Computing" }, { "paperId": "04b0873fdf91e4ef79090eda55181afe434456c6", "title": "Peer to Peer for Privacy and Decentralization in the Internet of Things" }, { "paperId": "0356360ce4e31a901f5cc48b090af30f56bb3f2d", "title": "Blockchains everywhere - a use-case of blockchains in the pharma supply-chain" }, { "paperId": "33d0d59b44a93a985140a746a9300ebcb843d4a9", "title": "CONNECT: CONtextual NamE disCovery for blockchain-based services in the IoT" }, { "paperId": "7400152e6e475eaa2f88a91182df7991c9519156", "title": "Towards Better Availability and Accountability for IoT Updates by Means of a Blockchain" }, { "paperId": "a2acc206b32850719a6d4903f7a649d84d365033", "title": "Integration of Color and Texture Features in CBIR System" }, { "paperId": "ba022e0f1b3abbb3e74ef84c7f8e3be08c806e77", "title": "An Overview of Risk Estimation Techniques in Risk-based Access Control for the Internet of Things" }, { "paperId": "28fe6a3fab2f2097a6f9aac5ae9799577badf883", "title": "Blockchain for IoT security and privacy: The case study of a smart home" }, { "paperId": "c04cfb8074797c8fdb688da0b64f1dd19c49773a", "title": "IoT and Blockchain Convergence: Benefits and Challenges" }, { "paperId": "631cc57858eb1a94522e0090c6640f6f39ab7e18", "title": "Blockchain as a Service for IoT" }, { "paperId": "451729b3faedea24771ac4aadbd267146688db9b", "title": "Blockchain in internet of things: Challenges and Solutions" }, { "paperId": "c998aeb12b78122ec4143b608b517aef0aa2c821", "title": "Blockchains and Smart Contracts for the Internet of Things" }, { "paperId": "139cfb65d375bba4ca59acc19efb0b7ac99247dc", "title": "CoinParty: Secure Multi-Party Mixing of Bitcoins" }, { "paperId": "148f044225ce7433e5fcf2c214b3bb48d94f37ef", "title": "Mastering Bitcoin: Unlocking Digital Crypto-Currencies" }, { "paperId": "24711b2a7a4dc4d0dad74bbbfeea9140abab047b", "title": "Managing IoT devices using blockchain platform" }, { "paperId": "447c7fb53a6528cabb445209e03e59114e8bfc0b", "title": "The Impact of Cloud Forensic Readiness on Security" }, { "paperId": null, "title": "“4 Challenges to Blockchain Adoption From Fidelity CEO,”" }, { "paperId": "7e07327ce9bdd65791968f59b508e2047fb7d1b9", "title": "Analyzing IoT Reference Architecture Models" }, { "paperId": "3ed0db58a7aec7bafc2aa14ca550031b9f7021d5", "title": "A Case Study for Blockchain in Healthcare : “ MedRec ” prototype for electronic health records and medical research data" }, { "paperId": null, "title": "“Decentralized Autonomous Organization to Automate Governance,”" }, { "paperId": null, "title": "“Adopting Blockchain Technology for Electronic Health Record Interoperability,”" }, { "paperId": null, "title": "“Using the Bitcoin Blockchain as a Botnet Resilience Mechanism,”" }, { "paperId": null, "title": "“The Internet of Things: Network and Security Architecture,”" }, { "paperId": null, "title": "ADEPT: An IoT Practitioner Perspective" }, { "paperId": "539091726bef30c8837dc197e89ae7f5f0d6dbf8", "title": "Overview of the Internet of Things" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": "4ea759dc35b564d4d795c554f407fdb8652b8bed", "title": "That ‘Internet of Things’ Thing" }, { "paperId": null, "title": "Speed : A blockchain transaction is distributed across the network in minutes and will be processed at any time throughout the day" }, { "paperId": null, "title": "Security : Blockchain has the ability to provide a secure network over untrusted parties which is needed in IoT with numerous and heterogeneous devices" }, { "paperId": null, "title": "Lack of skills : The blockchain technology is still new. Therefore, a few people have large knowledge and skills about the blockchain, especially in banking" }, { "paperId": null, "title": "Processing Power and Time : The processing power and time needed to achieve encryption for all the objects included in a blockchain system" } ]
10,813
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0118bd5580dc6d23a0c0b67241c735b25e405e09
[ "Computer Science" ]
0.886177
Communication Requirements and Deployment Challenges of Cloudlets in Smart Grid
0118bd5580dc6d23a0c0b67241c735b25e405e09
International Conference on Smart Communications and Networking
[ { "authorId": "73771375", "name": "Stephen Ugwuanyi" }, { "authorId": "9136586", "name": "Jidapa Hansawangkit" }, { "authorId": "145678109", "name": "Rabia Khan" }, { "authorId": "39158735", "name": "Kinan Ghanem" }, { "authorId": "153758455", "name": "Ross McPherson" }, { "authorId": "143667221", "name": "J. Irvine" } ]
{ "alternate_issns": null, "alternate_names": [ "SmartNets", "Int Conf Smart Commun Netw" ], "alternate_urls": null, "id": "661ae2a4-ae88-4ce1-a7a3-73bb04510f92", "issn": null, "name": "International Conference on Smart Communications and Networking", "type": "conference", "url": null }
Intelligent and distributed power networks are becoming more complex with the addition of cloudlets infrastructure. This paper is the initial output of a Low Latency Edge Containerisation and Virtualisation project at PNDC. This project investigates the performance and communication requirements of deploying edge computing devices to enable low-latency applications. The edge network illustrates how additional intelligent network resources can help realise better-distributed automation and remote configuration in smart grids. With cloudlets, the performance of smart grid networks will be enhanced in terms of low latency, higher bandwidth, and other network requirements. This paper presents cloudlets technology's current state of the art, including network design requirements, implementation techniques, and integration challenges with the legacy power networks. It also explored the main challenges of processing smart grid data by constrained IoT devices. The feasibility of using edge containers to provide low latency communication to several critical end applications in the smart grid could improve the performance of the networks. Furthermore, the key use cases of cloudlets in critical end applications for power utility networks were identified
# q p y Challenges of Cloudlets in Smart Grid 1Stephen Ugwuanyi _Electrical and Electronic Engineering_ _University of Strathclyde_ Glasgow, United Kingdom Stephen.ugwuanyi@strath.ac.uk 2Kinan Ghanem _Power Networks Demonstration Centre_ _University of Strathclyde_ Glasgow, United Kingdom kinan.ghanem@strath.ac.uk 3Jidapa Hansawangkit _Electrical and Electronic Engineering_ _University of Strathclyde_ Glasgow, United Kingdom jidapa.hansawangki@strath.ac.uk 4Ross McPherson _Electrical and Electronic Engineering_ _University of Strathclyde_ Glasgow, United Kingdom ross.mcpherson@strath.ac.uk 5Rabia Khan _Power Networks Demonstration Centre_ University of Strathclyde Glasgow, United Kingdom rabia.khan@strath.ac.uk 6James Irvine _Electrical and Electronic Engineering_ _University of Strathclyde_ Glasgow, United Kingdom j.m.irvine@strath.ac.uk **_Abstract—Intelligent and distributed power networks are_** **becoming more complex with the addition of cloudlets** **infrastructure. This paper is the initial output of a Low Latency** **Edge Containerisation and Virtualisation project at PNDC. This** **project investigates the performance and communication** **requirements of deploying edge computing devices to enable** **low-latency applications. The edge network illustrates how** **additional intelligent network resources can help realise better-** **distributed automation and remote configuration in smart grids.** **With cloudlets, the performance of smart grid networks will be** **enhanced in terms of low latency, higher bandwidth, and other** **network** **requirements.** **This** **paper** **presents** **cloudlets** **technology’s current state of the art, including network design** **requirements, implementation techniques, and integration** **challenges with the legacy power networks. It also explored the** **main challenges of processing smart grid data by constrained** **IoT devices. The feasibility of using edge containers to provide** **low latency communication to several critical end applications** **in the smart grid could improve the performance of the** **networks. Furthermore, the key use cases of cloudlets in critical** **end applications for power utility networks were identified** **_Keywords—Cloudlets,_** **_Containerisation,_** **_Communication_** **_Requirements, Deployment, Security, Smart Grid., Virtualisation_** I. INTRODUCTION Overcoming service availability and delay is one of the deployment challenges of the Internet of Things (IoT) technology in power utility networks without centralised cloud infrastructures. Cloud computing, as defined by the National Institute of Standards and Technology (NIST), is a model for facilitating ubiquitous and on-demand access to a shared pool of network resources in three different service models; Software as a Service (SaaS); Platform as a Service (PaaS); and Infrastructure as a Service (IaaS) [1]. While each service model can be provisioned with network characteristics such as broad network access, on-demand self-service, resource pooling, rapid elasticity, and measured service, they could be used in private or public networks or a combination of both. For instance, PaaS resources can be rented to install Supervisory and Data Acquisition System (SCADA) systems to be shared by distributed generation sources and Distributed Network Operators (DNOs) [2]. IoT-based smart grid networks generate massive data from heterogeneous edge devices that require high computing resources to process the data in the cloud and deliver optimal network performance. Strategically placed cloud resources closer to the edge devices will help to offload the massive dynamic traffic generated by the distributed IoT devices to guarantee high quality of services. It will free the smart grid from Wide Area Network (WAN) related delays, jitter, congestion and network failures [3]. Cloud process and compute time required for such high numbers of IoT devices is a bottleneck in today’s networks without technology such as cloudlets. It is essential in IoT networks to move data processing points closer to the data sources to meet the communication requirements of real-time applications. It will ensure fast response time and reduce the amount of unnecessary data migrated to the centralised data centre. Moreover, power utilities avoid using public clouds for critical applications due to latency and security issues. However, they seek an alternative private cloud infrastructure to process the generated data locally and securely before the fine-grained data is sent to the centralised private cloud for other processing that may involve machine learning and artificial intelligence algorithms. This process will provide computing power support for IoT devices when deployed correctly using adequate communication technology such as 5G cellular networks in good coverage [4]. Reducing network response time in mobile utility IoT applications may be challenging as traffic may not be offloaded optimally. To offload traffic optimally requires many factors, like ensuring that cloudlets are optimally positioned with adequate coverage of reliable and bandwidthefficient communication technology and that the edge devices have enough processing, storage and memory capabilities. An entropy-weight-based proof of concept algorithm found to be optimal, cost-effective and can meet network delay requirements has been proposed to tackle the cloudlets placement problems [5]. Similarly, a dynamic clustering algorithm-based cloudlets deployment is another approach to solving latency issues in cloudlets due to edge device mobility in smart grids [4]. Cloudlets concepts in the smart grid are seen as a data processing approach to moving cloud computing capabilities closer to intelligent field devices and serve limited and localised utility assets like Remote Terminal Units (RTUs) and smart transformers than wider utility IoT resources. As a mini data centre designed to provide cloud computing services to IoT field devices within a close geographical area, it will facilitate edge virtualisation and intelligence. Both enhancements require the proper hardware and software with sufficient processing, memory and real-time operating capability to enable accurate grid data synchronisation for controlled functions, efficiency and the effectiveness of the ----- Cloud infrastructure could be either on-premise or offpremises, but both play essential roles in seamlessly integrating utility assets for easy computation of smart grid data. Cloudlets are the intermediate layers between the cloud infrastructure and utility assets [6]. Introducing a data centre closer to the edge devices would facilitate the virtualisation of substation devices such that edge devices can easily instantiate Virtual Machine (VM) software on the cloudlets. II. RELATED WORK Cloud computing is on-demand storage, data processing and information exchange model for enabling global and continuous access to network resource management systems. In [7], it is seen as a trusted cluster of computers connected to the Internet and designed to deliver cloud computing services to IoT devices within a specific geographical neighbourhood. Cloud computing is one solution to reducing resourceconstrained IoT devices’ impacts on network performance. This can be achieved via VMs, as each connecting user or end application is associated with VM instances created within cloudlets. In smart grid ecosystems, cloudlet is an evolving cloud computing infrastructure used to federate the associated processing of networking logics embedded in the edge and wireless cloud [2], [8]. Its performance is mainly affected by communication and processing factors. Cloudlet performance in a rural area has been investigated and likened to locations of utility assets. In this study [9], cloudlet is noted to reduce the lack of communication infrastructure barrier in hard-toreach locations and make power networks more open, inexpensive, adaptable, and an extensible platform only when implemented with higher compute devices. Cloudlets have been deployed in many sectors, including the smart grid [2]. Cloudlet has been used to reduce the execution end time of workflow applications in metropolitan area networks [10]. Well-deployed cloudlets in the secondary or primary substations will reduce the substation-to-substation communication latency between substation devices in Figure 1. The intermediate mini cloud data centre will ensure efficient and secure communication between the grid entities. As shown in Figure **1, a resource-rich cloudlet will act as an** intermediates layer between the cloud resources and utility assets to deliver time-critical applications with minimal hop path and bandwidth. However, cloudlets have key deployment challenges that must be tackled. These challenges will include questions about how smart grid communication protocols could support virtualisation and integrate with new and existing network resources. It will also involve identifying locations for cloudlets placement with adequate network coverage, compatibility, interoperability, scalability, trust, and security. Notwithstanding that the use of cloudlets in the industry is continuing to grow especially for achieving low latency communication and reducing costs, there are still open research questions on using cloudlets in smart grids. Studies on cloudlets’ performance in real smart grid networks are limited. Navantia industrial AR (IAR) network architecture is designed to leverage cloud, cloudlets and fog computing to deliver traffic-efficient industrial IoT networks [11]. The findings indicate that cloudlet’s response rate outperformed cloud and fog computing at payloads greater than 128 KB. While this value is four times greater than fog computing when many applications were served, fog computing achieved the fastest response rate for small payloads. Both the cloud and applications that decrease network latency response. Similarly, in a cloudlets-based Wireless Local Area Network (WLAN), offloading the MAC layer processing from the access points to the cloudlets allowed flexibility in service provisioning at reduced costs [12]. The capital and operational costs of running a cloudlets-based network in a smart grid will reduce as the network operators could easily implement new services without procuring more expensive equipment. Network Function Virtualisation (NFV) is another aspect of cloudlets that will simplify network management and remote service provisioning, reduce access latency, and make deployment easier to implement. Figure 1. Proposed Cloudlets Supported Smart Grid to Reduced Substation-to-Substation Latency Figure 1 shows a three-level smart grid infrastructure with integrated IoT systems. The IoT devices provide the field measurements transmitted to the cloud centre for processing through the gateway. In the smart grid, collecting field data measurement, communicating, monitoring, and controlling the distrusted end devices are part of the IoT system’s objectives. Field data are better processed in the cloudlets with adequate resources because IoT devices have the lowest computing, storage, and power capabilities to support connectivity solutions for data delivery and analysis. These limitations are responsible for technologies that drive data processing towards the edge to reduce communication latency, especially for time-critical applications. Cloudlets reduce network latency as a localised cloud/data processing point. The unique nature of the power network as an Operational Technology (OT) infrastructure means that connectivity solutions for data transmission between smart grid IoT devices and the cloud centre are not straightforward solutions. It will require open communication frameworks such as OpenFMB for network integration. Hence, deploying cloudlets in the smart grid will allow DNOs to have more network control and quickly provision service functions such as security and privacy. Cloudlets also need to be coordinated during design and implementation. Coordinated cloudlets are described as small clouds in network infrastructures that are interconnected [13], and each server is discoverable, localised and stateless with one or more VM in operation. An end-to-end direct connection through a Software-Defined Wide Area Network (SD-WAN) could facilitate real-time applications for geographically distributed cloudlets. Installing a cloudlet closer to the IoT field devices will open the door for complete local processing and reduces the time to migrate data to the central cloud or data centre. The cloudlet will be able to host virtual access points, which will also avoid the complexity of a physical access point common in legacy networks [12]. ----- As shown in Figure 2, the proposed implementation architecture is a new test setup regime at PNDC in collaboration with DNOs for investing cloudlet’s performance in smart grids. The RTDS Simulator generates IEC 61850 GOOSE and Active Network Management (ANM) packets, whereas SCADA Simulator generates Modbus and DNP3 used to investigate IT/OT convergence edge container that converts field protocols to OPC/UA. Both simulators exchange data with external hardware or software devices in real time through many different communication protocols. Input and output via Ethernet (though standard-compliant data packets) allow the closed-loop testing of digital substations and other non-wires alternatives. The edge container (cloudlet) has virtualisation and real-time operating system capabilities to process IEC 61850, NDP3, and Modbus protocols while communicating with the ANM system. The development stage of the test platform is shown in Figure 2, along with the design requirements needed to develop the IEC 61850 protocol adaptor in the OT/IT convergence testbed. Figure 2. High-Level Cloudnet Test Network Architecture to Facilitate IT/OT Convergence at PNDC. IV. CLOUDLETS REQUIREMENTS IN SMART GRID Power networks consist of generation, transmission, and distribution subsystems that deliver electricity to consumers. Its implementation requires dedicated, secure and reliable technologies for monitoring, communicating and controlling grid assets in a two-way fashion. Cloudlets could help deliver security [14] - [15], computing [7], communication [16], and data storage network requirements in smart grids. Some of these performance specifications are better defined by the DNOs, regulators, policy-makers and vendors. They include: _A._ _Communication Technology_ Connectivity for cloudlets-based critical applications such as teleprotection requires very low latency delivered by LTE or 5G networks to maintain data synchronisation among multiple devices and systems [17]. The frequent exchange of Phasor Measurement Unit (PMU) data and control commands between substations and the control centre requires a robust, reliable, low-latency communication link across the network. The same communication requirements are needed for critical IEC 61850 Sample Values (SV) and GOOSE messages. Cloudlets could be seen as a proper technique for providing low-latency communication if the field measurement data are processed at the network edge [18]. For some critical applications, the required end-to-end latency to enable the functionality of PMU synchrophasor is 100 ms and around 1 ms for critical control commands, which can be challenging to satisfy. In the smart grid, fibre optics, LTE, and 5G will play an important role in completing the path to full digitalisation of power networks. Critical end applications such as synchrophasor data and any routable GOOSE messages will require redundancy to ensure resilience in case of any problem in the connection to the data centre or with another cloudlet. As fully identified in the meshed cloudlets in Figure **5, smart grid resources must be integrated with robust,** reliable, low latency communication technology. With cloudlets, smart grid assets must support network virtualisation technologies such as Network Function Virtualisation (NFV) and Software Defined Network (SDN) to simplify such complex and meshed networks. _C._ _Privacy and Security_ Cloudlets may face privacy and security requirements in smart grids, especially when data is distributed and shared across multiple entities with less computational and storage capabilities. When moving the processing capabilities to the edge, the security approach used at the main data centre could be applied in the cloudlets. Implementing end-to-end encryption at the edge without compromising security and privacy has always been challenging. In [14], a TLS-based secure protocol extension proposed could allow edge functions to process encrypted traffic at the edge of an IoT network. However, our previous study identified that implementing encryption techniques such as TLS and IPsec within the utility assets will increase the data overhead by a significant fold [19]. Encrypting the exchanged data among different systems and devices using strong encryption keys will raise an issue about the bandwidth requirements needed to enable low latency communication [20]. Cloudlets in the OT environment could become challenging to manage from a security perspective. Smart grid networks could be made less secure due to the limited number of Information Technology (IT) security layers and its interoperability issues with the OT security layers in resource-constrained devices. Because cloudlets are more accessible and interfaced with less secure IoT devices, their use in smart grids will require more protection to secure the local data storage systems and the communication link between the cloud, cloudlets and the field devices. Ensuring that cloudlets technology satisfies cyber security standards like NIST/ENA/IEC guidelines and policies, including supporting utility protocols such as Modbus, DNP3, IEC 60870-5-104, 60870-5-101 and IEC61850 securely, reliably and at low costs are essential. Smart grid communication protocols carry a large amount of sensitive time-critical utility data prone to intrusion, subversion, or spoofing attacks. Because cloudlets have the capability to create meshed networks with a shared medium of high availability when deployed in smart grids, their vulnerabilities, threats, and impacts need to be risk assessed. _D._ _Storage_ Smart grid network resources like PMUs and Intelligent Electronic Devices (IEDs) generate crucial time-sensitive data requiring appropriate data storage systems. According to the Utility AMI Working Group, utility data storage systems must be long-term and able to store data within the device securely. In the case of a smart meter, the data must be stored for 45 days [21]. The data must be accessible remotely and with a standard model that allows it to be exchanged between multiple vendors’ equipment. The rise of the Smart Storage ----- reliable storage systems. Cloudlets will enable fast implementation of Virtual Power Plant (VPP) and microgrids for controlling the DERs. It will also help DNOs identify and locate DERs for energy usage measurements and analysis. _E._ _Disaster Recovery_ Cloudlets introduce redundancy to the power networks as critical processing at the edge can be continued through cloudlets when the main cloud infrastructure fails [18]. The capability of the smart grid to function as a standalone in disaster and blackout scenarios is essential to enable basic functions like communicating with the secondary data centre when required. However, cloudlets functionalities may be lost in natural disasters, including processing, storage and power sources. Cloudlet outage must be avoided as critical services could be interrupted, frozen and QoS violated. To analyse the impact of these network characteristics on smart grid performance, we have discussed three different scenarios in which cloudlets architecture could be configured. The first is a multi-access cloudlets architecture shown in Figure **3. Multi-access cloudlets enable the integration of** several smart grid assets and deliver services and computing functions needed by the end applications. With the massive amount of devices at the edge and high variation of dataintensive applications, 5G/6G networks have been noted to meet such demands in the future cloud and communications networks [22]. This will improve application response, enhance outage control, and support new applications in smart grids. Figure 3. Multi-Access Cloudlets Architecture In this paper, we see cloudlets deployment in power networks as a standalone edge processing box within the secondary or primary substations with sufficient power backup, where direct connectivity is provided to restore service outages. As shown in Figure **4, connectivity to the** private cloud infrastructure is not always needed to provide the required services, and data synchronisation with the private cloud could be initiated when cloud connectivity becomes available. Figure 4. High-Level Standalone Cloudlet Topology at the Edge. of a cloudlet s architecture in a smart grid is a fully meshed network, as shown in Figure 5, it must be designed following the European Telecommunication Standards Institute (ETSI) Multi-Access Edge Computing (MEC) standard [22]. The operation of fully meshed cloudlets in power networks requires a private network fully operational and managed by the power utilities. Accurate data synchronisation and the number of hops between cloudlets and IoT field devices, connectivity and security are a few challenges that might be seen in a fully meshed scenario. Figure 5. Fully Meshed Multi-Access Cloudlet Architecture V. CLOUDLETS CHALLENGES IN POWER NETWORKS Significant benefits can be obtained by deploying cloudlets in critical infrastructures such as power networks, as they can give the DNOs full control over their distributed infrastructure in terms of implementation, management and security. However, its deployment faces the following challenges: _A._ _Lack of Skilled Professionals_ Cloudlet is a cloud computing technology which can deliver hosted services to IoT devices over a network. It is advantageous to deploy cloudlets on the power network’s edge compared to the public cloud, where the enterprise cloud operators manage it. With cloudlet technology at the edge of power networks, DNOs can easily deploy and manage their infrastructure if adequate skilled network professionals are available. The lack of skilled professionals is challenging to many existing DNOs, and relying on a third party to operate the cloudlet may not apply to some DNOs for security and privacy reasons. Notwithstanding, skilled IT professionals will be needed to define the most appropriate cloudlet technology and its operational requirements in power networks. Recent developments are expanding the use of OpenFMB and OpenADR to meet the needs of utilities, but their implementation requires skilled professionals. _B._ _Communication Frameworks Integration_ One challenge facing utility operators is communicating with more distributed energy resources (PVs, batteries, EVs, etc.) without improving the network architecture. Open Field Message Bus (OpenFMB) and Open Automated Demand Response Communications Specification (OpenADR) are two widely used open communication frameworks in smart grids for network integration [23]. OpenFMB is a framework that functions with new and existing standards, such as the IEC 61968 for distributed edge intelligence designed to drive interoperability and facilitate data exchange between field devices. It has an agile and evolving architecture that is flexible enough to handle data models and publish/subscribe protocols like MQTT and AMQP. OpenADR is designed to ----- continuous dynamic price signals such as hourly day-ahead or day-of real-time pricing. Today, utilities worldwide are investigating OpenADR to manage the growing demand for electricity and the peak capacity of electric systems. Demandside resource aggregation by Pearlstone Energy and National grid [24] and Project ELBE [25] are good examples. _C._ _Data Synchronisation Challenges_ Smart grid networks use precision timing of grid assets’ information to manage and maintain the operation of the power network. Connecting IEDs to a cloudlet requires precise synchronisation of various data types and several sources of measurement. Synchronising data measurements is crucial for critical end applications such as synchrophasors and protection systems. The required synchronisation accuracy varies based on the criticality of the end applications. Achieving local synchronisation is easier than remote synchronisation, as it is easier to coordinate precisely with other systems’ components in real time with fewer errors and duplications. Moreover, ensuring synchronisation among multiple cloudlets is significant for any future development of the cloudlets in the power networks. Losing such synchronisation limits cloudlet usage in power networks and creates a less efficient power network management system that performs below expectations. Measurements such as outputs from synchrophasors (Sampled Values (SV) and GOOSE) are synchronised with precise timing stamps. Deviations from the synchronisation could create disruption and instability in the power networks. In today’s networks, DNOs need accurate time synchronisation in phasor measurement units, merging units and IEDs to coordinate the electrical grid and monitor protection functions. Deploying cloudlets in smart grids requires precise time synchronisation to utilise the cloudlet capability fully. _D._ _Systems Integration and Legacy Challenge_ Legacy systems in power networks may not directly communicate with new utility IoT devices without protocol conversions. These devices could create several issues affecting the power grid’s digitalisation. Legacy hardware is seen as a severe bottleneck to enhancing the future operating capacity of power networks. Upgrading such field devices may not be the right solution for digitalisation, where some old devices may not be upgradeable. Assuming that some of the distributed assets can be upgraded, the time and cost of procuring and installing new devices will be saved. Systems integration is another challenge for DNOs to overcome during the digital transition period. Dealing with too many different protocols and the lack of interoperability will add more complexity to the network architecture that the DNO’s whole system could affect cloudlets integration and field implementations. _E._ _Connectivity and Power_ A significant part of the distributed power networks is located in hard-to-reach areas where affording simple connectivity can be challenging. Lack of connectivity does not just limit the ability to deploy cloudlet but also slows the transition into a smarter digital grid. This is a common issue for many power utilities across the world. Another point to consider is that cuts, power outages and blackouts. Ensuring a sufficient source of backup power (i.e., from battery storage for instant or renewable energy sources) will allow the power networks to rely on the communication network to bring the electricity back in case of an unexpected power loss or black start scenario. Full backup power will be required to enable the full functionality of cloudlets. Reliable backup power for distributed cloudlets is a must to maintain any mission-critical applications. A real concern in the existing power networks is the proprietary interfaces and protocols operated by legacy software. This will significantly affect any future deployment of intelligence at the edge. Converting several field protocols into a unified platform transparent to cloudlet can be seen as a tool to mitigate such challenges. However, there is a need to check the power system performance in terms of data handling and data polling mechanisms. _F._ _Security_ Security and the trust environment used for data storage is a prime issue for cloudlets integration in smart grid. This is because compromised cloudlets cannot attain the missioncritical requirements of power networks. An example of a physical security challenge is protecting a cluster of digital substations connected using cloudlets, where they could share and exchange data locally without involving the centralised data centre. Data storage in the distributed local cloudlets could make them vulnerable to physical attack, as the location of some cloudlet boxes in the rural areas will make it easier for physical access. Cloudlet’s physical proximity to the edge is also very essential to achieving end-to-end response time, low-latency, one-hop, high bandwidth wireless access to the cloud [7]. Physical security is a less explored area in cloudlets, according to a study that investigated cloudlets deployment options in rural and remote areas to improve service availability and support community-based local services during network or power outages [9]. Collaboration Intrusion and Detection System proposed for incorporating security, data sharing, and interruption discovery in cloudlets to protect network privacy and the distribution of cloud and cloud medical applications could offer the needed security for smart grid [15]. To deploy cloudlets in the smart grid, security is one factor of high interest to the DNOs. The requirement is that cloudlets have to be remotely managed and provisioned adequately to enhance security and performance [13]. As new security measures can be provided on-demand and customisable through NFV, security functions like next-generation firewalls, gateways, and access policies could easily be implemented. A good example is the proof of concept described in [26] that would allow device-to-device and device-to-infrastructure communication in a cloudletssupported network and can ensure the reliability, security and privacy of peer-to-peer communication in an intelligent transportation system. VI. BENEFITS OF DEPLOYING CLOUDLETS IN POWER NETWORKS As an emerging technology, the use of cloudlets in smart grids is accelerating daily and has several benefits, such as rapid response, disaster recovery, outage control, and new technologies [18]. The edge devices and the centralised cloud introduce intelligence to support new utility technology. This ----- security. It can also open the door for more real-time applications at the edge, where low latency distributed functions and applications are needed. Cloudlets technology benefits various sections of the power network. In the primary substations, the data generated from the distributed IoT devices (i.e., distributed IEDs such as Merging Units (MUs) and protections relays) could be processed locally, thereby helping to meet the strict latency requirements for different real-time applications and at the same time ensure that the generated data are kept on-premise securely. It could also facilitate using real-time applications such as Virtual Reality (VR), Augmented Reality (AR), and live machine learning models at the edge. The cloudlets at the primary substation should have more processing capabilities than those installed at the secondary substation because the required processing capability is higher at the primary substation level rather than at the secondary substation level. This will allow sychrophasors, SV and GOOSE messages to be processed quicker than MMS and SCADA messages generated at the secondary substation level. Another key benefit for energy utilities from implementing cloudlets in smart grids is horizontal and vertical network extendibility. Cloudlets allow the smart grid utilities to timely respond to distribution, generation, market or regulatory changes across various services. Implementing cloudlets with rich computational capability will help deploy several security approaches and techniques between the cloudlet and the end devices as larger key sizes will be supported. VII. POSSIBLE FUTURE RESEARCH The next phase of this low latency edge container and virtualisation project at PNDC will be testing and data analysis to demonstrate how such intelligence at the edge could help improve the performance of the power networks communications in terms of latency and bandwidth to ensure the reliability and resilience of smart grid networks. The cost implications of deploying cloudlets in a secondary substation will also be considered. IX. CONCLUSION This paper has evaluated how the intelligence at the edge of a smart grid can help achieve network performance improvements in distributed automation and remote configuration of smart grid assets. We, therefore, conclude that deploying cloudlets in a smart grid comes with challenges such as systems integration, connectivity, security and the absence of standards-based field bus protocol to enable the interoperability of distributed field devices and data exchange. The challenges are summarised as service description, management and orchestration, monitoring and optimisation, VM placement, and elasticity and scalability-related problems similar to challenges identified in [12]. With the possibility of implementing OpenFMB and OpenADR standards inside the cloudlets, devices from different vendors and utilities will be able to interoperate directly and exchange data. Additionally, if configured correctly, the cloudlets will process data locally and respond to any field requests, facilitating remote-oriented functions that are very useful in an unexpected event or black start scenario. ACKNOWLEDGMENT The authors acknowledge the contributions of PNDC tier 1 members (mainly - Scottish Power Energy Networks, Scottish and Southern Electricity Networks, UK Power REFERENCES [1] P. Mell and T. Grance, “The NIST Definition of Cloud Computing Recommendations of the National Institute of Standards and Technology,” 2011. [2] M. Muzakkir Hussain, M. Saad Alam, and M. M. Sufyan Beg, “Fog Computing for Smart Grid Transition: Requirements, Prospects, Status Quos, and Challenges,” EAI/Springer Innov. Commun. Comput., pp. 47–61, 2021. [3] S. Bouzefrane, A. F. B. Mostefa, F. Houacine, and H. Cagnon, “Cloudlets authentication in nfc-based mobile computing,” Proc. - 2nd IEEE Int. Conf. _Mob. Cloud Comput. Serv. Eng. MobileCloud 2014, pp. 267–272, 2014._ [4] X. Jin, F. Gao, Z. Wang, and Y. Chen, “Optimal deployment of mobile cloudlets for mobile applications in edge computing,” J. Supercomput., vol. 78, no. 6, pp. 7888–7907, Apr. 2022. [5] C. Guo et al., “Optimal Placement of Cloudlets Considering Electric Power Communication Network and Renewable Energy Resource,” _Proc. - 4th_ _IEEE Int. Conf. Smart Cloud, SmartCloud 2019 3rd Int. Symp. Reinf. Learn._ _ISRL 2019, pp. 199–203, Dec. 2019._ [6] I. Stojmenovic, “Fog computing: A cloud to the ground support for smart things and machine-to-machine networks,” _2014 Australas. Telecommun._ _Networks Appl. Conf. ATNAC 2014, pp. 117–122, Jan. 2015._ [7] M. Satyanarayanan, P. Bahl, R. Cáceres, and N. Davies, “The case for VMbased cloudlets in mobile computing,” IEEE Pervasive Comput., vol. 8, no. 4, pp. 14–23, Oct. 2009. [8] S. Mehmi, H. K. Verma, and A. L. Sangal, “Comparative Analysis of Cloudlet Completion Time in Time and Space Shared Allocation Policies During Attack on Smart Grid Cloud,” Procedia Comput. Sci., vol. 94, pp. 435–440, Jan. 2016. [9] S. Helmer Claus Pahl Julian Sanin Lorenzo Miori Stefan Brocanelli Filippo Cardano Daniele Gadler Daniel Morandini Alessandro Piccoli Saifur Salam Alam Mahabub Sharear Angelo Ventura, P. Abrahamsson, and T. Daniel Oyetoyan, “Bringing the Cloud to Rural and Remote Areas via Cloudlets,” 2016. [10] X. Zhao, C. Lin, and J. Zhang, “Cloudlet deployment for workflow applications in a mobile edge computing-wireless metropolitan area network,” Peer-to-Peer Netw. Appl., vol. 15, no. 1, pp. 739–750, Jan. 2022. [11] T. M. Fernández-Caramés, P. Fraga-Lamas, M. Suárez-Albela, and M. Vilar-Montesinos, “A Fog Computing and Cloudlet Based Augmented Reality System for the Industry 4.0 Shipyard,” 2018. [12] F. Ben Jemaa, G. Pujolle, and M. Pariente, “Cloudlet- and NFV-based carrier Wi-Fi architecture for a wider range of services,” _Ann. des Telecommun._ _Telecommun., vol. 71, no. 11–12, pp. 617–624, Dec. 2016._ [13] A. Alsaleh, “Can cloudlet coordination support cloud computing infrastructure?,” J. Cloud Comput., vol. 7, no. 1, pp. 1–12, Dec. 2018. [14] K. Bhardwaj, M.-W. Shih, A. Gavrilovska, T. Kim, and C. Song, “SPX: Preserving End-to-End Security for Edge Computing,” Sep. 2018. [15] M. P. Reddy, A. M. F. Anwar, A. Sahithi, and A. K. Shravani, “Data Security and Vulnerability Prevention for Cloudlet-Based Medical Data Sharing,” _Proc. 5th Int. Conf. Electron. Commun. Aerosp. Technol. ICECA 2021, pp._ 1477–1481, 2021. [16] M. Rihan and M. Rihan, “Applications and Requirements of Smart Grid,” pp. 47–79, 2019. [17] K. Ghanem, S. Ugwuanyi, R. Asif, and J. Irvine, “Challenges and Promises of 5G for Smart Grid Teleprotection Applications,” _2021 Int. Symp._ _Networks, Comput. Commun. ISNCC 2021, 2021._ [18] M. Babar, M. S. Khan, F. Ali, M. Imran, and M. Shoaib, “Cloudlet Computing: Recent Advances, Taxonomy, and Challenges,” IEEE Access, vol. 9, pp. 29609–29622, 2021. [19] K. Ghanem, J. Hansawangkit, R. Asif, S. Ugwuanyi, R. McPherson, and J. Irvine, “Bandwidth efficient secure authentication and encryption techniques on IEC-60870-5-104 for remote outstations,” _2021 Int. Conf._ _Smart Appl. Commun. Networking, SmartNets 2021, Sep. 2021._ [20] K. Ghanem, R. Asif, S. Ugwuanyi, and J. Irvine, “Bandwidth and security requirements for smart grid,” in _IEEE PES Innovative Smart Grid_ _Technologies Conference Europe, 2020, vol. 2020-Octob._ [21] T. Sato _et al., “Smart Grid Standards: Specifications, Requirements, and_ Technologies,” Smart Grid Stand. Specif. Requir. Technol., pp. 1–463, Feb. 2015. [22] T. Koketsurodrigues, J. Liu, and N. Kato, “Offloading Decision for Mobile Multi-Access Edge Computing in a Multi-Tiered 6G Network,” IEEE Trans. _Emerg. Top. Comput., 2021._ [23] S. Laval, “Duke Energy Emerging Technology Office Open Field Message Bus (OpenFMB): Enabling Distributed Intelligence,” 2019. [24] OpenADR, “Demand-Side Resource Aggregation.” 2020. [25] H. Energie and S. Hamburg, “OPENADR EUROPEAN CASE STUDY PROJECT ELBE.” 2020. [26] M. Gupta, J. Benson, F. Patwa, and R. Sandhu, “Secure V2V and V2I Communication in Intelligent Transportation using Cloudlets,” IEEE Trans. _Serv. Comput., pp. 1–1, Sep. 2020._ -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/SmartNets55823.2022.9993993?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/SmartNets55823.2022.9993993, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://strathprints.strath.ac.uk/82542/1/Ugwuanyi_etal_ICSACN_2022_Communication_requirements_and_deployment_challenges_of_cloudlets_in_smart_grid.pdf" }
2,022
[ "JournalArticle", "Conference" ]
true
2022-11-29T00:00:00
[]
8,932
en
[ { "category": "Environmental Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/011c4c415a99465d1d643d4f3f5371f3d885637c
[]
0.855318
Edge AI and Blockchain for Smart Sustainable Cities: Promise and Potential
011c4c415a99465d1d643d4f3f5371f3d885637c
Sustainability
[ { "authorId": "48527054", "name": "E. Badidi" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "http://mdpi.com/journal/sustainability", "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-172127" ], "id": "8775599f-4f9a-45f0-900e-7f4de68e6843", "issn": "2071-1050", "name": "Sustainability", "type": "journal", "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-172127" }
Modern cities worldwide are undergoing radical changes to foster a clean, sustainable and secure environment, install smart infrastructures, deliver intelligent services to residents, and facilitate access for vulnerable groups. The adoption of new technologies is at the heart of implementing many initiatives to address critical concerns in urban mobility, healthcare, water management, clean energy production and consumption, energy saving, housing, safety, and accessibility. Given the advancements in sensing and communication technologies over the past few decades, exploring the adoption of recent and innovative technologies is critical to addressing these concerns and making cities more innovative, sustainable, and safer. This article provides a broad understanding of the current urban challenges faced by smart cities. It highlights two new technological advances, edge artificial intelligence (edge AI) and Blockchain, and analyzes their transformative potential to make our cities smarter. In addition, it explores the multiple uses of edge AI and Blockchain technologies in the fields of smart mobility and smart energy and reviews relevant research efforts in these two critical areas of modern smart cities. It highlights the various algorithms to handle vehicle detection, counting, speed identification to address the problem of traffic congestion and the different use-cases of Blockchain in terms of trustworthy communications and trading between vehicles and smart energy trading. This review paper is expected to serve as a guideline for future research on adopting edge AI and Blockchain in other smart city domains.
## sustainability _Review_ # Edge AI and Blockchain for Smart Sustainable Cities: Promise and Potential **Elarbi Badidi** Department of Computer Science and Software Engineering, College of Information Technology, UAE University, Al-Ain P.O. Box 15551, United Arab Emirates; ebadidi@uaeu.ac.ae; Tel.: +971-3-713-5552 [����������](https://www.mdpi.com/article/10.3390/su14137609?type=check_update&version=1) **�������** **Citation: Badidi, E. Edge AI and** Blockchain for Smart Sustainable Cities: Promise and Potential. _Sustainability 2022, 14, 7609._ [https://doi.org/10.3390/](https://doi.org/10.3390/su14137609) [su14137609](https://doi.org/10.3390/su14137609) Academic Editors: Miguel de Simón-Martín, Stefano Bracco, Enrique Rosales-Asensio, Alberto González-Martínez and Marc A. Rosen Received: 28 March 2022 Accepted: 18 June 2022 Published: 22 June 2022 **Publisher’s Note: MDPI stays neutral** with regard to jurisdictional claims in published maps and institutional affil iations. **Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **Abstract: Modern cities worldwide are undergoing radical changes to foster a clean, sustainable and** secure environment, install smart infrastructures, deliver intelligent services to residents, and facilitate access for vulnerable groups. The adoption of new technologies is at the heart of implementing many initiatives to address critical concerns in urban mobility, healthcare, water management, clean energy production and consumption, energy saving, housing, safety, and accessibility. Given the advancements in sensing and communication technologies over the past few decades, exploring the adoption of recent and innovative technologies is critical to addressing these concerns and making cities more innovative, sustainable, and safer. This article provides a broad understanding of the current urban challenges faced by smart cities. It highlights two new technological advances, edge artificial intelligence (edge AI) and Blockchain, and analyzes their transformative potential to make our cities smarter. In addition, it explores the multiple uses of edge AI and Blockchain technologies in the fields of smart mobility and smart energy and reviews relevant research efforts in these two critical areas of modern smart cities. It highlights the various algorithms to handle vehicle detection, counting, speed identification to address the problem of traffic congestion and the different use-cases of Blockchain in terms of trustworthy communications and trading between vehicles and smart energy trading. This review paper is expected to serve as a guideline for future research on adopting edge AI and Blockchain in other smart city domains. **Keywords: edge computing; edge intelligence; Blockchain; smart grids; smart mobility; smart energy** **1. Introduction** Many countries have created strategies to transform their cities into smart cities to exploit the opportunities arising from urbanization. Smart cities enable operational efficiencies, maximize environmental sustainability, and develop new services for citizens. For example, the United Arab Emirates has launched its initiative to transform its cities into smart cities. The UAE government has also outlined its overall Blockchain strategy for increased security, immutability, resilience, and transparency. With the climate change issues that have surfaced in the last few years, cities and civil society are increasingly demanding a more sustainable future for their citizens and communities [1,2]. The long-term sustainability of cities requires new, innovative, and disruptive solutions and services that are good for people, the planet, and businesses [3]. Building sustainable cities and environments will not be possible without the right technologies to digitize all city and business processes and obtain and share insights from data [4]. The advancements in sensing and communication technologies, the proliferation of mobile devices, and the widespread use of social media networks have resulted in an exponential growth in the information generated and exchanged. The phenomenon of big data refers to this exponential growth in data volume. It is made up of a set of technologies and algorithms that allow processing massive amounts of data in real time to derive insights from it. The processed information and the resulting insights are made available to decision-makers. Therefore, the reliability of the data is of the utmost importance to permit their exchange and facilitate transactions between businesses. ----- _Sustainability 2022, 14, 7609_ 2 of 30 Billions of edge devices are connected to the Internet and generate zettabytes of data. Extracting value from these massive volumes of data at the required speed of the applications remains the main problem to be solved [5]. For many applications, the processing power offered by cloud computing is often used to process data. However, sending data to cloud servers for processing reveals limitations due to increased communication delays and network bandwidth consumption. Therefore, using cloud computing is not the best solution for real-time and latency-sensitive applications [6–8]. There is a growing trend towards using edge and fog computing to process data and extract value for these latency-sensitive applications. The use of streaming data analytics, machine learning, and deep learning for data processing at the edge resulted in the emergence of a new interdisciplinary technology known as edge AI that enables distributed intelligence with edge devices [9,10]. Research on edge AI and commercial solutions of this new technology are still relatively new. The execution of transactions generally depends on many intermediaries who authenticate the information exchanged to establish “trust” between the parties in the transaction. A typical example is banking, where banks are responsible for validating financial transactions, and building trust between the parties in the transaction [11]. The essence of trusted intermediaries, such as banks, notaries, lawyers, and the government, is to facilitate a transaction that does not force the parties to trust each other. In today’s digital age, reliance on these trusted intermediaries is just the result of a fundamental “lack of faith.” The recent years have witnessed the emergence of Blockchain technology to address this issue of trust [12–14]. A blockchain creates a source of truth that allows peer-to-peer (P2P) transactions to get rid of the need for trusted intermediaries. Its distributed ledger securely stores transaction information across multiple computer systems on the blockchain. Each block in the chain contains information concerning several transactions. Each time a new transaction occurs between two peers on the blockchain network, the ledger of each participant appends a record of that transaction with a hash, which is an immutable cryptographic signature. A change in a block of a chain means tampering with the block. To corrupt a blockchain system, hackers would have to change every block in the chain, and in all versions of the chain distributed across the blockchain network [15]. Blockchain is poised to revolutionize the way businesses, as well as governments, conduct all types of transactions [16]. It will significantly impact everyone (logistics, industry, government, banking, real estate, health, education, and citizen services). Blockchain technology has the potential to improve government services, streamline government processes and provide secure yet efficient information sharing [17,18]. Moreover, by using Blockchain technology, governments can finally offer different services, eliminate bureaucracy and the lack of transparency, prevent tax evasion and reduce waste. _1.1. Contributions_ Although edge computing and blockchain have been extensively studied in the literature, very few works survey the integration of edge AI and blockchain in smart cities. This article reviews recent research efforts on edge AI and blockchain for enabling intelligent and secure edge applications and networks in two fundamental areas of smart cities—smart mobility and smart energy. Beginning with an introduction to edge AI and blockchain, we then review research efforts to integrate these two emerging technologies, including training learning models at the edge, security, privacy, scalability, and model sharing. Mainly, we provide a survey on the use of edge AI in various applications in smart mobility, such as traffic monitoring and management in intelligent transport systems, and smart energy, such as optimized energy management in smart buildings, green energy management, and energy efficiency in smart cities. Furthermore, we review recent research efforts on the use of Blockchain in various applications in smart mobility, including distributed credential management, reputation systems, key and trust management, and smart energy, including distributed energy management and energy trading. Possible research challenges and future directions are also outlined. The key contributions of this article are highlighted as follows: ----- _Sustainability 2022, 14, 7609_ 3 of 30 1. It provides an overview of edge AI and blockchain fundamentals. 2. It analyzes the opportunities brought by edge AI in smart mobility and smart energy. 3. It analyzes the opportunities brought by Blockchain in smart mobility and smart energy. 4. It reviews some efforts to integrate these two emerging technologies in the context of smart cities. 5. Finally, it outlines key open research issues and future directions toward the full realization of edge AI and Blockchain in smart cities. For the reader’s convenience, the studies discussed in this review are shown in Figure 1. **Figure 1. Classification of the studies of this review.** _1.2. Structure of the Review_ The remainder of this review is organized as follows: Section 2 unfolds the challenges facing smart cities. Sections 3 and 4 present the fundamentals of edge AI, federated learning, and Blockchain technology, and describe their potential to support smart city operations. The methodology used in this review is described in Section 5. Section 6 describes the transformative potential and applications of edge AI and Blockchain in two vital areas of smart cities, smart mobility and smart energy. Section 7 highlights some efforts showing the convergence of these two technologies. The open research issues and future directions are highlighted in Section 8. Finally, Section 9 concludes this review. **2. Smart City Systems and Key Challenges** As the world population grows, small and large cities are witnessing large migratory waves that pressure local governments and officials to deal with many social issues. These issues essentially concern ensuring a steady supply of water and electricity, providing appropriate healthcare services for all citizens, building and maintaining road infrastructure, providing adequate public transportation, ensuring security and safety throughout the city, and offering adequate education services [19]. The future of cities looks bright as many local governments start to build on smart city initiatives and embrace new digital technologies and innovations to tackle all of these issues, maximize the use of resources, provide a better quality of life for residents and a favorable investment climate for business [20,21]. For companies, smart city initiatives offer ----- _Sustainability 2022, 14, 7609_ 4 of 30 many innovation opportunities to develop new services and provide smart solutions for the cities. The vast amounts of data obtained by smart city systems and advancements in data stream processing, machine learning, and artificial intelligence enable entrepreneurs to develop new smart solutions and new business models [22]. Smart cities such as Dubai, Barcelona, Amsterdam, Singapore, New York, and Stockholm, to name a few, are enticing other cities to jump on the bandwagon [23]. Smart cities are complex entities that integrate various systems to support the human life cycle. These systems include smart healthcare, smart transportation, smart manufacturing, smart buildings, smart energy, and smart farming, among others. _2.1. Smart Healthcare_ Smart healthcare is a set of technologies that are harnessed to actively manage healthcare data and respond to the needs of the medical ecosystem intelligently to increase longevity and improve the quality of life for citizens. These technologies include mobile devices, Internet of Things (IoT) devices, and mobile Internet, which enable dynamic access to information, connecting people, materials, and health-related institutions. Smart healthcare aims to foster interaction between all entities in health care, including hospitals, pharmacies, healthcare insurers, help them make informed decisions, ensure that participants have access to the services they need, and facilitate the rational allocation of resources [24,25]. _2.2. Smart Transportation_ With the emergence of intelligent transportation systems, the proliferation of IoT-based solutions, and advances in artificial intelligence, smart cities are entering a new era of a development called smart transportation. Smart city traffic management and smart transportation are revolutionizing the way cities approach mobility and emergency response while solving traffic problems by reducing congestion and the number of accidents on the streets and roads of cities [26,27]. Smart transportation relies on the deployment and use of sensors, advanced communication technologies, high-speed networks, and automation [28]. _2.3. Smart Manufacturing_ Smart manufacturing is a technology-driven approach for monitoring the production process using machines connected to the Internet. Its main goal is to present opportunities for automating operations using data analytics to boost manufacturing and energy efficiency, enhance labor security, and reduce environmental pollution levels [29]. Smart manufacturing deployments involve integrating IoT devices into manufacturing machinery to collect operational status and performance data. In addition, many technologies are being used to help enable smart manufacturing, including data streams processing, edge and fog computing, artificial intelligence, robotics, driverless vehicles, blockchain, and digital twins [30,31]. _2.4. Smart Buildings_ Smart buildings are buildings in the tertiary sector or residential buildings for which high-tech tools, such as sensors and sophisticated control systems, make it possible to adapt the settings according to the needs of the occupants [32]. The proliferation of new information and communication technologies now makes it possible to considerably improve our living environment by managing and controlling lighting, ventilation, and air conditioning, in short, the entire infrastructure of a modern building. The implementation of intelligent buildings brings more comfort and convenience to its residents, reduces energy consumption, and mitigates our negative impact on the environment. _2.5. Smart Energy Systems_ Smart energy systems represent one of the most attractive smart city opportunities. Unlike smart grids, which primarily focus on the electricity sector, smart energy systems focus on the comprehensive integration of more sectors, including electricity, cooling, ----- _Sustainability 2022, 14, 7609_ 5 of 30 heating, buildings, manufacturing, and transportation. They aim to transform existing solutions into future renewable and sustainable energy solutions [33]. _2.6. Smart Farming_ Smart farming is an emerging concept in modern agriculture that refers to managing farms using digital technologies such as IoT, soil scanning, drones, robots, edge and cloud data management solutions, and AI [34,35]. It aims to increase the quantity and improve the quality of crops and agricultural products while optimizing the human labor required for production. When equipped with these technologies, farmers can remotely monitor crops and field conditions without going into the field. In addition, they will be able to make strategic decisions for their farms based on data collected from various devices. Despite the promising and potential benefits that digital technologies bring to smart cities, there are many challenges in the way of a successful digital transformation [36]. These challenges mainly relate to the aging infrastructure, which hampers the development of many cities, security and privacy concerns with the proliferation of digital technologies, and social inclusion, which requires the design of solutions that address all categories of citizens and not only tech-savvy people. Addressing these challenges and concerns requires the use of new technologies and the development of new data-driven urban planning methods that challenge traditional models of urban development. The technology and innovative spirit of the new generation of entrepreneurs are the main catalysts for smart cities to be sustainable, safer, and more livable. These technologies and innovations are dramatically changing the way residents, businesses and government entities interact with each other for the benefit of all. Two promising technologies that are starting to make their way into several smart city projects are Blockchain and edge AI, which can potentially disrupt many of the areas above related to smart cities. They can make the various smart city operations and initiatives safer, transparent, efficient, smart, and resilient, resulting in more efficient and productive cities. **3. Edge AI and Federated Learning to Support Smart Cities** _3.1. Edge AI Overview_ Edge computing, sometimes referred to as IoT, is proliferating and is becoming an essential component in most business strategies over the last few years [5,37–39]. IoT devices, sensors, and smartphones transform many businesses from top to bottom. Furthermore, the emergence of artificial intelligence has been phenomenally stunning in its ability to impact the operations at the network edge. Increased computing power at the edge combined with the light deployment of machine learning and deep learning help make edge devices extremely smart [10,40]. Edge AI enables devices to deliver real-time insights and predictive analytics without sending data to remote cloud servers. Many businesses are now taking advantage of this by deploying intelligent solutions in production. With the various industrial IoT devices deployed in modern factories, manufacturers can be alerted with issues in their supply chain and proactively avoid unplanned downtime [41]. Additionally, a small device on a street radar can now instantly recognize a car that is speeding, the passengers in the vehicle, and whether the driver has a license or not [42]. Artificial intelligence (AI) with pre-trained models has the potential to empower smart cities by permitting decision-makers to make informed decisions, which will benefit both the city and citizens [43]. For instance, many smart city sectors will benefit from two typical vision-based image processing tasks, image classification and object detection, which arise in many edge-based AI applications [44–46]. AI continues to enter new segments with great promise at a high rate. Currently, digital industries such as finance, retail, advertising, and multimedia have been the sectors that have exploited AI the most. AI has created real value in these fields. However, the significant and vital problems in several other areas remain unresolved. The solution to the problems of cities concerning transport, energy and water supply, citizen security, healthcare, and many others is to replace or upgrade old and ineffective technologies. ----- _Sustainability 2022, 14, 7609_ 6 of 30 New and AI-driven technologies have the potential to enable efficient transport systems, clean energy, and efficient health systems and industry [47]. A critical element in these areas is introducing and deploying intelligence “at the network edge” of high-speed and broadband networks. The edge is the bulk of our world at present. Bringing intelligence to the edge means that even the smallest devices deployed everywhere are capable of detecting, learning from, and reacting to their environments. AI enables, for example, devices on certain streets or public spaces in the city to make higher-level decisions, act autonomously, and report significant flaws or improvements to affected users or the cloud. Edge AI means that AI algorithms are executed locally on a hardware edge device [48,49]. The AI device can process its local data and make decisions independently without requiring a connection to function correctly. The device must have sensors connected to a small microcontroller unit (MCU) to use edge AI. The MCU is loaded with specific machine learning models that have been pre-trained on certain typical scenarios that the device will encounter. The learning process can also be continuous, allowing the device to learn as soon as it faces new situations. The AI reaction can be a physical actuation on the device’s immediate environment or a notification to a specific user or the cloud for further analysis and assistance. Recently, special-purpose hardware has emerged to accelerate specific compute- or I/O-intensive operations at the edge. These edge hardware accelerators include Google’s edge Tensor Processing Unit (TPU) [50,51], Nvidia’s Jetson Nano and TX2 edge Graphical Processing Units (GPUs) [52,53], Intel’s Movidius Vision Processing Unit (VPU) [54], and Apple’s Neural Engine, which have emerged recently. They are explicitly designed for edge computing to support edge AI applications such as visual and speech analytics, face recognition, object detection, and deep learning inference. Edge computing and edge AI encompass operations such as data collection, parsing, aggregation, and forwarding, as well as rich and advanced analytics that involve machine learning and event processing and actions at the edge. Edge AI will enable real-time operations, including data creation, decision making, and reaction time in milliseconds. These operations are essential for monitoring public spaces with crowds of people, self-driving cars, robots, monitoring machines in a factory, and many other areas. Edge AI will reduce data communication costs and power consumption as edge devices process data locally and transmit fewer data to the cloud, improving battery life, which is extremely important. Smart cities are ideal for the use of edge computing and edge AI. Indeed, sensors and actuators can receive commands based on local decisions without waiting for decisions made in another distant place. Cities can use edge computing for video surveillance applications and getting up-to-date data concerning the conditions of roads, intersections, and buildings to take remedial actions before accidents occur. They also can use it for controlling lighting, energy and power management, water consumption, and many more. Municipalities and local governments can push the processing of urban IoT data streams from the cloud to the edge, reducing network traffic congestion and shortening end-to-end latency. By processing the data generated by edge devices locally, urban facilities can avoid the problem of streaming and storing large amounts of data in the cloud, which impact privacy and make them vulnerable. _3.2. Federated Learning at the Edge_ Machine learning techniques typically rely on centrally managed training data, even when the training process is performed on a cluster of machines. This process often takes advantage of the characteristics of the overall training data set and the availability of validation data sets to adjust several parameters. However, centralizing data management for training is often not feasible or practical because of data privacy, confidentiality, and regulatory compliance. Privacy regulatory frameworks require that data holders maintain the privacy of personal information and limit how to use the data. Examples of these frameworks include the European Union’s General Data Protection Regulation (GDPR) [55] and the Health ----- _Sustainability 2022, 14, 7609_ 7 of 30 Insurance Portability and Accountability Act 1996 (HIPAA) [56]. These restrictions make the management of central data repositories very expensive and a burden for data holders. Federated learning (FL) is a learning approach that aims to solve the issues mentioned above of centralized training data management and data privacy. It allows collaboratively building a learning model without having to move the data beyond the firewalls of the participating organizations [57,58]. Instead, as shown in Figure 2, an initial AI model, hosted in a central server, is transferred to multiple organizations. Each organization trains the AI mode with its data to obtain new weight parameters sent back to the central server. The central server then uses any new weight settings from the participating organizations to create an updated single model. Several iterations of this process may be necessary to obtain an AI model good enough to be used in production. Several research efforts have evaluated the performance of models trained by FL. They have found that they achieve performance levels comparable to models trained on centrally hosted data sets and superior to models that only use isolated data from a single organization [59,60]. **Figure 2. Federated learning architecture.** **4. Blockchain to Support Smart Cities’ Operations** _4.1. Blockchain Overview_ Blockchains are essentially shared databases that enable the participants, called nodes in a network, to confirm, reject, and view transactions. They facilitate recording transactions and tracking asset movements in a business network. Assets can be tangible, such as property, cars, land, or intangible, such as patents and copyrights. Transaction data are stored in a block-based structure, where blocks are linked to each other through a method known as cryptographic hashing. Combined with the distributed and decentralized nature of the blockchain ledger, this method makes each block of data virtually impossible to change once it is added to the chain. Therefore, the blockchain distributed ledger is cryptographically secure and immutable. It works in append-only mode and can only be updated by consensus or peer-to-peer agreement. Blockchain is often viewed as a specific subset of the larger universe of distributed ledger technology (DLT) [61]. The distributed ledger makes Blockchain technology resilient since the network does not have a single point of vulnerability. In addition, each block uniquely connects to previous blocks via a digital signature. Making a change to a record without disrupting earlier records in the chain is impossible, making the information tamper-proof. Allowing its participant to ----- _Sustainability 2022, 14, 7609_ 8 of 30 transfer assets over the Internet without a centralized third party is the essential innovation in Blockchain technology. Blockchain technology emerged over the last few years as the underlying technology for Bitcoin. The consequences of the subprime crisis in 2008 reduced confidence in the existing financial system [62]. Satoshi Nakamoto wrote a white paper describing the “bitcoin protocol”, which used a distributed ledger and consensus to compute algorithms in the same year. The protocol was authored to facilitate direct P2P transactions and disintermediate traditional financial intermediaries [63]. Since the birth of the Internet, many attempts to create virtual currencies have failed due to the double-spending problem. The current solution to eliminate the double-spending problem is introducing “trusted intermediaries” such as banks. Blockchain technology solves the double-spending problem without these trusted intermediaries, making it easier to securely move assets such as virtual currencies over the Internet. Other areas other than currencies could benefit from this concept, making Blockchain technology very promising. As illustrated in Figure 3, the blockchain architecture allows participants in a business network, for example, to share an updated ledger using peer-to-peer replication each time a transaction occurs. Each participant acts as a publisher and subscriber and can receive or send transactions to other participants, and data are synchronized across the network. The blockchain network eliminates duplication of effort and reduces the need to use the services of intermediaries, making it economical and efficient. Using consensus models to validate transaction information also makes the network less vulnerable. Transactions are secure, authenticated, and verifiable. **Figure 3. Network of business parties and intermediaries without and with Blockchain. (a) Trans-** actions between Org. A, B, and C involve intermediaries. (b) Participants share an updated ledger using P2P replication each time a transaction occurs. _4.2. Blockchain Benefits_ The blockchain network stores data in a tamper-proof form, and it permits valid users only to append data to the blockchain. Understanding the primary attributes, depicted in Figure 4, of Blockchain that make this technology unique is essential to comprehend its full potential. - **Distributed shared ledger: This is a distributed append-only system shared across** the corporate or business network, making the system more resilient by eliminating the centralized database, which is a single point of failure. - **Consensus: A transaction is only committed and appended to the ledger when all** validating parties consent to a network verified transaction. - **Provenance: The entire history of an asset is available over a blockchain.** ----- _Sustainability 2022, 14, 7609_ 9 of 30 - **Immutability: Records are indelible and cannot be tampered with once committed to** the shared ledger, thereby making all information trustworthy. - **Finality: Once a transaction is completed over a blockchain, it can never be reverted.** - **Smart contracts: Code is built within a blockchain that computers/nodes execute based** on a triggering event. Essentially, an “if this then that” statement can be auto-executed. **Figure 4. Blockchain benefits.** Blockchain has the potential to disrupt any form of transaction that requires information to be trusted. With the advent of Blockchain technology, all trusted intermediaries are the subject of disruption in one form or another, and Blockchain technology solves the problems associated with the way information-related transactions occur today. Blockchain creates a permanent and unalterable ledger of information by validating transactions through its distributed network of peers. _4.3. Types of Blockchain Networks_ Blockchain networks are either public or private. A public blockchain network operates in a decentralized open environment with no restriction on the number of people joining the network, and the private blockchain network functions within limits defined by a control entity. The intrinsic technology of both networks remains the same; however, the dynamics and utility of closed and open networks are different. This difference plays out based on the incentives for nodes to remain a part of the network. The key idea is that in a public blockchain, the consensus mechanism rewards each participant for staying a part of the network, and in a private blockchain, the need for creating this incentive does not exist. A genuinely transparent public registry’s democratized nature may not be helpful to an organization or corporate network since the parties are known, and there is a level of understanding of the members who can participate in the network and transactions. The consensus is that while public blockchains work well for specific applications such as cryptocurrency (bitcoin) based transactions, the most important application of Blockchain ----- _Sustainability 2022, 14, 7609_ 10 of 30 technology as an enterprise solution would not be possible than with the increased regulatory control associated with a private Blockchain ecosystem. Blockchain technology is still emerging, and therefore its different applications evolve continuously and iteratively. An ecosystem where multiple private blockchains interact with each other on a publicly distributed network may address the issue of public vs. private blockchain networks. In that shared ecosystem, public and private blockchains work in symbiosis in the same way private networks interact with the Internet. Blockchain technology is being applied in numerous domains of smart cities, such as healthcare, power grid, transportation, supply chain management, education, manufacturing, the construction industry, and many others. Several works survey and describe the application of Blockchain in these areas [64–66]. _4.4. Blockchain Suitability_ Blockchain technology is only suitable when multiple parties share data and need a common information view. However, sharing data is not the only qualifying criteria for Blockchain to be a viable solution. The following situations make Blockchain a viable and efficient solution: - A transaction depends on several intermediaries whose presence increases the transaction’s time, cost, and complexity. - Reducing delays and speeding up a transaction is incredibly advantageous for the business. - Transactions created by the business participants depend on each other. - Actions undertaken by multiple participants should be recorded and involved validated data updated. - Building trust between the participants is necessary for the business. To sum up, Blockchain technology is certainly not a solution to all transaction issues. **5. Methodology** This review paper uses a qualitative research approach to synthesize the relevant literature on the article’s subject. Given the descriptive nature of the present study, the qualitative approach allows for reviewing and synthesizing a large amount of pertinent literature. A systematic review strategy was adopted without claiming to be exhaustive in pursuing this objective. _5.1. Search Criteria Formulation_ The search criteria used were: - C1: (“Edge AI” OR “edge intelligence”) AND “Blockchain”; - C2: (“Edge AI” OR “edge intelligence”) AND (“smart mobility” OR “smart transportation”); - C3: “Blockchain” AND (“smart mobility” OR “smart transportation”); - C4: (“Edge AI” OR “edge intelligence”) AND “smart energy”; - C5: “Blockchain” AND “smart energy”. The purpose of this review paper is to answer the following research questions. - RQ-1: What are the applications of edge AI and Blockchain regarding smart mobility and smart energy? This research question intends to identify the state-of-the-art research regarding the applications of edge AI and Blockchain technology in these two key areas of a smart city. - RQ-2: What are the potential open research issues and future directions in edge AI and Blockchain implementation in these two vital areas of a smart city? This question aims to define the open questions and research directions for the wide adoption of edge AI and Blockchain to address the challenges in implementing smart mobility and smart energy. Consequently, answering this question encourages researchers to understand the current research findings and trends in edge intelligence and Blockchain. ----- _Sustainability 2022, 14, 7609_ 11 of 30 _5.2. Source Selection and Approach_ The review included articles published between 2017 and 2021. A search for relevant research on the topic of this review was conducted using the following databases and search engines: (i) Scopus, (ii) ScienceDirect, and (iii) Google Scholar, which provide excellent coverage of the study topics. The search used the search criteria above and revolved around the terms “Edge AI” and “Blockchain” while including synonyms as additional terms such as “edge intelligence” and “distributed ledger” to increase the search results. Most of the papers reviewed are journal articles, with some conference papers also included. Papers were selected based on the quality of the journal, relevance to the topic, and filtered by date of publication. Edge intelligence and Blockchain are still in their infancy and are evolving rapidly. Article selection was based on titles, keywords, abstracts, and conclusions relevant to the topic. References cited in this review paper published before 2017 mainly concern the background and literature review on smart city areas and challenges, edge computing, and Blockchain. The initial search for the five search criteria (C1–C5) found 417 references from Scopus, 533 from ScienceDirect, and 931 from Google Scholar (review articles). However, the total number of papers was reduced to 150 after the title and abstract screening, excluding, and eliminating duplicates. Afterwards, the papers were classified into four main classes: background and fundamentals, edge AI and Blockchain convergence, applications of edge AI in smart mobility and smart energy, and applications of Blockchain in smart mobility and smart energy. **6. Transformative Potential of Edge AI and Blockchain in Smart Cities** Modern cities struggle to automate many of their processes and coordinate them with various stakeholders. Citizens expect their governments and smart city entities to respond quickly to their demands and needs while ensuring transparency, fairness, and accountability to the public. Success in these endeavors, especially in the digital age, requires that up-to-date data be collected and processed in near real-time. Much of the challenge is in the management and processing of data. Unfortunately, traditional centralized databases and data management tools are not enough to meet the new challenges that smart cities face. The data exchanged between the various city actors can be tampered with. The single point of failure of the standard database client-server model compromises data security, making transparency challenging to achieve when city databases are centralized. Additionally, using centralized databases results in slow and inefficient operations such as registering identifications (IDs) and electoral voting. Smart cities and government entities can address the above issues by taking advantage of the recent advances in edge AI and using an innovative data management structure. This data structure uses distributed ledgers and cryptography. Furthermore, they can offer citizens smart on-demand services while ensuring data privacy and security, unprecedented transparency, fairness, and accountability [67,68]. Here, we discuss the potential of these two technologies in two crucial subsystems of a smart city, smart mobility, and smart energy management, and review relevant research works on their usage in these areas. _6.1. Smart Mobility_ Modern cities suffer from major issues such as traffic congestion, emissions, and safety. Without innovative solutions, mobility problems will intensify due to the continued growth of the population, which leads to an increase in the number of vehicles on the roads, the kilometers traveled, and consequently the increase in emissions. In response, the mobility industry is developing a fascinating range of innovations designed for urban roads, such as intelligent traffic and parking management systems, mobility as a service, and carpooling solutions. “Smart transport” often refers to the use of new digital technologies and data-driven management techniques in transport systems to address the mobility problems [28,69]. The phenomenal technological developments in recent years, which have brought about significant changes in all aspects of our life, promise to improve transport ----- _Sustainability 2022, 14, 7609_ 12 of 30 in cities in all its forms. Smart transport, being a dream, is becoming more and more a reality. We are seeing more and more applications that integrate live data and feedback from multiple sources to gain a holistic and real-time view of the traffic status, helping stakeholders better manage road traffic and deliver quality services to road users. Other innovations that contribute to smart transport and mobility include: - The development of new models of shared mobility; - The development of more reliable and convenient public transport; - The development of applications allowing to alert drivers of hazardous situations quickly; - The development of navigation applications that allow drivers to find in real-time the best route possible; - The ability to adjust road signals and speed limits in real-time based on current traffic conditions; - The development of new concepts of electric, connected, and autonomous vehicles. Because of the costly computations of traffic management systems, the improvement of the real-time processing of data is one of the best ways to optimize traffic management systems [27]. Traffic data are obtained from various sensors and IoT devices deployed on urban roads and vehicles by transportation systems. Intelligent transport systems are evolving towards intensive use of edge computing and edge AI technologies, especially for traffic management processes [70]. Gigabytes of sensory data are analyzed, filtered, and compressed locally before being transmitted through IoT edge gateways to multiple systems for later use. Edge processing for traffic management solutions allows one to save on storage, network expenses, and operating costs. 6.1.1. Edge AI for Traffic Monitoring and Management Traffic management is an undeniable component of smart mobility, which combines different measures to preserve traffic capacity, reduce congestion at roads and intersections, and improve the safety and reliability of the overall road transport system. Modern traffic management systems are composed of advanced sensing and monitoring technologies, management tools, and a set of intelligent applications to achieve these goals. These technological solutions prepare smart cities for future cutting-edge technological developments, in particular the proliferation of autonomous vehicles, connected vehicles, and the largescale deployment of Fifth Generation (5G) cellular networks and edge AI systems [71]. Several works investigated edge computing-based solutions for traffic management in smart cities. Barthélemy et al. [70] designed a visual sensor for monitoring the flow of bicycles, vehicles, and pedestrians traffic. Their complete edge-computing-based solution aims to deploy multiple visual sensors and collect data through a framework called Agnosticity. The visual sensor hardware uses the NVIDIA Jetson TX2 on-board computing platform to perform all computations onboard. Its software pairs YOLOv3 [72], a popular convolutional deep neural network, with Simple Online and Realtime Tracking (SORT) [73], a real-time tracking algorithm. The metadata are then extracted and transmitted using Ethernet or LoRaWAN protocols. The sensor provides a privacy-compliant tracking solution by transmitting only metadata instead of raw or processed images. Municipalities can combine the sensors with the existing Closed-circuit television (CCTV) infrastructure, and this integration helps optimize infrastructure usage and add value to the network by leveraging the vast video data collected. Besides, the Long Range Wide Area Network (LoRaWAN) protocol facilitates the deployment of additional cameras in areas where conventional internet connectivity is not available. Dinh et al. [74] proposed an inexpensive and efficient edge-based system integrating object detection models to perform vehicle detection, tracking, and counting. They created a Video Detection Dataset (VDD) in Vietnam and then examined it on two different types of edge devices. They evaluated their proposed traffic counting system in a Coral Dev TPU Board and then a Jetson Nano GPU Board and implemented several models in the two boards. The MobileDet 320 × 320 SSD model implemented in the Coral Dev TPU Board ----- _Sustainability 2022, 14, 7609_ 13 of 30 for the vehicle detection context achieves an accuracy of 92.1%, and the proposed method achieves a maximum inference speed of around 26.8 Frames per second (FPS) on VDD. Additionally, Kumar et al. [75] investigated how to detect and track vehicles effectively. Their proposed method detects tracks and extracts vehicle parameters for speed estimation using a single camera. They used the Automatic Number Plate Recognition (ANPR) system to select keyframes where a speed limit violation occurs. The average detection accuracy obtained is approximately 87.7%. The proposed approach uses cropping operations to minimize the scope of any detection of false positives on both sides of the road. The average detection accuracy obtained is 87.7%. The proposed approach tracks vehicles moving in one direction but fails to detect vehicles coming from opposite directions. Likewise, Song et al. [76] proposed a vision-based vehicle detection and counting system for highways. The proposed method is not expensive, is highly stable, and does not require a significant investment in terms of monitoring equipment. They used a “Vehicle dataset” to train a YOLOv3 network to obtain the vehicle object detection model. Image segmentation and YOLOv3 allowed them to detect three types of vehicles: cars, buses, and trucks. A convolutional neural network and the Oriented FAST and Rotated BRIEF (ORB) algorithm [77] were used to extract the features of detected vehicles. The authors stated that vehicles’ detection speed is fast, and its accuracy is high. Traffic footages taken by highway surveillance video cameras have good adaptability to the YOLOv3 network. Multi-object tracking uses the object box detected in vehicle detection using YOLOv3. The ORB algorithm uses the Features from the Accelerated Segment Test (FAST) to detect feature points, and the Harris operator performs corner detection. In many cities, a segment of a public or private road can be used to load and unload goods at specific times or at any time. Parking signs and road markings are typically used to warn drivers of parking regulations. These areas are known as loading bays. Parking inspectors generally monitor these areas, and motorists found violating the rules can be fined. These restrictions on urban freight deliveries require establishing a loading bay system and dividing the last mile delivery into driving and walking segments. Loading bays are sometimes occupied, requiring rerouting delivery vehicles and searching for an alternative loading bay. The authors in [78] introduced a fuzzy clustering method to test different optimization approaches and make the system flexible enough to accommodate this problem. We believe that edge AI and computer vision can help address where and how many loading bays should be used to perform this transshipment and execute last-mile delivery most efficiently. 6.1.2. Blockchain for Smart Mobility With the population growth of cities and the rapid increase in demand for smart transport and mobility solutions, there is an urgent need for innovative solutions that use existing infrastructure in cities and on external roads and highways between cities. Smart mobility technologies aim to provide many new applications and perspectives for efficient and safe movement on roads while reducing Carbon dioxide (CO2) emissions and improving air quality [69]. Transportation systems management is a challenging endeavor in many modern cities [79]. Blockchain technology can improve information sharing between different stakeholders in cities, improve the robustness of the overall transport system and facilitate communication between vehicles, contact with road units, and transport traffic control centers. In addition, Blockchain in the transport sector also can reduce the processing time of transport-related transactions, approvals, and exchange of documents and speed up customs clearance. This section summarizes relevant work on Blockchain-based solutions for smart transportation and mobility. Figure 5 depicts the main areas where Blockchain has been used to contribute to the smart mobility goals, and Table 1 summarizes the focus area of each of the reviewed works and the Blockchain mechanisms they used. ----- _Sustainability 2022, 14, 7609_ 14 of 30 **Figure 5. Blockchain for smart mobility.** **Table 1. Summary of Blockchain-based smart mobility literature review.** **Ref.** **Focus** **Blockchain Used Mechanisms** Blockchain as the operating system of smart cities, with transportation [80] Etherium-like Blockchain, smart contracts management as one of the main focus areas Blockchain in vehicular communications, in particular, a sys [81] tem for revocation and accountability in a Security Credential Distributed Ledger, hierarchical consensus Management System. Distributed blockchain vehicular network, Miner [82] A blockchain-based vehicular network architecture in smart city. Vehicular Node, revocation authority, Block Node Controller Reputation systems in vehicular networks based on Blockchain Vehicular blockchain, Miner Vehicle, Trusted Au [83] technology. thority, distributed consensus Blockchain-based key management scheme to transfer security keys [84] between distributed security managers in heterogeneous Vehicular Communication Systems. Blockchain structure without the third-party authorities, Transaction format, Mining, and Proof algorithm. Blockchain-based decentralized Key Management Mechanism Vehicular blockchain network, Ethereum-based [85] for VANET. Smart contract, mining functions. Decentralized Trust Management system in vehicular networks based Vehicular blockchain, Miner Vehicle, Trusted Au [86] on Blockchain technology. thority, distributed consensus Decentralized Trust Management system in vehicular networks based Vehicular blockchain, Tendermint( consensus with [87] on Blockchain technology and the Tendermint consensus protocol. out mining), BFT based consensus. A location privacy protection system based on trust management in Vehicular blockchain to record the trustworthiness [88] Blockchain-based VANET. of vehicles, PBFT consensus algorithm. A Blockchain-based system combined with auctions to enable BEVs Blockchain to record trading contracts, Smart [89] to trade energy using day-ahead and real-time trading markets. contract. Roaming charging process of electric vehicles and Blockchain technol [90] ogy to support user identity management and record energy transac- Distributed ledge to record energy transactions tions securely. [91] Blockchain to mitigate trust Issues in Electric Vehicle Charging. Hyperledger Fabric, smart contract. ----- _Sustainability 2022, 14, 7609_ 15 of 30 Bagloee et al. [80] suggested that to reduce traffic congestion and achieve system equilibrium, traffic authorities may issue a limited number of mobility permits, distributed equally to all drivers, which may be tradable in an open market. Such a progressive scheme is now possible in light of the ever-increasing use of various kinds of sensors, cameras, RFIDs, radars, and lidars. Blockchain technology and smart contracts can be used as a valid, promising, and feasible solution for implementing the tradable part of this scheme. The authors also suggested that drivers and passengers use the Tradable Mobility Pass (TMP) equally to pay parking fees, public transport tickets, car registration fees, and highway tolls. An Ethereum-like blockchain and “smart contracts” can be used to program their mobility credits for trading in the open market and spending against the above payments and mileage. They can also be used to trade TMPs en-route by permitting vehicles to communicate with each other and place bids for faster routes at higher prices. Blockchain can also facilitate communication between connected vehicles and the road infrastructure by considering data exchange requests as transactions to be stored and retrieved from a blockchain database. Additionally, Blockchain can provide safe, secure, and well-informed access to driving behavior information for driving license agencies and insurance companies, which typically know little about driving behavior. Insurance companies’ predictions are based on claims history [92]. Access to data from connected vehicles can help them set insurance premiums commensurate with drivers’ risk levels. **Blockchain in vehicular communications. Some works proposed Blockchain-based** solutions to help create a secure, trusted, and distributed autonomous Intelligent Transportation System (ITS) capable of controlling and managing physical and digital assets. At the same time, most ITSs were centralized [93]. The authors in [81] described the design of a Blockchain-based decentralized alternative to existing security credential management systems, which aimed to get rid of the need of using the services of a centralized trusting authority. Vehicle-to-Everything (V2E) communications are an essential component in any ITS. They help provide information on road accidents, road conditions, traffic jams, allowing road drivers to be aware of critical situations, thus enhancing transport safety. Sharma et al. [82] proposed a distributed transport management system that allows vehicles to share their resources and create a network where value-added services, such as automatic gas refill and ride-sharing, can be produced. Additionally, Yang et al. [83] proposed reputation systems in vehicular networks based on Blockchain technology. Lei et al. [84] proposed a Blockchain-based key management scheme to transfer security keys between distributed security managers in heterogeneous Vehicular Communication Systems (VCS). The blockchain structure enables secure key transfer between participating network security managers and eliminates the need for a central manager or third-party authority. Likewise, the authors in [85] proposed a decentralized key management mechanism for Vehicular Ad-hoc Networks (VANETs) with Blockchain to automatically register, update, and revoke the user’s public key. They also described a lightweight mutual authentication and key agreement protocol based on the bivariate polynomial. Additionally, they analyzed the security of their proposed mechanism for managing distributed keys and have shown that it can prevent typical attacks, including insider attacks, public key tampering attacks, Denial-of-Service (DoS) attacks, and collusion attacks. Additionally, Yang et al. [86] proposed a decentralized Blockchain-based trust management system in vehicular networks. Vehicles can query the trust values of neighboring vehicles and assess the credibility of received messages. The RSUs aggregate the confidence values based on evaluations generated by the messages’ recipients. Using Blockchain, all RSUs contribute to maintaining a reliable database. Similarly, Arora et al. [87] proposed a Blockchain-based trust management system for VANETs based on the Tendermint protocol to eliminate the possibility of malicious nodes entering the network and reduce power consumption. Vehicles assess the messages ----- _Sustainability 2022, 14, 7609_ 16 of 30 received from neighboring vehicles using the gradient boosting technique (GBT). Based on the assessment results, the message source vehicle generates the ratings, uploads them to RSUs, and calculates the trust offset value. All RSUs maintain the trust blockchain, and each RSU adds its blocks to the trust blockchain. In another work, Luo et al. [88] proposed a location privacy protection system based on trust in Blockchain-based VANET. Their trust management approach uses Dirichlet distribution to allow requesters to cooperate only with vehicles they trust. In addition, they also developed the blockchain data structure to record the trustworthiness of vehicles on publicly accessible blocks promptly to allow any vehicle to access historical trust information of counterparties whenever necessary. **Blockchain for Electrical Vehicles. Battery Electric Vehicles (BEVs) are known for** their low operating costs because they have fewer moving parts that require maintenance. In addition, they are very environmentally friendly as they do not use fossil fuels. Modern BEVs use rechargeable lithium-ion batteries, which have a longer life and retain energy very well with a self-discharge rate of only 5% per month. In many cities around the world, Charging Stations (CSs) are increasingly deployed in various geographic locations, residential garages, and public/private parking lots to meet the energy needs of BEVs, increasing the load on electrical distribution systems. Intelligent car parking lots offer BEVs parking and recharging services during their parking time for a fee. Customers of these parking lots want fast charging services at low cost, while parking lot operators aim to maximize their profit. BEV owners increasingly tend to purchase power from other electric vehicles to reduce recharging costs and reliance on the primary electricity grid. Huang et al. [89] proposed a Blockchain-based system to enable BEVs to trade energy using day-ahead and real-time trading markets. Users of BEVs submit their price offers to participate in a double auction. Then, the operator of the charging system performs intelligent matching of the different offers to reduce the impact on the power grid by programming the charging and discharging behavior of electric vehicles taking into account the satisfaction of EV users and the social benefits. The operator of the charging system uploads the trading contract to the blockchain once the trading results are cleared. Case studies have demonstrated the effectiveness of the proposed model. Ferreira et al. [90] studied the roaming charging process of electric vehicles and used Blockchain technologies to support user identity management and record energy transactions securely. They used off-chain cloud storage to record transaction details. Blockchain-based digital identity management avoids charging cards used as an authentication process in charging systems. It can achieve interoperability between different countries, allowing a roaming process of BEV charging. In [91], Gorenflo et al. described a methodology for the design of Blockchainbased systems. They have demonstrated its usefulness in creating a system for recharging electric vehicles in a decentralized network of recharging stations. The proposed system aims to solve the problem of trust between the different actors of the system, including customers, providers of electric vehicle charging services, and property owners. Trust problems arise from the potential for tampering with transaction data. The blockchain ledger in the proposed solution contains a record of every transaction and acts as an immutable audit trail. _6.2. Smart Energy_ In recent years, the term “Smart Energy” has been used more and more to mean an approach that goes beyond the concept of “Smart Grid.” While the smart grid concept mainly focuses on the electricity sector, smart energy embodies a holistic approach that includes many sectors (electricity, heating, cooling, buildings, industry, and transport). It allows the development of affordable solutions for transforming existing systems into future renewable and sustainable energy solutions [33]. Smart energy solutions typically use various disruptive technologies, including artificial intelligence, deep learning, Blockchain ----- _Sustainability 2022, 14, 7609_ 17 of 30 and distributed ledger technologies, distributed sensing and actuation technologies, and, recently, edge computing and federated learning technologies. 6.2.1. Edge AI for Smart Energy Management Several research efforts are increasingly studying and developing smart energy solutions. Shah et al. [94] reviewed several research works that use different energy optimization techniques in smart buildings and rely on IoT solutions. Their study aimed to identify algorithms and methods for optimized energy use and edge and fog computing techniques used in smart home environments. From an initial batch of 3800 papers, they found only 56 articles relevant to their study. The detailed analysis of these papers revealed that many researchers had developed new optimization algorithms to optimize energy consumption in smart homes. Zhang et al. [95] proposed an IoT-based green energy management system to improve the energy management of power grids in smart cities. With the implementation of IoT, smart cities can control energy through ubiquitous monitoring and secure communications. The proposed system uses deep reinforcement learning. The authors’ results show that IoT sensors help detect energy consumption, predict energy demand in smart cities, and reduce costs. Aided by a systematic learning process, the energy management system can balance energy availability and demand by stably maintaining grid states. Abdel-Basset et al. [96] proposed a smart edge computing framework to achieve efficient energy management in smart cities. They reviewed relevant work on data-driven load forecasting (LF) techniques used in real-life scenarios such as smart buildings to predict the day’s energy demand in advance and make appropriate energy demands on smart grids. These short-term forecasts help to avoid energy shortages and promote fair consumption. They classified these techniques into two classes: statistical or machine learning-based techniques and deep learning-based techniques. They introduced a new deep learning architecture, called Energy-Net, to predict energy consumption by integrating the spatial and temporal learning capability. They validated the robustness of their proposed architecture through a comparative analysis of public datasets with recent cutting-edge approaches. According to the authors, the trained Energy-Net system is deployable on resource-limited edge devices to forecast potential energy needs sent as a request to the smart grid through cloud-fog servers. As a result, the smart grid supplies the demanded energy to different smart city sectors. Energy management is, therefore, performed efficiently. The authors in [97] studied and proposed an energy management framework based on edge computing for a smart city. They developed an energy scheduling scheme based on deep reinforcement learning to deal with the intermittency and uncertainty of energy supplies and demands in cities for a long-term goal. They analyzed the efficiency of the energy scheduling scheme in the cases with and without edge servers, respectively. Their results demonstrate that the proposed model can achieve low energy costs while exhibiting lower delays than traditional schemes. 6.2.2. Blockchain for Smart Energy Management Blockchain technology in the energy sector is up-and-coming. It can significantly reduce energy trading costs, increase process efficiency, and deliver customer cost benefits. It can establish direct interactions between all the actors involved, which guarantees the optimal use of existing production capacities while offering energy at the best price. The application of Blockchain in emerging smart energy systems in smart cities has recently received a great deal of attention. In addition to the BEV charging we mentioned, there is an increasing need for decentralized energy management, energy trading platforms development, and secure data and financial transactions between the different actors involved. This need arises from the proliferation of new devices, technologies, renewable energy resources, and electric vehicles. Additionally, there is a growing interest worldwide in using Blockchain technologies to create a secure and more resilient environment for the smart ----- _Sustainability 2022, 14, 7609_ 18 of 30 energy industry. Several research efforts investigated the opportunities, benefits, challenges, as well as drawbacks of Blockchain technologies in the context of smart energy [98–100]. This section reviews some efforts regarding the use of Blockchain in smart energy systems. We do not intend to provide a full survey. Andoni et al. [101] reviewed and ranked about 140 Blockchain-based projects in the energy sector. Additionally, the authors in [102] reviewed several research works regarding the applications of Blockchain technology in smart grids. They categorized them in decentralized energy management, energy trading, BEVs, financial transactions, cybersecurity, testbeds, environmental issues, and demand response (DR). A common aspect of most of the efforts is the usage of Blockchain to address decentralized energy management, energy trading, transparency, and its perceived benefits to system security. However, system security and user privacy are typically dependent on the type of blockchain used. Table 2 summarizes these efforts. **Table 2. Summary of Blockchain-based smart energy literature review.** **Ref.** **Focus** **Blockchain Used Mechanisms** Smart contracts, consensus-based DR validation [98] Distributed management of DR in smart grids approach [99] Smart energy trading Smart contracts pay-to-public-key-hash with multiple signatures to [100] P2P energy and carbon trading secure transaction [101] Review of challenges of Blockchain technology in the energy sector [102] Review of blockchain in future smart grids Review of blockchain applications in different areas of a smart city, [103] including smart energy Smart contracts, noncooperative game for consump [104] Automated energy DR, P2P energy trading tion strategy to reach consensus [105] Distributed energy system (short review) Smart contracts, consensus [106] Federated power plants with P2P energy trading Distributed energy management in a multi-energy market en [107] Smart contracts, consensus hanced with blockchain [108] Distributed energy exchange Smart contracts [109] Microgrid energy market, P2P energy trading [110] Electricity Trading for Neighborhood Renewable Energy P2P Blockchain network [111] Smart homes energy trading Ethereum’s smart contracts, consensus [112] P2P solar energy market Auction mechanism in the smart contracts. Ethereum-based blockchain, Smart contracts, Dis [113] P2P Energy Trading tributed consensus for verification and group management Federated Learning-based P2P Energy Sharing assisted with [114] smart contracts for energy demand prediction Blockchain Electrical energy transaction ecosystem between smart homes pro [115] Smart contracts (energy tags) sumers and consumers, P2P energy trading Review of applications of smart communities, including energy [116] Smart contracts, miners, consensus. trading in ITS using blockchain. **Decentralized Energy Management. The ever-growing deployment of renewable** energy systems in smart grids highlights the need to develop distributed energy management systems and trigger fundamental changes in energy trading [117,118]. A large body of literature has investigated the usage of Blockchain technologies to ease decentralized energy management according to the P2P model used by Blockchain [103–106,119]. ----- _Sustainability 2022, 14, 7609_ 19 of 30 Real-time energy management has the potential to resolve the impact of various uncertainties in the energy market, provide instant energy balance and improve business returns. Wang et al. [107] proposed a bidding strategy for the energy market, with multiple participants, which uses an adaptive learning process that incorporates a reserve price adjustment and a mechanism of dynamic compensation. Participants perform bid adjustments based on adaptive learning leveraging real-time market information to increase transaction rate and maximize profits. Blockchain technology guarantees the transparent and efficient performance of the presented bidding strategy. A decentralized Blockchain application showed that the system could achieve real-time energy management and dynamic trading in practice. **Energy trading. Recent years have seen the high penetration of renewable energy** systems in smart grids and homes. However, complex energy trading and complicated monitoring procedures are obstacles to developing renewable energies. Energy trading involves various actors, including residential consumers, renewable energy producers, BEVs, and energy storage, which can participate in a Blockchain-based market for energy trading with the roles of prosumer and consumer. Actors propose their energy costs due to their resources and capabilities, which leads to a competitive energy market. Therefore, the blockchain can facilitate energy trading and data transactions while guaranteeing transaction security, improving transparency, and easing financial transactions. The data flow between prosumers and consumers without human involvement [108]. A significant body of research has studied and proposed Blockchain-based networks to enable energy trading and related transactions. For example, the authors in [109,110] have studied renewable energy developments, including wind and solar power, in smart homes. They proposed to use Blockchain technology to trade energy between smart homes and increase their financial benefits. Additionally, Kang et al. [111] investigated energy trading between smart homes using Blockchain technology. Smart homes store energy in energy storage, and consumer nodes equipped with miners monitor energy consumption. Therefore, if the stored energy is not sufficient to power the loads, the additional energy is purchased from the prosumer nodes by having Ethereum smart contracts manage the energy trade according to the following rules: - Energy trading conditions should be specified to permit energy exchange between prosumers and consumers. - Prosumers and consumers should determine price and exchange procedures beforehand, and the prosumers should complete the proof-of-work. - If a consumer’s stored energy falls below a certain level, her home miners should send energy trading requests to appropriate prosumers. - Energy trading takes place when consumer requirements match prosumer conditions. It is widely expected that the global demand for clean and stable energy sources will continue to increase over the coming decades. With the recent penetration of distributed resources into energy trading, communities can take advantage of cheaper electricity prices while supporting green energy locally. However, this poses new challenges mainly in the auction process to ensure individual rationality and economic efficiency, mitigated with the help of Blockchain technology. Lin et al. [112] studied the application of P2P energy trading and Blockchain technology in the development of photovoltaic (PV) units. They proposed a P2P energy trading model using a Discriminatory and Uniform k-Double Auction (k-DA). They verified the financial benefits of the proposed model through simulation. The authors in [113] have exploited the opportunities offered by Blockchain in building the prosumer group in the context of P2P energy trading. They proposed a Blockchainassisted adaptive model, named SynergyChain, to improve the scalability and decentralization of the prosumer aggregation mechanism in the context of P2P energy trading. The model showed that the coalition of multiple energy prosumers through aggregation outperformed the case in which individual prosumers participated in the energy market. They implemented a reinforcement learning module that decides whether the system ----- _Sustainability 2022, 14, 7609_ 20 of 30 should act as a group or independently. The complete analysis using the hourly energy consumption dataset showed a substantial improvement in system performance and scalability compared to centralized systems. Furthermore, their system worked better with the learning module, in terms of cost-effectiveness and performance, than without it. In another work [114], the authors proposed FederatedGrids, a platform that uses federated learning and Blockchain for P2P energy trading and sharing. It creates a collaborative environment that maintains a good balance between the participants of the different microgrids. The blockchain helps to ensure trust and privacy between all participants. Smart contracts and federated learning allow the platform to predict future energy production and system load, thus allowing prosumers to make optimal decisions related to their energy sharing and exchange strategies. Smart cities can significantly benefit from Blockchain capabilities to maximize energy efficiency and improve energy resource planning and management. Blockchain-based networks can directly connect multiple energy resources and household appliances, thereby providing users with high-quality, inexpensive, and efficient energy [115]. They can help regulate the distribution and transformation of energy in smart grids, bringing more transparency to energy transactions [116]. **7. Edge AI and Blockchain Convergence** Several research efforts studied the convergence between Blockchain and edge computing without considering or giving details about the AI component at the edge [68,120–126]. However, as AI techniques further proliferate at the edge in various smart city systems (healthcare, transportation, power grid, etc.) and ensure huge benefits, they also introduce increased privacy and security threats. Therefore, robust security measures are needed to protect data and AI models at the edge. These measures include security features for data storage, encryption, data dissemination, and key/certificate management. As we discussed earlier, edge AI and Federated Learning are emerging technologies for building smart latency-sensitive services at the edge while protecting data privacy. On the other hand, Blockchain technology shows significant possibilities with its immutable, distributed, and auditable data recording for safeguarding against data breaches in a distributed environment. The convergence between Blockchain and AI is attracting much interest in academia and industry to solve many challenging problems to manage effectively a few years ago. The characteristics of blockchain technology and its decentralized architecture, which we discussed in Section 4.1, can help build robust and secure AI applications. Blockchain attributes of immutability, provenance, consensus, and transparency enable secure sharing of AI training data and pre-trained AI models using a permanent and unalterable record of AI data and models. Secure sharing of AI data and models is associated with increased trust in AI models and the data they work with. More and more research efforts study the convergence of edge AI and Blockchain. Table 3 summarizes those efforts. Jiang et al. [127] argued that conventional approaches for object detection that rely on classic and connectionist AI models are not adequate to support the large-scale deployment of the Visual Internet of Vehicles (V-IoV). On the other hand, edge intelligence, which integrates edge computing and AI, demonstrated a balance between efficiency and computational complexity. Edge AI involves training learning models and analyzing V-IoV data, reducing latency, improving time to action, and minimizing network bandwidth usage. Object detection tasks can be offloaded and executed on Roadside Units (RSUs) using the edge’s storage and computing power capabilities. The authors proposed an edge AI framework for object detection in the V-IoV system and a You Only Look Once (YOLO)-based abductive learning algorithm for robust and interpretable AI. The abductive model combines symbolic and connectionist AI to learn from data. Additionally, Blockchain complements edge AI with security, privacy, reliability, scalability, and enables model sharing. Lin et al. [128] consider that extracting knowledge, such as classification models, detection, and predictions from physical environments, from sensory data, could be achieved ----- _Sustainability 2022, 14, 7609_ 21 of 30 by introducing edge computing and edge AI into the Internet of Things. Since multiple nodes with heterogeneous Edge AI devices generate isolated knowledge, collaboration and data exchange between nodes are essential to building intelligent applications and services. The authors proposed a P2P knowledge marketplace to make knowledge tradable in edge AI-enabled IoT and a knowledge consortium blockchain for secure and efficient knowledge management and exchange in the market. The blockchain consortium includes a cryptographic knowledge coin, smart contracts, and a consensus mechanism as proof of trade. Rahman et al. [129] addressed in their work the challenge of bringing intelligent and cognitive processing to the edge where the massive amount of IoT data are generated and processed by mobile edge computing (MEC) nodes. Key transactions are anonymized and securely recorded in the blockchain, where big data are securely stored in the decentralized off-chain solutions with an immutable ledger. Qiu et al. [130] proposed AI-Chain, a Blockchain-based edge intelligence for Beyond Fifth-Generation (B5G) networks. AIChain is an immutable and distributed record of local learning outcomes that can lay a new foundation for sharing information between edge nodes. Leveraging the portability of deep learning, each node at the edge trains neural network components and applies AI-Chain to share its learning results. This process dramatically reduces the wastage of computing power and improves the learning power of the edge node through the learning power of other edge nodes. Du et al. [131] reviewed the existing literature on Blockchain-enabled edge intelligence in the IoT domain, identified emerging trends, and suggested open issues for further research, including transaction rejection, selfish learning, and fork issues. Fork problems arise when edge nodes disagree on the same learning model and alternative chains (i.e., forked chains) emerge. As a use case of the convergence of Blockchain and edge AI, we consider in the following some efforts in the context of smart mobility. IoV is an emerging technology that has the potential to alleviate traffic problems in smart cities. In an IoV network, the vehicles are equipped with modern communication and sensing technologies that allow the sharing and exchanging of data between the vehicles and the RSUs. The massive volume of data captured by vehicle sensors, including GPS and RADAR, favors data-driven AI models. Attacks against vehicles using polymorphic viruses cannot be easily recognized and predicted because their signatures continually change. The centralized ML paradigm is evolving towards a more decentralized and distributed learning framework, especially in a federated learning setup, to accommodate the increase in likely privacy and security issues. Several works proposed federated learning-based solutions for the IoV [132–135]. Although federated learning provides incredible security to learning structures, it faces several other security issues as it operates based on a centralized aggregator. For model training, federated learning relies on local workers, who may be vulnerable to cyber intrusions. If a local model is attacked, it can mislead other models, and therefore the global update is erroneous. Because of the likelihood of such possible attacks in federated learning, Blockchain is used with federated learning to give a decentralized arrangement to control incentives and reliably ensure security and protection. Due to the promising capability of federated learning, especially for building an ITS, and the requirement to alleviate potential attacks, some Blockchain-enabled federated learning schemes for IoV have been proposed over the last few years. The authors in [136] proposed a framework for knowledge sharing in IoV based on a hierarchical federated learning algorithm and a hierarchical blockchain. Vehicles and RSUs learn surrounding data through machine learning methods and share learning knowledge. The use of blockchain framework targets large-scale vehicle networks, and the hierarchical federated learning algorithm aims to meet the distributed model and privacy requirements of IoVs. They modeled knowledge sharing as a trading market process to drive sharing behaviors and formulated the trading process as a multi-leader, multi-player game. The authors stated that their simulation results showed that the proposed hierarchical algorithm improves sharing efficiency and learning quality and achieves approximately ----- _Sustainability 2022, 14, 7609_ 22 of 30 10% more accuracy than conventional federated learning algorithms. RSUs reach optimal utility during the sharing process. Moreover, the blockchain-enabled framework effectively protects against malicious workers during the sharing process. The authors in [137] proposed a blockchain-enabled federated learning framework to improve the performance and privacy of autonomous vehicles. The framework facilitates the efficient communication of autonomous vehicles, where on-board local learning modules exchange and verify their updates in a fully decentralized manner without any centralized coordination by leveraging the blockchain consensus mechanism. The framework extends the reach of its federation to untrustworthy public network vehicles via a validation process of local training modules. By offering rewards proportional to the usefulness of data sample sizes, the framework encourages vehicles with immense data samples to join the federated learning. In the IoV, exchanging messages between vehicles is essential to ensure road safety, and broadcasting is generally used for emergencies. To solve the low probability of receiving broadcast messages in high-density and vehicle mobility scenarios, the authors of [138] proposed a blockchain-assisted federated learning solution for message broadcasting. Similar to the Proof-of-Work (PoW) consensus used in several blockchains, vehicles compete to become a relay (minor) node by processing the proposed Proof-of-Federated-Learning (PoFL) consensus embedded in the smart contract of the blockchain. The Stackelberg game further analyzes the business model to incentivize vehicles to be involved in federated learning and message delivery. The authors stated that their solution outperforms the same solution without blockchain, allowing more vehicles to upload their local models and yield a more accurate aggregated model in less time. It also outperforms other blockchain-based approaches by reducing the consensus time by 65.2%, improving the message delivery rate by at least 8.2%, and more effectively maintaining the privacy of neighboring vehicles. Doku et al. [139] proposed a federated learning framework called iFLBC to bring artificial intelligence to edge nodes through a shared machine learning model powered by Blockchain technology. Their motivation is to filter relevant data from irrelevant data using a mechanism called Proof of Common Interest (PoCI). The relevant data of an edge node are used to train a model, which is then aggregated with models trained by other edge nodes to generate a shared model stored on the blockchain. Network members download the aggregated model to provide intelligent services to end-users. **Table 3. Summary of edge AI and Blockchain convergence literature review.** **Ref.** **Focus Area** **Edge AI Use Case** **Blockchain Use Case** Knowledge management [127] and exchange in the Internet of Vehicles (IoV) Making Knowledge [128] Tradable in Edge-AI Enabled IoT. Security, privacy, reliability, scalability, and model sharing. A knowledge consortium blockchain for secure and efficient knowledge management and exchange in the market. The blockchain consortium includes a cryptographic knowledge coin, smart contracts, and a consensus mechanism as proof of trade. Key transactions are anonymized and securely recorded in the blockchain, where big data are securely stored in the decentralized off-chain solutions with an immutable ledger. Object detection in and a YOLO-based abductive learning algorithm for robust and interpretable AI. Extracting knowledge, such as classification models, detection, and predictions from physical environments and sensory data at the edge. A P2P knowledge marketplace to make knowledge tradable in edge AI-enabled IoT Bringing intelligent and cognitive processing to the edge where the massive amount of IoT data are generated and processed by mobile edge computing (MEC) nodes. [129] Blockchain and IoT-Based Cognitive Edge Framework for Sharing Economy Services in a Smart City ----- _Sustainability 2022, 14, 7609_ 23 of 30 **Table 3. Cont.** **Ref.** **Focus Area** **Edge AI Use Case** **Blockchain Use Case** Blockchain Energized [130] Edge Intelligence for Beyond 5G Networks Edge Intelligence using a [139] federated learning Blockchain network [136] Knowledge sharing in IoV Federated Learning With [137] Blockchain for Autonomous Vehicles. Messages dissemination in [138] the IoV AI-Chain, a Blockchain-based edge intelligence for B5G networks. Each node at the edge trains neural network components and applies AI-Chain to share its learning results. iFLBC, a federated learning framework called to bring AI to edge nodes through a shared machine learning model. powered by Blockchain technology. The relevant data of an edge node is used to train a model, which is then aggregated with models trained by other edge nodes to generate a shared model. Hierarchical federated learning. Vehicles and RSUs learn surrounding data through machine learning methods and share learning knowledge. Aims to meet the distributed model and privacy requirements of IoVs. Federated learning framework. The framework extends the reach of its federation to untrustworthy public network vehicles via a validation process of local training modules. Blockchain-assisted federated learning solution for message broadcasting. The Stackelberg game further analyzes the business model to incentivize vehicles to be involved in federated learning and message delivery. An immutable and distributed record of local learning outcomes that lays the foundation for sharing information between edge nodes. The shared model is stored on the blockchain. Network members download the aggregated model to provide intelligent services to end-users. Hierarchical blockchain. Knowledge sharing is modeled as a trading market process to drive sharing behaviors. The trading process is formulated as a multi-leader, multi-player game. On-board local learning modules exchange and verify their updates in a fully decentralized manner without any centralized coordination by leveraging the blockchain consensus mechanism. Vehicles compete to become a relay (minor) node by processing the Proof-of-Federated-Learning (PoFL) consensus embedded in the smart contract of the blockchain. **8. Open Research Issues** The research initiatives reported above represent attempts to mitigate the challenges of implementing edge AI and Blockchain in two key areas of smart cities, smart mobility and smart energy. However, there remain unresolved challenges. This section examines four potential prospective research trends for future implementation. - **Collaboration and data exchange. As we described earlier, since multiple nodes with** heterogeneous edge devices generate isolated knowledge, collaboration and data exchange between nodes are essential to building intelligent applications and services for smart mobility and smart energy. Storing, sharing, querying, and exchanging data training models require additional security and privacy measures. Blockchain technology helps meet these requirements. However, edge devices with limited storage may not be able to store the training model or the blockchain structure that grows as transaction blocks are added to the blockchain. Moreover, it is common for edge devices to store distributed ledger data that are not even useful for their transactions. Therefore, cutting-edge blockchain-specific equipment or platforms to support decentralized blockchain data storage are required. ----- _Sustainability 2022, 14, 7609_ 24 of 30 - **Impact of edge connections on Blockchain-enabled smart mobility. In a smart mo-** bility scenario, edge devices on connected vehicles, for example, are often connected to other edge devices or cloud servers through unreliable wireless channels. As we discussed earlier, Blockchain can facilitate communication between connected vehicles and the road infrastructure by considering data exchange requests as transactions to be stored and retrieved from a blockchain database. Due to the inevitable network delays, a vehicle participating in the blockchain may not receive the most recent block. It may then create an alternative chain that branches off the main chain. This problem is known as the forking problem. It can also arise when edge nodes disagree on the same learning model and forked chains emerge. Such forking reduces throughput because only one chain survives, ultimately, while all other blocks in different chains are removed. Further research in this area is needed. - **Prediction of future energy production and system load. In P2P smart energy trad-** ing scenarios, the decentralization of prosumers brings many issues. Blockchain helps to ensure trust and privacy between all players in the energy market. Smart contracts and learning models at participating nodes should help predict future energy production and system load, allowing prosumers to make optimal decisions about sharing and pricing their energy. Further research on federated learning models for energy trading and pricing is needed. - **Energy efficiency. Incorporating AI in edge devices is challenging because of the** power-hungry features of deep learning algorithms, such as convolutional neural networks (CNNs). Therefore, energy efficiency is a critical issue for edge AI applications. Some research efforts investigated the usage of reservoir computing as an alternative, which promises to provide good performance while exhibiting low-power characteristics [140]. Additionally, with the growing calls for the application of rigid environmental standards and the rapidly rising energy costs, smart cities increasingly take the energy efficiency issue more seriously. However, some Blockchain consensus mechanisms such as PoW (Proof of Work) are computationally expensive as blockchain nodes perform complex computations to mine the next block. PoW is not an energy-efficient approach and consumes a large amount of electricity due to computation redundancy. Researchers are developing alternative less computationally expensive consensus mechanisms for blockchain systems. Although highly promising, these consensus mechanisms are still in their infancy and suffer from scalability issues, and their security has not been rigorously investigated. Therefore, further research is needed concerning the design of energy-efficient edge AI applications and consensus mechanisms for blockchain systems. **9. Conclusions** Smart cities face several challenges due to population growth and migratory waves. This article examines the current and potential contributions of edge AI and Blockchain technology in coping with smart city challenges through the lens of sustainability in two main areas, which are smart mobility and smart energy. It contributes to the sustainability literature by identifying and bringing together recent research on edge AI and Blockchain, highlighting their positive impacts and potential implications on smart cities. This review highlights the existing and potential convergence of edge AI and Blockchain. It shows that edge AI and Blockchain technology can help address the problem of traffic congestion and management by automating the detection, counting, and identification of vehicle speeds. Furthermore, these technologies can help establish trustworthy communications and energy trading between vehicles and reliable and secure distributed smart energy management. Finally, this article discusses potential research trends for future implementations of edge AI and Blockchain to provide innovative solutions in smart mobility and smart energy. It is expected that this review will serve as a guideline for future research on the adoption of edge AI and Blockchain in other areas of smart cities. ----- _Sustainability 2022, 14, 7609_ 25 of 30 **Funding: This work is supported by the UAEU Program for Advanced Research Grant N. G00003443.** **Institutional Review Board Statement: Not applicable.** **Informed Consent Statement: Not applicable.** **Data Availability Statement: Not applicable.** **Conflicts of Interest: The author declares no conflict of interest.** **References** 1. Hugo Priemus, S.D. Climate Change and Sustainable Cities; Taylor & Francis: Andover, UK, 2016. 2. Grimmond, C.S.B.; Roth, M.; Oke, T.R.; Au, Y.C.; Best, M.; Betts, R.; Carmichael, G.; Cleugh, H.; Dabberdt, W.; Emmanuel, R.; et al. Climate and More Sustainable Cities: Climate Information for Improved Planning and Management of Cities (Produc[ers/Capabilities Perspective). Procedia Environ. Sci. 2010, 1, 247–274. [CrossRef]](http://doi.org/10.1016/j.proenv.2010.09.016) 3. Albert, S. Innovative Solutions for Creating Sustainable Cities; Cambridge Scholars Publishing: Newcastle upon Tyne, UK, 2019. 4. Mondejar, M.E.; Avtar, R.; Diaz, H.L.B.; Dubey, R.K.; Esteban, J.; Gómez-Morales, A.; Hallam, B.; Mbungu, N.T.; Okolo, C.C.; Prasad, K.A.; et al. Digitalization to achieve sustainable development goals: Steps towards a Smart Green Planet. Sci. Total _[Environ. 2021, 794, 148539. [CrossRef] [PubMed]](http://dx.doi.org/10.1016/j.scitotenv.2021.148539)_ 5. [Shi, W.; Cao, J.; Zhang, Q.; Li, Y.; Xu, L. Edge Computing: Vision and Challenges. IEEE IoT J. 2016, 3, 637–646. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2016.2579198) 6. [Shi, W.; Dustdar, S. The Promise of Edge Computing. Computer 2016, 49, 78–81. [CrossRef]](http://dx.doi.org/10.1109/MC.2016.145) 7. [Satyanarayanan, M. The Emergence of Edge Computing. Computer 2017, 50, 30–39. [CrossRef]](http://dx.doi.org/10.1109/MC.2017.9) 8. [Abbas, N.; Zhang, Y.; Taherkordi, A.; Skeie, T. Mobile Edge Computing: A Survey. IEEE IoT J. 2017, 5, 450–465. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2017.2750180) 9. Xu, D.; Li, T.; Li, Y.; Su, X.; Tarkoma, S.; Jiang, T.; Crowcroft, J.; Hui, P. Edge Intelligence: Architectures, Challenges, and Applications. arXiv 2020, arXiv:2003.12172. 10. Wang, X.; Han, Y.; Leung, V.C.M.; Niyato, D.; Yan, X.; Chen, X. Convergence of Edge Computing and Deep Learning: [A Comprehensive Survey. IEEE Commun. Surv. Tutor. 2020, 22, 869–904. [CrossRef]](http://dx.doi.org/10.1109/COMST.2020.2970550) 11. Kowalski, M.; Lee, Z.W.Y.; Chan, T.K.H. Blockchain technology and trust relationships in trade finance. Technol. Forecast. Soc. _[Chang. 2021, 166, 120641. [CrossRef]](http://dx.doi.org/10.1016/j.techfore.2021.120641)_ 12. Werbach, K. The Blockchain and the New Architecture of Trust (Information Policy); The MIT Press: London, UK 2018. 13. Gaggioli, A.; Eskandari, S.; Cipresso, P.; Lozza, E. The Middleman Is Dead, Long Live the Middleman: The “Trust Factor” and the [Psycho-Social Implications of Blockchain. Front. Blockchain 2019, 2. [CrossRef]](http://dx.doi.org/10.3389/fbloc.2019.00020) 14. Shala, B.; Trick, U.; Lehmann, A.; Ghita, B.; Shiaeles, S. Blockchain and Trust for Secure, End-User-Based and Decentralized IoT [Service Provision. IEEE Access 2020, 8, 119961–119979. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.3005541) 15. Bashir, I. Mastering Blockchain: Distributed Ledger Technology, Decentralization, and Smart Contracts Explained, 2nd ed.; Packt Publishing: Mumbai, India, 2018. 16. Konstantinidis, I.; Siaminos, G.; Timplalexis, C.; Zervas, P.; Peristeras, V.; Decker, S. Blockchain for Business Applications: A [Systematic Literature Review. In Business Information Systems; Springer: Cham, Switzerland, 2018; pp. 384–399._28. [CrossRef]](http://dx.doi.org/10.1007/978-3-319-93931-5_28) 17. Ølnes, S.; Ubacht, J.; Janssen, M. Blockchain in government: Benefits and implications of distributed ledger technology for [information sharing. Gov. Inf. Q. 2017, 34, 355–364. [CrossRef]](http://dx.doi.org/10.1016/j.giq.2017.09.007) 18. [Shen, C.; Pena-Mora, F. Blockchain for Cities—A Systematic Literature Review. IEEE Access 2018, 6, 76787–76819. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2018.2880744) 19. Batty, M.J.; Axhausen, K.W.; Giannotti, F.; Pozdnoukhov, A.; Bazzani, A.; Wachowicz, M.; Ouzounis, G.K.; Portugali, Y. Smart [cities of the future. Eur. Phys. J. Spec. Top. 2012, 214, 481–518. [CrossRef]](http://dx.doi.org/10.1140/epjst/e2012-01703-3) 20. [Law, K.H.; Lynch, J.P. Smart City: Technologies and Challenges. IT Prof. 2019, 21, 46–51. [CrossRef]](http://dx.doi.org/10.1109/MITP.2019.2935405) 21. Silva, B.N.; Khan, M.; Jung, C.; Seo, J.; Muhammad, D.; Han, J.; Yoon, Y.; Han, K. Urban Planning and Smart City Decision [Management Empowered by Real-Time Data Processing Using Big Data Analytics. Sensors 2018, 18, 2994. [CrossRef]](http://dx.doi.org/10.3390/s18092994) 22. [Perätalo, S.; Ahokangas, P. Toward Smart City Business Models. J. Bus. Model. 2018, 6, 65–70. [CrossRef]](http://dx.doi.org/10.5278/ojs.jbm.v6i2.2466) 23. [Smart Cities Initiatives Around the World are Improving Citizens’ Lives. 2020. Available online: https://www.dronedek.com/](https://www.dronedek.com/news/smart-cities-initiatives-around-the-world-are-improving-citizens-lives/) [news/smart-cities-initiatives-around-the-world-are-improving-citizens-lives/ (accessed on 4 January 2022).](https://www.dronedek.com/news/smart-cities-initiatives-around-the-world-are-improving-citizens-lives/) 24. Tian, S.; Yang, W.; Grange, J.M.L.; Wang, P.; Huang, W.; Ye, Z. Smart healthcare: Making medical care more intelligent. Glob. _[Health J. 2019, 3, 62–65. [CrossRef]](http://dx.doi.org/10.1016/j.glohj.2019.07.001)_ 25. Zhu, H.; Wu, C.K.; Koo, C.H.; Tsang, Y.T.; Liu, Y.; Chi, H.R.; Tsang, K.F. Smart Healthcare in the Era of Internet-of-Things. IEEE _[Consum. Electron. Mag. 2019, 8, 26–30. [CrossRef]](http://dx.doi.org/10.1109/MCE.2019.2923929)_ 26. [Bodhani, A. Smart transport. Eng. Technol. 2012, 7, 70–73. [CrossRef]](http://dx.doi.org/10.1049/et.2012.0611) 27. Elsagheer Mohamed, S.A.; AlShalfan, K.A. Intelligent Traffic Management System Based on the Internet of Vehicles (IoV). J. Adv. _[Transp. 2021, 2021, 4037533. [CrossRef]](http://dx.doi.org/10.1155/2021/4037533)_ 28. Jimenez, J.A. Smart Transportation Systems. In Smart Cities Applications, Technologies, Standards, and Driving Factors; Springer: [Cham, Switzerland, 2017; pp. 123–133._8. [CrossRef]](http://dx.doi.org/10.1007/978-3-319-59381-4_8) 29. Tao, F.; Qi, Q.; Liu, A.; Kusiak, A. Data-driven smart manufacturing. _J. Manuf._ _Syst._ **2018, 48, 157–169.** [j.jmsy.2018.01.006. [CrossRef]](http://dx.doi.org/10.1016/j.jmsy.2018.01.006) ----- _Sustainability 2022, 14, 7609_ 26 of 30 30. Wang, J.; Ma, Y.; Zhang, L.; Gao, R.X.; Wu, D. Deep learning for smart manufacturing: Methods and applications. J. Manuf. Syst. **[2018, 48, 144–156. [CrossRef]](http://dx.doi.org/10.1016/j.jmsy.2018.01.003)** 31. Huang, Z.; Shen, Y.; Li, J.; Fey, M.; Brecher, C. A Survey on AI-Driven Digital Twins in Industry 4.0: Smart Manufacturing and [Advanced Robotics. Sensors 2021, 21, 6340. [CrossRef] [PubMed]](http://dx.doi.org/10.3390/s21196340) 32. Al Dakheel, J.; Del Pero, C.; Aste, N.; Leonforte, F. Smart buildings features and key performance indicators: A review. Sustain. _[Cities Soc. 2020, 61, 102328. [CrossRef]](http://dx.doi.org/10.1016/j.scs.2020.102328)_ 33. Lund, H.; Østergaard, P.A.; Connolly, D.; Mathiesen, B.V. Smart energy and smart energy systems. Energy 2017, 137, 556–565. [[CrossRef]](http://dx.doi.org/10.1016/j.energy.2017.05.123) 34. Farooq, M.S.; Riaz, S.; Abid, A.; Abid, K.; Naeem, M.A. A Survey on the Role of IoT in Agriculture for the Implementation of [Smart Farming. IEEE Access 2019, 7, 156237–156271. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2949703) 35. Zamora-Izquierdo, M.A.; Santa, J.; Martínez, J.A.; Martínez, V.; Skarmeta, A.F. Smart farming IoT platform based on edge and [cloud computing. Biosyst. Eng. 2019, 177, 4–17. [CrossRef]](http://dx.doi.org/10.1016/j.biosystemseng.2018.10.014) 36. Panori, A.; Kakderi, C.; Komninos, N.; Fellnhofer, K.; Reid, A.; Mora, L. Smart systems of innovation for smart places: Challenges [in deploying digital platforms for co-creation and data-intelligence. Land Use Policy 2021, 111, 104631. [CrossRef]](http://dx.doi.org/10.1016/j.landusepol.2020.104631) 37. Ai, Y.; Peng, M.; Zhang, K. Edge computing technologies for Internet of Things: A primer. Digit. Commun. Netw. 2018, 4, 77–86. [[CrossRef]](http://dx.doi.org/10.1016/j.dcan.2017.07.001) 38. [Sun, X.; Ansari, N. EdgeIoT: Mobile Edge Computing for the Internet of Things. IEEE Commun. Mag. 2016, 54, 22–29. [CrossRef]](http://dx.doi.org/10.1109/MCOM.2016.1600492CM) 39. Alnoman, A.; Sharma, S.K.; Ejaz, W.; Anpalagan, A. Emerging Edge Computing Technologies for Distributed IoT Systems. IEEE _[Netw. 2019, 33, 140–147. [CrossRef]](http://dx.doi.org/10.1109/MNET.2019.1800543)_ 40. Deng, S.; Zhao, H.; Fang, W.; Yin, J.; Dustdar, S.; Zomaya, A.Y. Edge Intelligence: The Confluence of Edge Computing and [Artificial Intelligence. IEEE IoT J. 2020, 7, 7457–7469. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2020.2984887) 41. Sun, W.; Liu, J.; Yue, Y. AI-Enhanced Offloading in Edge Computing: When Machine Learning Meets Industrial IoT. IEEE Netw. **[2019, 33, 68–74. [CrossRef]](http://dx.doi.org/10.1109/MNET.001.1800510)** 42. Jung, S.; Kim, Y.; Hwang, E. Real-time car tracking system based on surveillance videos. J. Image Video Proc. 2018, 2018, 1–13. [[CrossRef]](http://dx.doi.org/10.1186/s13640-018-0374-7) 43. Han, X.; Zhang, Z.; Ding, N.; Gu, Y.; Liu, X.; Huo, Y.; Qiu, J.; Zhang, L.; Han, W.; Huang, M.; et al. Pre-Trained Models: Past, [Present and Future. AI Open 2021, 2, 225–250. [CrossRef]](http://dx.doi.org/10.1016/j.aiopen.2021.08.002) 44. González García, C.; Meana-Llorián, D.; Pelayo G.; Bustelo, B.C.; Cueva Lovelle, J.M.; Garcia-Fernandez, N. Midgar: Detection of people through computer vision in the Internet of Things scenarios to improve the security in Smart Cities, Smart Towns, and [Smart Homes. Future Gener. Comput. Syst. 2017, 76, 301–313. [CrossRef]](http://dx.doi.org/10.1016/j.future.2016.12.033) 45. Ho, G.T.S.; Tsang, Y.P.; Wu, C.H.; Wong, W.H.; Choy, K.L. A Computer Vision-Based Roadside Occupation Surveillance System [for Intelligent Transport in Smart Cities. Sensors 2019, 19, 1796. [CrossRef]](http://dx.doi.org/10.3390/s19081796) 46. Mittal, V.; Bhushan, B. Accelerated Computer Vision Inference with AI on the Edge. In Proceedings of the 2020 IEEE 9th International Conference on Communication Systems and Network Technologies (CSNT), Gwalior, India, 10–12 April 2020; [pp. 55–60. [CrossRef]](http://dx.doi.org/10.1109/CSNT48778.2020.9115770) 47. Ullah, Z.; Al-Turjman, F.; Mostarda, L.; Gagliardi, R. Applications of Artificial Intelligence and Machine learning in smart cities. _[Comput. Commun. 2020, 154, 313–323. [CrossRef]](http://dx.doi.org/10.1016/j.comcom.2020.02.069)_ 48. Shi, Y.; Yang, K.; Jiang, T.; Zhang, J.; Letaief, K.B. Communication-Efficient Edge AI: Algorithms and Systems. IEEE Commun. _[Surv. Tutor. 2020, 22, 2167–2191. [CrossRef]](http://dx.doi.org/10.1109/COMST.2020.3007787)_ 49. Zhou, Z.; Chen, X.; Li, E.; Zeng, L.; Luo, K.; Zhang, J. Edge Intelligence: Paving the Last Mile of Artificial Intelligence With Edge [Computing. Proc. IEEE 2019, 107, 1738–1762. [CrossRef]](http://dx.doi.org/10.1109/JPROC.2019.2918951) 50. Hsu, K.C.; Tseng, H.W. Accelerating applications using edge tensor processing units. In SC ’21: Proceedings of the International _Conference for High Performance Computing, Networking, Storage and Analysis; Association for Computing Machinery: New York,_ [NY, USA, 2021; pp. 1–14. [CrossRef]](http://dx.doi.org/10.1145/3458817.3476177) 51. Edge TPU–Run Inference at the Edge|Google Cloud. 2021. Available online: [https://cloud.google.com/edge-tpu](https://cloud.google.com/edge-tpu) (accessed on 5 January 2022). 52. NVIDIA Jetson Nano For Edge AI Applications and Education. 2022. Available online: [https://www.nvidia.com/en-us/](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-nano) [autonomous-machines/embedded-systems/jetson-nano (accessed on 5 January 2022).](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-nano) 53. [NVIDIA Jetson TX2: High Performance AI at the Edge. 2022. Available online: https://www.nvidia.com/en-us/autonomous-](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-tx2) [machines/embedded-systems/jetson-tx2 (accessed on 5 January 2022).](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-tx2) 54. Intel[®] Movidius™ [Vision Processing Units (VPUs). 2022. Available online: https://www.intel.com/content/www/us/en/](https://www.intel.com/content/www/us/en/products/details/processors/movidius-vpu.html) [products/details/processors/movidius-vpu.html (accessed on 5 January 2022).](https://www.intel.com/content/www/us/en/products/details/processors/movidius-vpu.html) 55. [Voigt, P.; von dem Bussche, A. The EU General Data Protection Regulation (GDPR); Springer: Cham, Switzerland, 2017. [CrossRef]](http://dx.doi.org/10.1007/978-3-319-57959-7) 56. Annas, G.J. HIPAA Regulations: A New Era of Medical-Record Privacy? Sch. Commons Boston Univ. Sch. Law 2003, 348, 1486. [[CrossRef] [PubMed]](http://dx.doi.org/10.1056/NEJMlim035027) 57. Yang, Q.; Liu, Y.; Chen, T.; Tong, Y. Federated Machine Learning: Concept and Applications. ACM Trans. Intell. Syst. Technol. **[2019, 10, 1–19. [CrossRef]](http://dx.doi.org/10.1145/3298981)** 58. Li, T.; Sahu, A.K.; Talwalkar, A.; Smith, V. Federated Learning: Challenges, Methods, and Future Directions. IEEE Signal Process. _[Mag. 2020, 37, 50–60. [CrossRef]](http://dx.doi.org/10.1109/MSP.2020.2975749)_ ----- _Sustainability 2022, 14, 7609_ 27 of 30 59. Nilsson, A.; Smith, S.; Ulm, G.; Gustavsson, E.; Jirstrand, M. A Performance Evaluation of Federated Learning Algorithms. In _DIDL ’18: Proceedings of the Second Workshop on Distributed Infrastructures for Deep Learning; Association for Computing Machinery:_ [New York, NY, USA, 2018; pp. 1–8. [CrossRef]](http://dx.doi.org/10.1145/3286490.3286559) 60. Sarma, K.V.; Harmon, S.; Sanford, T.; Roth, H.R.; Xu, Z.; Tetreault, J.; Xu, D.; Flores, M.G.; Raman, A.G.; Kulkarni, R.; et al. Federated learning improves site performance in multicenter deep learning without data sharing. J. Am. Med. Inform. Assoc. 2021, _[28, 1259–1264. [CrossRef]](http://dx.doi.org/10.1093/jamia/ocaa341)_ 61. Rauchs, M.; Glidden, A.; Gordon, B.; Pieters, G.C.; Recanatini, M.; Rostand, F.; Vagneur, K.; Zhang, B.Z. Distributed ledger technol[ogy systems: A conceptual framework. 2018. Available online: https://ssrn.com/abstract=3230013 (accessed on 15 January 2022).](https://ssrn.com/abstract=3230013) [[CrossRef]](http://dx.doi.org/10.2139/ssrn.3230013) 62. [Sanders, A. The subprime crisis and its role in the financial crisis. J. Hous. Econ. 2008, 17, 254–261. [CrossRef]](http://dx.doi.org/10.1016/j.jhe.2008.10.001) 63. Nakamoto, S. Bitcoin: A peer-to-peer electronic cash system. Decentralized Business Review 2008; p. 21260. Available online: [https://www.debr.io/article/21260.pdf (accessed on 15 January 2022)](https://www.debr.io/article/21260.pdf) 64. Xie, J.; Tang, H.; Huang, T.; Yu, F.R.; Xie, R.; Liu, J.; Liu, Y. A survey of blockchain technology applied to smart cities: Research [issues and challenges. IEEE Commun. Surv. Tutor. 2019, 21, 2794–2830. [CrossRef]](http://dx.doi.org/10.1109/COMST.2019.2899617) 65. Upadhyay, A.; Mukhuty, S.; Kumar, V.; Kazancoglu, Y. Blockchain technology and the circular economy: Implications for [sustainability and social responsibility. J. Clean. Prod. 2021, 293, 126130. [CrossRef]](http://dx.doi.org/10.1016/j.jclepro.2021.126130) 66. Teisserenc, B.; Sepasgozar, S. Adoption of Blockchain Technology through Digital Twins in the Construction Industry 4.0: A [PESTELS Approach. Buildings 2021, 11, 670. [CrossRef]](http://dx.doi.org/10.3390/buildings11120670) 67. Xiong, Z.; Zhang, Y.; Niyato, D.; Wang, P.; Han, Z. When Mobile Blockchain Meets Edge Computing. IEEE Commun. Mag. 2018, _[56, 33–39. [CrossRef]](http://dx.doi.org/10.1109/MCOM.2018.1701095)_ 68. Guo, S.; Hu, X.; Guo, S.; Qiu, X.; Qi, F. Blockchain Meets Edge Computing: A Distributed and Trusted Authentication System. _[IEEE Trans. Ind. Inform. 2019, 16, 1972–1983. [CrossRef]](http://dx.doi.org/10.1109/TII.2019.2938001)_ 69. Benevolo, C.; Dameri, R.P.; D’Auria, B. Smart Mobility in Smart City. In Empowering Organizations; Springer: Cham, Switzerland, [2015; pp. 13–28._2. [CrossRef]](http://dx.doi.org/10.1007/978-3-319-23784-8_2) 70. Barthélemy, J.; Verstaevel, N.; Forehead, H.; Perez, P. Edge-Computing Video Analytics for Real-Time Traffic Monitoring in a [Smart City. Sensors 2019, 19, 2048. [CrossRef] [PubMed]](http://dx.doi.org/10.3390/s19092048) 71. de Souza, A.M.; Brennand, C.A.; Yokoyama, R.S.; Donato, E.A.; Madeira, E.R.M.; Villas, L.A. Traffic management systems: A [classification, review, challenges, and future perspectives. Int. J. Distrib. Sens. Netw. 2017, 13, 1550147716683612. [CrossRef]](http://dx.doi.org/10.1177/1550147716683612) 72. Redmon, J.; Farhadi, A. YOLOv3: An Incremental Improvement. arXiv 2018, arXiv:1804.02767. 73. Bewley, A.; Ge, Z.; Ott, L.; Ramos, F.; Upcroft, B. Simple online and realtime tracking. In Proceedings of the 2016 IEEE [International Conference on Image Processing (ICIP), Phoenix, AZ, USA, 25–28 September 2016; pp. 3464–3468. [CrossRef]](http://dx.doi.org/10.1109/ICIP.2016.7533003) 74. Dinh, D.L.; Nguyen, H.N.; Thai, H.T.; Le, K.H. Towards AI-Based Traffic Counting System with Edge Computing. J. Adv. Transp. **[2021, 2021, 1–15. [CrossRef]](http://dx.doi.org/10.1155/2021/5551976)** 75. Kumar, T.; Kushwaha, D.S. An Efficient Approach for Detection and Speed Estimation of Moving Vehicles. Procedia Comput. Sci. **[2016, 89, 726–731. [CrossRef]](http://dx.doi.org/10.1016/j.procs.2016.06.045)** 76. Song, H.; Liang, H.; Li, H.; Dai, Z.; Yun, X. Vision-based vehicle detection and counting system using deep learning in highway [scenes. Eur. Transp. Res. Rev. 2019, 11, 51. [CrossRef]](http://dx.doi.org/10.1186/s12544-019-0390-4) 77. Rublee, E.; Rabaud, V.; Konolige, K.; Bradski, G. ORB: An efficient alternative to SIFT or SURF. In Proceedings of the 2011 International Conference on Computer Vision, Barcelona, Spain, 6–13 November 2011; pp. 2564–2571. 78. Letnik, T.; Mencinger, M.; Peruš, I. Flexible Assignment of Loading Bays for Efficient Vehicle Routing in Urban Last Mile Delivery. _[Sustainability 2020, 12, 7500. [CrossRef]](http://dx.doi.org/10.3390/su12187500)_ 79. Liao, R. Smart Mobility: Challenges and Trends. In Toward Sustainable and Economic Smart Mobility; World Scientific (Europe): [London, UK, 2019; pp. 1–11._0001. [CrossRef]](http://dx.doi.org/10.1142/9781786347862_0001) 80. Bagloee, S.A.; Heshmati, M.; Dia, H.; Ghaderi, H.; Pettit, C.; Asadi, M. Blockchain: The operating system of smart cities. Cities **[2021, 112, 103104. [CrossRef]](http://dx.doi.org/10.1016/j.cities.2021.103104)** 81. van der Heijden, R.W.; Engelmann, F.; Mödinger, D.; Schönig, F.; Kargl, F. Blackchain: Scalability for resource-constrained accountable vehicle-to-x communication. In SERIAL ’17: Proceedings of the 1st Workshop on Scalable and Resilient Infrastructures for _[Distributed Ledgers; Association for Computing Machinery: New York, NY, USA, 2017; pp. 1–5. [CrossRef]](http://dx.doi.org/10.1145/3152824.3152828)_ 82. Sharma, P.K.; Moon, S.Y.; Park, J.H. Block-VN: A distributed blockchain based vehicular network architecture in smart city. J. Inf. _Process. Syst. 2017, 13, 184–195._ 83. Yang, Z.; Zheng, K.; Yang, K.; Leung, V.C. A blockchain-based reputation system for data credibility assessment in vehicular networks. In Proceedings of the 2017 IEEE 28th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC), Montreal, QC, Canada, 8–13 October 2017; pp. 1–5. 84. Lei, A.; Cruickshank, H.; Cao, Y.; Asuquo, P.; Ogah, C.P.A.; Sun, Z. Blockchain-Based Dynamic Key Management for Heteroge[neous Intelligent Transportation Systems. IEEE Internet Things J. 2017, 4, 1832–1843. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2017.2740569) 85. Ma, Z.; Zhang, J.; Guo, Y.; Liu, Y.; Liu, X.; He, W. An Efficient Decentralized Key Management Mechanism for VANET With [Blockchain. IEEE Trans. Veh. Technol. 2020, 69, 5836–5849. [CrossRef]](http://dx.doi.org/10.1109/TVT.2020.2972923) 86. Yang, Z.; Yang, K.; Lei, L.; Zheng, K.; Leung, V.C.M. Blockchain-Based Decentralized Trust Management in Vehicular Networks. _[IEEE Internet Things J. 2019, 6, 1495–1505. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2018.2836144)_ ----- _Sustainability 2022, 14, 7609_ 28 of 30 87. Arora, S.K.; Kumar, G.; Kim, T.H. Blockchain Based Trust Model Using Tendermint in Vehicular Adhoc Networks. Appl. Sci. **[2021, 11, 1998. [CrossRef]](http://dx.doi.org/10.3390/app11051998)** 88. Luo, B.; Li, X.; Weng, J.; Guo, J.; Ma, J. Blockchain Enabled Trust-Based Location Privacy Protection Scheme in VANET. IEEE _[Trans. Veh. Technol. 2020, 69, 2034–2048. [CrossRef]](http://dx.doi.org/10.1109/TVT.2019.2957744)_ 89. Huang, Z.; Li, Z.; Lai, C.S.; Zhao, Z.; Wu, X.; Li, X.; Tong, N.; Lai, L.L. A Novel Power Market Mechanism Based on Blockchain [for Electric Vehicle Charging Stations. Electronics 2021, 10, 307. [CrossRef]](http://dx.doi.org/10.3390/electronics10030307) 90. Ferreira, J.C.; Ferreira da Silva, C.; Martins, J.P. Roaming Service for Electric Vehicle Charging Using Blockchain-Based Digital [Identity. Energies 2021, 14, 1686. [CrossRef]](http://dx.doi.org/10.3390/en14061686) 91. Gorenflo, C.; Golab, L.; Keshav, S. Mitigating Trust Issues in Electric Vehicle Charging using a Blockchain. In e-Energy ’19: _Proceedings of the Tenth ACM International Conference on Future Energy Systems; Association for Computing Machinery: New York,_ [NY, USA, 2019; pp. 160–164. [CrossRef]](http://dx.doi.org/10.1145/3307772.3328283) 92. Löffler, M.; Mokwa, C.; Münstermann, B.; Wojciak, J. Shifting gears: Insurers adjust for connected-car ecosystems. In Digital _Mckinsey; Mckinsey & Company: Austin, TX, USA, 2016; pp. 1–13._ 93. Yuan, Y.; Wang, F.Y. Towards blockchain-based intelligent transportation systems. In Proceedings of the 2016 IEEE 19th International Conference on Intelligent Transportation Systems (ITSC), Rio de Janeiro, Brazil, 1–4 November 2016; pp. 2663–2668. [[CrossRef]](http://dx.doi.org/10.1109/ITSC.2016.7795984) 94. Shah, A.S.; Nasir, H.; Fayaz, M.; Lajis, A.; Shah, A. A Review on Energy Consumption Optimization Techniques in IoT Based [Smart Building Environments. Information 2019, 10, 108. [CrossRef]](http://dx.doi.org/10.3390/info10030108) 95. Zhang, X.; Manogaran, G.; Muthu, B. IoT enabled integrated system for green energy into smart cities. Sustain. Energy Technol. _[Assess. 2021, 46, 101208. [CrossRef]](http://dx.doi.org/10.1016/j.seta.2021.101208)_ 96. Abdel-Basset, M.; Hawash, H.; Chakrabortty, R.K.; Ryan, M. Energy-Net: A Deep Learning Approach for Smart Energy [Management in IoT-Based Smart Cities. IEEE Internet Things J. 2021, 8, 12422–12435. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2021.3063677) 97. Liu, Y.; Yang, C.; Jiang, L.; Xie, S.; Zhang, Y. Intelligent Edge Computing for IoT-Based Energy Management in Smart Cities. IEEE _[Netw. 2019, 33, 111–117. [CrossRef]](http://dx.doi.org/10.1109/MNET.2019.1800254)_ 98. Pop, C.; Cioara, T.; Antal, M.; Anghel, I.; Salomie, I.; Bertoncini, M. Blockchain Based Decentralized Management of Demand [Response Programs in Smart Energy Grids. Sensors 2018, 18, 162. [CrossRef] [PubMed]](http://dx.doi.org/10.3390/s18010162) 99. Pee, S.J.; Kang, E.S.; Song, J.G.; Jang, J.W. Blockchain based smart energy trading platform using smart contract. In Proceedings of the 2019 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), Okinawa, Japan, 11–13 [February 2019; pp. 322–325. [CrossRef]](http://dx.doi.org/10.1109/ICAIIC.2019.8668978) 100. Hua, W.; Sun, H. A Blockchain-Based Peer-to-Peer Trading Scheme Coupling Energy and Carbon Markets. In Proceedings of the 2019 International Conference on Smart Energy Systems and Technologies (SEST), Porto, Portugal, 9–11 September 2019; pp. 1–6. [[CrossRef]](http://dx.doi.org/10.1109/SEST.2019.8849111) 101. Andoni, M.; Robu, V.; Flynn, D.; Abram, S.; Geach, D.; Jenkins, D.; McCallum, P.; Peacock, A. Blockchain technology in the energy [sector: A systematic review of challenges and opportunities. Renew. Sustain. Energy Rev. 2019, 100, 143–174. [CrossRef]](http://dx.doi.org/10.1016/j.rser.2018.10.014) 102. Hasankhani, A.; Mehdi Hakimi, S.; Bisheh-Niasar, M.; Shafie-khah, M.; Asadolahi, H. Blockchain technology in the future smart [grids: A comprehensive review and frameworks. Int. J. Electr. Power Energy Syst. 2021, 129, 106811. [CrossRef]](http://dx.doi.org/10.1016/j.ijepes.2021.106811) 103. Casino, F.; Dasaklis, T.K.; Patsakis, C. A systematic literature review of blockchain-based applications: Current status, classification [and open issues. Telemat. Inform. 2019, 36, 55–81. [CrossRef]](http://dx.doi.org/10.1016/j.tele.2018.11.006) 104. Yang, X.; Wang, G.; He, H.; Lu, J.; Zhang, Y. Automated Demand Response Framework in ELNs: Decentralized Scheduling and [Smart Contract. IEEE Trans. Syst. Man, Cybern. Syst. 2019, 50, 58–72. [CrossRef]](http://dx.doi.org/10.1109/TSMC.2019.2903485) 105. Kumar, N.M. Blockchain: Enabling wide range of services in distributed energy system. Beni-Suef Univ. J. Basic Appl. Sci. 2018, _[7, 701–704. [CrossRef]](http://dx.doi.org/10.1016/j.bjbas.2018.08.003)_ 106. Morstyn, T.; Farrell, N.; Darby, S.J.; McCulloch, M.D. Using peer-to-peer energy-trading platforms to incentivize prosumers to [form federated power plants - Nature Energy. Nat. Energy 2018, 3, 94–101. [CrossRef]](http://dx.doi.org/10.1038/s41560-017-0075-y) 107. Wang, L.; Liu, J.; Yuan, R.; Wu, J.; Zhang, D.; Zhang, Y.; Li, M. Adaptive bidding strategy for real-time energy management in [multi-energy market enhanced by blockchain. Appl. Energy 2020, 279, 115866. [CrossRef]](http://dx.doi.org/10.1016/j.apenergy.2020.115866) 108. Mylrea, M.; Gourisetti, S.N.G. Blockchain for smart grid resilience: Exchanging distributed energy at speed, scale and security. In Proceedings of the 2017 Resilience Week (RWS), Wilmington, DE, USA, 18–22 September 2017; pp. 18–23. 109. Mengelkamp, E.; Gärttner, J.; Rock, K.; Kessler, S.; Orsini, L.; Weinhardt, C. Designing microgrid energy markets: A case study: [The Brooklyn Microgrid. Appl. Energy 2018, 210, 870–880. [CrossRef]](http://dx.doi.org/10.1016/j.apenergy.2017.06.054) 110. Xie, P.; Yan, W.; Xuan, P.; Zhu, J.; Wu, Y.; Li, X.; Zou, J. Conceptual Framework of Blockchain-based Electricity Trading for Neighborhood Renewable Energy. In Proceedings of the 2018 2nd IEEE Conference on Energy Internet and Energy System [Integration (EI2), Beijing, China, 20–22 October 2018; pp. 1–5. [CrossRef]](http://dx.doi.org/10.1109/EI2.2018.8581887) 111. Kang, E.S.; Pee, S.J.; Song, J.G.; Jang, J.W. A Blockchain-Based Energy Trading Platform for Smart Homes in a Microgrid. In Proceedings of the 2018 3rd International Conference on Computer and Communication Systems (ICCCS), Nagoya, Japan, [27–30 April 2018; pp. 472–476. [CrossRef]](http://dx.doi.org/10.1109/CCOMS.2018.8463317) 112. Lin, J.; Pipattanasomporn, M.; Rahman, S. Comparative analysis of auction mechanisms and bidding strategies for P2P solar [transactive energy markets. Appl. Energy 2019, 255, 113687. [CrossRef]](http://dx.doi.org/10.1016/j.apenergy.2019.113687) ----- _Sustainability 2022, 14, 7609_ 29 of 30 113. Ali, F.S.; Bouachir, O.; Ozkasap, O.; Aloqaily, M. SynergyChain: Blockchain-Assisted Adaptive Cyber-Physical P2P Energy [Trading. IEEE Trans. Ind. Inform. 2021, 17, 5769–5778. [CrossRef]](http://dx.doi.org/10.1109/TII.2020.3046744) 114. Bouachir, O.; Aloqaily, M.; Ozkasap, O.; Ali, F. FederatedGrids: Federated Learning and Blockchain-Assisted P2P Energy Sharing. _[IEEE Trans. Green Commun. Netw. 2022, 6, 424–436. [CrossRef]](http://dx.doi.org/10.1109/TGCN.2022.3140978)_ 115. Park, L.W.; Lee, S.; Chang, H. A Sustainable Home Energy Prosumer-Chain Methodology with Energy Tags over the Blockchain. _[Sustainability 2018, 10, 658. [CrossRef]](http://dx.doi.org/10.3390/su10030658)_ 116. Aggarwal, S.; Chaudhary, R.; Aujla, G.S.; Kumar, N.; Choo, K.K.R.; Zomaya, A.Y. Blockchain for smart communities: Applications, [challenges and opportunities. J. Netw. Comput. Appl. 2019, 144, 13–48. [CrossRef]](http://dx.doi.org/10.1016/j.jnca.2019.06.018) 117. Zhu, S.; Song, M.; Lim, M.K.; Wang, J.; Zhao, J. The development of energy blockchain and its implications for China’s energy [sector. Resour. Policy 2020, 66, 101595. [CrossRef]](http://dx.doi.org/10.1016/j.resourpol.2020.101595) 118. Mihaylov, M.; Razo-Zapata, I.; Nowé, A. NRGcoin—A Blockchain-based Reward Mechanism for Both Production and Consumption of Renewable Energy. In Transforming Climate Finance and Green Investment with Blockchains; Academic Press: Cambridge, MA, [USA, 2018; pp. 111–131. [CrossRef]](http://dx.doi.org/10.1016/B978-0-12-814447-3.00009-4) 119. Zahed Benisi, N.; Aminian, M.; Javadi, B. Blockchain-based decentralized storage networks: A survey. J. Netw. Comput. Appl. **[2020, 162, 102656. [CrossRef]](http://dx.doi.org/10.1016/j.jnca.2020.102656)** 120. Stanciu, A. Blockchain Based Distributed Control System for Edge Computing. In Proceedings of the 2017 21st International [Conference on Control Systems and Computer Science (CSCS), Bucharest, Romania, 29–31 May 2017; pp. 667–671. [CrossRef]](http://dx.doi.org/10.1109/CSCS.2017.102) 121. Casado-Vara, R.; de la Prieta, F.; Prieto, J.; Corchado, J.M. Blockchain framework for IoT data quality via edge computing. In Proceedings of the 1st Workshop on Blockchain-enabled Networked Sensor Systems, Shenzhen, China, 4 November 2018; pp. 19–24. 122. Shafagh, H.; Burkhalter, L.; Hithnawi, A.; Duquennoy, S. Towards blockchain-based auditable storage and sharing of IoT data. In Proceedings of the 2017 on Cloud Computing Security Workshop, Dallas, TX, USA, 3 November 2017; pp. 45–50. 123. Zhang, X.; Li, R.; Cui, B. A security architecture of VANET based on blockchain and mobile edge computing. In Proceedings of the 2018 1st IEEE International Conference on Hot Information-Centric Networking (HotICN), Shenzhen, China, 15–17 August 2018; pp. 258–259. 124. Mendki, P. Blockchain enabled iot edge computing: Addressing privacy, security and other challenges. In Proceedings of the 2020 The 2nd International Conference on Blockchain Technology, Hilo, HI, USA, 12–14 March 2020; pp. 63–67. 125. Tuli, S.; Mahmud, R.; Tuli, S.; Buyya, R. Fogbus: A blockchain-based lightweight framework for edge and fog computing. J. Syst. _[Softw. 2019, 154, 22–36. [CrossRef]](http://dx.doi.org/10.1016/j.jss.2019.04.050)_ 126. Zhaofeng, M.; Xiaochang, W.; Jain, D.K.; Khan, H.; Hongmin, G.; Zhen, W. A blockchain-based trusted data management scheme [in edge computing. IEEE Trans. Ind. Inform. 2019, 16, 2013–2021. [CrossRef]](http://dx.doi.org/10.1109/TII.2019.2933482) 127. Jiang, X.; Yu, F.R.; Song, T.; Leung, V.C. Edge Intelligence for Object Detection in Blockchain-Based Internet of Vehicles: [Convergence of Symbolic and Connectionist AI. IEEE Wirel. Commun. 2021, 28, 49–55. [CrossRef]](http://dx.doi.org/10.1109/MWC.201.2000462) 128. Lin, X.; Li, J.; Wu, J.; Liang, H.; Yang, W. Making Knowledge Tradable in Edge-AI Enabled IoT: A Consortium Blockchain-Based [Efficient and Incentive Approach. IEEE Trans. Ind. Inform. 2019, 15, 6367–6378. [CrossRef]](http://dx.doi.org/10.1109/TII.2019.2917307) 129. Rahman, M.A.; Rashid, M.M.; Hossain, M.S.; Hassanain, E.; Alhamid, M.F.; Guizani, M. Blockchain and IoT-Based Cognitive [Edge Framework for Sharing Economy Services in a Smart City. IEEE Access 2019, 7, 18611–18621. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2896065) 130. Qiu, C.; Yao, H.; Wang, X.; Zhang, N.; Yu, F.R.; Niyato, D. AI-Chain: Blockchain Energized Edge Intelligence for Beyond 5G [Networks. IEEE Netw. 2020, 34, 62–69. [CrossRef]](http://dx.doi.org/10.1109/MNET.021.1900617) 131. Du, Y.; Wang, Z.; Leung, V.C.M. Blockchain-Enabled Edge Intelligence for IoT: Background, Emerging Trends and Open Issues. _[Future Internet 2021, 13, 48. [CrossRef]](http://dx.doi.org/10.3390/fi13020048)_ 132. Lim, W.Y.B.; Huang, J.; Xiong, Z.; Kang, J.; Niyato, D.; Hua, X.S.; Leung, C.; Miao, C. Towards federated learning in uav-enabled internet of vehicles: A multi-dimensional contract-matching approach. IEEE Trans. Intell. Transp. Syst. 2021, 22, 5140–5154. [[CrossRef]](http://dx.doi.org/10.1109/TITS.2021.3056341) 133. Pokhrel, S.R.; Choi, J. Improving TCP performance over WiFi for internet of vehicles: A federated learning approach. IEEE Trans. _[Veh. Technol. 2020, 69, 6798–6802. [CrossRef]](http://dx.doi.org/10.1109/TVT.2020.2984369)_ 134. Manias, D.M.; Shami, A. Making a Case for Federated Learning in the Internet of Vehicles and Intelligent Transportation Systems. _[IEEE Netw. 2021, 35, 88–94. [CrossRef]](http://dx.doi.org/10.1109/MNET.011.2000552)_ 135. Peng, Y.; Chen, Z.; Chen, Z.; Ou, W.; Han, W.; Ma, J. Bflp: An adaptive federated learning framework for internet of vehicles. _[Mob. Inf. Syst. 2021, 2021, 6633332. [CrossRef]](http://dx.doi.org/10.1155/2021/6633332)_ 136. Chai, H.; Leng, S.; Chen, Y.; Zhang, K. A hierarchical blockchain-enabled federated learning algorithm for knowledge sharing in [internet of vehicles. IEEE Trans. Intell. Transp. Syst. 2020, 22, 3975–3986. [CrossRef]](http://dx.doi.org/10.1109/TITS.2020.3002712) 137. Pokhrel, S.R.; Choi, J. Federated learning with blockchain for autonomous vehicles: Analysis and design challenges. IEEE Trans. _[Commun. 2020, 68, 4734–4746. [CrossRef]](http://dx.doi.org/10.1109/TCOMM.2020.2990686)_ 138. Ayaz, F.; Sheng, Z.; Tian, D.; Guan, Y.L. A Blockchain Based Federated Learning for Message Dissemination in Vehicular [Networks. IEEE Trans. Veh. Technol. 2022, 71, 1927–1940. [CrossRef]](http://dx.doi.org/10.1109/TVT.2021.3132226) ----- _Sustainability 2022, 14, 7609_ 30 of 30 139. Doku, R.; Rawat, D.B. IFLBC: On the Edge Intelligence Using Federated Learning Blockchain Network. In Proceedings of the 2020 IEEE 6th Intl Conference on Big Data Security on Cloud (BigDataSecurity), IEEE International Conference on High Performance and Smart Computing, (HPSC) and IEEE International Conference on Intelligent Data and Security (IDS), Baltimore, [ML, USA, 25–27 May 2020; pp. 221–226. [CrossRef]](http://dx.doi.org/10.1109/BigDataSecurity-HPSC-IDS49724.2020.00047) 140. Morán, A.; Canals, V.; Galan-Prado, F.; Frasser, C.F.; Radhakrishnan, D.; Safavi, S.; Rosselló, J.L. Hardware-Optimized Reservoir [Computing System for Edge Intelligence Applications. Cogn. Comput. 2021, 1–9. [CrossRef]](http://dx.doi.org/10.1007/s12559-020-09798-2) -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/su14137609?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/su14137609, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/2071-1050/14/13/7609/pdf?version=1655902500" }
2,022
[ "Review" ]
true
2022-06-22T00:00:00
[ { "paperId": "417ceef82fbff346e4559fc2850e4f59a2f0d95d", "title": "Adoption of Blockchain Technology through Digital Twins in the Construction Industry 4.0: A PESTELS Approach" }, { "paperId": "8e76c41f4df07fdc512cb13f8e9b14e1692577ed", "title": "A Survey on AI-Driven Digital Twins in Industry 4.0: Smart Manufacturing and Advanced Robotics" }, { "paperId": "b62e98582bbbfc925ac0720c9f5aceb14aafd7f1", "title": "IoT enabled integrated system for green energy into smart cities" }, { "paperId": "7a46f1b84bb84f1ef33fc2d074eb56a51461b5c8", "title": "Blockchain technology in the future smart grids: A comprehensive review and frameworks" }, { "paperId": "9a1319e93f8fca656391491e81935fe62d03d7ad", "title": "Towards AI-Based Traffic Counting System with Edge Computing" }, { "paperId": "31feaa7ed2f37e021021c8c2725586fe067837fe", "title": "Accelerating Applications using Edge Tensor Processing Units" }, { "paperId": "0d8af7d81340934768a85a4c95369d602b99999c", "title": "Digitalization to achieve sustainable development goals: Steps towards a Smart Green Planet." }, { "paperId": "feba0c47bf12a02c3a725174bb53df78658a72a8", "title": "Pre-Trained Models: Past, Present and Future" }, { "paperId": "d78cf8ee82b2535a7700a18786706b9a7adf0297", "title": "Intelligent Traffic Management System Based on the Internet of Vehicles (IoV)" }, { "paperId": "4dfff13a5d469069cf50a39e75a0b862993aa400", "title": "Blockchain: The operating system of smart cities" }, { "paperId": "9d64cebb8033ff0db31d4e8f96d49d08bed337d0", "title": "Blockchain technology and trust relationships in trade finance" }, { "paperId": "c0c62eb787f72607575221accd1ef3476c430a56", "title": "Blockchain technology and the circular economy: Implications for sustainability and social responsibility" }, { "paperId": "4c9d81b73ff97e514ac0822980dbca83b7a3acb4", "title": "Roaming Service for Electric Vehicle Charging Using Blockchain-Based Digital Identity" }, { "paperId": "1f83c5f5570dd7f0c1129178f76916b476398be8", "title": "Blockchain-Enabled Edge Intelligence for IoT: Background, Emerging Trends and Open Issues" }, { "paperId": "b26df53a87692e1abc5eab1ee29b66665c9257e6", "title": "Hardware-Optimized Reservoir Computing System for Edge Intelligence Applications" }, { "paperId": "e688b468fcc8fea57d495c107696ae41f5ac02fa", "title": "Blockchain Based Trust Model Using Tendermint in Vehicular Adhoc Networks" }, { "paperId": "ef5abdbf50baa6f719338ffa8a4fe1e41fed0352", "title": "Federated learning improves site performance in multicenter deep learning without data sharing" }, { "paperId": "1e66cb1b5918603e4e75b77c15cf0eca16406d97", "title": "A Novel Power Market Mechanism Based on Blockchain for Electric Vehicle Charging Stations" }, { "paperId": "20afdc74da7374c90c5c148343a01b3e998c7bf4", "title": "Adaptive bidding strategy for real-time energy management in multi-energy market enhanced by blockchain" }, { "paperId": "085432ab4db6957e3e7517650b1198fedf6d2d46", "title": "Smart buildings features and key performance indicators: A review" }, { "paperId": "09d696e18a2425fd8a54ecf1abbb4ee2c630299b", "title": "Flexible Assignment of Loading Bays for Efficient Vehicle Routing in Urban Last Mile Delivery" }, { "paperId": "d1fdc4d37721f56a68738961406884766d379d94", "title": "Blockchain-based decentralized storage networks: A survey" }, { "paperId": "cf763cc175d8b1c29c517b3826c6c5a7ab5c5225", "title": "The development of energy blockchain and its implications for China's energy sector" }, { "paperId": "d9e3057a5855216d92ec2a900872c2463d06adf3", "title": "Smart systems of innovation for smart places: Challenges in deploying digital platforms for co-creation and data-intelligence" }, { "paperId": "0bbb17245edf92d08ff283e2acaec82df509ed74", "title": "Edge Intelligence: Architectures, Challenges, and Applications" }, { "paperId": "2e8d3c3eaf763246578b67e521f6df3d26b291ca", "title": "Applications of Artificial Intelligence and Machine learning in smart cities" }, { "paperId": "c1ec26b28edf355fd5c7094dcaa9a37e3095b84e", "title": "Blockchain Enabled IoT Edge Computing: Addressing Privacy, Security and other Challenges" }, { "paperId": "3bc5e91d3331120e0bfe3e97d9bbf3c3ec72ad97", "title": "Comparative analysis of auction mechanisms and bidding strategies for P2P solar transactive energy markets" }, { "paperId": "f45065ede28784c69371cfa51c8ea26ee8797ba8", "title": "Vision-based vehicle detection and counting system using deep learning in highway scenes" }, { "paperId": "8437fda2c8fa342a5fd874c73ee9b0686fa4260c", "title": "The Middleman Is Dead, Long Live the Middleman: The “Trust Factor” and the Psycho-Social Implications of Blockchain" }, { "paperId": "d2b5eec036d9b5094a1e242e0d9a7c3e26c00d4b", "title": "Blockchain for smart communities: Applications, challenges and opportunities" }, { "paperId": "3116671bbae44a1033171da733c4842a1268c174", "title": "Smart healthcare: making medical care more intelligent" }, { "paperId": "47923a27ece1661af7ec0833128fc0c04e2e2178", "title": "Smart Transport" }, { "paperId": "4512701ae56c6792416ff0c6eccdc4031c16c68e", "title": "Mitigating Trust Issues in Electric Vehicle Charging using a Blockchain" }, { "paperId": "a7221e8a4d5b833aa58d2b02d34eafae78d02934", "title": "Edge-Computing Video Analytics for Real-Time Traffic Monitoring in a Smart City" }, { "paperId": "7fcf0b09e89982b2f45594e088f2961c3f481384", "title": "A Computer Vision-Based Roadside Occupation Surveillance System for Intelligent Transport in Smart Cities" }, { "paperId": "d3c63f5a4bf5791dbb19afd776eca98d8db39931", "title": "A Review on Energy Consumption Optimization Techniques in IoT Based Smart Building Environments" }, { "paperId": "4c0945cb52d0734b25ecea49e3ae1c1b243fca66", "title": "A systematic literature review of blockchain-based applications: Current status, classification and open issues" }, { "paperId": "60be2610dba19761d6458bbac27527b744b0109e", "title": "Blockchain technology in the energy sector: A systematic review of challenges and opportunities" }, { "paperId": "62ccd99a65bfc7c735ae1f33b75b107665de95df", "title": "Federated Machine Learning" }, { "paperId": "3d95b2fdd4babe475134324f49c5453580fd2993", "title": "A Performance Evaluation of Federated Learning Algorithms" }, { "paperId": "d9c43d716a8a737b7a1d12bd7d14817df356e306", "title": "Blockchain: Enabling wide range of services in distributed energy system" }, { "paperId": "164e7be27409eff36ad4a8f2ebb355fa3bde3540", "title": "FogBus: A Blockchain-based Lightweight Framework for Edge and Fog Computing" }, { "paperId": "d5af8facee788b7799911e2a71bc9326bbf15c00", "title": "Real-time car tracking system based on surveillance videos" }, { "paperId": "5b8e9b3546d1785341003fbc6146ad1882c86130", "title": "Blockchain framework for IoT data quality via edge computing" }, { "paperId": "14fe35149aed6a47b6ebfd207deb7681b9446bb6", "title": "Urban Planning and Smart City Decision Management Empowered by Real-Time Data Processing Using Big Data Analytics" }, { "paperId": "c8061285333f8bb3ff10e23e4304ede05c1188d7", "title": "A security architecture of VANET based on blockchain and mobile edge computing" }, { "paperId": "354923bc917ac774438a048b87db9fabed4d757e", "title": "Deep learning for smart manufacturing: Methods and applications" }, { "paperId": "624cd8b4dac3b11e0ed97729582c1324c302b0be", "title": "Data-driven smart manufacturing" }, { "paperId": "ebc96892b9bcbf007be9a1d7844e4b09fde9d961", "title": "YOLOv3: An Incremental Improvement" }, { "paperId": "3a3ae82c0933f23221648978a23213a1f7c655c9", "title": "A Sustainable Home Energy Prosumer-Chain Methodology with Energy Tags over the Blockchain" }, { "paperId": "5a9a8a80ddbfd3169d92c5d7005377d237ce1c47", "title": "Using peer-to-peer energy-trading platforms to incentivize prosumers to form federated power plants" }, { "paperId": "8d3f008403e73226955699e31361f5ae02628dc8", "title": "Designing microgrid energy markets" }, { "paperId": "3e921b2a61649b70209bb93aeee73645beeec6c4", "title": "Blockchain Based Decentralized Management of Demand Response Programs in Smart Energy Grids" }, { "paperId": "ce21142d20575403293bcfa1eb962616339f505e", "title": "Blackchain: scalability for resource-constrained accountable vehicle-to-x communication" }, { "paperId": "fecc02639e25554c85337574a059da16a75652eb", "title": "Smart energy and smart energy systems" }, { "paperId": "14547ea3ab20720b201f899acfc1467616416dde", "title": "A blockchain-based reputation system for data credibility assessment in vehicular networks" }, { "paperId": "488ebe4db7190efe445c225aa67a10f70bc46d8d", "title": "Blockchain in government: Benefits and implications of distributed ledger technology for information sharing" }, { "paperId": "5d2eaf88f653dbf6157e80d7bda993bbfe79dc26", "title": "Blockchain for smart grid resilience: Exchanging distributed energy at speed, scale and security" }, { "paperId": "1a64721d1e9bbdcb7a9d7105e241935630992dde", "title": "The EU General Data Protection Regulation (GDPR)" }, { "paperId": "952d4d7515cc878a9aaee1fdd22779dc3c28b54c", "title": "Edge computing technologies for Internet of Things: a primer" }, { "paperId": "24747c62e558cad008e9a65c9a7e2d463cd9f2de", "title": "Towards Blockchain-based Auditable Storage and Sharing of IoT Data" }, { "paperId": "ef573a42c7f3fe05c64d373c74efa23f76faf8de", "title": "Traffic management systems: A classification, review, challenges, and future perspectives" }, { "paperId": "cbe2f7dd07d869dd1d0b2ba1cbc03b84dd695bb1", "title": "Block-VN: A Distributed Blockchain Based Vehicular Network Architecture in Smart City" }, { "paperId": "5369bd51a019355c983fc8979071b36bc7cbf017", "title": "Midgar: Detection of people through computer vision in the Internet of Things scenarios to improve the security in Smart Cities, Smart Towns, and Smart Homes" }, { "paperId": "976e29fe6c9baeb39732bca0e35f66f84d5bdd90", "title": "ORB: An efficient alternative to SIFT or SURF" }, { "paperId": "8d32405265c8d6bc3071f6a1f5f1a88a5f04de48", "title": "The subprime crisis and its role in the financial crisis" }, { "paperId": "a7639cae1b62c72082ccf40d33d90435a5564c9c", "title": "A conceptual framework" }, { "paperId": "36bf057681b70b07e7a1fed50728235a6f804cd6", "title": "BFLP: An Adaptive Federated Learning Framework for Internet of Vehicles" }, { "paperId": "0b02189d15ef82c23cc4305f4fceabc72807131e", "title": "Smart farming IoT platform based on edge and cloud computing" }, { "paperId": null, "title": "Shifting gears: Insurers adjust for connected-car ecosystems" }, { "paperId": "047191b3e9c6709efe7fbf1da6afe421a1dde6a7", "title": "An Efficient Approach for Detection and Speed Estimation of Moving Vehicles" }, { "paperId": "cc10f02972a26d06114340023be1835b74e31c91", "title": "Ucl Centre for Advanced Spatial Analysis Smart Cities of the Future" }, { "paperId": "ac3a978b51d42728456b8caad06ac0ac648e05b4", "title": "Climate and more sustainable cities: climate information for improved planning and management of cities (producers/capabilities perspective)" } ]
31,436
en
[ { "category": "Economics", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Economics", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/011d61920ae95bbe63193f2e73e7eab7bc116206
[ "Economics" ]
0.865556
ANALYSIS OF BITCOIN MARKET EFFICIENCY BY USING MACHINE LEARNING
011d61920ae95bbe63193f2e73e7eab7bc116206
CBU International Conference Proceedings
[ { "authorId": "119596549", "name": "Yukikazu Hirano" }, { "authorId": "1763214", "name": "L. Pichl" }, { "authorId": "144594961", "name": "Cheoljun Eom" }, { "authorId": "2001544", "name": "T. Kaizoji" } ]
{ "alternate_issns": [ "1805-997X" ], "alternate_names": [ "CBU Int Conf Proc" ], "alternate_urls": [ "https://ojs.journals.cz/index.php/CBUIC/issue/archive" ], "id": "11318c27-eea7-4f77-842c-5fce350e9794", "issn": "1805-9961", "name": "CBU International Conference Proceedings", "type": null, "url": "http://ojs.journals.cz/index.php/CBUConference2013/issue/archive" }
The issue of market efficiency for cryptocurrency exchanges has been largely unexplored. Here we put Bitcoin, the leading cryptocurrency, on a test by studying the applicability of the Efficient Market Hypothesis by Fama from two viewpoints: (1) the existence of profitable arbitrage spread among Bitcoin exchanges, and (2) the possibility to predict Bitcoin prices in EUR (time period 2013-2017) and the direction of price movement (up or down) on the daily trading scale.  Our results show that the Bitcoin market in the time period studied is partially inefficient. Thus the market process is predictable to a degree, hence not a pure martingale. In particular, the F-measure for XBTEUR time series obtained by three major recurrent neural network based machine learning methods was about 67%, i.e. a way above the unbiased coin tossing odds of 50% equal chance.
CBU I NTERNATIONAL C ONFERENCE ON I NNOVATIONS IN S CIENCE AND E DUCATION M ARCH 21-23, 2018, P RAGUE, C ZECH R EPUBLIC WWW . CBUNI . CZ, WWW . JOURNALS . CZ # **ANALYSIS OF BITCOIN MARKET EFFICIENCY BY USING MACHINE LEARNING ** Yuki Hirano [1], Lukáš Pichl [2], Cheoljun Eom [3], Taisei Kaizoji [4] **Abstract:** The issue of market efficiency for cryptocurrency exchanges has been largely unexplored. Here we put Bitcoin, the leading cryptocurrency, on a test by studying the applicability of the Efficient Market Hypothesis by Fama from two viewpoints: (1) the existence of profitable arbitrage spread among Bitcoin exchanges, and (2) the possibility to predict Bitcoin prices in EUR (time period 2013-2017) and the direction of price movement (up or down) on the daily trading scale. Our results show that the Bitcoin market in the time period studied is partially inefficient. Thus the market process is predictable to a degree, hence not a pure martingale. In particular, the F-measure for XBTEUR time series obtained by three major recurrent neural network based machine learning methods was about 67%, i.e. a way above the unbiased coin tossing odds of 50% equal chance. **UDC Classification:** 004.8, 33; **DOI:** http://dx.doi.org/10.12955/cbup.v6.1152 **Keywords** : Bitcoin, XBT, Neural Network, Gated Recurrent Unit, Long Short-Term Memory **Introduction** Bitcoin was the first open source distributed cryptocurrency released in 2009 after it was introduced in a paper “Bitcoin: A Peer-to-Peer Electronic Cash System” by a developer under the pseudonym Satoshi Nakamoto. It has been quickly followed by a number of alternative coins (altcoins), derivatives of the original concept, and other block-chain based cryptocurrencies of more or less sophisticated design, such as Ethereum. As of the writing of this article (Feb. 15, 2018), the market capitalization of all cryptocurrencies is about USD 475 billion, with Bitcoin share being around USD 166 billion, followed by Ethereum (USD 92 billion; Coinmarketcap 2018). Considering the fact that no cryptocurrency has become a regular means of payment in any national economy or global sector yet, cryptocurrencies present a remarkable speculative enterprise in cyberspace with a theoretical potential of disrupting financial systems by the emergent digital commodity aspiring to function as a global means of payment and value storage. The future of Bitcoin and other currencies appears at stake however, because of the following problems. First, large prices of Bitcoin made micropayments impractical as the transaction fees for each payment rocketed, in spite of the original concept. Second, from the viewpoint of a stable currency, daily price fluctuations as high as 10 percent up or down on Bitfinex exchange market are a way too high; on 16 December 2017 the price of Bitcoin was more than 20 times higher relative to the same date a year earlier – just to fall down to a half of that maximum value 5 weeks later. Third, the Bitcoin mining process that sustains the integrity of the block chain has an enormous carbon footprint, consuming as much electric power as the entire country of Nigeria, according to CBS News (November 27, 2017). Therefore it is quite questionable whether the Bitcoin payment system can be scaled up in order to take the role of a national or even global currency. Consequently, some authorities maintain that cryptocurrencies, including Bitcoin, are just a pyramid scheme scam, whereas others proclaim the emergence of a new global monetary system. Given the absence of a substantial economic sector behind Bitcoin, and the above mentioned volatility with abundant bubbles and crashes, it is an open question as to what extent is the Bitcoin market system efficient. The prices of Bitcoin are very sensitive to market making news, such as the recognition of Bitcoin as a legal payment method by Japan from April 1, 2017, or the ban of cryptocurrency exchanges in China effective from November 1, 2017. The central question addressed in this article is whether Bitcoin exchange markets are efficient. In an efficient market, all the available information including the entire price history is fully reflected in the current price of the asset. Thus the Efficient Market Hypothesis (EMH) introduced by Eugene Fama (Fama, 1970 and 1991) implies that asset prices should follow a random walk which is impossible to forecast; in general, the price dynamics is then a martingale process, in which the expectation of the next value equals the current value of the asset, and the direction of price change is impossible to predict. Since the EMH assumes complete information efficiency with regard to price formation, it rules out the 1 International Christian University, Mitaka, Tokyo, Japan, kyabaria17@gmail.com 2 International Christian University, Mitaka, Tokyo, Japan, lukas@icu.ac.jp 3 Pusan National University, Busan, Republic of Korea, shunter@pusan.ac.kr 4 International Christian University, Mitaka, Tokyo, Japan, kaizoji@icu.ac.jp 175 ----- CBU I NTERNATIONAL C ONFERENCE ON I NNOVATIONS IN S CIENCE AND E DUCATION M ARCH 21-23, 2018, P RAGUE, C ZECH R EPUBLIC WWW . CBUNI . CZ, WWW . JOURNALS . CZ possibility of arbitrage transactions. In other words, if a profit-making arbitrage transaction is possible among markets, then it is a certain manifestation of partial inefficiency of the market system. In what follows we will show that in the case of Bitcoin, profitable arbitrage windows may open among Bitcoin exchanges to various fiat currencies, and that the next-day price-change direction (sign of logarithmic return) for a single time series may be predicted to a certain degree by using machine learning methods trained on the past daily data and at a prediction level that is higher than the equal odds of fair coin tossing. Fi g ure 1 : Time series of XBT prices in E U R o v er a p eriod of 9 27 tradin g da y s . Source: Authors ( Plot g enerated using R- p acka g e q uantmod ) This paper is organized as follows. Following the literature review in the next section, in Section 3 we explain the dataset and outline the methods of its analysis. Section 4 wraps up our results and discussions, which are followed by the concluding section. **Literature Review** Scientific literature on Bitcoin has become abundant recently. Most of the papers are related to the statistical analysis of Bitcoin and other cryptocurrencies, using methods from econometrics and general data analysis. As of present, we are not aware of any article that would apply machine learning for the estimation of cryptocurrency market efficiency. In a recent research work, Gkillas & Katsiampa, (2018) have studied the behavior of returns of five major cryptocurrencies using extreme value analysis, finding out that “Bitcoin Cash is the riskiest, while Bitcoin and Litecoin are the least risky cryptocurrencies”. In a statistical study by Phillip et al., (2018) diverse stylized facts such as long memory and heteroscedasticity have been explored for 224 different cryptocurrencies, which are found to “exhibit leverage effects and Student- error distributions”. The design issues of Bitcoin are revisited by Ziegeldorf et al., (2018) in a study proposing a novel oblivious shuffle protocol “to improve reliance against malicious attackers”. Their method is claimed to be “scalable, increasing anonymity and enabling deniability”. In a study of market efficiency, AlwarezRamirez et al., (2018) analyzed Bitcoin to find that the “Bitcoin market is not uniformly efficient, and asymmetries and inefficiency are replicated over different time scales”. In contrast to our work, their method is based on the detrended fluctuation analysis estimating long-range correlations for price returns, thus not covering the relation among Bitcoin exchanges and nonlinear dynamics patterns. Corbet et al., (2018) applied a time and frequency domain analysis to estimate the relationships between 3 major cryptocurrencies and a variety of financial assets, arriving to the conclusion that “cryptocurrencies may offer diversification benefits for investors with short investment horizons”. In a work motivated by market efficiency reasons related to ours, Lahmiri et al., (2018) analyzed the Bitcoin time series in seven different exchanges, finding that “the values of measured entropy indicate a high degree of randomness in the series”. In contrary to this finding, they claim however, “strong evidence against the EMH”. Compared to the present approach, they do investigate nonlinear patterns in volatility dynamics, but the work is limited by broad assumption of the four diverse statistical 176 ----- CBU I NTERNATIONAL C ONFERENCE ON I NNOVATIONS IN S CIENCE AND E DUCATION M ARCH 21-23, 2018, P RAGUE, C ZECH R EPUBLIC WWW . CBUNI . CZ, WWW . JOURNALS . CZ distributions employed. The interdependence of Bitcoin and altcoin markets was studied on short- and long-term scales by Ciaian et al., (2018) who found the price relationship stronger in the short-term run. Bariviera et al., (2018) studied the statistical features and long-range dependence of Bitcoin returns, focusing on the behavior of the Hurst exponent computed in sliding windows, showing that it has a similar behavior at different time scales. Luther & Salter, (2017) examined the relationship of possible hedging in Bitcoin for countries with troubled financial systems, such as Cyprus, finding little significant evidence that would support such transitions. Price clustering of Bitcoin at round numbers is found in the work of Urquhart, (2017) who also studies this effect in volume distributions and market liquidity. Hendrickson & Luther, (2017) employed a monetary model with endogenous search and random consumption preferences, in which they show that governments of sufficient size are capable of banning Bitcoin without serious consequences. The degree of synchronization of prices of Bitcoin across exchanges is studied by Pieters & Vivanco, (2017) who claim that the law of one price does not apply due to a reason ascribed to market efficiency failure, especially for markets with anonymous trading accounts. In a search for the determinants of Bitcoin price Hayes, (2017) argues that it closely follows the cost of production, in particular predominantly the energy consumption, which drives the relative value formation at the cost margin. In summary, the above reviewed literature deals directly with the issue of market efficiency of Bitcoin only in two cases, Alwarez-Ramirez et al., (2018), and Lahmiri et al., (2018) neither of which consider arbitrage opportunities among Bitcoin exchanges or use machine learning algorithms to predict price movement direction. Thus the present work provides a novel complementary insight into the issue of Bitcoin market efficiency. **Data and Methods** The dataset for triangular arbitrage has been retrieved from Yahoo finance using the R-package quantmod (R Core Team 2018; Ryan and Ulrich, 2017). It contains all 822 closing Bitcoin prices for the selected fiat currencies of AUD, CAD, CNY, EUR, GBP, JPY, and USD between January 1, 2015 and February 16, 2018. In order to analyze the triangular arbitrage of the type USD-XBT-CRC-USD, we retrieved the closing values of the USDAUD, USDCAD, USDCNY, USDEUR, USDGBP, and USDJPY exchange rates, correspondingly diminishing the amount of data points by the holidays of each particular foreign exchange market. The profit rate of the trianguar arbitrage transaction, in which the USD currency is first used to buy one Bitcoin, which is then sold for CRC and converted by the exchange rate CRCUSD=1/USDCRC back to USD, normalized to the initial expense for 1 XBT (i.e. the value of XBTUSD), reads 𝑋𝐵𝑇𝐶𝑅𝐶 ρ = 𝑋𝐵𝑇𝑈𝑆𝐷 ⁄ [𝑈𝑆𝐷𝐶𝑅𝐶−1 ≡𝑈𝑆𝐷𝐶𝑅𝐶(𝑋𝐵𝑇) 𝑈𝑆𝐷𝐶𝑅𝐶−1] ⁄ (1) The dataset used for machine learning prediction of price trend of Bitcoin is the XBTEUR time series retrieved from Bloomberg, shown in Fig. 1. The three methods of machine learning applied for prediction are (1) a Recurrent Neural Network in Elman configuration (Elman, 1990), depicted in Fig. 2, Fi g ure 2: Recurrent neural network in Elman’s to p olo gy . S ource : Authors 177 ----- CBU I NTERNATIONAL C ONFERENCE ON I NNOVATIONS IN S CIENCE AND E DUCATION M ARCH 21-23, 2018, P RAGUE, C ZECH R EPUBLIC WWW . CBUNI . CZ, WWW . JOURNALS . CZ (2) a LSTM network depicted in Fig. 3, and (3) a GRU network shown in Fig. 4. Since these methods are standard in deep learning libraries, such as TensorFlow, which we applied, we do not repeat the equations, only briefly comment on the notation in the schematic figures. In particular, in Fig. 2, the input vector (components *i* taken from the time series as 20-element long moving window) at time *t* are fed to the hidden layer using a weight matrix of *W(in)* . Then hidden unit values are computed, which are fed back in a recurrent connection with parameter matrix *W* (recurrence shown by the bold arrow), and also passed over to the output layer with the weights *W(out)* . For the machine learning example, we assign 70% of the dataset to training, 15% of the dataset for validation (using early stopping criterion), and 15% of the dataset to testing (our result data). Figure 3 shows the far-more complicated design of the LSTM network. The recurrent unit is shown as the black circle. In addition to the input and output gates depicted on the right, there is an additional forget gate shown on the left, which regulates what data will be remembered and for how long. The addition and multiplication symbols are shared with Fig. 4. Finally, in Fig. 4 a design of the GRU network is presented, which is a simplification of the LSTM method that uses less parameters but is capable of producing results of similar accuracy as those by the LSTM algorithm in most cases. Reset and update gates regulate the flow of the neural signal through the network. Sigmoid and tangent-hyperbolic activation functions are used as shown in the legend. Fi g ure 3: Lon g S hort-Term Memory ( LSTM ) neural net w ork schematic Source: Authors **Results and Discussions** Table 1 shows the results for the triangular arbitrage. The medians of the distributions are very small, close to zero (i.e. never exceeding 1 percent, which, if transaction fees are considered, is probably not a profitable value). The main result are the standard deviation values, measuring the width of the distribution, which is often asymmetric and exhibits outliers. The minimum value, maximum value, mean, median, skewness and kurtosis parameters complement the standard-deviation based analysis. We can see that USD-EUR based Bitcoin arbitrage offers virtually no profit opportunities whereas the arbitrage window widens to almost 6% for the Chinese currency. It can be said that as the currency becomes minor, the arbitrage window broadens. Table 2 shows the information retrieval measures for the trend prediction results of the three ML algorithms, using 2 different predictors (prices and log returns). 178 ----- CBU I NTERNATIONAL C ONFERENCE ON I NNOVATIONS IN S CIENCE AND E DUCATION M ARCH 21-23, 2018, P RAGUE, C ZECH R EPUBLIC WWW . CBUNI . CZ, WWW . JOURNALS . CZ |Table 1: Summary of triangular arbitrage distributions (normalized profit rate) for the USD-XBT- CRC-USD scheme using 6 different currencies as CRC|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10| |---|---|---|---|---|---|---|---|---|---| ||Currency|Min|Max|Mean|Median|St. Dev.|Skewness|Kurtosis|| ||AUD|-0.2405|0.3389|0.0308|0.0162|0.0478|1.6583|10.0418|| ||CAD|-0.1655|0.3953|0.0232|0.0093|0.0510|2.4611|12.8804|| ||CNY|-0.4321|0.3998|0.0053|0.0059|0.0585|0.5499|16.0649|| ||EUR|-0.0817|0.0706|0.002|0.0021|0.0101|0.1192|14.3642|| ||GBP|-0.1616|0.6654|0.0085|0.0065|0.0298|13.3933|290.9942|| ||JPY|-0.1208|0.267|0.0219|0.0122|0.0407|3.3021|16.6164|| |Source: Authors|||||||||| Fi g ure 4: Gated Recurrent Unit ( GRU ) neural network schematic Source: Authors Table 2 : Machine Learnin g Algorithm results for binar y trend p rediction of XBTE U R time series ( a ) Trainin g b y p rices ( b ) Trainin g b y lo g arithmic return |Method|Accuracy|Recall|Precision|F-measure|Accuracy|Recall|Precision|F-measure| |---|---|---|---|---|---|---|---|---| |RNN|0.58|0.69|0.66|0.67|0.60|0.95|0.60|0.73| |LSTM|0.57|0.69|0.64|0.67|0.54|0.73|0.58|0.65| |GRU|0.58|0.69|0.66|0.67|0.56|0.85|0.58|0.69| Source: Authors **Conclusion** We have established partial information inefficiency of the Bitcoin market by means of triangular arbitrage between USD-XBT-CRC-USD where CRC stands for one of 6 major currencies. Whereas on the daily data trading scale, the profit window is very narrow in case of major currencies such as EUR, it widens up for currencies such as AUD, CAD and CNY, beyond the standard transaction fee levels. In addition, by using three machine learning algorithms, the RNN, LSTM and GRU methods, we have proved that machine learning algorithms are capable of predicting the direction of the price change for 179 ----- CBU I NTERNATIONAL C ONFERENCE ON I NNOVATIONS IN S CIENCE AND E DUCATION M ARCH 21-23, 2018, P RAGUE, C ZECH R EPUBLIC WWW . CBUNI . CZ, WWW . JOURNALS . CZ the next day based on the past data with the F-measure in the range of 67% to 73%. When compared to the USDEUR exchange rate values, the Bitcoin market shows a substantially greater deal of inefficiency. These results present a significant argument to question the validity of the EMH in case of Bitcoin exchanges. **Acknowledgement** This research was supported by JSPS Grants-in-Aid Nos. 2538404, 2628089. **References** Alvarez-Ramirez, J., Rodriguez, E., Ibarra-Valdez, C. (2018) Long-range correlations and asymmetry in the Bitcoin market, Physica A: Statistical Mechanics and its Applications vol. 492, pp. 948-955. Bariviera, A. F., M. J. Basgall, W. Hasperue, and M. Naiouf (2017) Some Stylized Facts of the Bitcoin Market, Physica A vol. 484, pp. 82-90. Ciaian, P., Rajcaniova, M., Kancs d'A (2018) Virtual relationships: Short- and long-run evidence from BitCoin and altcoin markets, Journal of International Financial Markets, Institutions and Money vol. 52, pp. 173-195. Coinmarketcap (2018) Cryptocurrency Market Capitalizations, https://coinmarketcap.com/, Accessed 2018/02/15. Corbet, S., Meegan, A., Larkin, C., Lucey, B., Yarovaya, L. (2018) Exploring the dynamic relationships between cryptocurrencies and other financial assets, Economics Letters, vol. 165 pp. 28-34. Elman, J. L. (1990) Finding Structure in Time, Cognitive Science vol. 14 No. 2, pp. 179-211. Fama, E.F. (1970) Efficient Capital Markets: A Review of Theory and Empirical Work, The Journal of Finance vol. 25, pp. 383-417. Fama, E.F. (1991) Efficient Capital Markets: II, The Journal of Finance vol. 46, pp. 1575-1617. Gkillas K., Katsiampa, P. (2018) An application of extreme value theory to cryptocurrencies, Economics Letters vol. 164, pp. 109-111. Hayes, A. S. (2017) Cryptocurrency value formation: An empirical study leading to a cost of production model for valuing bitcoin, Telematics and Informatics vol. 34 No. 7, pp. 1308-1321. Hendrickson, J. R., Luther, W. J. (2017) Banning bitcoin, Journal of Economic Behavior & Organization vol. 141, pp. 188195. Lahmiri, S., Bekiros, S.. Salvi A. (2018) Long-range memory, distributional variation and randomness of bitcoin volatility, Chaos, Solitons & Fractals vol. 107, pp. 43-48. Luther, W. J., Salter, A. W. (2017) Bitcoin and the bailout, The Quarterly Review of Economics and Finance vol. 66, pp. 5056. Phillip, A., Chan, J. S. K., Peiris, S. (2018) A new look at Cryptocurrencies, Economics Letters vol. 163, pp. 6-9. Pieters, G., Vivanco, S. (2017) Financial regulations and price inconsistencies across Bitcoin markets, Information Economics and Policy vol. 39, pp. 1-14. R Core Team (2018). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/. Ryan, J. A., Ulrich, J. M. (2017). quantmod: Quantitative Financial Modelling Framework. R package version 0.4-12. https://CRAN.R-project.org/package=quantmod Urquhart, A. (2017) Price clustering in Bitcoin, Economics Letters vol. 159, pp. 145-148. Ziegeldorf, J. H., Matzutt, R., Henze, M., Grossmann, F., Wehrle, K. (2018) Secure and anonymous decentralized Bitcoin mixing, Future Generation Computer Systems vol. 80, pp. 448-466. 180 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.12955/CBUP.V6.1152?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.12955/CBUP.V6.1152, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://ojs.journals.cz/index.php/CBUIC/article/download/1152/pdf/" }
2,018
[ "Conference" ]
true
2018-09-24T00:00:00
[]
5,521
en
[ { "category": "Economics", "source": "external" }, { "category": "Business", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/011e06a3b06257e73ee37540df71cd6a595d3c03
[ "Economics" ]
0.878632
Blockchain for good
011e06a3b06257e73ee37540df71cd6a595d3c03
[ { "authorId": "49827510", "name": "B. Kewell" }, { "authorId": "144118186", "name": "R. Adams" }, { "authorId": "38464485", "name": "G. Parry" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
**Blockchain for Good?[1]** Beth Kewell, University of Surrey, Surrey, UK. Richard Adams, University of Surrey, Surrey, UK. Glenn Parry, University of the West of England, UK. Explores key areas of Blockchain innovation that appear to represent viable catalysts for achieving global Sustainable Development targets. Projects and initiatives seeking to extend the reach of Distributed Ledger Technology (DLTs), seem mostly intended for the benefit of for-profit businesses, governments, and consumers. DLT projects, devised for the public good, could aim, in theory, to fulfil the United Nation’s current Sustainable Development Goals (SDG). Our overview of these initiatives suggests that blockchain technology is being applied in ways that could transform this ambition for good into a practical reality. Current examples of blockchain deployment are being specified within a value-creation remit that is most likely to benefit for-profit businesses, governments, and consumers (Ng, 2013; Bohme et al., 2015; Swan, 2015; Potts et al., 2016; McWaters et al., 2016; Walport, 2016). Received ideas about what blockchain can and should be used for are based on perceptions that the key role of this technology is to unlock cost savings and secure efficiency gains, whilst also enabling widespread business model transformation (Walport, 2016). Within this scenario, blockchain affordances (Gibson, 1978) are principally seen to ‘do good’ by resolving longstanding obstacles to profitability and value-capture (Walport, 2016). The aim of this paper is to consider how blockchain solutions could be used to achieve good outcomes for the sustainable development agenda by, for example, helping to fulfil the UN’s Sustainable Development Goals (UN, 2015). Kranzberg’s first law of 1 JEL Codes: O38; O39; D20. Acknowledgements: The authors gratefully acknowledge the support of a BA/Leverhulme Small Research Grant (SG160335) in the preparation of this work. Page 1 of 32 ----- technology avers that ‘Technology is neither good nor bad; nor is it neutral’ (Kranzberg’s, 1986, p.545). In doing so, Kranzberg reminds us that innovations are morally and ethically instantiated. To date, research has tended to focus on the technical characteristics, efficiency gains - and profits - to be yielded from blockchain projects and experimental Distributed Ledger Technology (DLTs) and ‘permissioned ledgers’ being run by private consortia (Ng, 2013; Bohme et al., 2015; Swan, 2015; Potts et al., 2016; McWaters et al., 2016; Walport, 2016). While initially fixed on the commercial and consumer benefits to be drawn from blockchain innovation, attention is beginning to shift toward the appropriation of socially and environmentally beneficial use cases that aim to tackle global challenges such as, for example, financial exclusion (CTPM, 2016). Drawing on affordance theory, this exploratory paper reflects on innovative applications of blockchain projects that could help deliver socially and environmentally beneficial outcomes by challenging existing business models and providing new opportunities for value creation that also serve a philanthropic purpose (Botsman and Rogers, 2010). We call this ‘Blockchain for Good’, where ‘Good’ can be framed in terms of the UN’s Sustainable Development Goals (SDG) (UN, 2015). The SDGs provide a vision for governmental, corporate and civic action leading the way towards ‘development that meets the needs of the present without compromising the ability of future generations to meet their own needs’ (WCED, 1987, para 27). Page 2 of 32 ----- The paper proceeds as follows: First, we describe our approach to this exploratory research. Second, we offer a brief overview of the technological characteristics of DLT. Third, we examine the notion that DLTs have unique affordances rendering them appropriate solutions to the SDGs. Consequently, in this article we begin to explore the impact of DLTs on the UN’s Sustainable Development Goals which is the contribution of the paper. **Affordances** The repositioning of blockchain technologies as a device for mobilising good causes, including those positioned at a global level, represents a considerable departure from their original remit as payments reconciliation systems which may be utilised without the need for banks and clearing houses (Ng, 2013; Bohme et al., 2015; Swan, 2015; Welch, 2015; Potts et al., 2016; McWaters et al., 2016; Walport, 2016). The identification of such an important ‘change of use’ draws attention to a concomitant shift in perceptions of blockchain affordances – that is to say, discernment of what the software can do for sustainable development and environmental protection in parallel with an appreciation of what novel deployment could realise for vulnerable and impoverished communities (Seidel et al. 2013). Key organisations, such as the UN, are actively focused on establishing blockchain’s capacity for achieving SDGs in, for example, identity provision and financial inclusion (CTPM, 2016). Scoping exercises, focused on pinpointing blockchain’s potential contribution to the sustainability field may represent a first step towards developing a future ‘affordance taxonomy’ (Conole and Dyke, 2004) to guide the deployment of blockchain-for-good in Page 3 of 32 ----- the third sector and among social enterprises, including those already taking advantage of crowd-funding and other charitable activities made possible by virtual platforms (Choy and Schlagwein, 2016). Affordances are bestowed upon artefacts – they are the qualities users perceive objects, places, contexts, and constructs, uphold and encompass (Gaver, 1991; Zammuto et al., 2007; Maier and Fadel, 2009; Faraj et al., 2011; Withagen, 2012; Majchrzak and Markus, 2012; Xenakis and Arnellos, 2013; Lankton et al., 2015; Ciavola and Gershenson, 2016; Choy and Schlagwein, 2016; Beynon-Davies and Lederman, 2017). Affordances are bound to expectations of what artefacts can be/can do and thus reputational information (Kewell, 2007) that tells us whether the actions they ought to assist Ciavola and Gershenson (2016, p.252) or facilitate are worthwhile, valuable, risky or unwise (Zammuto et al., 2007). Affordances can therefore be seen to possess an implicit moral imperative (Dierksmeier and Seele, 2016). Artefacts by themselves have no power; they do nothing (Geels, 2005). Affordance theory suggests that an artefact is fundamentally perceived in terms of its ‘action _possibilities’ (Withagen, 2012, p.521). Drawing on Gibson’s (1978) work on the_ ecology of perception, Pea (1993, p. 51) describes an ‘affordance’ as the perceived and _actual properties of a thing, primarily those functional properties that determine just_ _how the thing could possibly be used. An affordance, then, is what an object or_ technology offers, provides or furnishes in the context of use: depending on the user, a chair ‘affords’ sitting or an improvised ladder (Maier and Fadel, 2009); a bicycle ‘affords’ travel, or exercise or the delivery of health benefits (Conole and Dyke, 2004; Page 4 of 32 ----- Maier and Fadel, 2009; Volkoff and Strong, 2013; Lankton et al., 2015; Ciavola and Gershenson, 2016). Affordance theory subsequently delineates between intended uses built into the design process and consequential affordances, which avail themselves as prototypes are tested, and end-products are evaluated by potential users and consumers leading to the development of ‘sequential’ and ‘nested’ affordances (Gaver, 1991, p.4). Original construals of affordance can change markedly by the end of this learning curve (Gaver, 1991). Dual affordances can also emerge over time, once an artefact, prototype or design has acquired up-take (Beynon-Davies and Lederman, 2017). Thus, the recycling movement has also recently shown how the original meaning of an artefact’s affordances can be usurped or overturned by, for example, making objects with established or traditional affordance perform tasks for which they were not originally intended. A good illustration of the latter is provided by the current trend in urban cities for ‘container living,’ whereby redundant sea freight containers are converted into sustainable homes. A relatively new area of technological affordance theory examines the development of simulated computer technologies and the impact perceptions of what they can do has on human and organizational relations in the ‘real world’ (Zammuto et al., 2007; Boyd, 2010; Faraj et al., 2011). The affordances of software products can be altered by multiple designers and mass users (Boyd, 2010, p.7). By itself, this capacity rewrites existing conceptions of the interaction between designers, users, artefacts, and the environment in which they are embedded (Zammuto et al., 2007; Faraj et al., 2011; Page 5 of 32 ----- Ciavola and Gershenson, 2016). When positioned within social networks (Boyd, 2010), the affordances of simulated technology (such as platforms and blockchains) are percieved to foster new forms of communitarian action and social exchange (Faraj et _al., 2011; Choy and Schlagwein, 2016). When considered within an organisational_ context, these technologies are said to change perceptions of what may be afforded by systems, structures and processes, for example, those illustrated in workflow visualisation software (Beynon-Davies and Lederman, 2017), allowing previously hidden sources of value to become more self-evident (Zammuto et al., 2007; Faraj, et _al., 2011; Ciavola and Gershenson, 2016)._ The discovery of affordances related to blockchain technology is following patterns identified within inter and intraorganisational contexts. Blockchain is part of a thriving ecosystem, populated, en masse, by designers and users, who are continually improvising new affordances, as they tweak the technology for use in different settings. The advent of usecases for intraorganizational and consortia based blockchain deployment (as DLTs and permissioned ledgers), suggests that it will not be very long before companies begin to perceive blockchains as instruments of change. In what follows, we consider how perceptions of blockchain affordances are likely to change understandings of what can be achieved in the sustainable development field, as an ecosystem that must address mutiple requirements (from disaster relief to microfinance), using extremely complex networks of interactions. Could these interactions be placed on blockchains? Could this placement deliver better outcomes for aspects of society and the environment that are most in need and generate new sources of good? Page 6 of 32 ----- **Affordances for good** It is important to consider what we mean by good before addressing these questions. The western philosophical tradition has, for millennia, distinguished between intrinsic and extrinsic good (Smith, 1948) the former, good for its own sake, the latter being derivatives of the former – that is, an extrinsic good is good not for its own sake, but because its enactment leads back to an intrinsic good. The debates about the ontological status of intrinsic and extrinsic good, what constitutes them, the sorts of things that are or have intrinsic or extrinsic good(ness) and how these might be assessed or computed are beyond the scope of this paper. Frankena (1973) provides a comprehensive list of those things which are intrinsically good – as deemed by other authors to be good or rational to desire for their own sakes. Others, for example George Moore (1903), reject the notion of intrinsic good and take a more consequentialist view that things are good when they are perceived to be good, where their consequences are in some sense better than those of alternatives. There is a substantial literature on ethical issues surrounding ICTs, much of it framed around what constitutes ‘better’ and how that might be evaluated, including: the impact of technological progress on society (Lee, 2005) and the influence of technology on the development of virtuous interactions (Benkler and Nissenbaum, 2006). Arguing that ICT’s beneficial impact can be evaluated by distinguishing between local and systemic levels, the difference between content and process, the implication of Taddeo and Vaccaro’s (2011) framing is that an ethical understanding of technologies can be gained Page 7 of 32 ----- through an interrogation of how the ways in which they work enable new beneficial actions and outcomes. In ascribing a DLT initiative as being good, we are undertaking an evaluation. Value is said to be the measure of goodness (Ng, 2013) and pragmatically we seek to make a judgment of what is good in our case. Our evaluation of DLTs is not based on judgement of an intrinsic or extrinsic goodness. Rather the judgement is based on the decisions made by the people who invent, develop, distribute and use them (Argandoña, 2003) in relation to the consequences of those technologies for the UN’s 17 SDGs and 169 targets which, on September 25th, 2015 the 193 Member States of the United Nations unanimously adopted. DLTs have, in some quarters, received an unfavourable press largely grounded in the observation that the DLT enabled cryptocurrencies – notably Bitcoin –have been associated with illicit and illegal activities such as drug dealing and arms trading (it should be said, a critique that applies equally to cash). Leading financial institutions and banking consortia are currently looking for ways to create their own permissioned or private cryptocurrency ecosystems, (as seen in the example of the ‘r3’ consortium[2]). Blockchain use cases focus, typically, on mapping the affordances DLTs might convey within largescale finance services (European Central Bank, 2012; Ali et al., 2014; McWaters et al., 2016).With little consensus about the potential impact of DLTs for good or ill, it is clear that the subject requires serious analysis. To focus on a single application or specific usage of the technology is to overlook its possible significance 2 http://r3members.com/ Page 8 of 32 ----- for ethical impacts at a global level. To ensure that the opportunities for ethical action potentially engrained in new technologies such as DLTs may be realized, it is important that the wider significance of the so-called ‘Blockchain for Good’ (B4G) is understood. The blockchain first appeared, largely unheralded, in 2008. Attention, instead, was directed toward the application whose existence blockchain technology made possible. The focal application and the first to run on a blockchain was the crypto-currency Bitcoin (Nakamoto, 2008; Lemieux, 2013). The significance of the underlying DLT is that it enables the digital transfer of value between two unknown entities without the need for a trusted third party. Simply put, DLT allows anyone to transact with anyone anywhere on a P2P basis. DLTs enhance the transparency of information exchanges (including payments and deposits), making trust obligations much easier to discharge between transacting parties. The service of value transfer is normally provided by intermediaries such as banks. DLT reallocates the responsibilities of transfer management to computers and algorithms (Ali et al., 2014; Welch, 2015; McWaters et al., 2016). Because of the way in which the technology is configured to allow P2P digital exchange of value, the blockchain, to many observers, represents a revolutionary and disruptive innovation (Swan, 2015; Zuberi and Levin, 2016). Fundamentally, a blockchain is a ledger of transactions of digital assets: of who owns what, who transacts what, of what is transacted and when. Transactions are not recorded on a single database but distributed on the computers of the network of users (nodes) of Page 9 of 32 ----- the system. No single entity owns or controls the ledger, and so network members can view the recorded transactions. Transactions are recorded and stored in ‘blocks,’ and each block linked chronologically (hence chain) and cryptographically to those which precede it to create an immutable, tamper-resistant record. All transactions are time stamped to provide a record of when transactions occurred and in what order: this assures against ‘double spending’ and tampering with previous transaction records (Reber and Feuerstein, 2014). The ledger is ‘kept honest’ by network consensus, a transaction validation process undertaken by network users, which includes checking that digital signatures are correct through a process known as ‘mining’: mining is incentivised by reward systems. Once a block is accepted by the network and added to the chain, it cannot be changed: it is a permanent, transparent and immutable record. Consequently, DLTs may be characterised as globally distributed, P2P, open ledgers of exchange providing an immutable and verifiable record and encrypting the identities of users that is hard to tamper with. Davidson et al. (2016) describe DLTs as a new general purpose technology which are, by definition, highly pervasive and can impact entire economies giving rise to creative destruction (Schumpeter, 1934; Jovanovic and Rousseau, 2005) with the potential to disrupt any centralized system that coordinates valuable information (Wright and De Filippi, 2015). DLTs represent a fundamental change in the way in which humans can exchange value, and two important implications follow. First, because the technology provides the required trust to give peers the confidence to exchange value directly, the requirement for socially-constructed institutional third-party providers of trust is significantly Page 10 of 32 ----- reduced: they become disintermediated. The second implication is that the blockchain presages a new functionality for the internet: it moves from an internet of information to _an internet of value (Swan, 2015). It means, that for objects that can be expressed in_ code, multiple novel application possibilities are opened up, and raises the question, how can blockchain technology that creates immutable, tamper-resistant distributed records of transactions of digital assets be applied in the service of SDGs? Mattila (2016) points out that the technology stack components of DLTs is diverse and can be configured in a variety of ways, resulting in different DLT architectures, implying the need for design decisions. Blockchains can be categorized as, for example, Permissioned/Permissionless and Specific Purpose Blockchains optimized for the management of assets and General Purpose Blockchains designed to allow users to write their own programmes to be stored on the blockchain and automatically executed in a distributed manner. Notwithstanding these divergences, DLTs share certain characteristics which may be more or less attenuated depending on the context of the application, in particular: the distributed (decentralized) consensus mechanism, immutability, algorithmic trust, resilience against manipulation, and secure information sharing. Nakamoto’s (2008) white paper describes what might be considered to be a pure form of DLT, that is to say a permissionless blockchain encompassing a network of participants that are not known to one another and each of them can access the blockchain with complete freedom to read or write to it, no actor can prevent any other actor from contributing content nor can any actor remove any previously validated Page 11 of 32 ----- contribution; and consensus is incentivised through economic mechanisms. Permissionless Blockchains are therefore highly censorship resistant and can provide an immutable[3], network-validated global record of transaction histories – right up to the present moment. On the other hand, anyone[4] may have a copy of the ledger in a permissioned blockchain, but only certain authorised parties may write to it and the consensus process is determined by the owner(s) of that blockchain, usually carried out by trusted actors in the network (CPTM, 2016). Assuming that chosen actors honestly and disinterestedly validate transactions, then permissioned blockchains can offer certain advantages, in at least two respects: first, they can be designed with specific functionality in mind and, second, alternatives to economically-incentivized validation mechanisms (proof-of work) can be incorporated. As a result, permissioned blockchains can be more efficient and faster than unpermissioned versions (CPTM, 2016) but at the cost of reduced security, immutability and censorship-resistance (Mattila, 2016). A sub-category of the permissioned blockchain is the private blockchain in which only certain authorised users have access to the database, whether for reading or writing, which tend to exist behind some organizational firewall but offer within-group transparency, privacy, and control, for a defined set of users. Whether or not they truly are DLTs continues to be debated, but the permissioned blockchain does have a role in helping deliver the SDG agenda. In the following, we explore some of these further and consider their affordance in terms of the SDGs. 3 Immutable to the extent that that particular blockchain continues to be maintained. It is not clear what happens in the circumstance that a particular blockchain ceases to be maintained by a network. 4 Anyone, subject to, of course, the nature of the permissions. Page 12 of 32 ----- _Blockchain mining_ In the Bitcoin blockchain, transactions are validated by network members (nodes) in a process known as mining. This distributed, network-member-driven process, performs the function of the centralized trusted third party intermediary model. Network participants compete with each other using computer power (known as proof-of-work) to validate blocks of transactions every 10 minutes or so. The proof-of-work is difficult to produce but easy for other nodes to verify and so transaction validity is established by majority consensus of network members. The miner that first successfully validates a block is rewarded with newly minted Bitcoins[5]. That network members commit resources to validating transactions, which in turn contributes to the cryptographic security and fraud resilience of the Bitcoin blockchain. The network is configured in such a way that it makes more sense for would-be attackers to participate as miners (greater opportunity for reward at lesser cost), thus increasing the resilience of the blockchain (Doguet, 2013; Fox-Brewster, 2015; Welch, 2015). However, the computationally intensive method of proof-of-work has been described as costly and wasteful (McWaters et al., 2016). As miners around the world competitively dedicate resources to validate transactions, Aste (2016) estimates about a billion Watts of electricity are consumed globally every second to produce a valid proof of work for Bitcoin. In light of this, alternative validation mechanisms are being investigated, some 5 For more details on mining, see Antonopoulos, A.M. (2014). Mastering Bitcoin: unlocking digital cryptocurrencies, O'Reilly Media, Inc.; Swan, M. (2015). Blockchain: Blueprint for a New Economy, O'Reilly Media, Inc., and: http://www.coindesk.com/information/how-Bitcoin-mining-works/ Page 13 of 32 ----- of which resonate with the SDG agenda but also relax some of the communitarian properties of the proof-of-work approach (such as openness to the whole community). Dierksmeier and Seele (2016) argue that it should be possible to promote ethical goals in society, by for example, hitching the ‘mining’ to the creation of ecological or social benefits. Certainly, reducing energy consumption in the process would ameliorate ecological harms and a small number of initiatives have emerged in this area. SolarCoin[6], for example, rewards generators of solar energy with new coin; another, GridCoin (Halford, 2014) introduces a novel algorithm based on work done in BOINC (Berkeley Open Infrastructure for Network Computing) projects: miners are incentivized to participate in scientific projects (as in healthcare and space exploration) aiming to provide benefit to humanity. In the CureCoin blockchain, the Bitcoin validation calculations are replaced by (useful) protein folding tasks: mining CureCoin helps science through simulating protein behaviour and providing these data to research scientists. _The internet of value(s)_ The previous section describes how social or ecological benefit can be linked to the production of alt-currencies. This section focuses on how these benefits can be related to currency use. The notion of coloured coins (Bradbury, 2013) is used to denote a small part of a coin with specific attributes which may represent anything from physical assets to a community’s values. By moving coloured coins through the network, asset ownership can be securely transferred. Similarly, coins coloured with values, in which 6 https://solarcoin.org/ Page 14 of 32 ----- morals, principles or ethics are embedded in the code, can allow individuals to align their spending closely with their values. Taghiyeva et al. (2016) describe a proof-of-concept pilot for a blockchain-based Islamic crypto-currency in which transactions and Muslim values, including a blended anti radicalisation agenda, are aligned: a currency with a community’s desirable social principles engineered-in. This resonates with Helbing’s (2013, 2014) concept of _Qualified Money where values can be embedded in DLTs. CarbonCoin[7] claims to be the_ first digital currency with a conscience, designed to engage the environmentally conscious community. Such possibilities raise important questions about whose values are embedded into a currency and who does the engineering. In terms of assets, DLTs provide a mechanism both for their registration and transfer. A number of commentators have argued that this may prove a boon in developing or politically unstable economies for the registration of individual’s property rights. Where there is a lack of trust in central authorities to maintain uncorrupted registers of assets, such as property title, these may be recorded immutably, transparently, and verifiably on a blockchain. A number of pilots and trial projects are underway: Bitland[8] use DLT to map land title in Ghana providing a registry of ownership which subsequently facilitates the mobilization of capital as well as a transparent property market. Similar initiatives can be found in Honduras (Alejandro, 2016), Sweden (Rizzo, 2016) and Georgia (Shin, 2016). Progress has been slow and success mixed (ODI, 2016), attesting to the still emergent nature of the technology. Indeed, it is too easy to get carried away 7 http://carboncoin.cc/ 8 http://bitlandglobal.com Page 15 of 32 ----- by the theoretical potential of DLTs. While a blockchain based registry of assets may be transparent and immutable, for it to be meaningful in terms of economic participation and activity it must exist within a stable infrastructure: armed aggressors, for example, may still unlawfully seize property regardless of whether or not it is recorded on the blockchain. However, the existence and immutability of the record may act as a deterrent against such behaviour. _Supply chains_ Assets can be registered to the blockchain using unique keys. This provides a register of ownership as well as tracking and pattern of ownership over time. Initiatives that have leveraged this affordance include Everledger[9], a permanent ledger for diamond certification and related transaction history transparently recording ownership history and reducing crime, and Provenance[10] who provide a system for tracking materials and products in a manner that is public, secure and inclusive. For the SDGs, this means that claims (albeit excluding those blood diamonds or sustainably fished tuna) can be demonstrated to be authentic right through the supply chain, shifting the value system towards origin and provenance (Greenspan, 2015). DLT applications are also being explored in the energy market both as a system enabling individuals to sell excess solar-generated electricity to each other without going through third parties (such as PowerLedger[11] and TransActive[12]) as well as developing a market infrastructure for carbon trading, an independent ledger of the 9 http://www.everledger.io/ 10 https://www.provenance.org/ 11 http://powerledger.io/ 12 http://transactivegrid.net/ Page 16 of 32 ----- permits to emit Earth’s allowance of greenhouse gases (Casalotti, 2016). One scenario is that, within a short time, every individual on the planet, for example, be issued with an annual carbon allocation that may be traced via the DLT network. _Innovations in governance_ Within the DLT code substitutes for trust and allows for new types of commerce. Appropriately designed, these can be the building blocks of new forms of economic and social governance that meet the objectives of the SDGs. Smart contracts are computer protocols that facilitate, verify and enforce the performance of a contract: self-executing code. They are the automation of the performance of contracts which only execute when pre-specified conditions are met, thus removing the need for third party resolution. This is an assured and low-cost mechanism that can offer for Bottom of the Pyramid economic actors increased speed, efficiency, and trust that the contract will be executed as agreed, thus enabling arm’s length transactions and payments triggered on receipt of goods. A further application is in the realm of providing more secure and inclusive voting and elections. The danger, of course, is that the contract performs no matter what: this raises questions about who writes them (Quis custodiet ipsos custodies?), how to write-in flexibility to respond to and incorporate external events, and individual’s free will in connecting with them. It is a small step from smart contracts to Decentralized Autonomous Organizations (DAOs) which are similarly executed by code but, unlike smart contracts, may include a potentially unlimited number of participants (Buterin, 2014). DAOs remain largely Page 17 of 32 ----- untested and use cases relating to SDGs are hard to find: nevertheless, indicative of the infancy of the technology, one major DAO initiative fell victim to misappropriation of approximately $80m (Price, 2016), indicating the need for further developmental work. One area where the concept has been developed is in the creation of DLT mediated organisations made of people but where the governance structure is encoded directly into the technical infrastructure stipulating and enabling the rules and procedures of the organisation that every member of the organisation will have to abide by such design propositions may help to eliminate fraud and corruption. _Sharing economy_ The sharing economy has been heralded as one solution to the challenges of sustainability by promoting environmentally sensitive forms of consumption, encouraging different models of ownership and addressing issues such as the under utilisation of assets. However, some scholars recognise a Dark Side (Malhotra and Van Alstyne, 2014), partly for its tendency to reinforce the contemporary unsustainable economic paradigm (Martin, 2016), partly because some providers’ business models are argued to be as much about evading regulations as about sharing, partly for spreading precarity throughout the workforce, for middlemen sucking profits out of previously un-monetized interactions (Scholz, 2016) and for being unavailable to disadvantaged groups, those of low socioeconomic status and users from emerging regions (Thebault-Spieker et al., 2015). DLTs address some of these criticisms by decentralising and disintermediating. Embedding sensors into existing assets, our ‘things’ can collect and share data. By Page 18 of 32 ----- integrating these data into the blockchain, we can keep an immutable ledger of shared transactions without the need for middlemen (Huckle et al., 2016). La’Zooz[13] is a decentralized transportation platform owned by the community and utilising vehicles’ unused space, enabling people with private cars to share their drive with others travelling the same route: a decentralized Uber. La’Zooz generates new tokens from ‘Proof of Movement’ not ‘Proof of Work.' As they drive, drivers earn Zooz, passengers pay using Zooz and can also earn Zooz by providing route advice to drivers. La’Zooz offers to provide a ride-sharing service that is based on truer sharing economy principles, rather than monetary incentives (Bheemaiah, 2015). The business model moves from rent extraction to value creation in networks: value is distributed amongst those who created it, offering a greater reward and opportunity for inclusion. _Financial inclusion_ The opportunity for wider financial inclusion is held up as one of the great promises for SDGs of DLTs. Through automation, disintermediation, low cost and security of transfer comes the opportunity for transactions involving low-value units and for remote, disenfranchised, peripheral and marginal communities to connect in new ways either amongst themselves or with activities in the wider world. DLTs allow the almost instantaneous transfer of digital tokens, if not at zero cost then at a significantly cheaper rate than established services. This makes the transfer of small amounts of currency economically viable, enabling new actors to enter the field and new opportunities for e 13 http://lazooz.net/ Page 19 of 32 ----- commerce (Athey, 2015). It might be anticipated, then, that reductions in the cost of financial transactions through DLTs will result in widening financial inclusion. One critical factor in enabling greater financial inclusion is identity which, it is argued (Birch, 2014) will underpin future digital transactions and lies at the heart of realising the potential of DLT. The question of what defines identity is challenging, not least because it ‘does not lend itself easily to definition nor does it remain unchangeable’ (Ajana, 2010, p.5). Identities are made up of multiple attributes: date and place of birth, parents’ names, school, criminal record, employment record, biometrics, papers published, etc. These attributes reflect who we are and are configurable depending on whom we need to identify ourselves to and for what purpose. For most, it is relatively straightforward to assemble authenticated attributes of identity (passport, utility bill, etc.), but approximately 1.8bn of the world’s population have no legally recognised identity (Dahan and Gelb, 2015). The reasons are various, but the consequence is that the ‘identityless’ exist on the margins of society unable formally to participate in democratic, educative, healthcare and economic activity. Part of the problem of identitylessness is the extent to which identity has been a centralised phenomenon, something that, to a large extent, is given to people by some authority. The affordances of DLTs offer an alternative approach to building identities from the bottom up, as the gradual accretion of different attributes of identity. This way, an individual’s identity is not under the control or the gift of any central authority, nor is it vulnerable to tampering or theft from malicious third parties. Further, individuals are Page 20 of 32 ----- able to control which attributes may/may not be made public depending on the authentication need. This is currently an area of intense DLT development including initiatives from ID2020[14], BitNation[15], BlockchainBorderBank[16], BanQu[17], and NevTrace[18]. **Conclusion** In 2013 Nobel Prize-winning economist, Paul Krugman declared that ‘Bitcoin is evil.’ Others, too, have been critical (Lemieux, 2013; Doguet, 2013; Fox-Brewster, 2015; Welch, 2015; Böhme et al., 2015). Despite these criticisms, DLTs have also been heralded as an incremental innovation with the potential for inducing efficiency gains _and ethically empowering business, or disruptive innovations (triggering the emergence_ of new economic systems), that may prove to be more socially and environmentally responsible (Swan, 2015; Davidson et al., 2016; Walport, 2016). This paper has explored, through affordance theory, how DLTs might contribute to the sustainability agenda. On the face of it, the potential appears significant. DLTs provide a technical basis for a degree of change that many observers have found exciting. The way we relate to DLTs is not merely a technical matter but strongly relates to the ways in which we configure our social world (Reijers and Coeckelbergh, 2016). Consequently, we propose the notion of Blockchain for Good as an emergent phenomenon or shared interpretative schema that is being co-constructed by a wide ecosystem of actors as a means of giving direction and catalyzing actions, choices, and 14 http://id2020.org/ 15 https://bitnation.co/ 16 http://law.mit.edu/blockchainborderbank 17 http://www.banquap.com/ 18 http://nevtrace.com/ Page 21 of 32 ----- behaviours (Ranson et al., 1980). Crucially, this approach unlocks the potential for more detailed examinations of the moral and ethical impetus behind blockchain projects. Within this limited space, we have presented a rather one-sided perspective and are aware that DLTs are not a universal panacea. The notion of Blockchain for Good inevitably raises questions about its counter, ‘Blockchain for Bad’ and there exists, beyond the scope of this paper, a body of cautionary literature. Analysing crypto currencies through the lens of ethical impact, Dierksmeier and Seele (2016) also find detrimental outcomes, such as the facilitation of nefarious consumption. Physicist Stephen Hawking, Elon Musk and, as of 12 November 2016, 8,749 others have signed an open letter counselling against the incautious application of artificial intelligence and DAOs (Russell et al., 2015). DLTs feel no guilt, regret or remorse. This raises questions about who will do the coding. As yet, there is little regulation specific to DLT. Still, might DLTs yet be subsumed by incumbent organizations and authorities as another tool of control and surveillance, or can they really deliver a more democratic, egalitarian, collaborative and sustainable society? DLTs are still at an early stage of development, and it remains unclear in which direction they will go. The essential premise of technology affordance is that to understand the uses and consequences of technologies, they must be considered in the context of their dynamic interactions between people and organizations (Majchrzak and Markus, 2012), DLTs are a case-in-point. Dozens of crypto-currencies now exist, each optimized for different purposes, each idiosyncratic in terms of its operation, uptake, Page 22 of 32 ----- exchange rate and convertibility. Similarly, others are exploring DLT applications that are not currency-oriented. Given this variety, further research is required to understand which type works best in which circumstances and why, as well as the extent to which they can deliver on the sustainability agenda. **References** Ajana B. 2010. Recombinant identities: Biometrics and narrative bioethics. Journal of _Bioethical Inquiry_ **7: 237-258.** Alejandro J. 2016. Blockchain for good – a beginner’s guide. Available at: http://tech.newstatesman.com/guest-opinion/blockchain-for-good-beginners-guide Ali R, Barrdear J, Clews R, Southgate J. 2014. Innovations in payment technologies and the emergence of digital currencies. Bank of England Quarterly Bulletin **54(3): 262-275.** Argandoña A. 2003. The New Economy: Ethical issues. Journal of Business Ethics **44:** 3-22. Aste T. 2016. The Fair Cost of Bitcoin Proof of Work. Available at: https://ssrn.com/abstract=2801048 Athey S. 2015. 5 ways digital currencies will change the world. Available at: https://www.weforum.org/agenda/2015/01/5-ways-digital-currencies-will-change-the world/. Bateman A. 2015. Tracking the value of traceability. Supply Chain Management Review (November): 8-10. Benkler Y, Nissenbaum H. 2006. Commons‐based peer production and virtue. Journal _of Political Philosophy_ **14: 394-419.** Page 23 of 32 ----- Bheemaiah K. 2015. Why business schools need to teach about the blockchain. Available at: http://ssrn.com/abstract=2596465. Birch D. 2014. Identity Is The New Money. London: London Publishing Partnership. Böhme R, Christin N, Edelman B, Moore T. 2015. Bitcoin: Economics, technology, and governance. The Journal of Economic Perspectives **29: 213-238.** Botsman R, Rogers R. 2010. What's Mine Is Yours: How Collaborative Consumption is _Changing the Way We Live. HarperBusiness: London._ Beynon-Davies, P. and Lederman, R., 2017. Making sense of visual management through affordance theory. Production and Planning, 28(2): 142-157. Boyd, D., 2010. Social networked sites as networked publics: Affordances, dynamics and implications. In: Z. Papacharissi, ed. Networked Self: Identity, Community, and _Culture on Social Networking Sites. Routledge: New York; 39-58._ Bradbury D. 2013. Colored Coins Paint Sophisticated Future for Bitcoin. Available at: http://www.coindesk.com/colored-coins-paint-sophisticated-future-for-Bitcoin/. Buterin V. 2014. DAOs, DACs, DAS and more: An incomplete terminology guide. Available at: https://blog.ethereum.org/2014/05/06/daos-dacs-das-and-more-an incomplete-terminology-guide/ Choy K, Schlagwein, D.2016. Crowdsourcing for a better world: On the relation between IT affordances and donor motivations in charitable crowdsourcing. Information _Technology and People 29(1): 221-247._ Ciavola B, Gershenson, J. 2016. Affordance theory for engineering design. Research _Engineering Design 27: 251-263._ Page 24 of 32 ----- Conole G, Dyke M. 2004. What are the affordances of information and communication technologies? ALT-J Research in Learning Technology **12(2): 113-124.** Casalotti A. 2016. Global Carbon Trading on the Blockchain. Bitcoin and Blockchain Leadership Forum. Available at: http://cptm.org/documents/CFMM_Brief_percent202016.pdf CPTM. 2016. Commonwealth Partnership for Technology Management (CPTM) Brief on Adaptive Flexibility Approaches to Financial Inclusion in a Digital Age: Recommendations and Proposals. London: CPTM Smart Partners’ Hub. Dahan M, Gelb A. 2015. The Role of Identification in the Post-2015 Development Agenda. World Bank Working Paper. Davidson S, De Filippi P, Potts J. 2016. Economics of Blockchain. Available at: http://ssrn.com/abstract=2744751. Dierksmeier C, Seele P. 2016 Cryptocurrencies and Business Ethics. Journal of _Business Ethics 1-14._ Doguet J. 2013. The nature of the form: Legal and regulatory issues surrounding Bitcoin digital currency system. Louisiana Law Review **73(4):1118-1153.** European Central Bank. 2012. Virtual Currency Schemes. Available at: https://www.ecb.europa.eu/pub/.../virtualcurrencyschemes201210en.pdf Fox-Brewster T. 2015. How hackers abused Tor to rob blockchain, steal Bitcoin, target private email and get away with it. Forbes **24 (02):7.** Frankena WK.1973. Ethics. Prentice Hall: Englewood Cliffs. Page 25 of 32 ----- Faraj J, Jarvenpaa S, Majchrzak A. 2011. Knowledge collaboration in online communities. Organization Science 22(5): 1224-1239. Gaver W. 1991. Technology Affordances. ACM: New Orleans, Lousiana, New York: 79-84. Geels FW. 2005. Technological Transitions and System Innovations: A Co _Evolutionary and Socio-Technical Analysis. Edward Elgar Publishing: Cheltenham UK:_ Gibson JJ. 1978. The ecological approach to the visual perception of pictures. Leonardo **11: 227-235.** Greenspan G. 2015. MultiChain Private Blockchain — White Paper. Available at: http://www.multichain.com/white-paper/: Multichain Halford R. 2014. Gridcoin - Crypto-Currency using Berkeley Open Infrastructure Network Computing Grid as a Proof Of Work. Available at: http://www.gridcoin.us/images/gridcoin-white-paper.pdf Helbing D. 2013. Economics 2.0: The natural step towards a self-regulating, participatory market society. Evolutionary and Institutional Economics Review **10: 3-** 41. Helbing D. 2014. Qualified Money-A Better Financial System for the Future. Available at: 2526022. Huckle S, Bhattacharya R, White M, Beloff N. 2016. Internet of things, blockchain and shared economy applications. Procedia Computer Science **98: 461-466.** Jovanovic B, Rousseau PL. 2005. General purpose technologies. Handbook of _Economic Growth_ **1: 1181-1224.** Page 26 of 32 ----- Kewell B. 2007. Linking risk and reputation: A research agenda and methodological analysis.’ Risk Management: An International Journal **9(4): 238-254.** Kranzberg M. 1986. Technology and History: ‘Kranzberg's Laws.’ Technology and _Culture_ **27: 544-560.** Krugman PR. 2013. Bitcoin is evil. Available at: http://krugman.blogs.nytimes.com/2013/12/28/Bitcoin-is-evil/?_r=1: New York Times. Lee E. 2005. The ethics of innovation: P2P software developers and designing substantial noninfringing uses under the Sony Doctrine. Journal of Business Ethics **62:147-162.** Lemieux P. 2013. Who is Satoshi Nakamoto? Regulation (Fall):14-15. Lankton N, McKnight D, Tripp J. 2015. Technology, humanness, and trust: rethinking trust in technology. Journal of the Association for Information Systems 16(10): 880-918. Maier J, Fadel G. 2009. Affordance-based design: A relational theory of design. _Research Design Engineering 20: 13-27._ Majchrzak A, Markus L. 2012. Technology Affordances and Constraint Theory of MIS. Thousand Oaks: CA. Sage Publishing. Malhotra A, Van Alstyne M. 2014. The dark side of the sharing economy… and how to lighten it. Communications of the ACM **57: 24-27.** Martin CJ. 2016. The sharing economy: A pathway to sustainability or a nightmarish form of neoliberal capitalism? Ecological Economics **121: 149-159.** Martin, K.E, Freeman, R.E. 2004. The separation of technology and ethics in Business Ethics. Journal of Business Ethics **53: 353-364.** Page 27 of 32 ----- Mattila J. 2016. The Blockchain Phenomenon: The Disruptive Potential of Distributed Consensus Architectures. Available at: http://brie.berkeley.edu/BRIE/: Berkeley Roundtable on the International Economy (BRIE), University of California, Berkeley. McWaters R, Galaski R, Chaterjee S. 2016. The Future of Financial Infrastructure: An Ambitious Look at How Blockchain can Reshape Financial Services. World Economic _Forum Available at: https://www.weforum.org/reports/the-future-of-financial-_ infrastructure-an-ambitious-look-at-how-blockchain-can-reshape-financial-services/ Nakamoto S. 2008. Bitcoin: A peer-to-peer electronic cash system. Consulted, 1, 28. Ng I. 2013. Value and Worth: Creating New Markets In The Digital Economy. Innovorsa: Cambridge. ODI. 2016. Applying Blockchain Technology in Global Data Infrastructure. Available at: http://theodi.org/technical-report-blockchain-technology-in-global-data infrastructure: Open Data Institute Pea RD. 1993. Practices for distributed intelligence and designs for education. In Salomon G (ed) Distributed Cognitions: Psychological and Educational _Considerations. Cambridge University Press: Cambridge, UK; 47-87._ Potts J, Davidson S, De Filippi P. 2016. Disrupting governance: The new institutional _economics of Distributed Ledger Technology. Available at:_ http://ssrn.com/abstract=1295507 Price R. 2016. Digital currency Ethereum is cratering amid claims of a $50 million hack. Business Insider (17 June) Available at: http://uk.businessinsider.com/dao hacked-ethereum-crashing-in-value-tens-of-millions-allegedly-stolen-2016-6 Page 28 of 32 ----- Ranson S, Hinings B, Greenwood R. 1980. The structuring of organizational structures. _Administrative Science Quarterly_ **25: 1-17.** Reber D, Feuerstein S. 2014. Bitcoins-hype or real alternative? Internet Economics VIII (81). Reijers W, Coeckelbergh, M. 2016. The blockchain as a narrative technology: Investigating the social ontology and normative configurations of cryptocurrencies. _Philosophy of Technology (October):1-28._ Rizzo P. 2016. Sweden tests blockchain smart contracts for Land Registry. CoinDesk June 16. Russell SV, Dewey D, Tegmark M. 2015. Research priorities for robust and beneficial artificial intelligence: An open letter. Association for the Advancement of Artificial Intelligence. Scholz T. 2016. Platform Cooperativism: Challenging the Corporate Sharing Economy. New York. Available at : www.rosalux-nyc.org, Rosa Luxemburg Stiftung Schumpeter JA. 1934. The Theory of Economic Development: An Inquiry into Profits, _Capital, Credit, Interest and the Business Cycle. London, UK: Oxford University Press._ Seidel S, Recker J, Vom Brocke J. 2013. Sensemaking and sustainable practicing: Functional affordances of information systems in green transformations. MIS Quarterly **37:** 1275-1299. Shin L. 2016. Republic of Georgia to develop blockchain Land Registry system. _Forbes, 21 April._ Smith JW. 1948. Intrinsic and extrinsic good. Ethics **58: 195-208.** Page 29 of 32 ----- Swan M. 2015. Blockchain: Blueprint for a New Economy. O'Reilly Media, Inc.: Beijing, Cambrdige, Farnham, Koln, Sebastopol, Tokyo. Taddeo M, Vaccaro A. 2011. Analyzing peer-to-peer technology using information ethics. Information Society **27: 105-112.** Taghiyeva M, Mellish B, Ta'eed O. 2016. Currency of Intangible Non-Financial Value. Available at: https://github.com/seratio/whitepaper: IoV Blockchain Alliance for Good (www.bisgit.org) Thebault-Spieker J, Terveen LG, Hecht B. 2015. Avoiding the South Side and the Suburbs: The Geography of Mobile Crowdsourcing Markets. Proceedings of the 18th ACM Conference on Computer Supported Cooperative Work and Social Computing; 265-275. UN. 2015. Transforming our world: The 2030 Agenda for Sustainable Development. Available at: https://sustainabledevelopment.un.org/post2015/transformingourworld. Volkoff O, Strong DM. 2013. Critical realism and affordances: Theorizing IT associated organizational change processes. MIS Quarterly **37: 819-834.** Xenakis I, Arnellos A. 2013. The relation between interaction aesthetics and affordances. Design Studies, 34 (1): 57-73. Walport. 2016. Distributed Ledger Technology: Beyond Blockchain. Government Office for Science. Available at: https://www.gov.uk/government/publications/distributed-ledger-technology-blackett review WCED. 1987. Our Common Future. World Commission on Environment and _Development. Oxford University Press: Oxford, UK._ Page 30 of 32 ----- Welch A. 2015. The Bitcoin blockchain as financial market infrastructure: A consideration of operational risk. Legislation and Public Policy **8: 837-893.** Withagen R, Chemero, A. 2012. Affordances and classification: On the significance of a sidebar in James Gibson's last book. Philosophical Technology 25(4): 521-537. Wright A, De Filippi P.2015. Decentralized Blockchain Technology and the Rise of Lex _Cryptographia. Available at:_ https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2580664 Zuberi M, Levin R. 2016. Schumpeter's Revenge: The gale of creative destruction. _Banking and Financial Services Policy Report_ **35(5):1-8.** Zammuto R, Griffith T, Majchrak A, Dougherty D, Faraj S. 2007. Information technology and the changing fabric of organization. Organizational Science 18(5): 749 762. **Biographical Notes** Beth Kewell is a Research Fellow at Surrey University Business School’s Centre for the Digital Economy (CoDE), where she specialises in interpretative research, positioned at the boundary between innovation management, Science and Technology Studies (STS), and risk analysis. Correspondence to: Surrey Centre for the Digital Economy, University of Surrey, Surrey GU2 7XH, UK. [Email: e.kewell@surrey.ac.uk](mailto:e.kewell@surrey.ac.uk) Richard Adams is a Senior Research Fellow at Surrey University Business School’s Centre for the Digital Economy (CoDE). His research interests lie at the intersection of (responsible) innovation, digital disruption, and sustainability and business models. Correspondence to: Surrey Centre for the Digital Economy, University of Surrey, Surrey GU2 7XH, UK. [Email: r.adams@surrey.ac.uk](mailto:r.adams@surrey.ac.uk) Page 31 of 32 ----- Glenn Parry is Professor of Strategy and Operations Management at Bristol Business School, University of the West of England. He is primarily interested in what 'Good' means for an organisation, exploring value as a measurement of 'goodness'. He uses business models as a framework to understand value co-creation between provider and client in context. Correspondence to: Faculty of Business and Law, UWE Frenchay Campus, Coldharbour Lane, Bristol, BS16 1QY, UK. [Email: Glenn.Parry@uwe.ac.uk](mailto:Glenn.Parry@uwe.ac.uk) Page 32 of 32 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1002/JSC.2143?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1002/JSC.2143, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://openresearch.surrey.ac.uk/view/delivery/44SUR_INST/12139402000002346/13140709980002346" }
2,017
[]
true
2017-09-01T00:00:00
[]
13,312
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/011f7eb8584458d0676216165e63dbefce1d393a
[ "Computer Science" ]
0.887264
MINERVA infinity : A Scalable Efficient Peer-to-Peer Search Engine.
011f7eb8584458d0676216165e63dbefce1d393a
[ { "authorId": "2241164441", "name": "Sebastian Michel" }, { "authorId": "47463435", "name": "P. Triantafillou" }, { "authorId": "1751591", "name": "G. Weikum" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
# MINERVA : A Scalable Efficient Peer-to-Peer ∞ Search Engine Sebastian Michel[1], Peter Triantafillou[2], and Gerhard Weikum[1] 1 Max-Planck-Institut f¨ur Informatik, 66123 Saarbr¨ucken, Germany _{smichel, weikum}@mpi-inf.mpg.de_ 2 R.A. Computer Technology Institute and University of Patras, 26500 Greece peter@ceid.upatras.gr **Abstract. The promises inherent in users coming together to form data** sharing network communities, bring to the foreground new problems formulated over such dynamic, ever growing, computing, storage, and networking infrastructures. A key open challenge is to harness these highly distributed resources toward the development of an ultra scalable, efficient search engine. From a technical viewpoint, any acceptable solution must fully exploit all available resources dictating the removal of any centralized points of control, which can also readily lead to performance bottlenecks and reliability/availability problems. Equally importantly, however, a highly distributed solution can also facilitate pluralism in informing users about internet content, which is crucial in order to preclude the formation of information-resource monopolies and the biased visibility of content from economically-powerful sources. To meet these challenges, the work described here puts forward MINERVA, a novel search _∞_ engine architecture, designed for scalability and efficiency. MINERVA _∞_ encompasses a suite of novel algorithms, including algorithms for creating data networks of interest, placing data on network nodes, load balancing, top-k algorithms for retrieving data at query time, and replication algorithms for expediting top-k query processing. We have implemented the proposed architecture and we report on our extensive experiments with real-world, web-crawled, and synthetic data and queries, showcasing the scalability and efficiency traits of MINERVA . _∞_ ## 1 Introduction The peer-to-peer (P2P) approach facilitates the sharing of huge amounts of data in a distributed and self-organizing way. These characteristics offer enormous potential benefit for the development of internet-scale search engines, powerful in terms of scalability, efficiency, and resilience to failures and dynamics. Additionally, such a search engine can potentially benefit from the intellectual input (e.g., bookmarks, query logs, click streams, etc.) of a large user community participating in the sharing network. Finally, but perhaps even more importantly, a P2P web search engine can also facilitate pluralism in informing users about internet content, which is crucial in order to preclude the formation of information-resource monopolies and the biased visibility of content from economically powerful sources. G. Alonso (Ed.): Middleware 2005, LNCS 3790, pp. 60–81, 2005. _⃝c_ IFIP International Federation for Information Processing 2005 ----- MINERVA : A Scalable Efficient P2P Search Engine 61 _∞_ Our challenge therefore was to exploit P2P technology’s powerful tools for efficient, reliable, large-scale content sharing and delivery to build a P2P web search engine. We wish to leverage DHT technology and build highly distributed algorithms and data infrastructures that can render P2P web searching feasible. The crucial challenge in developing successful P2P Web search engines is based on reconciling the following high-level, conflicting goals: on the one hand, to respond to user search queries with high quality results with respect to precision/recall, by employing an efficient distributed top-k query algorithm, and, on the other hand, to provide an infrastructure ensuring scalability and efficiency in the presence of a very large peer population and the very large amounts of data that must be communicated in order to meet the first goal. Achieving ultra scalability is based on precluding the formation of central points of control during the processing of search queries. This dictates a solution that is highly distributed in both the data and computational dimensions. Such a solution leads to facilitating a large number of nodes pulling together their computational (storage, processing, and communication) resources, in essence increasing the total resources available for processing queries. At the same time, great care must be exercised in order to ensure efficiency of operation; that is, ensure that engaging greater numbers of peers does not lead to unnecessary high costs in terms of query response times, bandwidth requirements, and local peer work. With this work, we put forward MINERVA, a P2P web search engine _∞_ architecture, detailing its key design features, algorithms, and implementation. MINERVA features offer an infrastructure capable of attaining our scalability _∞_ and efficiency goals. We report on a detailed experimental performance study of our implemented engine using real-world, web-crawled data collections and queries, which showcases our engine’s efficiency and scalability. To the authors’ knowledge, this is the first work that offers a highly distributed (in both the data dimension and the computational dimension), scalable and efficient solution toward the development of internet-scale search engines. ## 2 Related Work Recent research on structured P2P systems, such as Chord [17], CAN [13], SkipNets [9] or Pastry [15] is typically based on various forms of distributed hash tables (DHTs) and supports mappings from keys to locations in a decentralized manner such that routing scales well with the number of peers in the system. The original architectures of DHT-based P2P networks are typically limited to exact-match queries on keys. More recently, the data management community has focused on extending such architectures to support more complex queries [10,8,7]. All this related work, however, is insufficient for text queries that consist of a variable number of keywords, and it is absolutely inappropriate for full-fledged Web search where keyword queries should return a ranked result list of the most relevant approximate matches [3]. Within the field of P2P Web search, the following work is highly related to our efforts. Galanx [21] is a P2P search engine implemented using the Apache HTTP ----- 62 S. Michel, P. Triantafillou, and G. Weikum server and BerkeleyDB. The Web site servers are the peers of this architecture; pages are stored only where they originate from. In contrast, our approach leaves it to the peers to what extent they want to crawl interesting fractions of the Web and build their own local indexes, and defines appropriate networks, structures, and algorithms for scalably and efficiently sharing this information. PlanetP [4] is a pub/sub service for P2P communities, supporting content ranking search. PlanetP distinguishes local indexes and a global index to describe all peers and their shared information. The global index is replicated using a gossiping algorithm. This system, however, appears to be limited to a relatively small number of peers (e.g., a few thousand). Odissea [18] assumes a two-layered search engine architecture with a global index structure distributed over the nodes in the system. A single node holds the complete, Web-scale, index for a given text term (i.e., keyword or word stem). Query execution uses a distributed version of Fagin’s threshold algorithm [5]. The system appears to create scalability and performance bottlenecks at the single-node where index lists are stored. Further, the presented query execution method seems limited to queries with at most two keywords. The paper actually advocates using a limited number of nodes, in the spirit of a server farm. The system outlined in [14] uses a fully distributed inverted text index, in which every participant is responsible for a specific subset of terms and manages the respective index structures. Particular emphasis is put on minimizing the bandwidth used during multi-keyword searches. [11] considers content-based retrieval in hybrid P2P networks where a peer can either be a simple node or a directory node. Directory nodes serve as super-peers, which may possibly limit the scalability and self-organization of the overall system. The peer selection for forwarding queries is based on the Kullback-Leibler divergence between peerspecific statistical models of term distributions. Complementary, recent research has also focused into distributed top-k query algorithms [2,12] (and others mentioned in these papers which are straightforward distributed versions/extensions of traditional centralized top-k algorithms, such as NRA [6]). Distributed top-k query algorithms are an important component of our P2P web search engine. All these algorithms are concerned with the efficiency of top-k query processing in environments where the index lists for terms are distributed over a number of nodes, with index lists for each term being stored in a single node, and are based on a per-query coordinator which collects progressively data from the index lists. The existence of a single node storing a complete index list for a term undoubtedly creates scalability and efficiency bottlenecks, as our experiments have showed. The relevant algorithms of MINERVA ensure high degrees of distribution for index lists’ data and _∞_ distributed processing, avoiding central bottlenecks and boosting scalability. ## 3 The Model In general, we envision a widely distributed system, comprised of great numbers of peers, forming a collection with great aggregate computing, communication, ----- MINERVA : A Scalable Efficient P2P Search Engine 63 _∞_ and storage capabilities. Our challenge is to fully exploit these resources in order to develop an ultra scalable, efficient, internet-content search engine. We expect that nodes will be conducting independent web crawls, discover ing documents and computing scores of documents, with each score reflecting a document’s importance with respect to terms of interest. The result of such activities is the formation of index lists, one for each term, containing relevant documents and their score for a term. More formally, our network consists of a set of nodes N, collectively storing a set D of documents, with each document having a unique identifier docID, drawn from a sufficiently large name space (e.g., 160 bits long). Set T refers to the set of terms. The notation _S_ denotes the cardi_|_ _|_ nality of set S. The basic data items in our model are triplets of the form (term, _docID, score). In general, nodes employ some function score(d, t) : D_ (0, 1], _→_ which for some term t, produces the score for document d. Typically, such a scoring function utilizes tdf*idf style statistical metadata. The model is based on two fundamental operations. The Post(t, d, s) op eration, with t _T, d_ _D, and s_ (0, 1], is responsible for identifying a _∈_ _∈_ _∈_ network node and store there the (t, d, s) triplet. The operation Query(Ti, k) : _return(Lk), with Ti ⊆_ _T, k an integer, and Lk = {(d, T otalScore(d)) : d ∈_ _D, T otalScore(d)_ _RankKscore_, is a top-k query operation. T otalScore(d) _≥_ _}_ denotes the aggregate score for d with respect to terms in Ti. Although there are several possibilities for the monotonic aggregate function to be used, we employ summation, for simplicity. Hence, T otalScore(d) = [�]t∈Ti _[score][(][d, t][). For a]_ given term, RankKscore refers to the k-th highest TotalScore, smin (smax) refers to the minimum (maximum) score value, and, given a score s, next(s) (prev(s)) refers to the score value immediately following (preceding) s. All nodes are connected on a global network G. G is an overlay network, modeled as a graph G = (N, E), where E denotes the communication links connecting the nodes. E is explicitly defined by the choice of the overlay network; for instance, for Chord, E consists of the successor, predecessor, and finger table (i.e., routing table) links of each node. In addition to the global network G, encompassing all nodes, our model employs term-specific overlays, coined Term Index Networks (TINs). I(t) denotes the TIN for term t and is used to store and maintain all (t, d, s) items. TIN I(t) is defined as I(t) = (N (t), E(t)), N (t) _N_ . Note that nodes in N (t) have _⊆_ in addition to the links for participating in G, links needed to connect them to the I(t) network. The model itself is independent of any particular overlay architecture. _I(t).n(si) defines the node responsible for storing all triplets (t, d, s) for which_ _score(d, t) = s = si. When the context is well understood, the same node is_ simply denoted as n(s). ## 4 Design Overview and Rationale The fundamental distinguishing feature of MINERVA is its high distribu_∞_ tion both in the data and computational dimensions. MINERVA goes far _∞_ ----- 64 S. Michel, P. Triantafillou, and G. Weikum beyond the state of the art in distributed top-k query processing algorithms, which are based on having nodes storing complete index lists for terms and running coordinator-based top-k algorithms [2,12]. From a data point of view, the principle is that the data items needed by top-k queries are the triplets (term, docID, score) for each queried term (and not the index lists containing them). A proper distributed design for such systems then should appropriately distribute these items controllably so to meet the goals of scalability and efficiency. Thus, data distribution in MINERVA is at the level of this, much finer _∞_ data grain. From a system’s point of view, the design principle we follow is to organize the key computations to engage several different nodes, with each node having to perform small (sub)tasks, as opposed to assigning single large task to a single node. These design choices, we believe, will greatly boost scalability (especially under skewed accesses). Our approach to materializing this design relies on the employment of the novel notion of Term Index Networks (TINs). TINs may be formed for every term in our system, and they serve two roles: First, as an abstraction, encapsulating the information specific to a term of interest, and second, as a physical manifestation of a distributed repository of the term-specific data items, facilitating their efficient and scalable retrieval. A TIN can be conceptualized as a virtual node storing a virtually global index list for a term, which is constructed by the sorted merging of the separate complete index lists for the term computed at different nodes. Thus, TINs are comprised of nodes which collectively store different horizontal partitions of this global index list. In practice, we expect TINs to be employed only for the most popular terms (a few hundred to a few thousand) whose accesses are expected to form scalability and performance bottlenecks. We will exploit the underlying network G[′]s architecture and related algo rithms (e.g., for routing/lookup) to efficiently and scalably create and maintain TINs and for retrieving TIN data items, from any node of G. In general, TINs may form separate overlay networks, coexisting with the global overlay G[1]. The MINERVA algorithms are heavily influenced by the way the well_∞_ known, efficient top-k query processing algorithms (e.g., [6]) operate, looking for docIDs within certain ranges of score values. Thus, the networks’ lookup(s) function, will be used using scores s as input, to locate the nodes storing docIDs with scores s. A key point to stress here, however, is that top-k queries Q({t1, ..., tr}, k) can originate from any peer node p of G, which in general is not a member of any I(ti), i = 1, ..., r and thus p does not have, nor can it easily acquire, the necessary routing state needed to forward the query to the TINs for the query terms. Our infrastructure, solves this by utilizing for each TIN a fairly small number (relative to the total number of data items for a term) of nodes of G 1 In practice, it may not always be necessary or advisable to form full-fledged separate overlays for TINs; instead, TINs will be formed as straightforward extensions of G: in this case, when a node n of G joins a TIN, only two additional links are added to the state of n linking it to its successor and predecessor nodes in the TIN. In this case, a TIN is simply a (circular) doubly-linked list. ----- MINERVA : A Scalable Efficient P2P Search Engine 65 _∞_ which will be readily identifiable and accessible from any node of G and can act as gateways between G and this TIN, being members of both networks. Finally, in order for any highly distributed solution to be efficient, it is cru cial to keep as low as possible the time and bandwidth overheads involved in the required communication between the various nodes. This is particularly challenging for solutions built over very large scale infrastructures. To achieve this, the algorithms of MINERVA follow the principles put forward by top-performing, _∞_ resource-efficient top-k query processing algorithms in traditional environments. Specifically, the principles behind favoring sequential index-list accesses to random accesses (in order to avoid high-cost random disk IOs) have been adapted in our distributed algorithms to ensure that: (i) sequential accesses of the items in the global, virtual index list dominate, (ii) they require either no communication, or at most an one-hop communication between nodes, and (iii) random accesses require at most O(log _N_ ) messages. _|_ _|_ To ensure the at-most-one-hop communication requirement for successive se quential accesses of TIN data, the MINERVA algorithms utilize an order pre_∞_ _serving hash function, first proposed for supporting range queries in DHT-based_ data networks in [20]. An order preserving hash function hop() has the property that for any two values v1, v2, if v1 > v2 then hop(v1) > hop(v2). This guarantees that data items corresponding to successive score values of a term t are placed either at the same or at neighboring nodes of I(t). Alternatively, similar functionality can be provided by employing for each I(t) an overlay based on skip graphs or skip nets [1,9]. Since both order preserving hashing and skip graphs incur the danger for load imbalances when assigning data items to TIN nodes, given the expected data skew of scores, load balancing solutions are needed. The design outlined so far, leverages DHT technology to facilitate efficiency and scalability in key aspects of the system’s operation. Specifically, posting (and deleting) data items for a term from any node can be done in O(log _N_ ) time, _|_ _|_ in terms of the number of messages. Similarly, during top-k query processing, the TINs of the terms in the query can be also reached in O(log _N_ ) messages. _|_ _|_ Furthermore, no single node is over-burdened with tasks which can either require more resources than available, or exhaust its resources, or even stress its resources for longer periods of time. In addition, as the top-k algorithm is processing different data items for each queried term, this involves gradually different nodes from each TIN, producing a highly distributed, scalable solution. ## 5 Term Index Networks In this section we describe and analyze the algorithms for creating TINs and populating them with data and nodes. **5.1** **Beacons for Bootstrapping TINs** The creation of a TIN has these basic elements: posting data items, inserting nodes, and maintaining the connectivity of nodes to ensure the efficiency/scalability properties promised by the TIN overlay. ----- 66 S. Michel, P. Triantafillou, and G. Weikum As mentioned, a key issue to note is that any node p in G may need to post (t, d, s) items for a term t. Since, in general, p is not a member of I(t) and does not necessarily know members of I(t), efficiently and scalably posting items to _I(t) from any p becomes non-trivial. To overcome this, a bootstrapping process_ for I(t) is employed which initializes an TIN I(t) for term t. The basic novelty lies in the special role to be played by nodes coined beacons, which in essence become gateways, allowing the flow of data and requests between the G and I(t) networks. In the bootstrap algorithm, a predefined number of “dummy” items of the form (t, ⋆, si) is generated in sequence for a set of predefined score values si, _i = 1, ..., u. Each such item will be associated with a node n in G, where it_ will be stored. Finally, this node n of G will also be made a member of I(t) by randomly choosing a previously inserted beacon node (i.e., for the one associated with an already inserted score value sj, 1 ≤ _j ≤_ _i −_ 1) as a gateway. The following algorithm details the pseudocode for bootstrapping I(t). It utilizes an order-preserving hash function hop() : T × (0, 1] → [m], where m is the size of the identifiers in bits and [m] denotes the name space used for the overlay (e.g., all 2[160] ids, for 160-bit identifiers). In addition, a standard hash function h() : (0, 1] [m], (e.g. SHA-1) is used. The particulars of the order _→_ preserving hash function to be employed will be detailed after the presentation of the query processing algorithms which they affect. The bootstrap algorithm selects u “dummy” score values, i/u, i = 1, ..., u, finds for each such score value the node n in G where it should be placed (using hop()), stores this score there and inserts n into the I(t) network as well. At first, the I(t) network contains only the node with the dummy item with score zero. At each iteration, another node of n is added to I(t) using as gateway the node of G which was added in the previous iteration to I(t). For simplicity of presentation, the latter node **Algorithm 1. Bootstrap I(t)** 1: input: u: the number of “dummy” items (t, ⋆, si), i = 1, ..., u 2: input: t: the term for which the TIN is created 3: p = 1/u 4: for i = 1 to u do 5: _s = i × p_ 6: _lookup(n.s) = hop(t, s) { n.s in G will become the next beacon node of I(t) }_ 7: **if s = p then** 8: _N_ (t) = {n.s} 9: _E(t) = ∅{Initialized I(t) with n.s with the first dummy item}_ 10: **end if** 11: **if s ̸= p then** 12: _n1 = hop(t, s −_ _p) {insert n(s) into I(t) using node n(s −_ _p) as gateway}_ 13: call join(I(t), n1, s) 14: **end if** 15: store (t, ⋆, s) at I(t).n(s) 16: end for ----- MINERVA : A Scalable Efficient P2P Search Engine 67 _∞_ can be found by simply hashing for the previous dummy value. A better choice for distributing the load among the beacons is to select at random one of the previously-inserted beacons and use it as a gateway. Obviously, a single beacon per TIN suffices. The number u of beacon scores is intended to introduce a number of gateways between G and I(t) so to avoid potential bottlenecks during TIN creation. u will typically be a fairly small number so the total beacon-related overhead involved in the TIN creation will be kept small. Further, we emphasize that beacons are utilized by the algorithm posting items to TINs. Post operations will in general be very rare compared to query operations and query processing does not involve the use of beacons. Finally, note that the algorithm uses a join() routine that adds a node n(s) storing score s into I(t) using a node n1 known to be in I(t) and thus, has the required state. The new node n(s) must occupy a position in I(t) specified by the value of hop(t, s). Note that this is ensured by using h(nodeID), as is typically done in DHTs, since these node IDs were selected from the order-preserving hash function. Besides the side-effect of ensuring the order-preserving position for the nodes added to a TIN, the join routine is otherwise straightforward: if the TIN is a full-fledged DHT overlay, join() is updating the predecessor/successor pointers, the O(log _N_ ) routing state of the new node, and the routing state of _|_ _|_ each I(t) node pointing to it, as dictated by the relevant DHT algorithm. If the TIN is simply a doubly-linked list, then only predecessor/successor pointers are the new node and its neighbors are adjusted. **5.2** **Posting Data to TINs** The posting of data items is now made possible using the bootstrapped TINs. Any node n1 of G wishing to post an item (t, d, s) first locates an appropriate node of G, n2 that will store this item. Subsequently, it inserts node n2 into I(t). To do this, it randomly selects a beacon score and associated beacon node, from all available beacons. This is straightforward given the predefined beacon score values and the hashing functions used. The chosen beacon node has been made a member of I(t) during bootstrapping. Thus, it can “escort” n2 into I(t). The following provides the pseudocode for the posting algorithm. By design, the post algorithm results in a data placement which introduces two characteristics, that will be crucial in ensuring efficient query processing. First, (as the bootstrap algorithm does) the post algorithm utilizes the order-preserving hash function. As a result, any two data items with consecutive score values for the same term will be placed by definition in nodes of G which will become one-hop neighbors in the TIN for the term, using the join() function explained earlier. Note, that within each TIN, there are no ‘holes’. A node n becomes a member of a TIN network if and only if a data item was posted, with the score value for this item hashing to n. It is instructing here to emphasize that if TINs were not formed and instead only the global network was present, in general, any two successive score values could be falling in nodes which in G could be many hops apart. With TINs, following successor (or predecessor) links always leads to ----- 68 S. Michel, P. Triantafillou, and G. Weikum **Algorithm 2. Posting Data to I(t)** 1: input: t, d, s: the item to be inserted by a node n1 2: n(s) = hop(t, s) 3: n1 sends (t, d, s) to n(s) 4: if n(s) /∈ _N_ (t) then 5: _n(s) selects randomly a beacon score sb_ 6: _lookup(nb) = hop(t, sb) { nb is the beacon node storing beacon score sb }_ 7: _n(s) calls join(I(t), nb, s)_ 8: end if 9: store ((t, d, s) nodes where the next (or previous) segment of scores have been placed. This feature in essence ensures the at-most-one-hop communication requirement when accessing items with successive scores in the global virtual index list for a term. Second, the nodes of any I(t) become responsible for storing specific segments (horizontal partitions) of the global virtual index list for t. In particular, an I(t) node stores all items for t for a specific (range of) score value, posted by any node of the underlying network G. **5.3** **Complexity Analysis** The bootstrapping I(t) algorithm is responsible for inserting u beacon items. For each beacon item score, the node n.s is located by applying the hop() function and routing the request to that node (step 5). This will be done using G’s lookup algorithm in O(log _N_ ) messages. The next key step is to locate the previously _|_ _|_ inserted beacon node (step 11) (or any beacon node at random) and sending it the request to join the TIN. Step 11 again involves O(log _N_ ) messages. The _|_ _|_ actual join() routine will cost O(log[2] _N_ (t) ) messages, which is the standard _|_ _|_ join() message complexity for any DHT of size N (t). Therefore, the total cost is _O(u_ (log _N_ + log[2] _N_ (t) ) messages. _×_ _|_ _|_ _|_ _|_ The analysis for the posting algorithm is very similar. For each post(t, d, s) operation, the node n where this data item should be stored is located and the request is routed to it, costing O(log _N_ ) messages (step 3). Then a random _|_ _|_ beacon node is located, costing O(log _N_ ) messages, and then the join() routine is _|_ _|_ called from this node, costing O(log[2] _N_ (t) ) messages. Thus, each post operation _|_ _|_ has a complexity of O(log _N_ ) + O(log[2] _N_ (t) ) messages. _|_ _|_ _|_ _|_ Note that both of the above analysis assumed that each I(t) is a full-blown DHT overlay. This permits a node to randomly select any beacon node to use to join the TIN. Alternatively, if each I(t) is simply a (circular) doubly-linked list, then a node can join a TIN using the beacon storing the beacon value that is immediately preceding the posted score value. This requires O(log _N_ ) hops _|_ _|_ to locate this beacon node. However, since in this case the routing state for each node of a TIN consists of only the two (predecessor and successor) links, the cost to join is in the worst case O( _N_ (t) ), since after locating the beacon node with _|_ _|_ the previous beacon value, O( _N_ (t) ) successor pointers may need to be followed _|_ _|_ in order to place the node in its proper order-preserving position. Thus, when ----- MINERVA : A Scalable Efficient P2P Search Engine 69 _∞_ TINs are simple doubly-linked lists, the complexity of both the bootstrap and post algorithms are O(log _N_ + _N_ (t) ) messages. _|_ _|_ _|_ _|_ ## 6 Load Balancing **6.1** **Order-Preserving Hashing** The order preserving hash function to be employed is important for several reasons. First, for simplicity, the function can be based on a simple linear transform. Consider hashing a value f (s) : (0, 1] _I, where f_ (s) transforms a score s into _→_ an integer; for instance, f (s) = 10[6] _× s. Function hop() can be defined then as_ _f_ (s) − _f_ (smin) _hop(s) = (_ _f_ (smax) − _f_ (smin) _[×][ 2][m][)][ mod][ 2][m]_ (1) Although such a function is clearly order-preserving, it has the drawback that it produces the same output for items of equal scores of different terms. This leads to the same node storing for all terms all items having the same score. This is undesirable since it cannot utilize all available resources (i.e., utilize different sets of nodes to store items for different terms). To avoid this, hop() is refined to take as input the term name, which provides the necessary functionality, as follows. _f_ (s) − _f_ (smin) _hop(t, s) = (h(t) +_ _f_ (smax) − _f_ (smin) _[×][ 2][m][)][ mod][ 2][m]_ (2) The term h(t) adds a different random offset for different terms, initiating the search for positions of term score values at different, random, offsets within the namespace. Thus, by using the h(t) term in hop(t, s) the result is that any data items having equal scores but for different terms are expected to be stored at different nodes of G. Another benefit stems from ameliorating the storage load imbalances that result from the non-uniform distribution of score values. Assuming a uniform placement of nodes in G, the expected non-uniform distribution of scores will result in a non-uniform assignment of scores to nodes. Thus, when viewed from the perspective of a single term t, the nodes of I(t) will exhibit possibly severe storage load imbalances. However, assuming the existence of large numbers of terms (e.g., a few thousand), and thus data items being posted for all these terms over the same set of nodes in G, given the randomly selected starting offsets for the placement of items, it is expected that the severe load imbalances will disappear. Intuitively, overburdened nodes for the items of one term are expected to be less burdened for the items of other terms and vice versa. But even with the above hash function, very skewed score distributions will lead to storage load imbalances. Expecting that exponential-like distributions of score values will appear frequently, we developed a hash function that is order-preserving and handles load imbalances by assigning score segments of exponentially decreasing sizes to an exponentially increasing number of nodes. For instance, the sparse top 1/2 of the scores distribution is to be assigned to a single node, the next 1/4 of scores is to be assigned to 2 nodes, the next 1/8 of scores to 4 nodes, etc. The details of this are omitted for space reasons. ----- 70 S. Michel, P. Triantafillou, and G. Weikum **6.2** **TIN Data Migration** Exploiting the key characteristics of our data, MINERVA can ensure further _∞_ load balancing with small overheads. Specifically, index lists data entries are small in size and are very rarely posted and/or updated. In this subsection we outline our approach for improved load balancing. We require that each peer posting index list entries, first computes a (equi width) histogram of its data with respect to its score distribution. Assuming a targeted _N_ (t) number of nodes for the TIN of term t, it can create _N_ (t) equal_|_ _|_ _|_ _|_ size partitions, with lowscorei, highscorei denoting the score ranges associated with partition i, i = 1, ..., _N_ (t) . Then it can simply utilize the posting algorithm _|_ _|_ shown earlier, posting using the lowscorei scores for each partition. The only exception to the previously shown post algorithm is that the posting peer now posts at each iteration a complete partition of its index list, instead of just a single entry. The above obviously can guarantee perfect load balancing. However, subse quent postings (typically by other peers) may create imbalances, since different index lists may have different score distributions. Additionally, when ensuring overall load balancing over multiple index lists being posting by several peers, the order-preserving property of the placement must be guaranteed. Our approach for solving these problems is as follows. First, again the posting peer is required to compute a histogram of its index list. Second, the histogram of the TIN data (that is, the entries already posted) is stored at easily identifiable nodes. Third, the posting peer is required to retrieve this histogram and ‘merge’ it with his own. Fourth, the same peer identifies how the total data must now be split into _N_ (t), equal-size partitions of consecutive scores. Fi_|_ _|_ nally, it identifies all data movements (from TIN peer to TIN peer) necessary to redistribute the total TIN data so that load balancing and order preservation is ensured. Detailed presentation of the possible algorithms for this last step and their respective comparison is beyond the scope of this paper. We simply mention that total TIN data sizes is expected to be very small (in actual number of bytes stored and moved). For example, even with several dozens of peers posting different, even large, multi-million-entry, index lists, in total the complete TIN data size will be a few hundred MBs, creating a total data transfer movement equivalent to that of downloading a few dozen MP3 files. Further, index lists’ data posting to TINs is expected to be a very infrequent operation (compared to search queries). As a result, ensuring load balancing across TIN nodes proves to be relative inexpensive. **6.3** **Discussion** The approaches to index lists’ data posting outlined in the previous two sections can be used competitively or even be combined. When posting index lists with exponential score distributions, by design the posting of data using the orderpreserving hash function of Section 5.1, will be adequately load balanced and nothing else is required. Conversely, when histogram information is available and ----- MINERVA : A Scalable Efficient P2P Search Engine 71 _∞_ can be computed by posting peers, the TIN data migration approach will yield load balanced data placement. A more subtle issue is that posting with the order-preserving hash function also facilitates random accesses of the TIN data, based on random score values. That is, by hashing for any score, we can find the TIN node holding the entries with this score. This becomes essential if the web search engine is to employ top-k query algorithms which are based on random accesses of scores. In this work, our top-k algorithms avoid random accesses, by design. However, the above point should be kept in mind since there are recently-proposed distributed top-k algorithms, relying on random accesses and more efficient algorithms may be proposed in the future. ## 7 Top-k Query Processing The algorithms in this section focus on how to exploit the infrastructure presented previously in order to efficiently process top-k queries. The main efficiency metrics are query response times and network bandwidth requirements. **7.1** **The Basic Algorithm** Consider a top-k query of the form Q({t1, ..., tr}, k) involving r terms that is generated at some node ninit of G. Query processing is based on the following ideas. It proceeds in phases, with each phase involving ‘vertical’ and ‘horizontal’ communication between the nodes within TINs and across TINs, respectively. The vertical communications between the nodes of a TIN are occuring in parallel across all r TINs named in the query, gathering a threshold number of data items from each term. There is a moving coordinator node, that will be gathering the data items from all r TINs that enable it to compute estimates of the top-k result. Intermediate estimates of the top-k list will be passed around, as the coordinator role moves from node to node in the next phase where the gathering of more data items and the computation of the next top-k result estimate will be computed. The presentation shows separately the behavior of the query initiator, the (moving) query coordinator, and the TIN nodes. **Query Initiator** The initiator calculates the set of start nodes, one for each term, where the query processing will start within each TIN. Also, it randomly selects one of the nodes (for one of the TINs) to be the initial coordinator. Finally, it passes on the query and the coordinator ID to each of the start nodes, to initiate the parallel vertical processing within TINs. The following pseudocode details the behavior of the initiator. **Processing Within Each TIN** Processing within a TIN is always initiated by the start node. There is one start node per communication phase of the query processing. In the first phase, the ----- 72 S. Michel, P. Triantafillou, and G. Weikum **Algorithm 3. Top-k QP: Query Initiation at node G.ninit** 1: input: Given query Q = {t1,.., tr}, k : 2: for i = 1 to r do 3: _startNodei = I(ti).n(smax) = hop(ti, smax)_ 4: end for 5: Randomly select c from [1, ..., r] 6: coordID = I(tc).n(smax) 7: for i = 1 to r do 8: send to startNodei the data (Q, coordID) 9: end for start node is the top node in the TIN which receives the query processing request from the initiator. The start node then starts the gathering of data items for the term by contacting enough nodes, following successor links, until a threshold number γ (that is, a batch size) of items has been accumulated and sent to the coordinator, along with an indication of the maximum score for this term which has not been collected yet, which is actually either a locally stored score or the maximum score of the next successor node. The latter information is critical for the coordinator in order to intelligently decide when the top-k result list has been computed and terminate the search. In addition, each start node sends to the coordinator the ID of the node of this TIN to be the next start node, which is simply the next successor node of the last accessed node of the TIN. Processing within this TIN will be continued at the new start node when it receives the next message from the coordinator starting the next data-gathering phase. Algorithm 4 presents the pseudocode for TIN processing. **Algorithm 4. Top-k QP: Processing by a start node within a TIN** 1: input: A message either from the initiator or the coordinator 2: tCollectioni = ∅ 3: n = startNodei 4: while |tCollectioni| < γ do 5: **while |tCollectioni| < γ AND more items exist locally do** 6: define the set of local items L = {(ti, d, s) in n} 7: send to coordID : L 8: _|tCollectioni| = |tCollectioni| + |L|_ 9: **end while** 10: n = succ(n) 11: end while 12: boundi = max score stored at node n 13: send to coordID : n and boundi Recall, that because of the manner with which items and nodes have been placed in a TIN, following succ() links, items are collected starting from the item with the highest score posted for this term and proceeding in sorted descending order based on scores. ----- MINERVA : A Scalable Efficient P2P Search Engine 73 _∞_ **Moving Query Coordinator** Initially, the coordinator is randomly chosen by the initiator to be one of the original start nodes. First, the coordinator uses the received collections and runs a version of the NRA top-k processing algorithm, locally producing an estimate of the top-k result. As is also the case with classical top-k algorithms, the exact result is not available at this stage since only a portion of the required information is available. Specifically, some documents with high enough TotalScore to qualify for the top-k result are still missing. Additionally, some documents may also be seen in only a subset of the collections received from the TINs so far, and thus some of their scores are missing, yielding only a partially known TotalScore. A key to the efficiency of the overall query processing process is the ability to prune the search and terminate the algorithm even in the presence of missing documents and missing scores. To do this, the coordinator first computes an estimate of the top-k result, which includes only documents whose TotalScores are completely known, defining the RankKscore value (i.e. the smallest score in the top-k list estimate). Then, it utilizes the boundi values received from each start node. When a score for a document d is missing for term i, it can be replaced with boundi to estimate the T otalScore(d). This is done for all such _d with missing scores. If RankKscore > T otalScore(d) for all d with missing_ scores then there is no need to continue the process for finding the missing scores, since the associated documents could never belong to the top-k result. Similarly, if RankKscore > [�]i=1,...,r _[bound][i][, then similarly there is no need to try to find]_ any other documents, since they could never belong to the top-k result. When both of these conditions hold, the coordinator terminates the query processing and returns the top-k result to the initiator. If the processing must continue, the coordinator starts the next phase, send ing a message to the new start node for each term, whose ID was received in the message containing the previous data collections. In this message the coordinator also indicates the ID of the node which becomes the coordinator in this next phase. The next coordinator is defined to be the node in the same TIN as the previous coordinator whose data is to be collected next in the vertical processing in this TIN (i.e., the next start node at the coordinator’s TIN). Alternatively, any other start node can be randomly chosen as the coordinator. Algorithm 5 details the behavior of the coordinator. **7.2** **Complexity Analysis** The overall complexity has three main components: the cost incurred for (i) the communication between the query initiator and the start nodes of the TINs, (ii) the vertical communication within a TIN, and (iii) the horizontal communication between the current coordinator and the current set of start nodes. The query initiator needs to lookup the identity of the initial start nodes for each one of the r query terms and route to them the query and the chosen coordinator ID. Using the G network, this incurs a communication complexity of _O(r_ _log_ _N_ ) messages. Denoting with depth the average (or maximum) number _×_ _|_ _|_ ----- 74 S. Michel, P. Triantafillou, and G. Weikum **Algorithm 5. Top-k QP: Coordination** 1: input: For each i: tCollectioni and newstartNodei and boundi 2: tCollection = [�]i _[tCollection][i]_ 3: compute a (new) top-k list estimate using tCollection, and RankKscore 4: candidates = {d|d /∈top-k list} 5: for all d ∈ _candidates do_ 6: _worstScore(d) is the partial TotalScore of d_ 7: _bestScore(d) := worstScore(d) +_ [�]j∈MT _[bound][j][ {][Where][ MT][ is the set of term]_ ids with missing scores _}_ 8: **if bestScore(d) < RankKscore then** 9: remove d from candidates 10: **end if** 11: end for 12: if candidates is empty then 13: exit() 14: end if 15: if candidates is not empty then 16: _coordIDnew = pred(n)_ 17: calculate new size threshold γ 18: **for i = 1 to r do** 19: send to startNodei the data (coordIDnew, γ) 20: **end for** 21: end if of nodes accessed during the vertical processing of TINs, overall O(r _depth)_ _×_ messages are incurred due to TIN processing, since subsequent accesses within a TIN require, by design, one-hop communication. Each horizontal communication in each phase of query processing between the coordinator and the r start nodes requires O(r _log_ _N_ ) messages. Since such horizontal communication takes _×_ _|_ _|_ place at every phase, this yields a total of O(phases _r_ _log_ _N_ ) messages. _×_ _×_ _|_ _|_ Hence, the total communication cost complexity is _cost = O(phases_ _r_ _log_ _N_ + r _log_ _N_ + r _depth)_ (3) _×_ _×_ _|_ _|_ _×_ _|_ _|_ _×_ This total cost is the worst case cost; we expect that the cost incurred in most cases will be much smaller, since horizontal communication across TINs can be much more efficient than O(log _N_ ), as follows. The query initiator can _|_ _|_ first resolve the ID of the coordinator (by hashing and routing over G) and then determine its actual physical address (i.e., its IP address), which is then forwarded to each start node. In turn, each start node can forward this from successor to successor in its TIN. In this way, at any phase of query processing, the last node of a TIN visited during the vertical processing, can send the data collection to the coordinator using the coordinator’s physical address. The current coordinator also knows the physical address of the next coordinator (since this was the last node visited in its own TIN from which it received a message with the data collection for its term) and of the next start node for all terms (since these are the last nodes visited during vertical processing of the TINs, ----- MINERVA : A Scalable Efficient P2P Search Engine 75 _∞_ from which it received a message). Thus, when sending the message to the next start nodes to continue vertical processing, the physical addresses can be used. The end result of this is that all horizontal communication requires one message, instead of O(log _N_ ) messages. Hence, the total communication cost complexity _|_ _|_ now becomes _cost = O(phases_ _r + r_ _log_ _N_ + r _depth)_ (4) _×_ _×_ _|_ _|_ _×_ As nodes are expected to be joining and leaving the underlying overlay network _G, occasionally, the physical addresses used to derive the cost of (4) will not be_ valid. In this case, the reported errors will lead to nodes using the high-level IDs instead of the physical addresses, in which case the cost is that given by (3). ## 8 Expediting Top-k Query Processing In this section we develop optimizations that can further speedup the performance of top-k query processing. These optimizations are centered on: (i) the ‘vertical’ replication of term-specific data among the nodes of a TIN, and (ii) the ‘horizontal’ replication of data across TINs. **8.1** **TIN Data Replication** There are two key characteristics of the data items in our model, which permit their large-scale replication. First, data items are rarely posted and even more rarely updated. Second, data items are very small in size (e.g. < 50 bytes each). Hence, replication protocols will not cost significantly either in terms of replica state maintenance, or in terms of storing the replicas. **Vertical Data Replication. The issue to be addressed here is how to appro-** priately replicate term data within TIN peers so to gain in efficiency. The basic structure of the query processing algorithm presented earlier facilitates the easy incorporation of a replication protocol into it. Recall, that in each TIN I(t), query processing proceeds in phases, and in each phase a TIN node (the current start node) is responsible for visiting a number of other TIN nodes, a successor at a time, so that enough, (i.e., a batch size of) data items for t are collected. The last visited node in each phase which collects all data items, can initiate a ‘reverse’ vertical communication, in parallel to sending the collection to the coordinator. With this reverse vertical communication thread, each node in the reverse path sends to its predecessor only the data items its has not seen. In the end, all nodes in the path from the start node to the last node visited will eventually receive a copy of all items collected during this phase, storing locally the pair (lowestscore, highestscore), marking its lowest and highest locally stored scores. Since this is straightforward, the pseudocode is omitted for space reasons. Since a new posting involves all (or most) of the nodes in these paths, each node knows when to initiate a new replication to account for the new items. ----- 76 S. Michel, P. Triantafillou, and G. Weikum **Exploiting Replicas. The start node selected by the query initiator no longer** needs to perform a successor-at-a-time traversal of TIN in the first phase, since the needed data (replicas) are stored locally. However, vertical communication was also useful for producing the ID of the next start node for this TIN. A subtle point to note here is that the coordinator can itself determine the new start node for the next phase, even without receiving explicitly this ID at the end of vertical communication. This can simply be done using the minimum score value (boundi) it has received for term ti; the ID of the next start node is found hashing for score prev(boundi). Additionally, the query initiator can select as start nodes the nodes responsi ble for storing a random (expected to be high score) and not always the maximum score, as it does up to now. Similarly, the coordinator when selecting the ID of the next start node for the next batch retrieval for a term, it can choose to hash for a score value that is lower than the score prev(boundi). Thus, random start nodes within a TIN are selected at different phases and these gather the next batch of data from the proper TIN nodes, using the TIN DHT infrastructure for efficiency. The details of how this is done, are omitted for space reasons. **Horizontal Data Replication. TIN data may also be replicated horizontally.** The simplest strategy is to create replicated TINs for popular terms. This involves the posting of data into all TIN replicas. The same algorithms can be used as before for posting, except now when hashing, instead of using the term t as input to the hash function, each replica of t must be specified (e.g., t.v, where _v stands for a version/replica number). Again, the same algorithms can be used_ for processing queries, with the exception that each query can now select one of the replicas of I(t), at random. Overall, TIN data replication leads to savings in the number of messages and response time speedups. Furthermore, several nodes are off-loaded since they no longer have to partake in the query processing process. With replication, therefore, the same number of nodes overall will be involved in processing a number of user queries, except that each query will be employing a smaller set of peers, yielding response time and bandwidth benefits. In essence, TIN data replication increases the efficiency of the engine, without adversely affecting its scalability. Finally, it should be stressed that such replication will also improve the availability of data items and thus replication is imperative. Indirectly, for the same reason the quality of the results with replication will be higher, since lost items inevitably lead to errors in the top-k result. ## 9 Experimentation **9.1** **Experimental Testbed** Our implementation was written in Java. Experiments were performed on 3GHz Pentium PCs. Since deploying full-blown, large networks is not an option, we opted for simulating large numbers of nodes as separate processes on the same PC, executing the real MINERVA code. A 10,000 node network was simulated. _∞_ ----- MINERVA : A Scalable Efficient P2P Search Engine 77 _∞_ A real-world data collection was used in our experiments: GOV. The GOV collection consists of the data of the TREC-12 Web Track and contains roughly 1.25 million (mostly HTML and PDF) documents obtained from a crawl of the .gov Internet domain (with total index list size of 8 GB). The original 50 queries from the Web Track’s distillation task were used. These are term queries, with each query containing up to 4 terms. The index lists contained the original document scores computed as tf * log idf. tf and idf were normalized by the maximum tf value of each document and the maximum idf value in the corpus, respectively. In addition, we employed an extended GOV (XGOV) setup, with a larger number of query terms and associated index lists. The original 50 queries were expanded by adding new terms from synonyms and glosses taken from the WordNet thesaurus (http://www.cogsci.princeton.edu/ wn). The expansion _∼_ yielded queries with, on average, twice as many terms, up to 18 terms. **9.2** **Performance Tests and Metrics** **Efficiency Experiments. The data (index list entries) for the terms to be** queried were first posted. Then, the GOV/XGOV benchmark queries were executed in sequence. For simplicity, the query initiator node assumed the role of a fixed coordinator. The experiments used the following metrics: _Bandwidth. This shows the number of bytes transferred between all the nodes_ involved in processing the benchmarks’ queries. The benchmarks’ queries were grouped based on the number of terms they involved. In essence, this grouping created a number of smaller sub-benchmarks. _Query Response Time. This represents the elapsed, “wall-clock” time for_ running the benchmark queries. We report on the wall-clock times per subbenchmark and for the whole GOV and XGOV benchmarks. _Hops. This reports the number of messages sent over our network infras-_ tructures to process all queries. For communication over the global DHT G, the number of hops was set to be log _N_ (i.e., when the query initiator contacts the _|_ _|_ first set of start nodes for each TIN). Communication between peers within a TIN requires, by design, one hop at a time. To avoid the overestimation of response times due to the competition be tween all processes for the PC’s disk and network resources, and in order to produce reproducible and comparable results for tests ran at different times, we opted for simulating disk IO latency and network latency. Specifically, each random disk IO was modeled to incur a disk seek and rotational latency of 9 ms, plus a transfer delay dictated by a transfer rate of 8MB/s. For network latency we utilized typical round trip times (RTTs) of packets and transfer rates achieved for larger data transfers between widely distributed entities [16]. We assumed a RTT of 100 ms. When peers simply forward the query to a next peer, this is assumed to take roughly 1/3 of the RTT (since no ACKs are expected). When peers sent more data, the additional latency was dictated by a “large” data transfer rate of 800Kb/s, which includes the sender’s uplink bandwidth, the ----- 78 S. Michel, P. Triantafillou, and G. Weikum receivers downlink bandwidth, and the average internet bandwidth typically witnessed.[2] **Scalability Experiments. The tested scenarios varied the query load to the** system, measuring the overall time required to complete the processing of all queries in a queue of requests. Our experiments used a queue of identical queries involving four terms, with varying index lists characteristics. Two of these terms had small index lists (with over 22,000 and over 42,000 entries) and the other two lists had sizes of over 420,000 entries. For each query the (different) query initiating peer played the role of the coordinator. The key here is to measure contention for resources and its limits on the pos sible parallelization of query processing. Each TIN peer uses his disk, his uplink bandwidth to forward the query to his TIN successor, and to send data to the coordinator. Uplink/downlink bandwidths were set to 256Kbps/1Mbps. Similarly, the query initiator utilizes its downlink bandwidth to receive the batches of data in each phase and its uplink bandwidth to send off the query to the next TIN start nodes. These delays define the possible parallelization of query execution. By involving the two terms with the largest index lists in the queries, we ensured the worst possible parallelization (for our input data), since they induced the largest batch size, requiring the most expensive disk reads and communication. **9.3** **Performance Results** Overall, each benchmark experiment required between 2 to 5 hours for its realtime execution, a big portion of which was used up by the posting procedure. Figures 1 and 2 show the bandwidth, response times, and hops results for the GOV and XGOV group-query benchmarks. Note, that different query groups have in general mutually-incomparable results, since they involve different index lists with different characteristics (such as size, score distributions etc). In XGOV the biggest overhead was introduced by the 8 7-term and 6 11-term queries. Table 1 shows the total benchmark execution times, network bandwidth consumption, as well as the number of hops for the GOV and XGOV benchmarks. Generally, for each query, the number of terms and the size of the corre sponding index list data are the key factors. The central insight here is that the choice of the NRA algorithm was the most important contributor to the overhead. The adaptation of more efficient distributed top-k algorithms within MINERVA (such as our own [12], which also disallow random accesses) can _∞_ reduce this overhead by one to two orders of magnitude. This is due to the fact that the top-k result can be produced without needing to delve deeply into the index lists’ data, resulting in drastically fewer messages, bandwidth, and time requirements. 2 This figure is the average throughput value measured (using one stream – one cpu machines) in experiments conducted for measuring wide area network throughput (sending 20MB files between SLAC nodes (Stanford’s Linear Accelerator Centre) and nodes in Lyon France [16] using NLANR’s iPerf tool [19]. ----- MINERVA : A Scalable Efficient P2P Search Engine 79 _∞_ **GOV** **1200** **GOV** **12000** **GOV** **1000** **10000** **800** **8000** **600** **6000** **400** **4000** **200** **2000** **3** **4** **0** **0** **Number of Query Terms** **2** **Number of Query Terms3** **4** **2** **Number of Query Terms3** **4** **Fig. 1. GOV Results: Bandwidth, Execution Time, and Hops** **120000** **100000** **80000** **60000** **40000** **20000** **0** **1800** **1600** **1400** **1200** **1000** **800** **600** **400** **200** **0** **25000** **20000** **15000** **10000** **5000** **0** |XGOV|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10| |---|---|---|---|---|---|---|---|---|---| ||||||||||| ||||||||||| ||||||||||| ||||||||||| ||||||||||| |XGOV|Col2|Col3|Col4|Col5|Col6|Col7|Col8| |---|---|---|---|---|---|---|---| ||||||||| ||||||||| ||||||||| ||||||||| ||||||||| |XGOV|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10| |---|---|---|---|---|---|---|---|---|---| ||||||||||| ||||||||||| ||||||||||| ||||||||||| **4** **5** **6** **7** **8** **9** **10** **11** **12** **13** **14** **15** **18** **Number of Query Terms** **4** **5** **6** **7** **8** **9** **10** **11** **12** **13** **14** **15** **18** **Number of Query Terms** **4** **5** **6** **7** **8** **9** **10 11 12 13 14 15 18** **Number of Query Terms** **Fig. 2. XGOV Results: Bandwidth, Execution Time, and Hops** **Table 1. Total GOV and XGOV Results** The 2-term queries introduced the biggest overheads. There are 29 2-term, 7 3-term, and 4 4-term queries in GOV. Figure 3 shows the scalability experiment results. Query loads tested rep resent queue sizes of 10, 100, 1000, and 10000 identical queries simultaneously arriving into the system. This figure also shows what the corresponding time would be if the parallelization contributed by the MINERVA architecture was _∞_ not possible; this would be the case, for example, in all related-work P2P search architectures and also distributed top-k algorithms, where the complete index lists at least for one query term are stored completely at one peer. The scalability results show the high scalability achievable with MINERVA . It is due _∞_ to the “pipelining” that is introduced within each TIN during query processing, where a query consumes small amounts of resources from each peer, pulling together the resources of all (or most) peers in the TIN for its processing. For comparison we also show the total execution time in an environment in which each complete index list was stored in a peer. This is the case for most related work on P2P search engines and on distributed top-k query algorithms. In this case, the resources of the single peer storing a complete index list are required ----- 80 S. Michel, P. Triantafillou, and G. Weikum **10000000** **1000000** **100000** **10000** **1000** **100** |Minerva|Col2|Col3| |---|---|---| ||Minerva|| ||Infinity no parallel processing|| |||| |||| |||| **1** **10** **100** **1000** **10000** **Query Load: Queue Size** **Fig. 3. Scalability Results** for the processing of all communication phases and for all queries in the queue. In essence, this yields a total execution time that is equal to that of a sequential execution of all queries using the resources of the single peers storing the index lists for the query terms. Using this as a base comparison, MINERVA is _∞_ shown to enjoy approximately two orders of magnitude higher scalability. Since in our experiments there are approximately 100 nodes per TIN, this defines the maximum scalability gain. ## 10 Concluding Remarks We have presented MINERVA, a novel architecture for a peer-to-peer web _∞_ search engine. The key distinguishing feature of MINERVA is its high-levels _∞_ of distribution for both data and processing. The architecture consists of a suite of novel algorithms, which can be classified into algorithms for creating Term Index Networks, TINs, placing index list data on TINs and of top-k algorithms. TIN creation is achieved using a bootstrapping algorithm and also depends on how nodes are selected when index lists data is posted. The data posting algorithm employs an order-preserving hash function and, for higher levels of load balancing, MINERVA engages data migration algorithms. Query processing _∞_ consists of a framework for highly distributed versions of top-k algorithms, ranging from simple distributed top-k algorithms, to those utilizing vertical and/or horizontal data replication. Collectively, these algorithms ensure efficiency and scalability. Efficiency is ensured through the fast sequential accesses to index lists’ data, which requires at most one hop communication and by algorithms exploiting data replicas. Scalability is ensured by engaging a larger number of TIN peers in every query, with each peer being assigned much smaller subtasks, avoiding centralized points of control. We have implemented MINERVA _∞_ and conducted detailed performance studies showcasing its scalability and efficiency. Ongoing work includes the adaptation of recent distributed top-k algorithms (e.g., [12]) into the MINERVA architecture, which have proved one to two _∞_ orders of magnitude more efficient than the NRA top-k algorithm currently employed, in terms of query response times, network bandwidth, and peer loads. ----- MINERVA : A Scalable Efficient P2P Search Engine 81 _∞_ ## References 1. J. Aspnes and G. Shah. Skip graphs. In Fourteenth Annual ACM-SIAM Symposium _on Discrete Algorithms, pages 384–393, Jan. 2003._ 2. P. Cao and Z. Wang. Efficient top-k query calculation in distributed networks, PODC 2004. 3. S. Chakrabarti. _Mining the Web: Discovering Knowledge from Hypertext Data._ Morgan Kaufmann, San Francisco, 2002. 4. F. M. Cuenca-Acuna, C. Peery, R. P. Martin, and T. D. Nguyen. PlanetP: Using Gossiping to Build Content Addressable Peer-to-Peer Information Sharing Communities. Technical Report DCS-TR-487, Rutgers University, Sept. 2002. 5. R. Fagin. Combining fuzzy information from multiple systems. J. Comput. Syst. _Sci., 58(1):83–99, 1999._ 6. R. Fagin, A. Lotem, and M. Naor. Optimal aggregation algorithms for middleware. _J. Comput. Syst. Sci., 66(4), 2003._ 7. P. Ganesan, M. Bawa, and H. Garcia-Molina. Online balancing of range-partitioned data with applications to peer-to-peer systems. In VLDB, pages 444–455, 2004. 8. A. Gupta, O. D. Sahin, D. Agrawal, and A. E. Abbadi. Meghdoot: content-based publish/subscribe over p2p networks. In Proceedings of the 5th _ACM/IFIP/USENIX international conference on Middleware, pages 254–273, New_ York, NY, USA, 2004. Springer-Verlag New York, Inc. 9. N. Harvey, M. Jones, S. Saroiu, M. Theimer, and A. Wolman. Skipnet: A scalable overlay network with practical locality properties. In USITS, 2003. 10. R. Huebsch, J. M. Hellerstein, N. Lanham, B. T. Loo, S. Shenker, and I. Stoica. Querying the internet with pier. In VLDB, pages 321–332, 2003. 11. J. Lu and J. Callan. Content-based retrieval in hybrid peer-to-peer networks. In _Proceedings of CIKM03, pages 199–206. ACM Press, 2003._ 12. S. Michel, P. Triantafillou, and G. Weikum. Klee: A framework for distributed top-k query algorithms. In VLDB Conference, 2005. 13. S. Ratnasamy, P. Francis, M. Handley, R. Karp, and S. Schenker. A scalable content-addressable network. In Proceedings of ACM SIGCOMM 2001, pages 161– 172. ACM Press, 2001. 14. P. Reynolds and A. Vahdat. Efficient peer-to-peer keyword searching. In Proceed _ings of International Middleware Conference, pages 21–40, June 2003._ 15. A. Rowstron and P. Druschel. Pastry: Scalable, decentralized object location, and routing for large-scale peer-to-peer systems. In IFIP/ACM International Confer_ence on Distributed Systems Platforms (Middleware), pages 329–350, 2001._ 16. D. Salomoni and S. Luitz. High performance throughput tuning/measurement. http://www.slac.stanford.edu/grp/scs/net/talk/High perf ppdg jul2000.ppt.2000. 17. I. Stoica, R. Morris, D. Karger, M. F. Kaashoek, and H. Balakrishnan. Chord: A scalable peer-to-peer lookup service for internet applications. In Proceedings of the _ACM SIGCOMM 2001, pages 149–160. ACM Press, 2001._ 18. T. Suel, C. Mathur, J. Wu, J. Zhang, A. Delis, M. Kharrazi, X. Long, and K. Shan mugasunderam. Odissea: A peer-to-peer architecture for scalable web search and information retrieval. Technical report, Polytechnic Univ., 2003. 19. A. Tirumala et al. iperf: Testing the limits of your network. http://dast.nlanr.net/ projects/iperf/. 2003. 20. P. Triantafillou and T. Pitoura. Towards a unifying framework for complex query processing over structured peer-to-peer data networks. In DBISP2P, 2003. 21. Y. Wang, L. Galanis, and D. J. de Witt. Galanx: An efficient peer-to-peer search engine system. Available at http://www.cs.wisc.edu/ yuanwang. -----
{ "disclaimer": null, "license": null, "status": null, "url": "" }
2,005
[]
false
null
[]
17,535
en
[ { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/012060a1b33dbb659214465af5ad64974ca35f8e
[]
0.821707
Exploratory literature review of blockchain in the construction industry
012060a1b33dbb659214465af5ad64974ca35f8e
Automation in Construction
[ { "authorId": "79423048", "name": "D. Scott" }, { "authorId": "98555753", "name": "Tim Broyd" }, { "authorId": "2115504054", "name": "Ling Ma" } ]
{ "alternate_issns": null, "alternate_names": [ "Autom Constr" ], "alternate_urls": [ "http://www.sciencedirect.com/science/journal/09265805", "https://www.journals.elsevier.com/automation-in-construction" ], "id": "cbe2e2e0-f4d3-4923-8b48-a02259e5f89c", "issn": "0926-5805", "name": "Automation in Construction", "type": "journal", "url": "http://www.elsevier.com/wps/find/journaldescription.cws_home/523112/description#description" }
null
Contents lists available at ScienceDirect # Automation in Construction [journal homepage: www.elsevier.com/locate/autcon](https://www.elsevier.com/locate/autcon) #### Review ## Exploratory literature review of blockchain in the construction industry ### Denis J. Scott [*], Tim Broyd, Ling Ma _Bartlett School of Sustainable Construction, University College London (UCL), London, United Kingdom_ A R T I C L E I N F O _Keywords:_ Blockchain Smart contract Decentralised applications Construction industry Built environment Smart cities **1. Introduction** A B S T R A C T First academic publications on blockchain in construction instantiated in 2017, with three documents. Over the course of several years, new literature emerged at an average annual growth rate of 184%, surmounting to 121 documents at time of writing this article in early 2021. All 121 publications were reviewed to investigate the expansion and progression of the topic. A mixed methods approach was implemented to assess the existing environment through a literature review and scientometric analysis. Altogether, 33 application categories of blockchain in construction were identified and organised into seven subject areas, these include (1) procurement and supply chain, (2) design and construction, (3) operations and life cycle, (4) smart cities, (5) intelligent systems, (6) energy and carbon footprint, and (7) decentralised organisations. Limitations included using only one scientific database (Scopus), this was due to format inconsistencies when downloading and merging various bibliographic data sets for use in visual mapping software. Blockchain is the technology that enables triple entry accounting, which allows multiple parties to transact across a shared synchronous ledger. Each transaction is substantiated with a digital signature to provide proof of its authenticity [1]. Blockchain includes several key features, such as decentralised, distributed, and consensus [2]. A typical public blockchain is comprised of thousands of computer nodes con­ nected through a decentralised network, and it does not require a central power of authority to manage the system [3]. Blockchain is a selfsustaining network that rewards users for participating in mining, which is the process of creating new blocks and distributing them across all nodes on the network [4]. Whenever transactions are sent to the network, they are placed in a pool of unverified transactions, where they are periodically collected and validated by miners before they are placed into a block [5]. Miners apply a consensus mechanism to check each other’s results prior to the inclusion of new blocks, this is to ensure that there is only one version of the ledger in existence at any moment in time [6]. Bitcoin was the first blockchain which came into existence in 2009, since then, its protocol has proved immutable to hacks and has not suffered accounting errors, such as double spending [7]. Ethereum was the second blockchain to come into existence, which emerged in 2015 and introduced smart contracts, which allowed transacting parties to codify and deploy peer-to-peer agreements without the reliance of a trusted third party [8]. Smart contracts include a unique property, in that they cannot be changed once deployed, which mitigates against users unfairly withdrawing from signed agreements [9]. Smart contracts disallow external entities from interfering with peer-to-peer contracts and enables atomic transactability. The codified terms of a smart con­ tracts are transparent and open for auditing, which allows transacting parties to verify agreements for consistency. The timescale of this review spans from 2017 to 2021 and in­ corporates 121 academic documents. A bottom-up method was imple­ mented to assess the existing environment through a literature review, which includes an exploratory investigation of the progression of the topic across a wide range of application categories. The document types used in the review are comprise of journal articles, conference papers, and book chapters. Non-academic sources such as company reports were not included into the review as they do not include the same level of scientific rigor as peer-reviewed content, furthermore, the quantity of documents attainable from academic sources were of sufficient quantity. Two search queries were conducted on the Scopus scientific data­ base, which was used to obtain all of the reviewed documents. The _research method_ chapter displays the structure of these queries dia­ grammatically; furthermore, the search string for retrieving the results is available in the appendix, which allows users to replicate the search results. Other scientific databases that were considered include Web of Science (WoS), IEEE Xplore, Science Direct, Directory of Open Access Journals (DOAJ), and JSTOR [10]. Based on the topic of blockchain in construction, Scopus and WoS included the largest quantity of results by - Corresponding author. _[E-mail addresses: Denis.scott.19@ucl.ac.uk (D.J. Scott), Tim.broyd@ucl.ac.uk (T. Broyd), L.ma@ucl.ac.uk (L. Ma).](mailto:Denis.scott.19@ucl.ac.uk)_ [https://doi.org/10.1016/j.autcon.2021.103914](https://doi.org/10.1016/j.autcon.2021.103914) Received 8 June 2021; Received in revised form 22 July 2021; Accepted 19 August 2021 [0926-5805/© 2021 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/) ----- a substantial margin. In a comparison of these, it revealed Scopus with 53% more content, and with 85% of the WoS data already existent in Scopus. Both databases included a balanced range of top tier journals (top 25% based on Scientific Journal Ranking indicator), while Scopus included a larger number of mid to lower tier journals. The first academic literature on blockchain in construction emerged in 2017 within the categories of Building Information Modelling (BIM) [11], smart cities [12], and peer-to-peer energy markets [13]. The quantity of new publications on topic increased at an annual growth rate of 184% each year since 2017. The quantitative aspect of this article provides data on the expansion of the topic through statistics and sci­ entometrics. VOS-Viewer was used to present scientometric data through visual mapping. The literature review chapter was structured around application categories of blockchain in construction. Each category was substantiated by a minimum of three documents to ensure a level of academic consensus was achieved. Altogether, 33 application categories were investigated and organised into seven subject areas, which are (1) procurement and supply chain, (2) design and construc­ tion, (3) operations and life cycle, (4) smart cities, (5) intelligent sys­ tems, (6) energy and carbon footprint, and (7) decentralised organisations. An exploratory method was implemented to encapsulate a wide range of categories to investigate the existing environment through a macro-orientated approach. This method aligns with the quantitative analysis that was conducted as part of this review. _1.1. Related works_ From the 121 reviewed documents in this article, six included re­ views of similar nature and are displayed in Table 1. From these, four delimited their results to academic documents, while two incorporated a combination of academic and non-academic sources. The Non-academic material included company and organisation reports [14]. An expansive literature review of 121 documents on blockchain in construction from academic publications have only recently been feasible since 2021, as there is now an established body of work on the topic. Blockchain is a fast-evolving technology, and this article builds upon the work displayed in Table 1 to provide an updated review on the contemporary state of the topic. Bhushan et al., conducted a comparative literature review of block­ chain in smart cities, published in Sustainable Cities and Society journal, which outlined six subject areas and eight categories [15]. Hunhevicz & Hall, produced a literature review of blockchain in construction, pub­ lished in Advanced Engineering Informatics journal, which included seven categories and 24 use-cases [16]. Kiu, et al., composed a sys­ tematic review of blockchain in construction, published in the Interna­ tional Journal of Construction Management, and outlined six subject areas [14]. Li et al., composed a systematic literature review published in Automation in Construction, which extrapolated seven built envi­ ronment application categories; furthermore, three use-cases were sub­ stantiated through interviews with academics and industry practitioners, such as “automated project bank accounts”, “regulation and compliance”, and “single shared-access BIM model” [17]. Perera, **Table 1** Related works. Author Year Categories Use- Ref. count Review type cases Bhushan, et al., 2020 6 10 42 Literature Hunhevicz & Hall 2020 7 24 15[a ] Literature Kiu, et al., 2020 6 b 57[a ] Systematic Li, et al., 2019 7 3 75 Systematic Perera, et al., 2020 18 b 27[a ] Systematic Yang, et al., 2020 4 b 83 Literature Note: a Includes content from non-peer reviewed sources (e.g., reports). b Includes many use-cases that were not individually itemised by its author. et al., produced a literature review article on blockchain in construction published in the Journal of Industrial Information Integration, and identified 18 categories, extracted from academic and non-academic sources [7]. Yang et al., included a literature review in their block­ chain proof of concept article published in Automation in Construction, which summarised four subject categories for managing business pro­ cesses [18]. **2. Research method** Content was collected from journal articles, book chapters, and conference proceedings. Scopus was selected as the scientific database for extracting documents, as it contained the largest bibliographic index of academic literature on the topic, and is reputably owned by pub­ lishing organisation Elsevier [19]. Reason for using only one scientific database is due to format inconsistencies when merging data sets from various databases. When conducting a parallel search on Scopus and Web of Science (WoS) (the top two largest academic indexes on the topic) [20], it revealed Scopus with 53% more content, and with 85% of WoS documents already existent in Scopus, thus Scopus was selected as the database of choice. Fig. 1 displays the two search queries. Search one incorporated inputting the ISSN and ISBN numbers of journals and books within the subject category of architecture, building and construction, and civil and structural engineering, followed by the key words shown in the search query column in Table 1. The ISSN and ISBN number is a unique iden­ tifier given to each journal and book, which can be downloaded from [https://www.scimagojr.com. The SCImago web portal provides an](https://www.scimagojr.com) index of academic publishers for each specific subject area [21]. The Scopus web portal allows users to search for documents according to a predefined list of subject areas, in this case SUBJAREA(engi) was implemented into query two, with key terms such as blockchain and construction. Two queries were used to increase the accuracy of results from Scopus, which returned to a combined total of 412 documents. Upon removing duplicates and filtering content for suitability, the final result surmounted to 121 publications. **3. Quantitative analysis** Fig. 2 displays the quantity of documents published each year, doc­ uments types, and scientific journal rankings (SJR). SJR is the impact factor of each journal, which is calculated through a network analysis of citations [22]. SJR is measured in quartiles, whereby, Q1 represents the top 25% of journals, while Q4 is the lowest 25% [22]. The statistics in Fig. 2 were obtained through conducting a search using the queries listed in Fig. 1. The results in Fig. 2 are based on full complete years, in this case 2017–2020. This article was written in 2021, thus results from that year were not included. The subject areas and categories of the literature review are dis­ played in Fig. 3. Each category was substantiated by a minimum of three documents to ensure a level of academic consensus was achieved. These categories were further organised into seven subject areas for the pur­ pose of adding structure when organising correlating categories together. Fig. 4 displays a timeline showing when each of the reviewed cate­ gories emerged in literature. The colours in Fig. 4 are assigned in conjunction with Fig. 3. The first publications on blockchain in con­ struction instantiated in 2017 with three documents and six categories, 2018 included 9 new publications (200% increase from previous year) with nine new categories, 2019 displayed 33 new publications (267% increase) with 13 new categories, while 2020 included 69 new publi­ cations (109% increase) with five new categories. Altogether totalling the 33 categories. At time of writing this article in early 2021, there were no new additions to the category list. The category with the highest number of publications include building information modelling (BIM) with 39 documents. Joint second ----- **Fig. 1. Search query process that was used to obtain the results from Scopus.** **Fig. 2. Quantity of published content each year, document types, and SJR rankings.** **Fig. 3. The 33 reviewed categories organised into seven subject areas.** ----- **Fig. 4. Timeline showing the emergence of each category from 2017 to 2020.** with 28 documents each includes internet of things (IoT), supply chain management, and smart grids. While peer-to-peer energy markets is third place with 27 documents. The newest categories which emerged in 2020 included machine learning, water management, physical waste management, geospatial, and Integrated Project Delivery (IPD). The topical coverage of each of the 121 reviewed documents were manually recorded and transferred into visual mapping software VOSviewer, to produce the Fig. 6 visual map. VOS-viewer algorithmically maps data using natural language processing techniques [23]. Fig. 6 is broken down into three parts, which includes categories (shown as cir­ cular nodes), colour clusters (shown as the groups of nodes displayed in blue, green, yellow, or red), and links (which are the lines that connect the nodes together). Each of the reviewed documents typically covered a range of categories. Illustrating the overlap/co-occurence of these cat­ egories is the purpose of the Fig. 6 co-occurrence map. Colour clusters are assigned when a group of categories frequently co-occur in the reviewed documents. Categories with a high number of shared links naturally gravitate to the centre, as a central position has greater equi­ distance with its shared links. However, categories also gravitate to each other based on their link strength, whereby, if two categories appear frequently together in literature, they will be positioned close to each other on the Fig. 6 map. Blockchain was positioned most centrally as it shares links with all of the 33 categories. BIM was also positioned cen­ trally as it shared links with 32 out of the 33 total categories. Whereas _IPD, carbon accounting, fintech and off-site construction were all positioned_ in the outskirt, due to their low number of shared links with the overall categories. Table 2 displays the results from Fig. 6. The table is sorted from largest to smallest according to links, followed by link strength, then oc­ _currences. The Link strength is calculated by the number of times each_ category co-occurs with another. While the occurrences is calculated by the number of times each category appears in literature irregardless of its link strength. The results show that 89% of the reviewed documents included multiple categories in their paper, while 11% focused their attention solely on one category. Fig. 7 displays which blockchain platforms were most utilised in the reviewed documents. 18 documents developed solutions for Ethereum [8], while 14 developed solutions for Hyperledger [24]; additionally, one publication investigated utilising both platforms [18]. Ethereum emerged in 2015 as a public blockchain platform; furthermore, it is currently the leading platform for decentralised applications and in­ cludes the largest population of blockchain developers [25]. Hyper­ ledger, by the Linux Foundation, instantiated their own variant in the same year (2015) using a private blockchain protocol [26]. Less popular platforms in the reviewed material include Multiledger [27], Bitcoin [28], Corda [29], and IOTA [30]. Fig. 8 displays the various types of data collection implemented in the reviewed documents. A conceptual framework was incorporated in 46% of documents, which was used as a foundation to formulate highlevel ideas [31]. Case studies were also a popular method used in 27% of documents, which included joint ventures between academia and industry [32]. Literature reviews were used in 26% of the documents, which were typically implemented as a prerequisite to support the development of conceptual frameworks [33], such as with the Brooklyn micro-grid project, which used a literature review to assess the existing environment prior to the implementation of a case study [34]. Statistics was incorporated in 23% of documents, such as with measuring the performance of blockchain-based network systems [35]. The types of data collection which appeared less frequently included systematic re­ views (12%), proof of concepts (12%), interviews (7%), surveys (7%), ----- and questionnaires (1%). **Fig. 5. Quantity of publications published for each category from 2017 to 2020.** depth, such as with a PoC. Fig. 9 displays a visual map showing the co-occurrences of the data collection types shown in Fig. 8. Fig. 9 displays links shown in red nu­ merals and _link strength_ shown in blue numerals. From analysing the diagram, the top three data collection types which co-occurred most frequently in the reviewed literature included conceptual frameworks, statistics, and case studies, demonstrated through their high link strength count shown in blue numerals. The outer position of systematic reviews revealed that it co-occurred less frequently than literature reviews, however, this particular statistic can be misleading, as both systematic and _literature_ reviews are terms used interchangeably throughout research; however, the author ensured not to interfere with the termi­ nologies provided in the reviewed documents. 12 publications con­ ducted a proof of concept (PoC), which surmounts to 10% of the reviewed documents. The data collection types with the least number of co-occurrences included questionnaires, systematic reviews, and sur­ veys. Altogether, 55% of the reviewed documents incorporated multiple data collection types in their research, while 45% included only one. Through conducting this review, the author noticed that papers which included higher numbers of data collection types were typically less technical overall, such as literature/systematic reviews. While papers which included only one data collection type were typically more in Table 3 is to be read in conjunction with Fig. 9, and is organised according to link count, _total link strength, and_ _occurrences. Link count_ refers to the quantity times a particular type of data collection co-occurs with another; however, it does not take into account the weight if each link. Whereas _link strength_ factors in the weight, which refers to the cumulative total of when each link co-occurred with another. The oc­ _currences column represents the quantity of times each data collection_ type occurred in literature regardless of its links or link strength. **4. Literature review** The literature review is broken down into seven sections, which is represent by the seven subject areas listed in Fig. 3, these are (1) pro­ curement and supply chain; (2) design and construction; (3) operations and lifecycle; (4) smart cities; (5) intelligent transport; (6) energy and carbon footprint; and (7) decentralised organisations. Each subject area includes several application categories, these were grouped according to their correlation. The subject areas and categories were selected following a bottom-up approach. This was conducted without a pre­ defined or systematic strategy on which topics to cover, provided that it was in conjunction with the construction industry or built environment. ----- **Fig. 6. Co-occurrence map of the 33 reviewed categories.** **Table 2** Presents the values of the categories displayed in Fig. 6. The colours labelled in the ’Clusters’ column is representative of the colour clusters shown in Fig. 6. Categories Link Total link Occurrences Cluster strength BIM 32 146 37 Blue Supply chain 29 132 31 Green IoT & cyber-physical 27 131 27 Red Intelligent transport 27 79 15 Red Smart cities 25 73 15 Red Cybersecurity 25 54 12 Red Logistics & scheduling 24 81 16 Green Cash flow & payments 24 56 12 Blue Smart grids 23 84 29 Yellow Digital contracts 22 70 14 Green Cloud, ERP & EDMS 22 61 13 Red FinTech 21 57 9 Green Standards 21 40 9 Blue Real estate 20 48 10 Green AI 20 47 8 Red Physical waste 20 28 3 Green Water mgmt 20 28 3 Green P2P energy 19 95 31 Yellow Citizen participation 19 26 4 Red ID & certificate 18 29 5 Green Big data & analytics 17 39 6 Red Smart homes 17 26 4 Blue Facility mgmt 16 25 5 Green Life cycle & circular 14 44 10 Green Procurement 14 36 11 Yellow Geospatial 14 25 4 Red Machine learning 14 21 4 Red Off-site const. 11 27 5 Red Decentralised Autonomous 10 12 3 Blue Organisation Carbon accounting 8 19 7 Yellow IFC 8 16 5 Blue Renewable energy 6 8 3 Yellow IPD 4 6 3 Blue The process followed an organic progression through manually making note on a spreadsheet the topical coverage of each of the review docu­ ments, as shown in the shared Google spreadsheet following the link provided below. [https://docs.google.com/spreadsheets/d/1V4UICRdoyWycaGENH9](https://docs.google.com/spreadsheets/d/1V4UICRdoyWycaGENH9rnuxukRNQJFIArQ-feV7NM0a4/edit?usp=sharing) [rnuxukRNQJFIArQ-feV7NM0a4/edit?usp=sharing](https://docs.google.com/spreadsheets/d/1V4UICRdoyWycaGENH9rnuxukRNQJFIArQ-feV7NM0a4/edit?usp=sharing) _4.1. Procurement and supply chain_ This section is comprised of six application categories grouped into the procurement and supply chain subject area. Altogether, this subject area was discussed in 57 out for the 121 documents and is focused on pre-construction activities. _Procurement, bid, and tender (discussed in 12 documents). In a survey_ conducted by Kim, et al., based on theme of lifecycle, project manage­ ment, and blockchain, and from respondents in construction industry, the top three applications for blockchain emerged as bidding, procure­ ment, and change management [36]. Lack of trust is particularly evident in procurement, and current management practices requires innovating to improve the ability to track provenance of faults, trace contract al­ terations, and drawing revisions, while minimising information asym­ metry during the tender process [37]. Based on a questionnaire and survey by Isikdag, of 64 industry practitioners in construction industry, consisting of architects, engineers, contractors, and subcontractors, eprocurement appeared to offer very few benefits compared to its nonelectronic counterpart, furthermore, the primary barrier to e-procure­ ment includes a lack of trust in supply chain, unsatisfactory legal infrastructure, and inadequate cybersecurity for storing confidential data [38]. Moreover, Isikdag, stated that blockchain can potentially be used to provide the vital infrastructure required to support privacy without the risks associated with centralised storage; furthermore, he discussed how e-procurement lacks standardisation from regulatory bodies [38]. _Logistics, scheduling and programme_ (discussed in 16 documents). Logistics management has become increasingly complex due to global­ isation [39]. Kifokeris, et al., performed a case study of seven Swedish logistics consultancies, which outlined that “delivery failure, imprecise ----- **Fig. 7. Utilisation of blockchain platforms in the reviewed documents** **Fig. 8. Data collection types existent in the reviewed documents.** **Fig. 9. Co-occurrence map of the data collection types.** data, delays in time, inefficient flows and data transfers between sys­ tems” are limitations in existing logistics processes, and discussed the lack of cyber-physical systems integration and analytics in managing on site assets [39]. Moreover, he proposed a blockchain solution for logis­ tics, using a crypto-economic model to incentivise collaboration [39]. Lanko, et al., considered that existing centralised computer systems are ----- **Table 3** Presents the values of the data collection types displayed in Fig. 9. The numerals highlighted in bold in the ’Total link strength’ column are the same values as the blue numerals shown in Fig. 9. Data collection type Link count Total link strength Occurrences Conceptual framework 7 **48** 52 Case study 6 **43** 32 Interview 6 **9** 8 Survey 6 **7** 5 Statistics 5 **39** 27 Literature review 5 **16** 31 Proof of concept 4 **12** 13 Systematic review 2 **3** 12 Review 2 **2** 7 Questionnaire 1 **1** 1 susceptible to data manipulation, and proposed a framework which incorporated blockchain with RFID for managing logistics of readymixed concrete on-site, whereby, RFID tags are used to record stages of delivery, such as pouring, transportation, handling, quality in­ spections, and mould forming, with all data exchanges recorded on the blockchain [40]. Blockchain in logistics provides opportunities in of­ fering improved service to clients through automating the process of storing and authenticating data with increased trust, furthermore, decentralised applications potentially reduce the resource requirements in managing systems efficiently [41]. _Cash flow and payments (discussed in 14 documents). Chong & Dia­_ mantopoulos conducted a case study on a project in Melbourne, Australia, that used blockchain to automate payments; Furthermore, works included the delivery of 5000 building façade panels tracked with Bluetooth sensors to monitor live location of each panel from factory in China to on-site, with BIM used to monitor installation of each panel, while smart contracts executed payments at delivery checkpoints [42]. Additionally, this integrated with a mobile phone application which allowed project participants to view progress of installation in real-time [42]. Ahmadisheykhsarmast developed an add-in for Microsoft Project using programming language C-sharp and Visual Studio, which allowed smart contracts to integrate with mainstream project management software; furthermore, blockchain platform Ethereum, with its native programming language Solidity, was used to link the front and back-end functions of the user application that connected blockchain to Microsoft Project [43]. Late payments is a major problem in construction, caused by con­ tractors performing cash farming, which is the process of withholding supply chain payments to sustain positive cashflow while aggressively investing in new work [44]. Das, et al., proposed a conceptual frame­ work that enabled smart contracts to control the release of payments to supply chain which includes integration with banking systems, furthermore, he discussed the potential to integrate with strategies such as Project Bank Account (PBA) [45]. _Digital and automated contracts (discussed in 14 documents). McNa­_ mara & Sepasgozar conducted an interview of industry practitioners in the construction industry and derived that trust, risk, and dispute management were ubiquitous concerns in almost all projects, with main contractors exerting dominance through unfair contract conditions [46]. In a survey conducted by Badi et al., of 104 respondents in the UK construction industry, regarding the use of smart contracts, the main factors which determined its adoption in enterprise is competitive edge and commercial value [47]. Hunhevicz, et al., proposed a digital con­ tracting framework which simulated the decision points of a typical design-bid-build project in Switzerland, which included the client, owner, planner, contractor, and supplier, all interacting with smart contracts to control the approvals and validations process of contract activities, such as project definition, design coordination, tendering, supplier selection, and contract signing; furthermore, this was proto­ typed through a web-based application connected to the Ethereum blockchain [48]. _Supply chain management_ (discussed in 30 of documents) Qian & Papadonikolaki conducted interviews of industry practitioners in the construction industry that are knowledgeable in supply chain and blockchain, and identified that blockchain can potentially be used to mitigate the trust problem in construction, through data traceability, non-repudiation, and disintermediation; furthermore, it was projected that blockchain can save up to 70% on costs associated with data pro­ cessing and management, through automating compliance checking, payments, and analytics on project performance [49]. Sheng, et al., proposed a framework which allowed project participants to assess compliance to standards and monitor information exchanges through a user application, where project participants would upload data associ­ ated with contract documents, project schedule, and cost; furthermore, the application would autonomously notify users of their responsibilities to upload or approve works, which automated the processing of pay­ ments and completion certificates [50]. Dutta, et al., conducted a sys­ tematic review of blockchain in supply chain and identified several key attributes where blockchain can improve performance, such as eviden­ tiary trail of delivered works substantiated with immutable data, resil­ ience from network disruption, improved data synchronicity, data trust in cyber-physical systems, business process automation through smart contracts, and improved tracking of product revisions [51]. _Standards, regulation, and compliance_ (discussed in 10 documents). The transparent and irrefutable properties of the blockchain make it a suitable technology for trialling whether smart contracts can be used to automate the compliance checking of objects in BIM models [52]. Nawari & Ravindran proposed an automated regulation and compliance checking framework for BIM, whereby, modelling elements are scanned and cross-checked with client specifications, which autonomously no­ tifies designers of their obligations to make design alterations [53]. Blockchain can also be used as a decentralised authority to provide BIM objects with copyright verification, through a lookup service that checks the intellectual property signature of a BIM object, and cross-checks it with data stored in a distributed database; furthermore, designers and contractors working on a BIM model can be instantaneously notified of any copyright infringement of model objects [54]. _4.2. Design and construction_ The design and construction subject area consists of five application categories discussed in 44 of the review documents. This section is fo­ cuses on the capital expenditure stage of construction projects. _Building Information Modelling (BIM)_ (discussed in 41 documents). One of the fundamental reasons for the slow adoption of BIM is a lack of traceability in model revisions, as the current systems is based on manual data entry and relies on trust from designers to keep track of changes [55]. The ability for multiple users in a project to update a BIM model simultaneously is extremely challenging using existing central­ ised cloud systems, furthermore, the coupling of BIM with blockchain further creates bandwidth limitations, which is due to blockchain’s consensus properties, whereby, majority of the nodes on the network need to agree on changes before data can be revised [56]. Zheng, et al., proposed a mobile device application which allowed users to verify on their portable computing device (e.g., phone, tablet, laptop) whether a BIM model is the most recent version, whereby, a hash of the BIM model is stored on the blockchain which allows a lookup service to cross-check the hash of a downloaded model with the hash stored on-chain, after­ wards, the application would provide users with a verification receipt stating the model’s validity [57]. On another note, a case study by Mason, et al., discussed how the effective logging of geometry and volume in BIM models can transition effectively into computable code for smart contracts [58]. _IFC-based interoperability_ (discussed in 6 documents). IFC is a data standard format registered by the International Standards for Organi­ sation (ISO), which is used for saving BIM model files [59]. ----- BuildingSmart is an organisation that promotes digital workflow through utilisation of IFC, while OpenBIM is a set of common agreed workflow standards for BIM projects, for the purpose of increasing supply chain collaboration and standardising data exchange processes [59]. Hunhevicz, et al., produced a prototype which incentivised users to produce high quality data sets following the OpenBIM standard, this incorporated the use of smart contracts to provide financial rewards based on the quality of data provided by its users [48]. Ye, et al., pro­ duced a prototype which incorporated an IFC model that interoperated with smart contracts, which executed payments autonomously based on elements quantified within the BIM model; furthermore, readable text was maintained as it transferred into smart contracts, which allowed users the ability to intuitive cross-reference IFC data in blockchain code [60]. A study was conducted by Xue & Lu which investigated whether IFC semantics can be substantially minimised to allow for potential storage of IFC code on-chain, and whether small portions of the IFC code can be partitioned away from its original syntax while still remaining readable for purpose of isolating model revisions, which resulted in a semantic reduction of 99.98% of its original size; however, the consensus properties of blockchain proved to be problematic due to its low throughout with data processing, even when tested on a private blockchain network [56]. _Integrated Project Delivery (IPD)_ (discussed in 3 documents). IPD operates through onboarding the construction supply chain with a shared risk and reward contract for the purpose of promoting collabo­ rative workflow [61]. Hunhevicz, et al., discussed how the character­ istics of IPD integrate effectively with the ideologies of common pool resource (CPR) and the Ostrom principles for flat organisational struc­ tures, which incorporates mutual and economical benefit for project participants who work together to achieve a common goal, whereby, projects which implement blockchain in IPD contracts include potential to reward participants with tokenised and non-tokenised incentives, such as financial rewards for collaborative delivery, transparent agree­ ments, and automated payments upon validated completion of works [62]. Elghaish, et al., conducted a simulated proof of concept which incorporated blockchain in an IPD contract for managing supply chain payments, using private the blockchain platform Hyperledger Fabric (HLF); Whereby, financial operations such as reimbursed cost, profit pool, cost saving pool, and risk pool, were programmed into smart contracts which automated the dispensation of funds according preagreed terms, such as validated completion of works from appointed authorities and project milestone dates [61]. _Off-site construction (discussed in 4 documents). Off-site construction_ includes strong topical overlap with Internet of Things, blockchain, BIM, AI, robotics, and 3-D printing [63]. According to Turk, R. Klinc, the primary application for blockchain in off-site construction is supply chain management, with a projected average saving of 70% through reduced processing costs, which is amassed through improved systems integration, automation through smart contracts, and real-time data traceability [63]. Wang et al., proposed a framework using blockchain platform Hyperledger Fabric for the management of precast construc­ tion activities through a user interface, which allowed real-time querying of scheduling, production, and transportation [64]. Additive manufacturing, synonymously called 3-D printing, includes potential to integrate with off-site construction and blockchain for the production, cataloguing, and copyrighting of customised building components [65]. _Geospatial, 3-D scanning, and point cloud (discussed in 4 documents)._ Geospatial technologies such as “remote sensing, LiDAR, internet map­ ping, GPS and GIS” have strong implications working in conjunction with autonomous vehicles due to their rapid response in scanning geographical landscapes; furthermore, it interoperates effectively with BIM models, smart infrastructure, and cyber-physical systems [66]. 3-D scanning allows assets and geographical locations to be imported into BIM models; however, there is currently a lack of technological capacity for scanned objects to be autonomously cross-referenced with registered objects in a database [63]. Copeland and Bilec proposed a conceptual framework which integrated assets with geospatial sensors and block­ chain to produce what they called “buildings as material banks”, which utilises sensors affixed to building components which records metadata regarding its condition for reusability, using blockchain as the trusted system for authenticating components and materials within built assets [67]. _4.3. Operations and lifecycle_ The operations and lifecycle subject area is comprised of four cate­ gories and consists of 24 documents. This section is focused on the operational expenditure stage of an asset’s lifecycle. _Facilities management and maintenance (discussed in 6 documents). Li,_ et al., proposed a framework for the semi-automated procurement of replacement parts during the operations phase of a built asset, which includes the integration of Internet of Things (IoT) sensors and a com­ puter aided facilities management system (CAFM) for the automated identification of faulty parts; furthermore, a request for replacement parts is processed through a decentralised autonomous organisation, while an e-marketplace handles the bidding and appointment of pro­ spective contractors [68]. Blockchain includes the ability to transact on and off-chain for the purpose of increasing the performance of data exchanges in a decentralised network. Bai, et al., proposed a framework for managing the communications between IoT and blockchain for asset maintenance, which uses on-chain for immutable hash storage and smart contracts, while off-chain handles data storage, computational processing, and analytics [69]. Integrating off-chain applications with blockchain allows for greater transaction throughput, lower transaction fees, and greater control over system operations such as privacy controls. _Life cycle and circular economy (discussed in 11 documents). Shojaei_ discussed how metadata recorded of raw materials extracted from source can be appended onto the blockchain for end-to-end lifecycle assessment, which allows for a complete and uninterrupted data stream from each handling merchant to end-user to provide proof of prove­ nance from source to construction [70]. Asset data such as specifica­ tions, standards, and contract agreements include potential to integrate with blockchain for post-occupancy evaluation, utilising BIM as the data repository for the built environment asset and blockchain as its corre­ sponding data validator [71]. Copeland & Bilec proposed a framework which utilised RFID, BIM, and blockchain to provide components with an evidentiary trail of data throughout its lifecycle, through sensors periodically recording data at key stages, such as installation, decom­ mission, provenance, and metadata regarding supplier, manufacturer, and handling checkpoints [67]. This includes potential to integrate with a crypto-economic incentive scheme for the recycling of assets, with data verified by blockchain. _Construction waste management (discussed in 3 documents). Surplus_ waste generated by the construction industry is a global issue; further­ more, there is a lack of systems that can accurately account for material waste, which make it an acceptable by-product despite its carbon impact and incurred costs on projects [7]. However, blockchain includes po­ tential to increase the accountability of waste through its ability to verify its lifecycle from source to disposal [7]. Despite this, a proposed solution on who would supply the systems which allows supply chain to quan­ titatively account the unused material was not discussed in the reviewed papers. _Real estate and property registry (discussed in 10 documents). Dakhli,_ et al., conducted a case study of 56 residential properties and concluded that blockchain has potential to achieve construction cost savings of 8.3%, which is higher than a typical property developer’s net margin of 6%; furthermore, the projected cost savings were attributed to the use of smart contracts and a decentralised autonomous organisations (DAO) to manage and automate business processes [72]. The management of land registries in many developing countries is an unnecessarily complicated process which is prone to fraud and ----- manipulation [73]. Land management was identified in the World Bank’s Ease of Doing Business report as a one of the main services that affects the economic growth of a country, furthermore, blockchain was discussed as having the potential to provide a single source of truth to land records, thus reducing administrative overheads in data processing and alleviating risk of fraud [73]. _4.4. Smart cities_ The smart cities subject area is comprised of four categories and consists of 27 documents. This section is focused on how city infra­ structure networks can interoperate to provide a data-rich ecosystem of connect devices for managing built environment assets. _Smart cities (discussed in 16 documents). Ahad, et al., conducted a_ literature review on the topic of smart cities and suggested that they are driven by network-based technologies that integrate to support the de­ livery of industry 4.0 [66]. These technologies include Internet of Things (IoT), big data, cyber-physical systems, 5-G technology, artificial intel­ ligence (including machine learning and deep learning), blockchain, cloud/edge computing, and geospatial technologies [66]. The inter­ connected network of devices in a smart city increases the demand for trusted data, therefore, a new business model is required that is more resilient to hacks and central point of failure [74]. This can potentially be supported through the traceable, immutable, and decentralised properties of blockchain [74]. Fu & Zhu proposed a conceptual frame­ work which integrated technologies such as cloud platforms, block­ chain, and IoT to form a trusted platform for monitoring live data from infrastructure services, such as geographic information systems (GIS), safety devices, and weather monitoring systems that relay information to city infrastructure services such as transport, communication, and utility [75]. _Smart homes and buildings (discussed in 4 documents). Moretti, et al.,_ proposed a conceptual framework that incorporated the use of ultra­ sonic sensors for the purpose of monitoring indoor activity of a building, which includes sensors placed in rooms to monitor usage, occupancy, and maintenance, which integrate with analytics to provide automated reporting of indoor activity; furthermore, the author discussed the po­ tential to incorporate a blockchain-based management system, through using smart contracts to provide automated payments upon successful delivery of maintenance works [76]. Roy, et al., proposed a prototype for a smart home ecosystem, which included the aggregation of a home device network, blockchain platform, and maintenance service system; furthermore, the home network was comprised of smart meters, IoT, and actuators; the blockchain was used to store and validate results received from the home devices; while the maintenance system provided facility management through identifying when replacement parts were required and provided credentials of prospective suppliers [77]. _Intelligent transport_ (discussed in 15 documents). Lopez ´ & Farooq proposed a smart mobility blockchain framework for managing trans­ portation data, which was comprised of five layers such as (1) privacy, which gives users control of their data when using location revealing applications such as Google maps; (2) contract layer, which controls how smart contracts use user data; (3) communication layer, which appends digital identifiers to communication channels between network nodes; (4) incentive layer, which rewards users for participating in the blockchain network; and (5) consensus layer, which allows nodes to upload data verified by its users [24]. Implications of this included privacy between users and transportation system hosted on a decen­ tralised network [24]. Supplying battery recharge to electric vehicles based on a fast-charge system is technologically challenging, as current recharge systems need to be designed for both intermittent and continual usage [78]. Zhang, et al., conducted a 15 month study at University of California, Los Angeles (UCLA), which implemented a blockchain platform that incentivised users to charge their electric vehicles at specific timescales, which mitigates energy providers having to store unused energy in batteries for extended periods of time; moreover, a user interface pro­ vided users with a ranking system based on their record of renewable energy consumption, which rewarded users with discounts and the ability to choose flexible recharge schedules [78]. _Water management (discussed in 3 documents). The infrastructure for_ wastewater management in cities is reaching the end of its lifespan in many countries, which is caused by old treatment plants and damaged pipes which excrete sewage into environmentally sensitive areas that cause health and safety and wildlife concerns [79]. Berglund, et al, discussed how the construction of new water management systems can potentially benefit through innovations such as Internet of Things (IoT), smart meters, and blockchain, to provide live data feed on the perfor­ mance of water management systems, with implications in improving lifecycle maintenance of infrastructure assets [79]. Perera, et al., dis­ cussed how WaterChain, a water utility blockchain network in the United States, allows their participants to invest in water recycling plants and allows them to benefit through the dividends supplied by its service; furthermore, the management of the plant is transparent and can be investigated by the community at any time and dividends are automated through smart contracts, this merges the boundary between consumer and producer and allows the opportunity for communities to self-sustain and self-own their utilities [7]. _4.5. Intelligent systems_ The intelligent systems subject area includes six categories and consists of 46 documents altogether. This section focuses on advanced computer systems, information processing, and the benefits of data-rich networks. _Big data (discussed in 6 documents). The amount of new data pro­_ duced each year is increasing exponentially, furthermore, the con­ struction industry is under additional pressure to exploit the benefits of data-driven economies whilst in a resource deficit caused by poor margins in construction projects [80]. Blockchain offers a new type of data model which reduces the resource requirements for storing data securely, through bypassing the need to use heavily centralised systems to authenticate data [66]. Network systems such as internet of things (IoT) and smart technologies include the potential to integrate with blockchain to provide increased trust in authenticating data, which is achieved without reliance on oversight from centralised technology companies [24]. Concerns regarding privacy is mitigated through pri­ vate blockchain protocols such as Hyperledger Fabric, which uses an enterprise-centric model that provides platform developers with control over the privacy features on their network [81]. Alternatively, public blockchain protocols, such as Ethereum, include advanced crypto­ graphic methods such as zero-knowledge-proofs which allow private data exchanges to occur on a public network [82]. Big data integrated with blockchain includes practical applications in off-site construction and supply chain management, through improved contract manage­ ment, compliance checking, traceability of data in project reports, and reliable data for use with analytics [63]. _Artificial intelligence (AI) (discussed in 8 documents). AI, alongside_ additive manufacturing (synonymously called 3-D printing), autono­ mous vehicles, blockchain, drones, and Internet of Things, are the fundamental components that form to create the emerging industry 4.0, which were points first discussed in the 2011 report by Germany’s economic development agency [65]. Car manufacturers use AI powered robots that work alongside humans in production plants; furthermore, companies such as General Electric and Caterpillar are developing AI solutions to equip workers with robotic exoskeletons to assist with la­ bour intensive jobs [65]. AI is progressively being used in industries to streamline workflow and improve decision making, such as with JP Morgan, who developed a software algorithm called COIN, that scans thousands of contract documents instantaneously to provide judgement on written agreements [83]. A practical use-case for blockchain in AI is the ability to safeguard its coding through placing it in a smart contract, ----- which mitigates the risk of unauthorised manipulation of the code without permission from authorised actors, effectively, creating un­ breakable codified laws which govern the functionality of AI; simulta­ neously, AI can also be used to debug smart contracts and improve blockchain’s protocol design [84]. _Cloud computing and electronic document management system (EDM)_ (discussed in 13 documents). EDM allows companies to manage, store, and process documents electronically [30]. EDM platforms are limited with their potential to interoperate with other technology suppliers, which is due to its centralised systems architecture; conversely, a blockchain-based EDM is built with interoperability as its core and is not financially driven by sales of its product [14]. Cloud computing is a fundamental driver of logistics 4.0 (a branch of industry 4.0), which encompasses global standards, digitisation of business processes, and cyber-physical systems that interoperate with supply chain and logistics networks [14]. Blockchain-based decentral­ ised cloud platforms provide the ability for users and enterprises to store data with greater privacy, this is achieved without risk of hacks or data mining from service providers; however, due to its nascency, decen­ tralised storage solutions may lack in its ability to modularise its func­ tions to suit business workflows [85]. Singh, et al., proposed a framework for managing the data flows of cyber-physical systems in a smart city network, which integrates cloud computing, software-defined networking, and blockchain for trusted data exchanges [29]. _Cybersecurity (discussed in 12 documents). The decentralised char­_ acteristics of blockchain puts the responsibility in the custody of its users to manage their digital keys competently, which requires users to keep their private-key secret and not reveal the personal identity behind their public-key [86]. Xiong, et al., proposed a “secret-sharing-based key protection” protocol which allows users with compromised or lost private-keys to retrieve access to their account, which involves a stepby-step multiparty verification process, whereby, each party anony­ mously and privately reveals a small portion of the key, which altogether combines to produce the entire lost private-key [28]. The immutable property of blockchain also comes at the cost of low scalability (measured in transactions per second) and limited capacity to store large amounts of data on-chain [87]. To mitigate this, Bai, et al., proposed a framework which consists of on-chain and off-chain func­ tionalities, which included a “smart predictive maintenance” and a “sharing service of equipment status data” model, whereby, the hashes (unique identifiers) of files are stored on-chain, while off-chain handles high-volume data storage and computational processing [69]. This in­ cludes the use of a lookup service which connects the hashes stored onchain to data repositories off-chain, which amalgamates the immutable properties of blockchain with large capacity data storage [69]. _Machine learning (ML) (discussed in 3 documents). The procurement_ and management process of road construction in India is challenged with political corruption and fraud, through lack of compliance checks, material fraud, and unsupervised labour that leads to incomplete works [88]. Shinde, et al., discussed how ML can be used to forecast material quantities, labour requirements, and delivery schedules, while block­ chain can be used as the trusted system to verify the authenticity of data sets without reliance on a trusted third party; furthermore, ML coding can be stored in a smart contract or decentralised repository, which can be designed to allow authorised parties to jointly contribute to updating and verifying the code through consensus [88]. ML is used in con­ struction for statistical decision making, irregularity detection, and deriving insight from historic records [89]. Woo, et al., identified five software applications that use ML in the construction industry, these are (1) GenMEP, by Building Systems Planning, which uses ML for the automation of mechanical, electrical, and plumbing data in a Revit model; (2) BIM 360 IQ, by Autodesk, which uses ML to forecast and calculate the impact of subcontractor risks in construction projects; (3) SmartTag, by Smartvid.io, uses ML to automate the process of labelling/ tagging of site assets from pictures and videos; (4) Smart Construction, by Komatsu and NVIDIA, uses ML to simulate the construction process for health and safety and programme analysis; followed by, (5) IBM Watson IoT, who uses ML for proposing energy efficiency and occupancy enhancing solutions in buildings [89]. _Internet of things (IoT)_ (discussed in 31 documents). Wang, et al., discussed how IoT and blockchain can potentially integrate with building information modelling (BIM) to provide a central hub for managing and authenticating data received from built environment sensors; furthermore, the BIM model can be used to map the position of each sensor in a digital model, which provides a 3-D map for mainte­ nance suppliers to utilise [63]. IoT can also be fitted onto the wearables of personnel on construction sites to provide quantitative insight on the environmental conditions and geographic positioning of on-site workers, with blockchain used to hash and timestamp data received from the IoT [30]. Fu & Zhu proposed a smart city framework which incorporates the use of IoT to provide a system which integrates and monitors geographic, safety, and weather, which altogether feed data to a user interface to provide live analytics for use in construction and asset management [75]. _4.6. Energy and carbon footprint_ The energy and carbon footprint subject area includes four categories and consists of 38 documents altogether. This section is focuses on how blockchain can integrate as part of a system to better manage energy, renewables, and carbon. _Peer-to-peer (P2P) energy markets (discussed in 30 documents). P2P_ energy markets are designed around homeowners buying and selling excess renewable electricity through a local network, which provides neighbourhoods with self-sufficiency and promotes decarbonisation [90]. Esmat, et al., proposed a conceptual framework for a P2P energy marketplace hosted on blockchain, which includes automated uniform pricing and real-time settlements [91]. Ableitner, et al., conducted a 4month field study of 37 households in Switzerland to assess the outcome of a micro-grid prototype, which was a joint effort between academia and industry; furthermore, each of the households were supplied with renewable energy production technologies, smart meters, and a P2P energy trading application hosted on the blockchain [92]. Afterwards, the results were analysed through questionnaires, interviews, and sta­ tistics, which displayed active involvement from the participants with the blockchain application and an eagerness from the households to continue with the study after it concluded [92]. Energy trading can also occur between machine-to-machine (M2M) for the purpose of achieving full automation without the reliance of appointing users to authorise the trade, as shown in a conceptual framework by Sikorski, et al., which included a study of two energy suppliers that operate in tandem to provide consumers with the most economically priced electricity [13]. Despite the immutable property of blockchain, P2P markets are at po­ tential risk from producers manipulating the power measurements recorded at connection points; However, to mitigate this, Saha, et al., proposed a blockchain-based distributed verification algorithm that penalises inconsistent measurements of current [93]. _Smart grids (discussed in 29 documents). ‘Peer-to-peer (P2P) energy’_ and ‘smart grids’ are discussed interchangeably; however, the former relates to trading markets, while the latter relates to energy infrastruc­ ture and smart meters. The integration between decentralised microgrids and the main power grid is made possible through a demandside management (DSM) application proposed by Noor et al., whereby consumers are able to supply their own smart energy appliances and battery storage and utilise the DSM application to connect their local grid to the main grid [94]. Christidis, et al., conducted a case study of 63 solar panel fitted homes, situated in Texas, United States, which compared the efficiency of a semi-centralised versus a decentralised energy grid market, which included the former consisting of high transactions speeds with lower security, while the latter included low transaction speeds with higher security, which resulted in the block­ chain approach being less efficient due to its high latency in processing ----- transactions [81]. A similar framework was proposed by Foti & Vavalis, which investigated how a blockchain-based smart grid would perform with 1000 participants transacting on the Ethereum blockchain testnetwork, which resulted in the centralised grid being efficient at providing lower cost electricity due to the mining fees associated with blockchain, however, when factoring in the lifecycle cost of managing systems, the decentralised approach was discussed as potentially being more cost-effective and resilient to external threats such as cyber-attacks [95]. _Renewable energy solutions (discussed in 3 documents). The energy_ industry is experimenting with new business models that transition from centralised to decentralised, which includes the integration of smart devices, micro-grids, blockchain, and energy recycling technologies [96]. A combined heat and power (CHP) system provides energy recy­ cling through combining electricity and heat generation into one system, which integrates fittingly with renewable production technologies such as photovoltaic and wind turbines for the purpose of reducing carbon footprint [97]. Furthermore, in the event of natural disasters such as flooding, high winds, earthquakes, wild fires, snow/ice, and extreme temperature, CHP maintained performance most consistently in com­ parison with photovoltaic, wind turbine, standby generators, and biogas [97]. The demand for renewable energy increases with the depletion of oil and rise in global warming. Perrons et al., stated that the geothermal energy sector has received pressure from stakeholders to innovate renewable production methods and management systems, with block­ chain discussed as a potential candidate to improve the software aspect of it [98]. Keivanpour investigated two off-shore wind farms in United Kingdom, called Robin Rigg, and Walney Phase 1, and concluded that the current delivery method of industrial scale renewables is unneces­ sarily expensive due to longstanding supply chain processes, and dis­ cussed the innovation potential with blockchain, Internet of Things, and big data [99]. _Carbon accounting and decarbonisation_ (discussed in 7 documents). Khaqqi, et al., proposed a carbon emission trading framework, where a government organisation would issue construction companies with a limited number of carbon credits to expend on a construction project, whereby, each credit is representative of a tonne of carbon emissions; furthermore, companies are able to buy or sell excess carbon credits to each other through a decentralised online marketplace, which incenti­ vises renewable companies, while at the same time penalises nonrenewable companies [27]. Rodrigo, et al., conducted an interview with three industry practitioners, each with over 13 years of experience in information technology, which concluded that the inherent properties of blockchain, such as auditability, security, and decentralisation, is a suitable tool for embodied carbon estimating [100]. Hua, et al., pro­ posed an energy trading framework that rewards carbon credits to prosumers of a micro-grid network, whereby, energy producing tech­ nologies are linked to the blockchain to record the carbon footprint at time of production; furthermore, each prosumer is provided a set quantity of carbon credits which their permitted to expend during pro­ duction, which incentivises prosumers to act sustainably [101]. _4.7. Decentralised organisations_ The decentralised organisations subject areas is comprised of four categories and consists of 19 documents altogether. This section is focused on decentralised services and autonomous organisations. Some of the topics in this section are more general purpose than the previous sections, nevertheless, they included strong overlap with the construc­ tion industry and each category was discussed several times in the reviewed documents. _Decentralised Autonomous Organisation (DAO) (discussed in 5 docu­_ ments). DAO is an autonomous blockchain entity with decentralised governance at its core, which rewards users with tokenised incentives for participating in the network and operates entirely through smart contracts [102]. The construction industry is particularly known for incurring change orders and programme alterations, which is prob­ lematic for smart contracts due to their unalterable properties once deployed; furthermore, translating written agreements into codified form creates linguistical challenges between contract managers and programmers, whereby, each party may not understand the industryspecific culture differences of the other, such as terminologies and processes [68]. Dounas, et al, produced a prototype which utilised DAO and smart contracts to automate the awarding of works for architectural design submissions, which involved a simulated study where stake­ holders submit a request for a built environment asset through a DAO platform, followed by submission of the designs from prospective con­ tractors or architects, and finally, the autonomous calculation of the winning proposal through a predefined scoring system and awarding of work through a smart contract [103]. Similarly, DAO also includes the potential to integrate with the construction or operations phase of a built asset, through semi-automating the procurement process for obtaining new materials or replacement parts, whereby, DAO is used as the me­ dium for connecting prospective suppliers to new work, managing payments, cross-checking compliance certificates, and quantitatively assessing the risk of each supplier through their track record of delivered works [104]. _Identity and certificate authentication (discussed in 5 documents). The_ fundamental properties of blockchain (traceability, transparency, and immutability) make it a suitable technology for incorporating identity authentication services, as centralised systems are prone to hacks and data manipulation [86]. Private blockchains include privacy controls as a fundamental feature to its protocol. Whereas, public blockchains include cryptographic functions such as zero-knowledge proofs which permit private transactions to occur on a public network, however, this incurs additional transaction fees added onto the existing mining fee [82]. Nawari & Ravindran discussed how private blockchain Hyper­ ledger is suited for identity management services in construction due to its modular architecture, which allows automated compliance checking of identities on the network [53]. Similarly, Shojaei, et al., discussed how Hyperledger’s certificate authority can be used to maintain an active lists of supply chain participants in a construction project, which can be reused across multiple projects [105]. Blockchain allows the creation of non-fungible tokens (NFTs), which can be used as a digital certificate that represents the ownership of a physical asset; furthermore, this NFT can hold additional data such as title deeds, lifecycle data, building certificates, and any other associated data [106]. Implications include substantial reductions in data retrieval for insurers, estate agents, facility managers, and building inspectors [72]. Due to the immutable properties of blockchain, data stored in the NFT is append only, thus leaving an intact evidentiary trail of data throughout its lifespan. _Financial technology & banks (discussed in 7 documents). The emer­_ gence of decentralised finance in 2020 allows banks to extend their portfolio to include additional commercial products for customers [12]. Yao, et al., proposed a conceptual framework which discussed the viability for banks to provide blockchain-based supply chain finance, through using blockchain to verify the regulatory compliance of their customers, track signed agreements, and trace pending invoices [107]. Blockchain can be used to maintain an accurate and irrefutable record of transactions without risk of ledger inconsistencies, such as reconcilia­ tion errors and double spending; furthermore, banks can potentially provide escrow services through smart contracts, which allows trans­ acting parties to formalise agreements amongst themselves while under oversight from regulatory controls, this ensures compliance to fair business terms and legal standards [15]. Smart contracts also include the potential to automate tax duties, such as with the legal movement of goods across international borders, whereby compliance certificates would be autonomously awarded upon payment of taxes [108]. _Crowdsourcing_ (discussed in 4 documents). Blockchain-based crowdsourcing is a decentralised alternative to acquiring project fund­ ing, which includes benefits such as providing opportunities for skilled ----- talent in economically disadvantaged nations, reduced intermediaries, and codified agreements with auditable terms for the purpose of sup­ porting fair contract executions [109]. Public blockchains provide free protocol infrastructure that allows users to develop platforms and raise funds through initial coin offerings (ICOs), which is similarly compared to the initial public offerings (IPOs) offered in stock markets when pri­ vate companies transition to PLC [110]. However, ICOs have been a target for criminal activity due to their ability to raise funds from anonymous users and lack of regulation checks, such as know your customer (KYC) and anti-money laundering (AML). Hassija, et al., dis­ cussed how the crowdfunding platform, BitFund, allows investors to propose a problem to a public community of programmers and include project-specific parameters such as budget, timescale, use-case, etc., afterwards, the awarding of works is conducted algorithmically through smart contracts to ensure a fair selection process of the development team [109]. **5. Discussion** An exploratory approach was implemented into this review for the purpose of understanding which categories in construction are most influenced by blockchain. This review explored 33 application cate­ gories of blockchain in construction. Each category was substantiated by a minimum of three documents. These categories were further organised into seven subject areas, which include (1) procurement and supply chain, (2) design construction, (3) operations and life cycle, (4) smart cities, (5) intelligent systems, (6) energy and carbon footprint, and (7) decentralised organisations. When assessing the types of data collection used in the reviewed documents (as shown in Fig. 8), synonymous data collection terminologies were merged together for simplicity, such as conceptual frameworks, which included conceptual models and theo­ retical frameworks. Similarly, proof of concepts (PoC) included pilot studies and prototypes. The first three subject areas of this review are sequential stages that occur in a construction project, such as subject area one _procurement and supply chain, which includes implementing_ blockchain in the digital tendering process [111], contract and cashflow management [43], and automated checking of compliance to standards [68]. Subject area two _design and construction, incorporated using_ blockchain for trusted data exchanges [112] and traceability of de­ liverables throughout the supply chain [113]. While subject area three _life cycle and circular economy, included how blockchain can be used as_ part of the assessment and management of a built asset during its operational expenditure stage [114]. Subject area four smart cities, and subject area five _intelligent systems, included a macro-orientated_ approach, assessing how multiple built environment assets and ser­ vices interact through a smart city network, which includes the inter­ operability of various systems such as utility [115], transport [116], Internet of Things (IoT) [117], and smart technologies [88]. Subject area six energy and carbon, focused attention on peer-to-peer energy trading models [118], sustainable technologies for the built environment [119], and carbon accounting strategies [120]. And finally, subject area seven _decentralised organisations, incorporated decentralised autonomous or­_ ganisations (DAO) and decentralised services [103]. DAO is difficult to precisely classify in the current environment, as its definition is dynamic in translation and its development is constantly evolving; however, in construction, many of its activities (for now) overlap with the re­ sponsibilities of a main contractor, therefore, for simplicity, DAO can be described as a decentralised contractor. The aforementioned 33 categories and seven subject areas were not distinctly siloed and included substantial topical crossovers. E.g., the supply chain management category overlapped with all of the subject areas, however, based on the scientometric analysis conducted (as per Fig. 6), it was positioned most quantitatively relative in the procurement subject area, due to its high number of shared link with other categories in that area [121,122]. IoT also strongly overlapped with several subject areas, which include smart cities [7], energy and carbon [123], design and construction [113], procurement [85], and decentralised organi­ sations [79]; however, IoT was placed in the intelligent systems subject area due to its strong correlation with the other categories in this area. The categories electronic document management systems (EDMS) and digital/automated contracts were placed in separate subject areas despite their similarities, as the former is characterised by the digital management of documents on a centralised system, while the latter utilises smart contracts on a decentralised protocol, thus dissimilar systems architecture [124]. A smart medical record system, which includes managing patient records and sharing healthcare data with hospitals, is a category sup­ ported by two authors [15,75]; however, blockchain for healthcare is an entirely different subject area and a vast topic suited for a separate literature review altogether [115]. Health and safety monitoring of site conditions and historic records of on-site accidents were discussed in two documents [79,102]; however, despite its practical applications in construction, it also lacked content for substantiation. Another topic that was excluded despite its interest in two documents is smart governance, which incorporates governmental organisations implementing block­ chain to automate the compliance checking and auditing of built envi­ ronment assets [17,115]. Multi-category applications of blockchain in construction that were not included due to its general-purpose nature include transaction immutability, digital notarisation, decentralised applications (dApps), smart contracts, and information sharing, as effectively, these topics are already integrated within all of the reviewed categories and do not require itemising [102]. As blockchain is a decentralised technology, appropriate incentiv­ isation techniques must be applied to encourage platform interaction through a crypto-economic model [102]. The integration of blockchain in enterprise in the current environment is reliant on dApps harmonising with existing centralised systems, however, as blockchain matures, the transition to complete decentralisation is likely to increase. This assumption is based on assessing the growth and expansion of block­ chain in construction since its emergence in academic literature, and the intensifying global interest in blockchain. In a report regarding impact of blockchain, it was identified as potentially transforming 58 industries globally, which includes the construction industry [125]. Business operations are entirely based on risk management activities, which includes economic risks through investments in new business models, social risk through job losses, legal risk through dispute reso­ lution and corporate liability, environmental risk through sustainability and ecological sensitivity, and technical risk through increased pressure to integrate systems and provide data-driven solutions [85]. Blockchain mitigates against centralised hacks, data manipulation, accounting er­ rors, and provides a foundation for trusted data without reliance on a trusted third party [126]. An area which lacked discussion from the review documents was the integration capabilities of blockchain with existing enterprise systems, as blockchain is considered a high-risk technology due to its decentralised design and lack of standards. Trust is a term that appeared most frequently in the reviewed literature when describing the characteristics of blockchain, such as “stakeholder trust” [122], “peer-to-peer trust” [127], “trust in collaboration” [128], “in­ formation trust” [26], “removal of trusted authority” [11], and “trusted distributed ledger” [129]. Other commonly used terms include trans­ parency, traceability, immutability, security, automation, auditability, decentralisation, and disintermediation [9,118–120,123,130,131]. Over the course of 2017–2020, the rate at which new documents were published on blockchain in construction was recorded at an average of 184%; however, the sample number of years is small, and this level of growth cannot be maintained long-term. A 10-year period would provide a more statistically comprehensive result. Fig. 4 documented the annual expansion of new categories on topic since its emergence in 2017, which displayed six new categories in 2017, nine in 2018, 13 in 2019, followed by five in 2020. It is likely that the expansion of new categories on the topic has almost reached a plateau, therefore, over the next consecutive years, it is envisaged that existing categories will ----- undergo maturity as more attention is focused on testing and developing earlier ideations. **6. Conclusion** New academic documents on blockchain in construction increased at an average of 184% each year since 2017, surmounting to an accumu­ lated total of 121 documents at time of writing this article in 2021. An exploratory approach was implemented to investigate all 121 publica­ tions to examine the contemporary environment of the topic. This re­ view identified 33 application categories, these were organised into seven subject areas and included (1) procurement and supply chain; (2) design and construction; (3) operations and life cycle; (4) smart cities; (5) intelligent systems; (6) energy and carbon footprint; and (7) decentralised organisations. To support the literature review, statistics and scientometrics were incorporated to display the progression of the topical area. This includes visual maps that display the co-occurrences of the categories (as shown in Fig. 6) and data collection types imple­ mented in the reviewed documents (shown in Fig. 9). A complete list of the 121 reviewed documents, along with their category coverage, document type, data collection type, and impact factor, is provided in the shared Google spreadsheet link provided below. [https://docs.google.com/spreadsheets/d/1V4UICRdoyWycaGENH9](https://docs.google.com/spreadsheets/d/1V4UICRdoyWycaGENH9rnuxukRNQJFIArQ-feV7NM0a4/edit?usp=sharing) [rnuxukRNQJFIArQ-feV7NM0a4/edit?usp=sharing](https://docs.google.com/spreadsheets/d/1V4UICRdoyWycaGENH9rnuxukRNQJFIArQ-feV7NM0a4/edit?usp=sharing) Limitations included using only one scientific database, Scopus, due to the inconsistencies which emerged when amalgamating information from various scientific databases for use in visual mapping software. In a comparison of the search results from seven scientific databases and based on the topic of blockchain in construction, Scopus overshadowed its competition by a large margin; furthermore, up to 85% of the doc­ uments indexed in other scientific databases were already existent in Scopus. Another limitation was the restricted capacity to conduct indepth investigation on one particular subject area within the topic, this was due to the exploratory nature of the study, which covered a wide range of application categories. Despite this, the findings provided a solid foundation for aggregating all of the research areas of blockchain in construction in the contemporary environment. Content for this exploratory review was obtained predominantly from documents published from 2017 to 2020, as this article was written in early 2021; however, further work includes an extended review following the progression of the topic over the next consecutive years. **Declaration of Competing Interest** The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. **Acknowledgments** The authors would like to acknowledge his sponsors, University College London (UCL) and Costain PLC, who are joint funding the PhD of the primary author. **Appendix** _Search query one_ Search query one isolated the ISSN numbers of all academic docu­ ments in the subject areas of architecture, building & construction, and civil & structural engineering, followed by specific key words, such as _TITLE-ABS-KEY ("blockchain*" OR "block chain*" OR "distributed ledger*"_ _OR "smart contract*",_ which was inputted into the Scopus search to obtain the results. The exact string of text for query one consists of: _ISSN(08950563 or 17433509 or 23848898 or 23639075 or 20361602_ _or 20299990 or 23352000 or 17246768 or 22150900 or 22150897 or_ _24208213 or 23851546 or 20966717 or 08602395 or 25448870 or_ _19401507 or 19401493 or 02658135 or 0142694X or 23527102 or_ _15417808 or 15417794 or 23520124 or 14356066 or 09349839 or_ _15583066 or 15583058 or 17527589 or 17452007 or 1365232X or_ _09699988 or 15710882 or 17453755 or 14770857 or 14714175 or_ _17481317 or 20952635 or 09603182 or 15731529 or 23635150 or_ _23635142 or 23537396 or 20952430 or 20952449 or 20755309 or_ _1000131X or 10760431 or 19435568 or 07181299 or 07188358 or_ _02632772 or 00038628 or 17589622 or 26316862 or 19387806 or_ _02663511 or 09560599 or 20598033 or 20297947 or 20297955 or_ _22133038 or 2213302X or 19434618 or 15526100 or 21952701 or_ _18864805 or 18877052 or 13001884 or 15224600 or 13472852 or_ _13467581 or 23034521 or 14370980 or 01715445 or 18269745 or_ _22832998 or 1450569X or 22178066 or 22882987 or 12268046 or_ _2239267X or 21753369 or 18818153 or 13404202 or 19461194 or_ _19461186 or 24751448 or 2475143X or 00200883 or 19883234 or_ _22321500 or 18234208 or 22502157 or 22502149 or 18558399 or_ _03536483 or 10067930 or 13602365 or 14664410 or 18285961 or_ _01466518 or 22390243 or 07182309 or 16744764 or 01682601 or_ _22546103 or 11336137 or 18818188 or 13419463 or 00379808 or_ _20612710 or 03055477 or 17496292 or 19895313 or 16952731 or_ _22889930 or 22347224 or 23321091 or 23321121 or 00139661 or_ _15882764 or 23202661 or 00448680 or 13591355 or 14740516 or_ _07187262 or 0718204X or 13028324 or 19346026 or 15499715 or_ _20696469 or 20690509 or 18085741 or 22147233 or 22123202 or_ _00392553 or 00038504 or 20507836 or 20507828 or 10464883 or_ _1531314X or 14665123 or 07380895 or 20455895 or 20455909 or_ _2325159X or 23251581 or 23870346 or 23410531 or 0066622X or_ _13300652 or 0007473X or 20137087 or 14929600 or 02585316 or_ _23251395 or 23251379 or 25317644 or 26117487 or 07160852 or_ _07176996 or 20505833 or 11239247 or 23251662 or 23251670 or_ _23251638 or 2325162X or 21736723 or 20419112 or 20419120 or_ _0951001X or 23413050 or 23412747 or 23090103 or 19360886 or_ _19346832 or 02677768 or 20390491 or 18751490 or 18751504 or_ _15263819 or 16068238 or 01696238 or 23093072 or 23074485 or_ _12282472 or 19357001 or 00038520 or 0003858X or 00038695 or_ _03899160 or 10934421 or 11249064 or 22546332 or 11385596 or_ _21716897 or 21731616 or 23409711 or 23867027 or 23322578 or_ _23322551 or 00012505 or 08950563 or 01932527 or 17433509 or_ _01642006 or 7314906 or 03622479 or 10412336 or 20361602 or_ _20299990 or 23352000 or 24208213 or 23851546 or 20966717 or_ _00088846 or 03062619 or 09589465 or 18736785 or 03605442 or_ _03787788 or 01674730 or 08867798 or 03601323 or 17499518 or_ _17499526 or 09265805 or 07339445 or 01407007 or 0143974X or_ _10900268 or 19435614 or 09500618 or 15452263 or 15452255 or_ _02638231 or 13595997 or 18716873 or 10840702 or 19435592 or_ _19401507 or 19401493 or 09056947 or 16000668 or 09613218 or_ _14664321 or 1570761X or 03931420 or 17448980 or 15732479 or_ _07339364 or 19435533 or 08991561 or 13632469 or 1559808X or_ _12299367 or 19968744 or 19963599 or 1751763X or 00249831 or_ _23527102 or 14644177 or 17517648 or 08893241 or 01446193 or_ _1466433X or 15417808 or 15417794 or 19435509 or 08873828 or_ _23520124 or 16713664 or 1993503X or 02194554 or 09517197 or_ _17517605 or 0889325X or 17527589 or 17452007 or 13694332 or_ _1365232X or 09699988 or 20962754 or 24679674 or 14371006 or_ _00059900 or 12254568 or 23767642 or 00138029 or 14006529 or_ _14036835 or 19883226 or 04652746 or 20714726 or 20710305 or_ _14770857 or 14714175 or 2374474X or 23744731 or 2041420X or_ _20414196 or 00056650 or 2287531X or 22875301 or 15623599 or_ _20952635 or 13468014 or 13473913 or 22049029 or 14770849 or_ _01436244 or 1816112X or 10002383 or 23635150 or 23635142 or_ _23537396 or 20755309 or 17550750 or 17550769 or 22973362 or_ _1000131X or 09650911 or 17517702 or 19375247 or 19375255 or_ _10760431 or 19435568 or 23984708 or 17512549 or 17562201 or_ _24705322 or 24705314 or 10006869 or 17442591 or 02632772 or_ _21628246 or 01430750 or 12266116 or 20369913 or 2533168X or_ ----- _18670520 or 18670539 or 23644176 or 23644184 or 02630923 or_ _20484046 or 14613484 or 08879672 or 15503984 or 14733315 or_ _14371049 or 00389145 or 10168664 or 22143998 or 03405044 or_ _09328351 or 14370999 or 17517664 or 14784637 or 22135812 or_ _22135820 or 17595916 or 17595908 or 18236499 or 21804222 or_ _1351010X or 20598025 or 10034722 or 1028365X or 19969015 or_ _21862990 or 21862982 or 17566932 or 17508975 or 18748368 or_ _2214398X or 02663511 or 09560599 or 20598033 or 15551369 or_ _1822427X or 18224288 or 09766308 or 09766316 or 14780771 or_ _15732487 or 17448999 or 22133038 or 2213302X or 19434618 or_ _15526100 or 22110844 or 22110852 or 20466102 or 20466099 or_ _22132031 or 2213204X or 10079629 or 24123811 or 14258129 or_ _13472852 or 13467581 or 20569459 or 20569467 or 10840680 or_ _14370980 or 01715445 or 21487847 or 20937628 or 2093761X or_ _16876261 or 1687627X or 07177925 or 18764037 or 18764029 or_ _15744078 or 00467316 or 00012491 or 03649962 or 07362501 or_ _21688710 or 19302991 or 19302983 or 17329353 or 18818153 or_ _13404202 or 19461194 or 19461186 or 16719379 or 22488723 or_ _01205609 or 13556207 or 00200883 or 19883234 or 22321500 or_ _18234208 or 16718879 or 22502157 or 22502149 or 10067930 or_ _12101389 or 01466518 or 07162952 or 07185073 or 22390243 or_ _12295515 or 22342842 or 19853807 or 16744764 or 15460118 or_ _01998595 or 00194565 or 18818188 or 13419463 or 26007959 or_ _21803242 or 09700137 or 22889930 or 22347224 or 07224397 or_ _09490205 or 07224400 or 03734331 or 20696469 or 20690509 or_ _22147233 or 22123202 or 00060208 or 02446014 or 00392553 or_ _08919976 or 14665123 or 03552705 or 13003453 or 03002721 or_ _03410552 or 11790776 or 20137087 or 00333840 or 03731995 or_ _00105317 or 09745904 or 15280187 or 09502289 or 15337316 or_ _00105333 or 02677768 or 00209384 or 23793244 or 23793252 or_ _10450343 or 09698213 or 08919526 or 09621784 or 0267825X or_ _01426168 or 15274055 or 0306400X or 25673742 or 13693999 or_ _20541236 or 02702932 or 18672442 or 08950563 or 10459065 or_ _01932527 or 17433509 or 08938717 or 01610589 or 13395629 or_ _21875103 or 20299990 or 23352000 or 25901982 or 20966717 or_ _03062619 or 0968090X or 00431354 or 18792448 or 01912615 or_ _15265447 or 00411655 or 13665545 or 08883270 or 10961216 or_ _18736785 or 03605442 or 07232632 or 09658564 or 03787788 or_ _01674730 or 12086010 or 00083674 or 14678667 or 10939687 or_ _03601323 or 00494488 or 15729435 or 17499518 or 17499526 or_ _02638223 or 09265805 or 0734743X or 13619209 or 07339445 or_ _18737323 or 01410296 or 00457949 or 0143974X or 10900268 or_ _19435614 or 15568334 or 15568318 or 09500618 or 15452263 or_ _15452255 or 00207403 or 02638231 or 13595997 or 18716873 or_ _22106707 or 10840702 or 19435592 or 20460430 or 20460449 or_ _02677261 or 09613218 or 14664321 or 1570761X or 13698478 or_ _02668920 or 18784275 or 22143912 or 09641726 or 1361665X or_ _07339496 or 17448980 or 15732479 or 00380806 or 07339364 or_ _19435533 or 08991561 or 13632469 or 1559808X or 09204741 or_ _15731650 or 12299367 or 01676105 or 14680629 or 21647402 or_ _18142079 or 00221686 or 08873801 or 16449665 or 21905479 or_ _21905452 or 1751763X or 00249831 or 07339429 or 19437900 or_ _23527102 or 07339453 or 14644177 or 17517648 or 08893241 or_ _15376532 or 15210596 or 15376494 or 00936413 or 10001964 or_ _15706443 or 15417808 or 15417794 or 05785634 or 17936292 or_ _19435509 or 08873828 or 23520124 or 19971400 or 19966814 or_ _15230406 or 14356066 or 09349839 or 1477268X or 10298436 or_ _19435460 or 0733950X or 23253444 or 23267186 or 22341315 or_ _19760485 or 11749857 or 22150986 or 16713664 or 1993503X or_ _21996679 or 21996687 or 10006915 or 07339437 or 19434774 or_ _02194554 or 0889325X or 13694332 or 19435584 or 10840699 or_ _1365232X or 09699988 or 24680672 or 20962754 or 24679674 or_ _07339488 or 00224502 or 12254568 or 10760342 or 1943555X or_ _15276988 or 19475411 or 1947542X or 23767642 or 10007598 or_ _00138029 or 14006529 or 14036835 or 20926219 or 2005307X or_ _14647141 or 14651734 or 21999260 or 21999279 or 16742370 or_ _22886605 or 22886613 or 20714726 or 20710305 or 10693629 or_ _14770857 or 14714175 or 23350164 or 03542025 or 17481317 or_ _20957564 or 00056650 or 21964386 or 21964378 or 2287531X or_ _22875301 or 09203796 or 13894420 or 21167214 or 19648189 or_ _20771312 or 18223605 or 13923730 or 03611981 or 14488353 or_ _1816112X or 10286608 or 10290249 or 17522706 or 00396265 or_ _23635150 or 23635142 or 07339402 or 0733947X or 23537396 or_ _20952430 or 20952449 or 20755309 or 12267988 or 19763808 or_ _24732893 or 24732907 or 15397742 or 15397734 or 1000131X or_ _15361055 or 10523928 or 16797825 or 16797817 or 09650911 or_ _17517702 or 19375247 or 19375255 or 10760431 or 19435568 or_ _20796439 or 23984708 or 08931321 or 17517710 or 0965092X or_ _19491190 or 19491204 or 10709622 or 18759203 or 20927614 or_ _20927622 or 13287982 or 24705322 or 24705314 or 10017372 or_ _17350522 or 10006869 or 22918752 or 22918744 or 07339372 or_ _19437870 or 17476526 or 17476534 or 23791357 or 21653984 or_ _05536626 or 15873773 or 25735438 or 19969465 or 19969457 or_ _12266116 or 20936311 or 15982351 or 2044124X or 20441258 or_ _20369913 or 2533168X or 18670520 or 18670539 or 20950349 or_ _23644176 or 23644184 or 09715010 or 21643040 or 02630923 or_ _20484046 or 14613484 or 08879672 or 2195268X or 21952698 or_ _14784629 or 17517680 or 14733315 or 03151468 or 12086029 or_ _18213197 or 14514117 or 14371049 or 00389145 or 10168664 or_ _22143998 or 21532648 or 03405044 or 09328351 or 14370999 or_ _17517664 or 14784637 or 13354205 or 25857878 or 10535381 or_ _10044523 or 22342184 or 22342192 or 17476518 or 1747650X or_ _10375783 or 17579872 or 17579864 or 18236499 or 21804222 or_ _03535320 or 16431618 or 2449769X or 20083556 or 20086695 or_ _21967202 or 21967210 or 1028365X or 19969015 or 02663511 or_ _09560599 or 20598033 or 15551369 or 1822427X or 18224288 or_ _18657362 or 18657389 or 16878086 or 16878094 or 09766308 or_ _09766316 or 22286160 or 12302945 or 16480627 or 18224202 or_ _22133038 or 2213302X or 2191916X or 21919151 or 17587328 or_ _17587336 or 10096582 or 19434618 or 15526100 or 01376365 or_ _16879724 or 16879732 or 22110844 or 22110852 or 17881994 or_ _16711637 or 20466102 or 20466099 or 22132031 or 2213204X or_ _10263098 or 23831359 or 23832525 or 10079629 or 00025968 or_ _03502465 or 13339095 or 23144912 or 23144904 or 17517672 or_ _0965089X or 14513749 or 18207863 or 18676944 or 18676936 or_ _24123811 or 18741495 or 20588305 or 20588313 or 13472852 or_ _13467581 or 17550807 or 17550793 or 19434162 or 19434170 or_ _15630854 or 10840680 or 13365835 or 21996512 or 17550785 or_ _17550777 or 09650903 or 17517699 or 21646457 or 21646473 or_ _18245463 or 23915439 or 18632351 or 17514312 or 17514304 or_ _18713033 or 20421338 or 20421346 or 2226809X or 22235329 or_ _23539003 or 17344492 or 07177925 or 15744078 or 00467316 or_ _19930461 or 2225157X or 2229838X or 19302991 or 19302983 or_ _01878336 or 20072422 or 17329353 or 17386225 or 22882235 or_ _16719379 or 10212019 or 00200883 or 19883234 or 17085284 or_ _1000582X or 23746793 or 22502157 or 22502149 or 1006754X or_ _21928253 or 07162952 or 07185073 or 17400694 or 18753507 or_ _17568404 or 13681494 or 12295515 or 22342842 or 18029876 or_ _16744764 or 00465828 or 19918747 or 22243429 or 14439255 or_ _00194565 or 14412713 or 18022308 or 26007959 or 21803242 or_ _09700137 or 22889930 or 22347224 or 23321091 or 23321121 or_ _22926062 or 23361182 or 1802680X or 16106199 or 13028324 or_ _00060208 or 12104027 or 18052576 or 00174653 or 01580728 or_ _00392553 or 09568700 or 20449283 or 02755823 or 14665123 or_ _02126389 or 0923666X or 10093443 or 10155856 or 13003453 or_ _03002721 or 00137308 or 12313726 or 21959870 or 21959862 or_ _0376723X or 20817738 or 23007591 or 18825974 or 13443755 or_ _00333840 or 17593433 or 23662565 or 23662557 or 03731995 or_ _00105317 or 09745904 or 17452058 or 00124419 or 22783075 or_ _23793244 or 23793252 or 00263982 or 01665766 or 00333735 or_ _00348619 or 16954408 or 0149337X or 09698213 or 00097853 or_ _08857024 or 03600556 or 08919526 or 0267825X or 01426168 or_ ----- _00284939 or 22113444 or 13693999 or 20541236) AND TITLE-ABS-KEY_ _("blockchain*" OR "block chain*" OR "distributed ledger*" OR "smart_ _contract*")_ Search query two: Search query two used a simpler method, which included using one of the predefine subject areas available on Scopus, followed by specific key words. The limitation to using this search query is the high number irrelevant documents that accompany the results. The string of text for query to consists of: _SUBJAREA(ENGI)_ _AND_ _TITLE-ABS-KEY("Blockchain"_ _AND_ _"Construction")._ **References** [[1] I. Grigg, Triple Entry Accounting. https://nakamotoinstitute.org/triple-entry-acc](https://nakamotoinstitute.org/triple-entry-accounting/) [ounting/, 2005 accessed 6th February 2020.](https://nakamotoinstitute.org/triple-entry-accounting/) [2] R.N. Chen, Y.N. Li, Y. Yu, H.L. Li, X.F. Chen, W. Susilo, Blockchain-based dynamic provable data possession for smart cities, Inst. Electr. Electron. Eng. Internet [Things J. 7 (5) (2020) 4143–4154, https://doi.org/10.1109/jiot.2019.2963789.](https://doi.org/10.1109/jiot.2019.2963789) [3] M. Foti, C. Mavromatis, M. Vavalis, Decentralized blockchain-based consensus for [optimal power flow solutions, Appl. Energy 283 (2020), https://doi.org/](https://doi.org/10.1016/j.apenergy.2020.116100) [10.1016/j.apenergy.2020.116100.](https://doi.org/10.1016/j.apenergy.2020.116100) [4] Y. Wang, C.H. Chen, A. Zghari-Sales, Designing a blockchain enabled supply [chain, Int. J. Prod. Res. 59 (5) (2020) 1450–1475, https://doi.org/10.1080/](https://doi.org/10.1080/00207543.2020.1824086) [00207543.2020.1824086.](https://doi.org/10.1080/00207543.2020.1824086) [5] S. Karale, V. Ranaware, Applications of blockchain technology in smart city development: a research, Int. J. Innov. Technol. Explor. Eng. 8 (11) (2019) [556–559, https://doi.org/10.35940/ijitee.K1093.09811S19.](https://doi.org/10.35940/ijitee.K1093.09811S19) [6] M. Hribernik, K. Zero, S. Kummer, D.M. Herold, City logistics: towards a blockchain decision framework for collaborative parcel deliveries in micro-hubs, [Transp. Res. Interdisc. Perspect. 8 (2020), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.trip.2020.100274) [trip.2020.100274.](https://doi.org/10.1016/j.trip.2020.100274) [7] S. Perera, S. Nanayakkara, M.N.N. Rodrigo, S. Senaratne, R. Weinand, Blockchain technology: is it hype or real in the construction industry? J. Ind. Inf. Integr. 17 [(2020) https://doi.org/10.1016/j.jii.2020.100125.](https://doi.org/10.1016/j.jii.2020.100125) [8] D. Han, C. Zhang, J. Ping, Z. Yan, Smart contract architecture for decentralized energy trading and management based on blockchains, Energy 199 (2020), [https://doi.org/10.1016/j.energy.2020.117417.](https://doi.org/10.1016/j.energy.2020.117417) [9] X. Ye, M. Konig, Framework for automated billing in the construction industry ¨ using BIM and smart contracts, in: Lecture Notes in Civil Engineering Vol. 98, [Springer, S˜ao Paulo, Brazil, 2021, pp. 824–838, https://doi.org/10.1007/978-3-](https://doi.org/10.1007/978-3-030-51295-8_57) [030-51295-8_57.](https://doi.org/10.1007/978-3-030-51295-8_57) [[10] Paperpile, The Top List of Academic Research Databases. https://paperpile.](https://paperpile.com/g/academic-research-databases/) [com/g/academic-research-databases/, 2021 accessed 23rd May 2021.](https://paperpile.com/g/academic-research-databases/) [11] Z. Turk, R. Klinc, Potentials of blockchain technology for construction [ˇ] management, in: Creative Construction Conference (CCC) 2017 Vol. 196, Elsevier [Ltd, Primosten, Croatia, 2017, pp. 638–645, https://doi.org/10.1016/j.](https://doi.org/10.1016/j.proeng.2017.08.052) [proeng.2017.08.052.](https://doi.org/10.1016/j.proeng.2017.08.052) [12] R. Coyne, T. Onabolu, Blockchain for architects: challenges from the sharing [economy, Archit. Res. Q. 21 (4, 2017) 369–374, https://doi.org/10.1017/](https://doi.org/10.1017/S1359135518000167) [S1359135518000167.](https://doi.org/10.1017/S1359135518000167) [13] J.J. Sikorski, J. Haughton, M. Kraft, Blockchain technology in the chemical industry: machine-to-machine electricity market, Appl. Energy 195 (2017) [234–246, https://doi.org/10.1016/j.apenergy.2017.03.039.](https://doi.org/10.1016/j.apenergy.2017.03.039) [14] M.S. Kiu, F.C. Chia, P.F. Wong, Exploring the potentials of blockchain application in construction industry: a systematic review, Int. J. Constr. Manag. (2020), [https://doi.org/10.1080/15623599.2020.1833436.](https://doi.org/10.1080/15623599.2020.1833436) [15] B. Bhushan, A. Khamparia, K.M. Sagayam, S.K. Sharma, M.A. Ahad, N. C. Debnath, Blockchain for smart cities: a review of architectures, integration [trends and future research directions, Sustain. Cities Soc. 61 (2020), https://doi.](https://doi.org/10.1016/j.scs.2020.102360) [org/10.1016/j.scs.2020.102360.](https://doi.org/10.1016/j.scs.2020.102360) [16] J.J. Hunhevicz, D.M. Hall, Managing mistrust in construction using DLT: a review of use-case categories for technical design decisions, in: European Conference on [Computing in Construction, Crete, Greece, 2019, pp. 100–109, https://doi.org/](https://doi.org/10.35490/EC3.2019.171) [10.35490/EC3.2019.171.](https://doi.org/10.35490/EC3.2019.171) [17] J. Li, D. Greenwood, M. Kassem, Blockchain in the built environment and construction industry: a systematic review, conceptual models and practical use [cases, Autom. Constr. 102 (2019) 288–307, https://doi.org/10.1016/j.](https://doi.org/10.1016/j.autcon.2019.02.005) [autcon.2019.02.005.](https://doi.org/10.1016/j.autcon.2019.02.005) [18] R. Yang, R. Wakefield, S. Lyu, S. Jayasuriya, F. Han, X. Yi, X. Yang, G. Amarasinghe, S. Chen, Public and private blockchain in construction business [process and information integration, Autom. Constr. 118 (2020), https://doi.org/](https://doi.org/10.1016/j.autcon.2020.103276) [10.1016/j.autcon.2020.103276.](https://doi.org/10.1016/j.autcon.2020.103276) [19] J.F. Burnham, Scopus database: a review, Biomed. Digit. Libr. 3 (1, 2006) 1–8, [https://doi.org/10.1186/1742-5581-3-1.](https://doi.org/10.1186/1742-5581-3-1) [20] A. Aghaei Chadegani, H. Salehi, M.M. Md Yunus, H. Farhadi, M. Fooladi, M. Farhadi, N. Ale Ebrahim, A comparison between two main academic literature collections: Web of Science and Scopus databases, Asian Soc. Sci. 9 (5, 2013) [18–26, https://doi.org/10.5539/ass.v9n5p18.](https://doi.org/10.5539/ass.v9n5p18) [[21] SCImago, Scimago Journal & Country Rank. https://www.scimagojr.com/](https://www.scimagojr.com/) accessed 24th May 2021. [22] ScimagoResearchGroup, Description of Scimago Journal Rank Indicator, Scimago [Research Group, 2007. https://www.scimagojr.com/SCImagoJournalRank.pdf.](https://www.scimagojr.com/SCImagoJournalRank.pdf) [23] G.B. Ozturk, Interoperability in building information modeling for AECO/FM [industry, Autom. Constr. 113 (2020), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.autcon.2020.103122) [autcon.2020.103122.](https://doi.org/10.1016/j.autcon.2020.103122) [24] D. Lopez, B. Farooq, A multi-layered blockchain framework for smart mobility ´ [data-markets, Transp. Res. C Emerg. Technol. 111 (2020) 588–615, https://doi.](https://doi.org/10.1016/j.trc.2020.01.002) [org/10.1016/j.trc.2020.01.002.](https://doi.org/10.1016/j.trc.2020.01.002) [25] H. Zhang, J. Wang, Y. Ding, Blockchain-based decentralized and secure keyless [signature scheme for smart grid, Energy 180 (2019) 955–967, https://doi.org/](https://doi.org/10.1016/j.energy.2019.05.127) [10.1016/j.energy.2019.05.127.](https://doi.org/10.1016/j.energy.2019.05.127) [26] W.N. Suliyanti, R.F. Sari, Blockchain-based building information modeling, in: 2nd International Conference on Applied Engineering (ICAE), Institute of [Electrical and Electronics Engineers Inc., Batam, Indonesia, 2019, https://doi.](https://doi.org/10.1109/ICAE47758.2019.9221744) [org/10.1109/ICAE47758.2019.9221744.](https://doi.org/10.1109/ICAE47758.2019.9221744) [27] K.N. Khaqqi, J.J. Sikorski, K. Hadinoto, M. Kraft, Incorporating seller/buyer reputation-based system in blockchain-enabled emission trading application, [Appl. Energy 209 (2018) 8–19, https://doi.org/10.1016/j.](https://doi.org/10.1016/j.apenergy.2017.10.070) [apenergy.2017.10.070.](https://doi.org/10.1016/j.apenergy.2017.10.070) [28] F. Xiong, R. Xiao, W. Ren, R. Zheng, J. Jiang, A key protection scheme based on secret sharing for blockchain-based construction supply chain system, Inst. Electr. [Electron. Eng. 7 (2019) 126773–126786, https://doi.org/10.1109/](https://doi.org/10.1109/ACCESS.2019.2937917) [ACCESS.2019.2937917.](https://doi.org/10.1109/ACCESS.2019.2937917) [29] S.K. Singh, Y.S. Jeong, J.H. Park, A deep learning-based IoT-oriented [infrastructure for secure smart City, Sustain. Cities Soc. 60 (2020), https://doi.](https://doi.org/10.1016/j.scs.2020.102252) [org/10.1016/j.scs.2020.102252.](https://doi.org/10.1016/j.scs.2020.102252) [30] S. Sun, X. Zheng, J. Villalba-Díez, J. Ordieres-Mer´e, Data handling in industry 4.0: interoperability based on distributed ledger technology, Sensors (Switzerland) 20 [(11) (2020), https://doi.org/10.3390/s20113046.](https://doi.org/10.3390/s20113046) [31] R. Talat, M. Muzammal, Q. Qu, W. Zhou, M. Najam-Ul-Islam, S.M.H. Bamakan, J. Qiu, A decentralized system for green energy distribution in a smart grid, [J. Energy Eng. 146 (1) (2020), https://doi.org/10.1061/(ASCE)EY.1943-](https://doi.org/10.1061/(ASCE)EY.1943-7897.0000623) [7897.0000623.](https://doi.org/10.1061/(ASCE)EY.1943-7897.0000623) [32] W. Hu, Y.W. Hu, W.H. Yao, W.Q. Lu, H.H. Li, Z.W. Lv, A blockchain-based smart contract trading mechanism for energy power supply and demand network, Adv. [Prod. Eng. Manag. 14 (3) (2019) 284–296, https://doi.org/10.14743/](https://doi.org/10.14743/apem2019.3.328) [apem2019.3.328.](https://doi.org/10.14743/apem2019.3.328) [33] A. Shojaei, I. Flood, H.I. Moud, M. Hatami, X. Zhang, An implementation of smart contracts by integrating BIM and blockchain, in: 4th Future Technologies Conference (FTC) 2019 Vol. 1069, Springer, San Francisco, USA, 2020, [pp. 519–527, https://doi.org/10.1007/978-3-030-32523-7_36.](https://doi.org/10.1007/978-3-030-32523-7_36) [34] E. Mengelkamp, J. G¨arttner, K. Rock, S. Kessler, L. Orsini, C. Weinhardt, Designing microgrid energy markets: a case study: the brooklyn microgrid, Appl. [Energy 210 (2018) 870–880, https://doi.org/10.1016/j.apenergy.2017.06.054.](https://doi.org/10.1016/j.apenergy.2017.06.054) [35] L. Wang, J. Liu, R. Yuan, J. Wu, D. Zhang, Y. Zhang, M. Li, Adaptive bidding strategy for real-time energy management in multi-energy market enhanced by [blockchain, Appl. Energy 279 (2020), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.apenergy.2020.115866) [apenergy.2020.115866.](https://doi.org/10.1016/j.apenergy.2020.115866) [36] K. Kim, G. Lee, S. Kim, A study on the application of blockchain technology in the [construction industry, KSCE J. Civ. Eng. 24 (9, 2020) 2561–2571, https://doi.](https://doi.org/10.1007/s12205-020-0188-x) [org/10.1007/s12205-020-0188-x.](https://doi.org/10.1007/s12205-020-0188-x) [37] S. Fong Piew, S. Sarip, A.Y. AbdFatah, H.M. Kaidi, Business sustainability through technology management approach for construction company, Int. J. Emerg. [Trends Eng. Res. 8 (1) (2020) 22–26, https://doi.org/10.30534/ijeter/2020/](https://doi.org/10.30534/ijeter/2020/0481.12020) [0481.12020.](https://doi.org/10.30534/ijeter/2020/0481.12020) [38] U. Isikdag, An evaluation of barriers to E-Procurement in Turkish construction [industry, Int. J. Innov. Technol. Explor. Eng. 8 (4, 2019) 252–259. https://www.](https://www.ijitee.org/wp-content/uploads/papers/v8i4/D2731028419.pdf) [ijitee.org/wp-content/uploads/papers/v8i4/D2731028419.pdf.](https://www.ijitee.org/wp-content/uploads/papers/v8i4/D2731028419.pdf) [39] D. Kifokeris, C. Koch, Blockchain in building logistics: emerging knowledge, and related actors in Sweden, in: 35th Annual Conference on Association of Researchers in Construction Management, Association of Researchers in Construction Management (ARCOM), Gothenburg, Sweden, 2019, pp. 426–435. [https://www.scopus.com/inward/record.uri?eid=2-s2.0-85077126273&partnerI](https://www.scopus.com/inward/record.uri?eid=2-s2.0-85077126273&amp;partnerID=40&amp;md5=e54aeb29511b92b64daa353918869523) [D=40&md5=e54aeb29511b92b64daa353918869523.](https://www.scopus.com/inward/record.uri?eid=2-s2.0-85077126273&amp;partnerID=40&amp;md5=e54aeb29511b92b64daa353918869523) [40] A. Lanko, N. Vatin, A. Kaklauskas, Application of RFID combined with blockchain technology in logistics of construction materials, in: 2017 International Science Conference on Business Technologies for Sustainable Urban Development Vol. [170, EDP Sciences, St. Petersburg, Russia, 2018, https://doi.org/10.1051/](https://doi.org/10.1051/matecconf/201817003032) [matecconf/201817003032.](https://doi.org/10.1051/matecconf/201817003032) [41] A. Sivula, A. Shamsuzzoha, P. Helo, Blockchain in logistics: mapping the opportunities in con-struction industry, in: 3rd North American Industrial Engineering and Operations Management (IEOM) Conference, 2018, IEOM [Society, Washington DC, USA, 2018, pp. 1954–1960. https://www.scopus.com/](https://www.scopus.com/record/display.uri?eid=2-s2.0-85065252883&amp;origin=inward&amp;txGid=bbac08638afd88dbff5f6ba202ab4a6e) [record/display.uri?eid=2-s2.0-85065252883&origin=inward&txGid=bbac08](https://www.scopus.com/record/display.uri?eid=2-s2.0-85065252883&amp;origin=inward&amp;txGid=bbac08638afd88dbff5f6ba202ab4a6e) [638afd88dbff5f6ba202ab4a6e.](https://www.scopus.com/record/display.uri?eid=2-s2.0-85065252883&amp;origin=inward&amp;txGid=bbac08638afd88dbff5f6ba202ab4a6e) [42] H.Y. Chong, A. Diamantopoulos, Integrating advanced technologies to uphold [security of payment: data flow diagram, Autom. Constr. 114 (2020), https://doi.](https://doi.org/10.1016/j.autcon.2020.103158) [org/10.1016/j.autcon.2020.103158.](https://doi.org/10.1016/j.autcon.2020.103158) [43] S. Ahmadisheykhsarmast, R. Sonmez, A smart contract system for security of [payment of construction contracts, Autom. Constr. 120 (2020), https://doi.org/](https://doi.org/10.1016/j.autcon.2020.103401) [10.1016/j.autcon.2020.103401.](https://doi.org/10.1016/j.autcon.2020.103401) [44] S.L. Gruneberg, G.J. Ive, The Economics of the Modern Construction Firm, [Macmillan Press Ltd, Hampshire, UK, 2000, ISBN 978-0-230-51043-2, https://](https://doi.org/10.1057/9780230510432) [doi.org/10.1057/9780230510432.](https://doi.org/10.1057/9780230510432) ----- [45] M. Das, H. Luo, J.C.P. Cheng, Securing interim payments in construction projects [through a blockchain-based framework, Autom. Constr. 118 (2020), https://doi.](https://doi.org/10.1016/j.autcon.2020.103284) [org/10.1016/j.autcon.2020.103284.](https://doi.org/10.1016/j.autcon.2020.103284) [46] A.J. McNamara, S.M.E. Sepasgozar, Developing a theoretical framework for [intelligent contract acceptance, Constr. Innov. 20 (3) (2020) 421–445, https://](https://doi.org/10.1108/CI-07-2019-0061) [doi.org/10.1108/CI-07-2019-0061.](https://doi.org/10.1108/CI-07-2019-0061) [47] S. Badi, E. Ochieng, M. Nasaj, M. Papadaki, Technological, organisational and environmental determinants of smart contracts adoption: UK construction sector [viewpoint, Constr. Manag. Econ. 39 (1) (2021) 36–54, https://doi.org/10.1080/](https://doi.org/10.1080/01446193.2020.1819549) [01446193.2020.1819549.](https://doi.org/10.1080/01446193.2020.1819549) [48] J. Hunhevicz, T. Schraner, D. Hall, Incentivizing high-quality data sets in construction using blockchain: a feasibility study in the Swiss industry, in: 37th International Symposium on Automation and Robotics in Construction (ISARC), [2020, https://doi.org/10.22260/ISARC2020/0177.](https://doi.org/10.22260/ISARC2020/0177) [49] X. Qian, E. Papadonikolaki, Shifting trust in construction supply chains through blockchain technology, Eng. Constr. Archit. Manag. 28 (2, 2020) 584–602, [https://doi.org/10.1108/ECAM-12-2019-0676.](https://doi.org/10.1108/ECAM-12-2019-0676) [50] D. Sheng, L. Ding, B. Zhong, P.E.D. Love, H. Luo, J. Chen, Construction quality [information management with blockchains, Autom. Constr. 120 (2020), https://](https://doi.org/10.1016/j.autcon.2020.103373) [doi.org/10.1016/j.autcon.2020.103373.](https://doi.org/10.1016/j.autcon.2020.103373) [51] P. Dutta, T.M. Choi, S. Somani, R. Butala, Blockchain technology in supply chain operations: applications, challenges and research opportunities, Transp. Res. E [Logist. Transp. Rev. 142 (2020), https://doi.org/10.1016/j.tre.2020.102067.](https://doi.org/10.1016/j.tre.2020.102067) [52] N.O. Nawari, S. Ravindran, Blockchain and the built environment: potentials and [limitations, J. Build. Eng. 25 (2019), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.jobe.2019.100832) [jobe.2019.100832.](https://doi.org/10.1016/j.jobe.2019.100832) [53] N.O. Nawari, S. Ravindran, Blockchain and Building Information Modeling (BIM): [review and applications in post-disaster recovery, Buildings 9 (6) (2019), https://](https://doi.org/10.3390/BUILDINGS9060149) [doi.org/10.3390/BUILDINGS9060149.](https://doi.org/10.3390/BUILDINGS9060149) [54] A. Adibfar, A. Costin, R.R.A. Issa, Design copyright in architecture, engineering, and construction industry: review of history, pitfalls, and lessons learned, J. Leg. [Aff. Disput. Resolut. Eng. Constr. 12 (3) (2020), https://doi.org/10.1061/(ASCE)](https://doi.org/10.1061/(ASCE)LA.1943-4170.0000421) [LA.1943-4170.0000421.](https://doi.org/10.1061/(ASCE)LA.1943-4170.0000421) [55] P.K. Wan, L. Huang, H. Holtskog, Blockchain-enabled information sharing within a supply chain: a systematic literature review, Inst. Electr. Electron. Eng. 8 (2020) [49645–49656, https://doi.org/10.1109/ACCESS.2020.2980142.](https://doi.org/10.1109/ACCESS.2020.2980142) [56] F. Xue, W. Lu, A semantic differential transaction approach to minimizing information redundancy for BIM and blockchain integration, Autom. Constr. 118 [(2020), https://doi.org/10.1016/j.autcon.2020.103270.](https://doi.org/10.1016/j.autcon.2020.103270) [57] R. Zheng, J. Jiang, X. Hao, W. Ren, F. Xiong, Y. Ren, BcBIM: a blockchain-based big data model for BIM modification audit and provenance in mobile cloud, Math. [Probl. Eng. (2019), https://doi.org/10.1155/2019/5349538.](https://doi.org/10.1155/2019/5349538) [58] J. Mason, BIM Fork: are smart contracts in construction more likely to prosper with or without BIM? J. Leg. Aff. Disput. Resolut. Eng. Constr. 11 (4) (2019) [https://doi.org/10.1061/(ASCE)LA.1943-4170.0000316.](https://doi.org/10.1061/(ASCE)LA.1943-4170.0000316) [59] O.V. Bukunova, A.S. Bukunov, Tools of data transmission at building information modeling, in: 2019 International Science and Technology Conference Institute of [Electrical and Electronics Engineers Inc., Vladivostok, Russia, 2019, https://doi.](https://doi.org/10.1109/Eastonf.2019.8725373) [org/10.1109/Eastonf.2019.8725373.](https://doi.org/10.1109/Eastonf.2019.8725373) [60] X. Ye, K. Sigalov, M. Konig, Integrating BIM- and cost-included information ¨ container with Blockchain for construction automated payment using billing model and smart contracts, in: Proceedings of 37th International Symposium on Automation and Robotics in Construction (ISARC) 2020, International Association for Automation and Robotics in Construction (IAARC), Kitakyushu, [Japan, 2020, pp. 1388–1395, https://doi.org/10.22260/ISARC2020/0192.](https://doi.org/10.22260/ISARC2020/0192) [61] F. Elghaish, S. Abrishami, M.R. Hosseini, Integrated project delivery with [blockchain: an automated financial system, Autom. Constr. 114 (2020), https://](https://doi.org/10.1016/j.autcon.2020.103182) [doi.org/10.1016/j.autcon.2020.103182.](https://doi.org/10.1016/j.autcon.2020.103182) [62] J.J. Hunhevicz, B. Pierre-Antoine, M.M.M. Bonanomi, D.M. Hall, Blockchain and smart contracts for Integrated Project Delivery: inspiration from the commons, in: Proceedings of 18th Annual Engineering Project Organization Conference (EPOC), Engineering Project Organization Society (EPOS), Virtual due to Covid[19, 2020, https://doi.org/10.3929/ethz-b-000452056.](https://doi.org/10.3929/ethz-b-000452056) [63] M. Wang, C.C. Wang, S. Sepasgozar, S. Zlatanova, A systematic review of digital technology adoption in off-site construction: current status and future direction [towards industry 4.0, Buildings 10 (11, 2020) 1–29, https://doi.org/10.3390/](https://doi.org/10.3390/buildings10110204) [buildings10110204.](https://doi.org/10.3390/buildings10110204) [64] Z. Wang, T. Wang, H. Hu, J. Gong, X. Ren, Q. Xiao, Blockchain-based framework for improving supply chain traceability and information sharing in precast [construction, Autom. Constr. 111 (2020), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.autcon.2019.103063) [autcon.2019.103063.](https://doi.org/10.1016/j.autcon.2019.103063) [65] C.S. Tang, L.P. Veelenturf, The strategic role of logistics in the industry 4.0 era, [Transp. Res. E Log. Transp. Rev. 129 (2019) 1–11, https://doi.org/10.1016/j.](https://doi.org/10.1016/j.tre.2019.06.004) [tre.2019.06.004.](https://doi.org/10.1016/j.tre.2019.06.004) [66] M.A. Ahad, S. Paiva, G. Tripathi, N. Feroz, Enabling technologies and sustainable [smart cities, Sustain. Cities Soc. 61 (2020), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.scs.2020.102301) [scs.2020.102301.](https://doi.org/10.1016/j.scs.2020.102301) [67] S. Copeland, M. Bilec, Buildings as material banks using RFID and building information modeling in a circular economy, in: 27th CIRP Life Cycle Engineering [Conference Vol. 90, Elsevier, 2020, pp. 143–147, https://doi.org/10.1016/j.](https://doi.org/10.1016/j.procir.2020.02.122) [procir.2020.02.122.](https://doi.org/10.1016/j.procir.2020.02.122) [68] J. Li, M. Kassem, R. Watson, A blockchain and smart contract-based framework to increase traceability of built assets, in: Proc. 37th CIB W78 Information Technology for Construction Conference, S˜ao Paulo, Brazil, 2020, pp. 347–362, [https://doi.org/10.46421/2706-6568.37.2020.paper025.](https://doi.org/10.46421/2706-6568.37.2020.paper025) [69] L. Bai, M. Hu, M. Liu, J. Wang, BPIIOT: a light-weighted blockchain-based platform for industrial IoT, Inst. Electr. Electron. Eng. 7 (2019) 58381–58393, [https://doi.org/10.1109/ACCESS.2019.2914223.](https://doi.org/10.1109/ACCESS.2019.2914223) [70] A. Shojaei, Exploring applications of blockchain technology in the construction industry, in: 10th International Structural Engineering and Construction Conference (ISEC) 2019, ISEC Press, University of Illinois at Chicago, 2019, [https://doi.org/10.14455/isec.res.2019.78.](https://doi.org/10.14455/isec.res.2019.78) [71] G.M. Di Giuda, G. Pattini, E. Seghezzi, M. Schievano, F. Paleari, Digital Transformation of the Design, Construction and Management Processes of the [Built Environment, Springer, 2020, pp. 27–36. ISBN: 21987300, https://doi.org/](https://doi.org/10.1007/978-3-030-33570-0_3) [10.1007/978-3-030-33570-0_3.](https://doi.org/10.1007/978-3-030-33570-0_3) [72] Z. Dakhli, Z. Lafhaj, A. Mossman, The potential of blockchain in building [construction, Buildings 9 (4) (2019), https://doi.org/10.3390/buildings9040077.](https://doi.org/10.3390/buildings9040077) [73] R. Gupta, M.N. Shah, S.N. Mandal, Emerging paradigm for land records in India, [Smart Sustain. Built Environ. (2020), https://doi.org/10.1108/SASBE-11-2019-](https://doi.org/10.1108/SASBE-11-2019-0152) [0152.](https://doi.org/10.1108/SASBE-11-2019-0152) [74] S. Dewan, L. Singh, Use of blockchain in designing smart city, Smart Sustain. Built [Environ. 9 (4, 2020) 695–709, https://doi.org/10.1108/SASBE-06-2019-0078.](https://doi.org/10.1108/SASBE-06-2019-0078) [75] Y. Fu, J. Zhu, Trusted data infrastructure for smart cities: a blockchain [perspective, Build. Res. Inf. 49 (1, 2020) 21–37, https://doi.org/10.1080/](https://doi.org/10.1080/09613218.2020.1784703) [09613218.2020.1784703.](https://doi.org/10.1080/09613218.2020.1784703) [76] N. Moretti, J.D. Blanco Cadena, A. Mannino, T. Poli, F. Re Cecconi, Maintenance service optimization in smart buildings through ultrasonic sensors network, [Intell. Build. Int. 13 (1, 2020) 4–16, https://doi.org/10.1080/](https://doi.org/10.1080/17508975.2020.1765723) [17508975.2020.1765723.](https://doi.org/10.1080/17508975.2020.1765723) [77] G.G.R. Roy, S. Britto Ramesh Kumar, A security framework for a sustainable smart ecosystem using permissioned blockchain: performance evaluation, Int. J. [Innov. Technol. Explor. Eng. 8 (10) (2019) 4247–4250, https://doi.org/](https://doi.org/10.35940/ijitee.J9954.0881019) [10.35940/ijitee.J9954.0881019.](https://doi.org/10.35940/ijitee.J9954.0881019) [78] T. Zhang, H. Pota, C.C. Chu, R. Gadh, Real-time renewable energy incentive system for electric vehicles using prioritization and cryptocurrency, Appl. Energy [226 (2018) 582–594, https://doi.org/10.1016/j.apenergy.2018.06.025.](https://doi.org/10.1016/j.apenergy.2018.06.025) [79] E.Z. Berglund, J.G. Monroe, I. Ahmed, M. Noghabaei, J. Do, J.E. Pesantez, M. A. Khaksar Fasaee, E. Bardaka, K. Han, G.T. Proestos, J. Levis, Smart infrastructure: a vision for the role of the civil engineering profession in smart [cities, J. Infrastruct. Syst. 26 (2) (2020), https://doi.org/10.1061/(ASCE)IS.1943-](https://doi.org/10.1061/(ASCE)IS.1943-555X.0000549) [555X.0000549.](https://doi.org/10.1061/(ASCE)IS.1943-555X.0000549) [80] R. Woodhead, P. Stephenson, D. Morrey, Digital construction: from point [solutions to IoT ecosystem, Autom. Constr. 93 (2018) 35–46, https://doi.org/](https://doi.org/10.1016/j.autcon.2018.05.004) [10.1016/j.autcon.2018.05.004.](https://doi.org/10.1016/j.autcon.2018.05.004) [81] K. Christidis, D. Sikeridis, Y. Wang, M. Devetsikiotis, A framework for designing and evaluating realistic blockchain-based local energy markets, Appl. Energy 281 [(2021), https://doi.org/10.1016/j.apenergy.2020.115963.](https://doi.org/10.1016/j.apenergy.2020.115963) [82] A. Banerjee, M. Clear, H. Tewari, Demystifying the role of zk-SNARKs in Zcash, in: 2020 IEEE Conference on Application, Information and Network Security (AINS), Institute of Electrical and Electronics Engineers Inc., Kota Kinabalu, Malaysia, [2020, pp. 12–19, https://doi.org/10.1109/AINS50155.2020.9315064.](https://doi.org/10.1109/AINS50155.2020.9315064) [83] J. Veuger, Trust in a viable real estate economy with disruption and blockchain, [Facilities 36 (1-2) (2018) 103–120, https://doi.org/10.1108/F-11-2017-0106.](https://doi.org/10.1108/F-11-2017-0106) [84] S. Singh, P.K. Sharma, B. Yoon, M. Shojafar, G.H. Cho, I.H. Ra, Convergence of blockchain and artificial intelligence in IoT network for the sustainable smart [city, Sustain. Cities Soc. 63 (2020), https://doi.org/10.1016/j.scs.2020.102364.](https://doi.org/10.1016/j.scs.2020.102364) [85] O. Kodym, L. Kub´aˇc, L. Kavka, Risks associated with Logistics 4.0 and their [minimization using Blockchain, Open Eng. 10 (1) (2020) 74–85, https://doi.org/](https://doi.org/10.1515/eng-2020-0017) [10.1515/eng-2020-0017.](https://doi.org/10.1515/eng-2020-0017) [86] G. Wang, X. Song, L. Liu, Application prospect of blockchain technology in power field, in: 5th IEEE Information Technology and Mechatronics Engineering Conference (ITOEC) 2020, Institute of Electrical and Electronics Engineers Inc., [Chongqing, China, 2020, pp. 407–412, https://doi.org/10.1109/](https://doi.org/10.1109/ITOEC49072.2020.9141546) [ITOEC49072.2020.9141546.](https://doi.org/10.1109/ITOEC49072.2020.9141546) [87] E.A. Parn, D. Edwards, Cyber threats confronting the digital built environment: common data environment vulnerabilities and block chain deterrence, Eng. [Constr. Archit. Manag. 26 (2, 2019) 245–266, https://doi.org/10.1108/ECAM-](https://doi.org/10.1108/ECAM-03-2018-0101) [03-2018-0101.](https://doi.org/10.1108/ECAM-03-2018-0101) [88] R. Shinde, O. Nilakhe, P. Pondkule, D. Karche, P. Shendage, Enhanced road construction process with machine learning and blockchain technology, in: 2020 International Conference on Industry 4.0 Technology (I4Tech), Institute of [Electrical and Electronics Engineers Inc., Pune, India, 2020, pp. 207–210, https://](https://doi.org/10.1109/I4Tech48345.2020.9102669) [doi.org/10.1109/I4Tech48345.2020.9102669.](https://doi.org/10.1109/I4Tech48345.2020.9102669) [89] J. Woo, S. Shin, A.T. Asutosh, J. Li, C.J. Kibert, An overview of state-of-the-art technologies for data-driven construction, in: Lecture Notes in Civil Engineering Vol. 98, Springer, Sao Paulo, Brazil, 2021, pp. 1323˜ [–1334, https://doi.org/](https://doi.org/10.1007/978-3-030-51295-8_94) [10.1007/978-3-030-51295-8_94.](https://doi.org/10.1007/978-3-030-51295-8_94) [90] Y. Li, W. Yang, P. He, C. Chen, X. Wang, Design and management of a distributed hybrid energy system through smart contract and blockchain, Appl. Energy 248 [(2019) 390–405, https://doi.org/10.1016/j.apenergy.2019.04.132.](https://doi.org/10.1016/j.apenergy.2019.04.132) [91] A. Esmat, M. de Vos, Y. Ghiassi-Farrokhfal, P. Palensky, D. Epema, A novel decentralized platform for peer-to-peer energy trading market with blockchain [technology, Appl. Energy 282 (2021), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.apenergy.2020.116123) [apenergy.2020.116123.](https://doi.org/10.1016/j.apenergy.2020.116123) [92] L. Ableitner, V. Tiefenbeck, A. Meeuw, A. Worner, E. Fleisch, F. Wortmann, User ¨ behavior in a real-world peer-to-peer electricity market, Appl. Energy 270 (2020), [https://doi.org/10.1016/j.apenergy.2020.115061.](https://doi.org/10.1016/j.apenergy.2020.115061) [93] S. Saha, N. Ravi, K. Hreinsson, J. Baek, A. Scaglione, N.G. Johnson, A secure distributed ledger for transactive energy: the Electron Volt Exchange (EVE) ----- [blockchain, Appl. Energy 282 (2021), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.apenergy.2020.116208) [apenergy.2020.116208.](https://doi.org/10.1016/j.apenergy.2020.116208) [94] S. Noor, W. Yang, M. Guo, K.H. van Dam, X. Wang, Energy demand side management within micro-grid networks enhanced by blockchain, Appl. Energy [228 (2018) 1385–1398, https://doi.org/10.1016/j.apenergy.2018.07.012.](https://doi.org/10.1016/j.apenergy.2018.07.012) [95] M. Foti, M. Vavalis, Blockchain based uniform price double auctions for energy [markets, Appl. Energy 254 (2019), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.apenergy.2019.113604) [apenergy.2019.113604.](https://doi.org/10.1016/j.apenergy.2019.113604) [96] Y. Jiang, K. Zhou, X. Lu, S. Yang, Electricity trading pricing among prosumers with game theory-based model in energy blockchain environment, Appl. Energy [271 (2020), https://doi.org/10.1016/j.apenergy.2020.115239.](https://doi.org/10.1016/j.apenergy.2020.115239) [97] M.K. McGowan, Integrating CHP systems, renewable energy to increase resilience, Am. Soc. Hea. Refrigerat. Air Condition. Eng. 61 (12) (2019) 38–43, [https://doi.org/10.22260/ISARC2019/0010.](https://doi.org/10.22260/ISARC2019/0010) [98] R.K. Perrons, T. Cosby, Applying blockchain in the geoenergy domain: the road to [interoperability and standards, Appl. Energy 262 (2020), https://doi.org/](https://doi.org/10.1016/j.apenergy.2020.114545) [10.1016/j.apenergy.2020.114545.](https://doi.org/10.1016/j.apenergy.2020.114545) [99] S. Keivanpour, A. Ramudhin, D. Ait Kadi, An empirical analysis of complexity management for offshore wind energy supply chains and the benefits of [blockchain adoption, Civ. Eng. Environ. Syst. 37 (3, 2020) 117–142, https://doi.](https://doi.org/10.1080/10286608.2020.1810674) [org/10.1080/10286608.2020.1810674.](https://doi.org/10.1080/10286608.2020.1810674) [100] M.N.N. Rodrigo, S. Perera, S. Senaratne, X. Jin, Potential application of blockchain technology for embodied carbon estimating in construction supply [chains, Buildings 10 (8) (2020), https://doi.org/10.3390/BUILDINGS10080140.](https://doi.org/10.3390/BUILDINGS10080140) [101] W. Hua, J. Jiang, H. Sun, J. Wu, A blockchain based peer-to-peer trading framework integrating energy and carbon markets, Appl. Energy 279 (2020), [https://doi.org/10.1016/j.apenergy.2020.115539.](https://doi.org/10.1016/j.apenergy.2020.115539) [102] J.J. Hunhevicz, D.M. Hall, Do you need a blockchain in construction? Use case categories and decision framework for DLT design options, Adv. Eng. Inform. 45 [(2020), https://doi.org/10.1016/j.aei.2020.101094.](https://doi.org/10.1016/j.aei.2020.101094) [103] T. Dounas, D. Lombardi, W. Jabi, Framework for decentralised architectural design BIM and Blockchain integration, Int. J. Archit. Comput. 19 (2, 2021) [157–173, https://doi.org/10.1177/1478077120963376.](https://doi.org/10.1177/1478077120963376) [104] Z. Ye, M. Yin, L. Tang, H. Jiang, Cup-of-Water theory: a review on the interaction of BIM, IoT and blockchain during the whole building lifecycle, in: 35th International Symposium on Automation and Robotics in Construction and International AEC/FM Hackathon: The Future of Building Things (ISARC) 2018, International Association for Automation and Robotics in Construction (IAARC), [Berlin, Germany, 2018, pp. 478–486, https://doi.org/10.22260/isarc2018/0066.](https://doi.org/10.22260/isarc2018/0066) [105] A. Shojaei, J. Wang, A. Fenner, Exploring the feasibility of blockchain technology as an infrastructure for improving built asset sustainability, Built Environ. Project [Asset Manag. 10 (2, 2019) 184–199, https://doi.org/10.1108/BEPAM-11-2018-](https://doi.org/10.1108/BEPAM-11-2018-0142) [0142.](https://doi.org/10.1108/BEPAM-11-2018-0142) [106] V. Hargaden, N. Papakostas, A. Newell, A. Khavia, A. Scanlon, The role of blockchain technologies in construction engineering project management, in: 25th IEEE International Conference on Engineering, Technology and Innovation, Institute of Electrical and Electronics Engineers Inc., Valbonne Sophia-Antipolis, [France, 2019, https://doi.org/10.1109/ICE.2019.8792582.](https://doi.org/10.1109/ICE.2019.8792582) [107] Y. Yao, L. Chu, L. Shan, Q. Lei, Supply chain financial model innovation based on block-chain drive and construction of cloud computing credit system, in: 4th Institute of Electrical and Electronics Engineers (IEEE) International Conference on Smart Internet of Things 2020, IEEE Inc., Beijing, China, 2020, pp. 249–255, [https://doi.org/10.1109/SmartIoT49966.2020.00044.](https://doi.org/10.1109/SmartIoT49966.2020.00044) [108] W. Lu, Blockchain technology and its applications in FinTech, in: 2nd International Conference on Intelligent, Secure and Dependable Systems in Distributed and Cloud Environments (ISDDC) 2018, Springer Verlag, Vancouver, [Canada, 2018, pp. 118–124, https://doi.org/10.1007/978-3-030-03712-3_10.](https://doi.org/10.1007/978-3-030-03712-3_10) [109] V. Hassija, V. Chamola, S. Zeadally, BitFund: a blockchain-based crowd funding platform for future smart and connected nation, Sustain. Cities Soc. 60 (2020), [https://doi.org/10.1016/j.scs.2020.102145.](https://doi.org/10.1016/j.scs.2020.102145) [110] K.M. San, C.F. Choy, W.P. Fung, The potentials and impacts of blockchain technology in construction industry: a literature review, in: 11th Curtin University Technology, Science and Engineering International Conference (CUTSE) 2018 Vol. 495, Institute of Physics Publishing, Sarawak, Malaysia, 2019, [https://doi.org/10.1088/1757-899X/495/1/012005.](https://doi.org/10.1088/1757-899X/495/1/012005) [111] G. Pattini, G.M.D. Giuda, L.C. Tagliabue, Blockchain application for contract schemes in the construction industry, in: 3rd European and Mediterranean Structural Engineering and Construction Conference 2020, International Structural Engineering and Construction Society (ISEC), Limassol, Cyprus, 2020, [pp. 1–6, https://doi.org/10.14455/ISEC.res.2020.7(1).AAE-21.](https://doi.org/10.14455/ISEC.res.2020.7(1).AAE-21) [112] E. Aleksandrova, V. Vinogradova, G. Tokunova, Integration of digital technologies in the field of construction in the Russian Federation, Eng. Manag. [Product. Serv. 11 (3, 2019) 38–47, https://doi.org/10.2478/emj-2019-0019.](https://doi.org/10.2478/emj-2019-0019) [113] C.Z. Li, Z. Chen, F. Xue, X.T.R. Kong, B. Xiao, X. Lai, Y. Zhao, A blockchain- and IoT-based smart product-service system for the sustainability of prefabricated [housing construction, J. Clean. Prod. 286 (2021), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.jclepro.2020.125391) [jclepro.2020.125391.](https://doi.org/10.1016/j.jclepro.2020.125391) [114] C.S. Gotz, P. Karlsson, I. Yitmen, Exploring applicability, interoperability and integrability of Blockchain-based digital twins for asset life cycle management, [Smart Sustain. Built Environ. (2020), https://doi.org/10.1108/sasbe-08-2020-](https://doi.org/10.1108/sasbe-08-2020-0115) [0115.](https://doi.org/10.1108/sasbe-08-2020-0115) [115] P.F. Wong, F.C. Chia, M.S. Kiu, E.C.W. Lou, Potential integration of blockchain technology into smart sustainable city (SSC) developments: a systematic review, [Smart Sustain. Built Environ. (2020), https://doi.org/10.1108/SASBE-09-2020-](https://doi.org/10.1108/SASBE-09-2020-0140) [0140.](https://doi.org/10.1108/SASBE-09-2020-0140) [116] O. Aljabri, O. Aldhaheri, H. Mohammed, A. Alkaabi, J. Abdella, K. Shuaib, Facilitating electric vehicle charging across the UAE using blockchain, in: 2019 International Conference on Electrical and Computing Technologies and Applications (ICECTA), Institute of Electrical and Electronics Engineers, Ras Al [Khaimah, United Arab Emirates, 2019, https://doi.org/10.1109/](https://doi.org/10.1109/ICECTA48151.2019.8959705) [ICECTA48151.2019.8959705.](https://doi.org/10.1109/ICECTA48151.2019.8959705) [117] A. Ghosh, D.J. Edwards, M.R. Hosseini, Patterns and trends in Internet of Things (IoT) research: future applications in the construction industry, Eng. Constr. [Archit. Manag. 28 (2) (2020), https://doi.org/10.1108/ecam-04-2020-0271.](https://doi.org/10.1108/ecam-04-2020-0271) [118] M.E. Wainstein, Blockchains as enablers of participatory smart grids, Technol. [Archit. Des. 3 (2, 2019) 131–136, https://doi.org/10.1080/](https://doi.org/10.1080/24751448.2019.1640521) [24751448.2019.1640521.](https://doi.org/10.1080/24751448.2019.1640521) [119] A. Lüth, J.M. Zepter, P. Crespo del Granado, R. Egging, Local electricity market designs for peer-to-peer trading: the role of battery flexibility, Appl. Energy 229 [(2018) 1233–1243, https://doi.org/10.1016/j.apenergy.2018.08.004.](https://doi.org/10.1016/j.apenergy.2018.08.004) [120] J. Woo, A.T. Asutosh, J. Li, W.D. Ryor, C.J. Kibert, A. Shojaei, Blockchain: a theoretical framework for better application of carbon credit acquisition to the building sector, in: Construction Research Congress 2020: Infrastructure Systems and Sustainability, American Society of Civil Engineers (ASCE), Tempe, Arizona, [2020, pp. 885–894, https://doi.org/10.1061/9780784482858.095.](https://doi.org/10.1061/9780784482858.095) [121] H. Hamledari, M. Fischer, Role of blockchain-enabled smart contracts in automating construction progress payments, J. Leg. Aff. Disput. Resolut. Eng. [Constr. 13 (1) (2021), https://doi.org/10.1061/(ASCE)LA.1943-4170.0000442.](https://doi.org/10.1061/(ASCE)LA.1943-4170.0000442) [122] D. Kifokeris, C. Koch, A conceptual digital business model for construction logistics consultants, featuring a sociomaterial blockchain solution for integrated economic, material and information flows, J. Inf. Technol. Constr. 25 (29) (2020) [500–521, https://doi.org/10.36680/j.itcon.2020.029.](https://doi.org/10.36680/j.itcon.2020.029) [123] T. Kobashi, T. Yoshida, Y. Yamagata, K. Naito, S. Pfenninger, K. Say, Y. Takeda, A. Ahl, M. Yarime, K. Hara, On the potential of “Photovoltaics + Electric vehicles” for deep decarbonization of Kyoto’s power systems: techno-economic-social [considerations, Appl. Energy 275 (2020), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.apenergy.2020.115419) [apenergy.2020.115419.](https://doi.org/10.1016/j.apenergy.2020.115419) [124] H. Luo, M. Das, J. Wang, J.C.P. Cheng, Construction payment automation through smart contract-based blockchain framework, in: 36th International Symposium on Automation and Robotics in Construction, ISARC 2019, International Association for Automation and Robotics in Construction (IAARC), Banff Alberta, Canada, [2019, pp. 1254–1260, https://doi.org/10.22260/isarc2019/0168.](https://doi.org/10.22260/isarc2019/0168) [125] CB_Insights, Banking is Only the Beginning: 58 Big Industries Blockchain Could [Transform. https://www.cbinsights.com/research/industries-disrupted-blockchai](https://www.cbinsights.com/research/industries-disrupted-blockchain/) [n/, 2021 accessed 4th May 2021.](https://www.cbinsights.com/research/industries-disrupted-blockchain/) [126] Z. Zhao, X. Song, Y. Xu, S. Tian, M. Gao, Y. Zhang, Research on the construction of data pool and the management model of data pool integrating block chain concept for ubiquitous electric power Internet of Things, in: 3rd Institute of Electrical and Electronics Engineers (IEEE) Conference on Energy Internet and Energy System Integration 2019, IEEE Inc., Changsha, China, 2019, [pp. 1828–1833, https://doi.org/10.1109/EI247390.2019.9062040.](https://doi.org/10.1109/EI247390.2019.9062040) [127] J. Lin, M. Pipattanasomporn, S. Rahman, Comparative analysis of auction mechanisms and bidding strategies for P2P solar transactive energy markets, [Appl. Energy 255 (2019), https://doi.org/10.1016/j.apenergy.2019.113687.](https://doi.org/10.1016/j.apenergy.2019.113687) [128] Z. Liu, L. Jiang, M. Osmani, P. Demian, Building Information Management (BIM) and blockchain (BC) for sustainable building design information management [framework, Electronics 8 (7) (2019), https://doi.org/10.3390/](https://doi.org/10.3390/electronics8070724) [electronics8070724.](https://doi.org/10.3390/electronics8070724) [129] J. Yang, M. Guo, F. Fei, H. Wang, L. Zhang, J. Xie, Enabling technologies for multinational interconnected smart grid, in: 3rd Institute of Electrical and Electronics Engineers (IEEE) International Electrical and Energy Conference [2019, IEEE Inc., Beijing, China, 2019, pp. 1106–1111, https://doi.org/10.1109/](https://doi.org/10.1109/CIEEC47146.2019.CIEEC-2019411) [CIEEC47146.2019.CIEEC-2019411.](https://doi.org/10.1109/CIEEC47146.2019.CIEEC-2019411) [130] G. van Leeuwen, T. AlSkaif, M. Gibescu, W. van Sark, An integrated blockchainbased energy management platform with bilateral trading for microgrid [communities, Appl. Energy 263 (2020), https://doi.org/10.1016/j.](https://doi.org/10.1016/j.apenergy.2020.114613) [apenergy.2020.114613.](https://doi.org/10.1016/j.apenergy.2020.114613) [131] Z. Zeng, Y. Li, Y. Cao, Y. Zhao, J. Zhong, D. Sidorov, X. Zeng, Blockchain technology for information security of the energy internet: fundamentals, [features, strategy and application, Energies 13 (4) (2020), https://doi.org/](https://doi.org/10.3390/en13040881) [10.3390/en13040881.](https://doi.org/10.3390/en13040881) -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1016/j.autcon.2021.103914?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1016/j.autcon.2021.103914, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GREEN", "url": "https://discovery.ucl.ac.uk/10134004/7/Scott_Exploratory%20literature%20review%20of%20blockchain%20in%20the%20construction%20industry.pdf" }
2,021
[ "Review" ]
true
2021-12-01T00:00:00
[ { "paperId": "954e2a53f706c2bf86819243eba2e65b5b620b3b", "title": "A blockchain- and IoT-based smart product-service system for the sustainability of prefabricated housing construction" }, { "paperId": "45fe706be2ee2fcc64c4c69eb95d0e5ab6a81081", "title": "Role of Blockchain-Enabled Smart Contracts in Automating Construction Progress Payments" }, { "paperId": "7e783f2105c295ee53049f8d281859d7f52e88f3", "title": "A novel decentralized platform for peer-to-peer energy trading market with blockchain technology" }, { "paperId": "3b66139f31adc6fd6405ceac61f9ff1e663ea6ff", "title": "A framework for designing and evaluating realistic blockchain-based local energy markets" }, { "paperId": "41f03b88fade44b6f9cbcf94d83f79d169a1be2f", "title": "Potential integration of blockchain technology into smart sustainable city (SSC) developments: a systematic review" }, { "paperId": "ac0854da422d603e725b49f0d6006d41214cbf7e", "title": "Convergence of blockchain and artificial intelligence in IoT network for the sustainable smart city" }, { "paperId": "b8e0bfc698f85c9fafd009bdf697d857dc40baeb", "title": "A blockchain based peer-to-peer trading framework integrating energy and carbon markets" }, { "paperId": "149c810f988c1293675fdad36b4c5c09f60d1831", "title": "Construction quality information management with blockchains" }, { "paperId": "20afdc74da7374c90c5c148343a01b3e998c7bf4", "title": "Adaptive bidding strategy for real-time energy management in multi-energy market enhanced by blockchain" }, { "paperId": "e674bce4e2c03210fec8429756f5317a8a850b87", "title": "A smart contract system for security of payment of construction contracts" }, { "paperId": "8971043be9d8433368db1561135e9745697746dc", "title": "Exploring applicability, interoperability and integrability of Blockchain-based digital twins for asset life cycle management" }, { "paperId": "2a37ff0519bcc64e8354429f65dcb35fcca2d3a0", "title": "Decentralized blockchain-based consensus for Optimal Power Flow solutions" }, { "paperId": "7ae0e6c84231f6466be466863c38c2604c9abf3a", "title": "A Secure Distributed Ledger for Transactive Energy: The Electron Volt Exchange (EVE) Blockchain" }, { "paperId": "22a15985ac8eb2ee4ba92f17d2a76039a603d482", "title": "A Systematic Review of Digital Technology Adoption in Off-Site Construction: Current Status and Future Direction towards Industry 4.0" }, { "paperId": "f20474a6a82384934d0786b4c1f578fec55919dc", "title": "Blockchain: A Theoretical Framework for Better Application of Carbon Credit Acquisition to the Building Sector" }, { "paperId": "10cd5e487b9cd3c248f9c459f36e8bd8e92fa7f6", "title": "City logistics: Towards a blockchain decision framework for collaborative parcel deliveries in micro-hubs" }, { "paperId": "1aee9eecac104f75d7581c0dfe78685968fcebbb", "title": "A conceptual digital business model for construction logistics consultants, featuring a sociomaterial blockchain solution for integrated economic, material and information flows" }, { "paperId": "fa967c210277d738a71490aee0169859ac0e9cf3", "title": "Exploring the potentials of blockchain application in construction industry: a systematic review" }, { "paperId": "a793ae647d8f895034ca139aa868b4cec9a4e775", "title": "Integrating BIM- and Cost-included Information Container with Blockchain for Construction Automated Payment using Billing Model and Smart Contracts" }, { "paperId": "f518e180bc735ec3970547582f9e41a6d4d201e3", "title": "Framework for decentralised architectural design BIM and Blockchain integration" }, { "paperId": "33fe20d7e2d68f02f9da8004c774132ece1d43d4", "title": "Incentivizing High-Quality Data Sets in Construction Using Blockchain: A Feasibility Study in the Swiss Industry" }, { "paperId": "04b58585d0d5d707c3b3d9afd996a884ef387fde", "title": "Enabling technologies and sustainable smart cities" }, { "paperId": "8d95dc5c1d78f54bd8ef442353fb203f6f35bddc", "title": "A semantic differential transaction approach to minimizing information redundancy for BIM and blockchain integration" }, { "paperId": "1a415ec978c083c81148ad9f29a7a807c1acd9f4", "title": "Blockchain for smart cities: A review of architectures, integration trends and future research directions" }, { "paperId": "8fc3594eef16f18f8e35f8e4387337fbf3675d8b", "title": "On the potential of “Photovoltaics + Electric vehicles” for deep decarbonization of Kyoto’s power systems: Techno-economic-social considerations" }, { "paperId": "10a743a1c175e11218c13ddddddb488a40716e7c", "title": "Public and private blockchain in construction business process and information integration" }, { "paperId": "cece44ce43909efcf5346c61da0dc184748c3e5a", "title": "Securing interim payments in construction projects through a blockchain-based framework" }, { "paperId": "a9e212ef2e0abaca7878194ab831e23d5b5296df", "title": "Blockchain technology in supply chain operations: Applications, challenges and research opportunities" }, { "paperId": "d99829f8ecf858e7d131527793d8062f3ec662c2", "title": "Designing a blockchain enabled supply chain" }, { "paperId": "1ab288b6f82a90ca34321acc3e4a823323a67bf3", "title": "Technological, organisational and environmental determinants of smart contracts adoption: UK construction sector viewpoint" }, { "paperId": "bbc71211666ff20eae2cdb6ef662de85a52eb5d9", "title": "Business Sustainability through Technology Management Approach for Construction Company" }, { "paperId": "c1830c7b2454a6495cf54cbb1fda95fa5cd8a4f3", "title": "BitFund: A blockchain-based crowd funding platform for future smart and connected nation" }, { "paperId": "10ed4a256d051166963970c97e8f8946df19fe82", "title": "A deep learning-based IoT-oriented infrastructure for secure smart City" }, { "paperId": "efa27b5cacac592d1946fb807f8b27f35007a3ba", "title": "Patterns and trends in Internet of Things (IoT) research: future applications in the construction industry" }, { "paperId": "14180128da9ce6884c72a4168d98cc77761d1000", "title": "An Overview of State-of-the-Art Technologies for Data-Driven Construction" }, { "paperId": "90b00d8d6537c020f10379c699009f2980b3fd12", "title": "Framework for Automated Billing in the Construction Industry Using BIM and Smart Contracts" }, { "paperId": "90b9a4b54d9eda92158109905e3f343c04a770fa", "title": "Potential Application of Blockchain Technology for Embodied Carbon Estimating in Construction Supply Chains" }, { "paperId": "2532be5c8981a1c691e09b601f4ee180f8cfde63", "title": "Demystifying the Role of zk-SNARKs in Zcash" }, { "paperId": "2748bf0b3d8efda356588884703d36fec6aebb16", "title": "Supply Chain Financial Model Innovation Based on Block-chain Drive and Construction of Cloud Computing Credit System" }, { "paperId": "90536f5a48aaecaeced36b50f6b615c2572b69d4", "title": "Electricity trading pricing among prosumers with game theory-based model in energy blockchain environment" }, { "paperId": "29a1007aa79bda5664064ea5e8893caa1eca390b", "title": "Design Copyright in Architecture, Engineering, and Construction Industry: Review of History, Pitfalls, and Lessons Learned" }, { "paperId": "f28039807e149df3ec8704201347803e7a33fcd3", "title": "BLOCKCHAIN APPLICATION FOR CONTRACT SCHEMES IN THE CONSTRUCTION INDUSTRY" }, { "paperId": "ca075c74994f5d7cdaafde22a5030a04e615fc2b", "title": "A Study on the Application of Blockchain Technology in the Construction Industry" }, { "paperId": "7671d772320995511e118a5978cea2d615eb722b", "title": "An empirical analysis of complexity management for offshore wind energy supply chains and the benefits of blockchain adoption" }, { "paperId": "407e78a3b55075a371f58737fbff387f9714907b", "title": "Trusted data infrastructure for smart cities: a blockchain perspective" }, { "paperId": "cd66a67c19df4e4bb121360bbd794730de5b7b3b", "title": "User behavior in a real-world peer-to-peer electricity market" }, { "paperId": "30b9cbc50765a3ef2830003fa23c461d26f1a86a", "title": "Maintenance service optimization in smart buildings through ultrasonic sensors network" }, { "paperId": "088afe931b4dfd9d8517a4f2815417b1c6f2ead9", "title": "Smart Infrastructure: A Vision for the Role of the Civil Engineering Profession in Smart Cities" }, { "paperId": "dd3eb01107beda6f3e0bf2838b3561f9f55c4d48", "title": "Integrating advanced technologies to uphold security of payment: Data flow diagram" }, { "paperId": "e50b2476b0e00039faa64462ddce1ace17c611ed", "title": "Application Prospect of Blockchain Technology in Power Field" }, { "paperId": "369a6957d0e9fcaba5ffe77183750c57feb3383f", "title": "Integrated project delivery with blockchain: An automated financial system" }, { "paperId": "504ecf4f705430e1358c3ab009f07c718f3a7a8f", "title": "Data Handling in Industry 4.0: Interoperability Based on Distributed Ledger Technology" }, { "paperId": "5c1fa5d2d7eaf93bf5023f3df57a06b160a00838", "title": "Smart contract architecture for decentralized energy trading and management based on blockchains" }, { "paperId": "fbf9573f3a108028c773a8289ee8240a842b270e", "title": "Interoperability in building information modeling for AECO/FM industry" }, { "paperId": "54592cde9a2891327ae533100d570ab10466c8f0", "title": "Shifting trust in construction supply chains through blockchain technology" }, { "paperId": "15c8ec6c787605652ee8aaeb71e1607587b8f174", "title": "A Blockchain and Smart Contract-Based Framework to Inrease Traceability of Built Assets" }, { "paperId": "dc7e25aaffb63642dd3ee0a5220bd6de2ca57110", "title": "Emerging paradigm for land records in India" }, { "paperId": "4964f2dc4a999f212427428d421bb95f4e6427ef", "title": "An integrated blockchain-based energy management platform with bilateral trading for microgrid communities" }, { "paperId": "91faadf3071a022907b4c73a68efacebded8f981", "title": "Do you need a blockchain in construction? Use case categories and decision framework for DLT design options" }, { "paperId": "2a236bced8c548f991c3b338b12842570284278a", "title": "Developing a theoretical framework for intelligent contract acceptance" }, { "paperId": "698087b9986b2651ac253345fcbaf492d7c57f6e", "title": "Applying blockchain in the geoenergy domain: The road to interoperability and standards" }, { "paperId": "8d0b554012e3db32b825f3f7d213c211fe2cbdd1", "title": "Blockchain-Enabled Information Sharing Within a Supply Chain: A Systematic Literature Review" }, { "paperId": "f6a925fbbc7494d76236465bb480b8da1ecaf0b9", "title": "Use of blockchain in designing smart city" }, { "paperId": "9f7af9b09c2854704f65a8de044d8f224917c29f", "title": "Blockchain technology: Is it hype or real in the construction industry?" }, { "paperId": "cf8dadbb09cc0c9997de4ab4977d5f0924839e9f", "title": "Blockchain-based framework for improving supply chain traceability and information sharing in precast construction" }, { "paperId": "cf8967a7e11aad6cf4dfcce6410f248655fb9220", "title": "Blockchain Technology for Information Security of the Energy Internet: Fundamentals, Features, Strategy and Application" }, { "paperId": "04655d5fdc4af929cf84ae4e2c583bd202b16789", "title": "A Decentralized System for Green Energy Distribution in a Smart Grid" }, { "paperId": "7e9756034d74d7d46661b1c4121a8263a0ae9d8f", "title": "Enhanced Road Construction Process with Machine Learning and Blockchain Technology" }, { "paperId": "8c333011616f05c449b1d8a7d938663636cf972c", "title": "Blockchain-Based Dynamic Provable Data Possession for Smart Cities" }, { "paperId": "0af503e56b0966757388737db73b922880299e30", "title": "Risks associated with Logistics 4.0 and their minimization using Blockchain" }, { "paperId": "3bc5e91d3331120e0bfe3e97d9bbf3c3ec72ad97", "title": "Comparative analysis of auction mechanisms and bidding strategies for P2P solar transactive energy markets" }, { "paperId": "26da370edcdc698971e519358bdce50fba80fa31", "title": "Blockchain based uniform price double auctions for energy markets" }, { "paperId": "e5cc84bf611b5eff714fe379234f0ae2fbf079fd", "title": "Exploring the feasibility of blockchain technology as an infrastructure for improving built asset sustainability" }, { "paperId": "c451c411ec28594d7a11825cee40483dc78dfe72", "title": "Research on the Construction of Data Pool and the Management Model of Data Pool Integrating Block Chain Concept for Ubiquitous Electric Power Internet of Things" }, { "paperId": "154f71b83da195fbeb8899397edd1b60b5dec8ca", "title": "Facilitating Electric Vehicle Charging Across the UAE Using Blockchain" }, { "paperId": "57ef1675afb2e1c69173d2289ce88103ada39770", "title": "BIM Fork: Are Smart Contracts in Construction More Likely to Prosper with or without BIM?" }, { "paperId": "51a0c86e7d156453a0cc19de99efa61d93798704", "title": "An Implementation of Smart Contracts by Integrating BIM and Blockchain" }, { "paperId": "9b61095876485dc4771cfe8eaf072656bca38b02", "title": "Applications of Blockchain Technology in Smart City Development: A Research" }, { "paperId": "1fe1761ee3ac07c6974efeacafc5d1ff48aa4f04", "title": "Blockchain-based Building Information Modeling" }, { "paperId": "56f00126a7f95ad4f2f606abc0ffd00be756b02a", "title": "The strategic role of logistics in the industry 4.0 era" }, { "paperId": "cad4018665a33dbbce206e4cb4e4dbda0af0c8d1", "title": "Enabling Technologies for Multinational Interconnected Smart Grid" }, { "paperId": "15556a7cecf78ce0b7293791da1ec54bc7afaec3", "title": "Integration of digital technologies in the field of construction in the Russian Federation" }, { "paperId": "e6d0ab0fd0ed94dca9f659f197597139cb11d276", "title": "Blockchain and the built environment: Potentials and limitations" }, { "paperId": "e9abe56a9b4aad075a91e746af7f38e1b22f16f3", "title": "A Key Protection Scheme Based on Secret Sharing for Blockchain-Based Construction Supply Chain System" }, { "paperId": "fdbd52bab0983cdae4c16c68984ce0a3d1ceb996", "title": "Design and management of a distributed hybrid energy system through smart contract and blockchain" }, { "paperId": "a262adfcbaf56d77212a13c9dbfd896e457bdcaf", "title": "A Security Framework for A Sustainable Smart Ecosystem using Permissioned Blockchain: Performance Evaluation" }, { "paperId": "60502ffacc366d5ab67706fea21a935294af0c10", "title": "Blockchain-based decentralized and secure keyless signature scheme for smart grid" }, { "paperId": "38f6f0b1d21c0000bf03cb9e2254e8ba6736d2e7", "title": "Managing mistrust in construction using DLT: a review of use-case categories for technical decisions" }, { "paperId": "2b19ff6cb3725e8959ec7b9e541c868ab856fb99", "title": "Blockchains as Enablers of Participatory Smart Grids" }, { "paperId": "61d99bfe3589904598f9e6d00ff589cf47bcc137", "title": "Building Information Management (BIM) and Blockchain (BC) for Sustainable Building Design Information Management Framework" }, { "paperId": "6d98b4786da46b50790ad5be5c89c80d0e7bde45", "title": "Blockchain and Building Information Modeling (BIM): Review and Applications in Post-Disaster Recovery" }, { "paperId": "d693b7492b20fc2b6a2aed29950059c4102d93bc", "title": "A multi-layered blockchain framework for smart mobility data-markets" }, { "paperId": "473a63d45bd5b73cb3d1b55d3b3ccf5e44378b3f", "title": "Blockchain in the built environment and construction industry: A systematic review, conceptual models and practical use cases" }, { "paperId": "b18859babf3cb062440051bd5098fd9baa30705a", "title": "The Role of Blockchain Technologies in Construction Engineering Project Management" }, { "paperId": "9ad2a3e08cfeec91d402cfda9eb8b5cb2390c0d0", "title": "Construction Payment Automation through Smart Contract-based Blockchain Framework" }, { "paperId": "55a8d41c05a4fb5af720cad55f06832ec31597a9", "title": "BPIIoT: A Light-Weighted Blockchain-Based Platform for Industrial IoT" }, { "paperId": "f726dfa8480cb4309b3f3bfb4f57e8d358f67776", "title": "EXPLORING APPLICATIONS OF BLOCKCHAIN TECHNOLOGY IN THE CONSTRUCTION INDUSTRY" }, { "paperId": "21ef6e3b2a233393a4dae08ebdaae3712c4bc97e", "title": "The Potential of Blockchain in Building Construction" }, { "paperId": "ea2be6a62e543f9316ea301d0b85d4922771ae60", "title": "The Potentials and Impacts of Blockchain Technology in Construction Industry: A Literature Review" }, { "paperId": "74bc40c88738ee13e1f65d21f5e65d9a80f30450", "title": "bcBIM: A Blockchain-Based Big Data Model for BIM Modification Audit and Provenance in Mobile Cloud" }, { "paperId": "a3aa399df2ace99ba521e372ac89d3c707d68002", "title": "Tools of Data Transmission at Building Information Modeling" }, { "paperId": "4fb1ba8f667616b98ad2c23f45fb3ba5249410be", "title": "Cyber threats confronting the digital built environment" }, { "paperId": "40eebcf479382c5269bbbee663eeeeb2331286e0", "title": "Blockchain Technology and Its Applications in FinTech" }, { "paperId": "0623fccd7d9bed096539c5fc1d45cc69e4d46929", "title": "Local electricity market designs for peer-to-peer trading: The role of battery flexibility" }, { "paperId": "f1abdb39568305750bf8682177c19079213d498f", "title": "Energy Demand Side Management within micro-grid networks enhanced by blockchain" }, { "paperId": "f52da5de09381f55a39be3e2c308b1e0bb1ee630", "title": "Digital construction: From point solutions to IoT ecosystem" }, { "paperId": "6535a6e20ad025d209f98a906c4413bcd0bf536e", "title": "Real-time renewable energy incentive system for electric vehicles using prioritization and cryptocurrency" }, { "paperId": "3e64d0ea58070d3715aad218499660b043628173", "title": "Cup-of-Water Theory: A Review on the Interaction of BIM, IoT and Blockchain During the Whole Building Lifecycle" }, { "paperId": "8d3f008403e73226955699e31361f5ae02628dc8", "title": "Designing microgrid energy markets" }, { "paperId": "06e9bfcddee862c55971dc25b359ca4d5990d79e", "title": "Blockchain for architects: challenges from the sharing economy" }, { "paperId": "70599dfc16ff4fbe3ce9bf1d77b74f870175614e", "title": "Trust in a viable real estate economy with disruption and blockchain" }, { "paperId": "eabb07994b757d329a434a082e72b81fca9f6237", "title": "Blockchain technology in the chemical industry: Machine-to-machine electricity market" }, { "paperId": "2689296a82d8aabcd477c21c2980b41d0cd04881", "title": "A Comparison between Two Main Academic Literature Collections: Web of Science and Scopus Databases" }, { "paperId": "c9e151ba8e59422320013d64307a17a94e018a98", "title": "Scopus database: a review" }, { "paperId": "cdaf426db242cd7f2b97a2dc79629b213420a8e0", "title": "The Economics of the Modern Construction Firm" }, { "paperId": "7d2a2f737e5aa492bbc757c19cb7b4ecace88d4a", "title": "Technological" }, { "paperId": "eed76365143d2e3f35d0cd37cb8f41fdebffe34d", "title": "Country" }, { "paperId": null, "title": "Paperpile, The Top List of Academic Research Databases" }, { "paperId": "612d9c441c72178b360adb60043c480c09d40b03", "title": "Buildings as material banks using RFID and building information modeling in a circular economy" }, { "paperId": null, "title": "Blockchain and smart contracts for Integrated Project Delivery: inspiration from the commons" }, { "paperId": "e9e89d1417cf008f3512ce34b291277fab0f0bcc", "title": "Blockchain in building logistics: emerging knowledge, and related actors in Sweden" }, { "paperId": "ef7f7ac224beaec1ad26f21e204c84a19146bf84", "title": "A blockchain-based smart contract trading mechanism for energy power supply and demand network" }, { "paperId": null, "title": "An evaluation of barriers to E-Procurement in Turkish construction industry" }, { "paperId": "cb70bbd86f76b85a519cbe44ede044d71d32d2cb", "title": "Digital Transformation Design" }, { "paperId": null, "title": "Integrating CHP systems, renewable energy to increase resilience" }, { "paperId": "22a604ea60646ebd0e76caa49d1f8e01d2a3781a", "title": "Application of RFID combined with blockchain technology in logistics of construction materials" }, { "paperId": "256553ef50899b0338398fdf90bcc7920f3a78ea", "title": "Incorporating seller/buyer reputation-based system in blockchain-enabled emission trading application" }, { "paperId": "35a038a3d077b489d27eff403901373f2448eb4c", "title": "Blockchain in Logistics: Mapping the Opportunities in Construction Industry" }, { "paperId": "392fe6f21d9b735c719a742ed987702b893824dd", "title": "Designing microgrid energy markets A case study: The Brooklyn Microgrid" }, { "paperId": "921f24569bcf47c3743e5bfad6536371deb35fcd", "title": "Potentials of Blockchain Technology for Construction Management" }, { "paperId": null, "title": "Description of Scimago Journal Rank Indicator" }, { "paperId": null, "title": "ScimagoResearchGroup , Description of Scimago Journal Rank Indicator , Scimago Research Group , 2007" }, { "paperId": null, "title": "Banking is Only the Beginning: 58 Big Industries Blockchain Could Transform" } ]
38,524
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/012123313b7676f86c331a3e62bd39dec6fa9771
[]
0.900354
Blockchain and Interplanetary File System (IPFS)-Based Data Storage System for Vehicular Networks with Keyword Search Capability
012123313b7676f86c331a3e62bd39dec6fa9771
Electronics
[ { "authorId": "2212473323", "name": "N. Sangeeta" }, { "authorId": "33509446", "name": "S. Nam" } ]
{ "alternate_issns": [ "2079-9292", "0883-4989" ], "alternate_names": null, "alternate_urls": [ "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-247562", "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-247562", "https://www.mdpi.com/journal/electronics" ], "id": "ccd8e532-73c6-414f-bc91-271bbb2933e2", "issn": "1450-5843", "name": "Electronics", "type": "journal", "url": "http://www.electronics.etfbl.net/" }
Closed-circuit television (CCTV) cameras and black boxes are indispensable for road safety and accident management. Visible highway surveillance cameras can promote safe driving habits while discouraging moving violations. According to CCTV laws, footage captured by roadside cameras must be securely stored, and authorized persons can access it. Footages collected by CCTV and Blackbox are usually saved to the camera’s microSD card, the cloud, or hard drives locally but there are concerns about security and data integrity. These issues may be addressed by blockchain technology. The cost of storing data on the blockchain, on the other hand, is prohibitively expensive. We can have decentralized and cost-effective storage with the interplanetary file system (IPFS) project. It is a file-sharing protocol that stores and distributes data in a distributed file system. We propose a decentralized IPFS and blockchain-based application for distributed file storage. It is possible to upload various types of files into our decentralized application (DApp), and hashes of the uploaded files are permanently saved on the Ethereum blockchain with the help of smart contracts. Because it cannot be removed, it is immutable. By clicking on the file description, we can also view the file. DApp also includes a keyword search feature to assist us in quickly locating sensitive information. We used Ethers.js’ smart contract event listener and contract.queryFilter to filter and read data from the blockchain. The smart contract events are then written to a text file for our DApp’s keyword search functionality. Our experiment demonstrates that our DApp is resilient to system failure while preserving the transparency and integrity of data due to the immutability of blockchain.
# electronics _Article_ ## Blockchain and Interplanetary File System (IPFS)-Based Data Storage System for Vehicular Networks with Keyword Search Capability **N. Sangeeta and Seung Yeob Nam *** Department of Information and Communication Engineering, Yeungnam University, Gyeongsan 38541, Republic of Korea *** Correspondence: synam@ynu.ac.kr** **Citation: N. Sangeeta; Nam, S.Y.** Blockchain and Interplanetary File System (IPFS)-Based Data Storage System for Vehicular Networks with Keyword Search Capability. _[Electronics 2023, 12, 1545. https://](https://doi.org/10.3390/electronics12071545)_ [doi.org/10.3390/electronics12071545](https://doi.org/10.3390/electronics12071545) Academic Editor: Hamed Taherdoost Received: 17 February 2023 Revised: 22 March 2023 Accepted: 22 March 2023 Published: 24 March 2023 **Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **Abstract: Closed-circuit television (CCTV) cameras and black boxes are indispensable for road safety** and accident management. Visible highway surveillance cameras can promote safe driving habits while discouraging moving violations. According to CCTV laws, footage captured by roadside cameras must be securely stored, and authorized persons can access it. Footages collected by CCTV and Blackbox are usually saved to the camera’s microSD card, the cloud, or hard drives locally but there are concerns about security and data integrity. These issues may be addressed by blockchain technology. The cost of storing data on the blockchain, on the other hand, is prohibitively expensive. We can have decentralized and cost-effective storage with the interplanetary file system (IPFS) project. It is a file-sharing protocol that stores and distributes data in a distributed file system. We propose a decentralized IPFS and blockchain-based application for distributed file storage. It is possible to upload various types of files into our decentralized application (DApp), and hashes of the uploaded files are permanently saved on the Ethereum blockchain with the help of smart contracts. Because it cannot be removed, it is immutable. By clicking on the file description, we can also view the file. DApp also includes a keyword search feature to assist us in quickly locating sensitive information. We used Ethers.js’ smart contract event listener and contract.queryFilter to filter and read data from the blockchain. The smart contract events are then written to a text file for our DApp’s keyword search functionality. Our experiment demonstrates that our DApp is resilient to system failure while preserving the transparency and integrity of data due to the immutability of blockchain. **Keywords: blockchain; Ethereum blockchain; decentralized application (DApp); interplanetary file** system (IPFS); smart contracts **1. Introduction and Background** CCTV camera images are a valuable source of traffic surveillance that supplements other traffic control measures. CCTV is aimed at helping in the detection and prevention of criminal activity. It can be helpful in protecting the citizens in the community. It is placed in public areas to provide evidence to appropriate law enforcement agencies. CCTV cameras can be found on busy roads, atop traffic lights, and at highway intersections. Operators detect and monitor traffic incidents using images from CCTV cameras. It may be possible to predict the duration of a traffic incident based on prior experience and traffic modeling techniques. Cameras are used to observe and monitor traffic, as well as to record traffic pattern data. Moving violation tickets are even issued using cameras. The vehicle’s event data recorder is constantly recording information in a loop while we are driving, at least until a collision occurs. Black boxes save data collected at the time of impact, as well as 5 s before and after the event. The black boxes will record all human contact with the vehicle. The data collected helps us understand the reasons for collisions and prevent them from happening again. ----- _Electronics 2023, 12, 1545_ 2 of 23 CCTV footage is being used in crime investigations by police officers and insurance companies [1] all over the world. Recorded footage is typically used by investigators to locate or confirm the identity of a suspect. Real-time surveillance systems allow employees or law enforcement officials to detect and monitor any threat in real time. Then, there’s the archival footage record, which can be reviewed later if a crime or other issue is discovered. In these cases, the recorded footage must be securely deposited and kept for future use, making video storage a critical component of any video camera security system. The vast majority of information collected by surveillance cameras and dashboard cameras is securely kept on hard drives as well as memory cards. The amount of storage on the MicroSD card of our security camera, on the other hand, is determined by the amount of activity recorded by our camera. This type of storage necessitates a large amount of storage space and exposes our data to risk if the device’s hard drive fails or is damaged. It is critical to securely store CCTV and black box footage in order for it to be available and unaltered at all times. In many cases, the introduction and popularity of IP camera cloud storage have reduced the importance of local storage to a secondary option. Cloud systems are an extremely good tool that offers us many advantages and functionalities. Cloud storage systems, on the other hand, have flaws such as problems with data safety [2,3], centralized data storage, and the requirement for trusted third parties. Owners are reassured of the burden of maintaining local data storage, but they end up losing direct control over storage reliability and protection. Every year, large database hacks cost millions of dollars. Furthermore, because the data is kept on an external device, the owners have no power over it; if the service provider disconnects or limits access, they will not be able to access it. Due to the centralized nature of cloud storage data, an intruder to servers is able to view and alter it. Cloud data is untrustworthy and can be altered or removed at any time. As a result, making sure data security [4] and safeguarding users’ privacy [5] are critical. Users are usually needed to cover the cost of any storage plan they select, even if they only use a portion of it. Even the finest cloud service providers can face such a challenging problem while retaining strong maintenance standards. Centralized storage service providers occasionally fail to deliver the security service as agreed. For example, a hack on Dropbox [6] which is among the world’s largest online storage companies, did result in the leak of 68 million usernames and passwords on the dark web. Well-known cloud services have started experiencing blackouts and security breaches. The mass email deletion event of Gmail [7], Amazon S3’s latest shutdown [8], and the post-launch interruption of Apple MobileMe’s [9] are other examples. Blockchain technology may be able to address these issues. A blockchain is made up of a cryptographic algorithm, a timestamp, and transaction information that connects it to the preceding block. As a result, every block links to the next, forming a “chain” of blocks and producing safe and immutable records. In comparison, the blockchain is not designed for the purpose of file storage. The cost of keeping data on the blockchain is exorbitantly high. We can have decentralized as well as low-cost storage with the IPFS project [10]. Peer-to-peer networks provide greater security than centralized networks. As an outcome, they are ideal for protecting sensitive information from malicious actors. We propose an IPFS-based distributed and decentralized storage application that offers more storage space compared to solely blockchain-based systems in this paper. Using distributed storage, information is kept on different nodes or servers on the Internet. To upload files, we use the Geth [11] software client to operate an Ethereum node and an IPFS Daemon server to operate our own IPFS node. Users will link to the DApp through the use of one‘s web browser as well as a blockchain wallet, Metamask [12], to connect to a blockchain in our proposed scheme. Since it is powered by Ethereum smart contracts, the decentralized application will interact with the blockchain, which will keep all the code of the application in addition to the data. The smart contracts keep track of all sources of information in IPFS files. A DApp can receive any kind of information. The hash value of the uploaded file is permanently saved on the Ethereum blockchain via smart contracts and ----- _Electronics 2023, 12, 1545_ 3 of 23 it cannot be changed or deleted. Whenever a file is uploaded, the DApp hears the event “File Upload” and updates the DApp’s user interface. We retrieve all of the smart contract events and reveal them on our DApp, which is called the “smart contract event log”. The smart contract event log contains data such as the file name, file summary (including the event and location of the file), file type, size of the file, time and date of upload, Ethereum account information of the user, and the hash value of the file once it has been uploaded to IPFS. Users can also view the file by clicking on its description. The user does not need to remember and save the hash value independently, which could be dangerous if another individual has access to it. Our DApp also includes a keyword search feature to assist you in quickly locating sensitive information. Figure 1 shows an example scenario where our proposed system can be applied. When an accident occurs, our proposed system might be used to save the video taken by the dashboard camera on IPFS and the hash value of the video on the blockchain to prevent the manipulation of the video using the immutability property of blockchain. **Figure 1. Example scenario to illustrate the application of the proposed system in the presence of** accidents on the road. The key contributions of our paper can be summarized as follows: - Our proposed distributed storage application supports the storage of various file types since uploaded files are stored on IPFS and their hash values are stored in smart contracts on the Ethereum blockchain. Users need not remember the hash values since they can be retrieved from the blockchain later. - DApp provides a keyword search feature to help users quickly find the necessary files based on Ethers.js’s smart contract event listener and contract.queryFilter. - Our experiment shows that our DApp is resilient to system failure, and our system provides better transparency than is possible with centrally managed applications. The rest of our paper is structured as follows: Section 2 contains related work. Section 3 contains preliminary information. The proposed scheme is described in Section 4. Section 5 goes over implementation. The performance evaluation results are described in Section 6. Finally, Section 7 brings the paper to a close. ----- _Electronics 2023, 12, 1545_ 4 of 23 **2. Related Work** Hao, J. et al. [13] studied a blockchain and IPFS-based storage scheme for agricultural product tracking. During the manufacturing, processing, and logistics processes, sensors collect real-time data on product quality as well as video and picture data, according to this study. The server parses and encapsulates the data before writing it to IPFS, and the hash address is then stored in the blockchain to complete the data storage. The collected data is not directly written to an IPFS. The authors employ a private data server, and data collected by sensors is first stored on the private data server before being directly stored on the IPFS. If the server experiences problems, such as server failure, the collected data is lost, and the server is unable to write data to IPFS. There is no keyword search function for quickly finding agricultural product information. Rajalakshmi et al. [14] proposed a framework for access control methods in research records that manages to combine blockchain, IPFS, as well as other traditional encryption methods. The system stores the verification metadata information acquired from the IPFS on the blockchain network using Ethereum smart contracts, resulting in tamper-proof record-keeping for further auditing. There is no keyword search functionality for searching information related to research records in this proposed scheme, which only stores PDF files. Vimal, S. et al. [15] proposed a method to improve the efficiency of the P2P file-sharing system by incorporating trustworthiness and proximity awareness during file transfer using IPFS and blockchain. Any of these hashed files can be retrieved by simply calling the hash of the file. Miners who collaborate to ensure the successful transfer of resources are compensated. This study discusses the file transfer service, as well as the security strength and some of the IPFS-based incentives. This system is built around IPFS and Blockchain. Yongle Chen et al. [16] proposed a more efficient P2P file system scheme. The authors pointed out the high-throughput problem for individual IPFS users by incorporating the responsibility of content service providers. A novel zigzag-based storage model is utilized to improve the IPFS block storage model by taking data reliability and availability, storage overhead, and other issues for service providers into account. Rong Wang et al. proposed a video surveillance system relying on permissioned blockchains (BCs) and edge computing in their paper [17]. Convolutional neural networks (CNN), edge computing, and permissioned BCs, as well as IPFS technology, were used in this system. Edge computing was utilized to collect and process large amounts of wireless sensor data, while the IPFS storage solution was utilized to enable huge video data storage. CNN technology was applied to real-time monitoring, and Edge computing was utilized to gather and analyze large amounts of wireless sensor data. Sun, J. et al. [18] proposed a blockchain-based secure storage and access scheme for electronic medical records in IPFS, which ensures necessary access to electronic medical data while preserving retrieval efficiency. IPFS is a file system used in order to store encrypted electronic medical data. After receiving the hash value and encrypted hash address, the physician needs to be encrypted using the hash value and encoded hash address with a random number, hash the health information and index with the SHA256 hash function, and broadcast the hash value and encoded hash address to the blockchain. Furthermore, the system offers targeted defense against relevant keyword attacks. Medical data is not directly stored on IPFS, and electronic health data is encrypted before being stored on IPFS. It also takes time for the IPFS value to be encrypted before even being kept on the blockchain. Most of the previous works lack a keyword search functionality for quickly locating relevant information. They do not mention how to retrieve the metadata from the blockchain. It is not possible to retrieve data from IPFS without the hash value of the file. Table 1 compares our proposed system with existing approaches. ----- _Electronics 2023, 12, 1545_ 5 of 23 **Table 1. Comparison of existing approaches with the proposed scheme.** **Constraints** **Hao, J. et al. [13]** **Rajalakshmi A. [14]** **Sun, J. et al. [18]** **Our Proposed Scheme** High delay Collected High delay Encryption Delay data is not directly Low delay of medical data written to an IPFS Tampering on the Possibilities of data No tampering No tampering stored data tampering More storage capacity Less Storage capacity Storage capacity More storage capacity as the data stored on Stored on data server IPFS Low delay in uploading files to IPFS and file hash is automatically stored on BC with help of Smart contract No tampering of data as data is stored on IPFS and hash on Blockchain More storage capacity as the data stored on IPFS Uploading only video Only electronic medical Heterogeneous data Heterogeneous data Uploading only PDF’s and images on IPFS record upload Keyword Search No Keyword search No Keyword search No Keyword search Supports Keyword function function function function search function **3. Preliminaries** _3.1. IPFS_ The interplanetary file system is a distributed file system protocol developed by Joan Bennett in 2015 and managed by Protocol Labs. The IPFS network consists of computers running the IPFS client software. Anyone can join the IPFS network, either as an IPFS node running the IPFS client or as a network user storing and retrieving files. Any type of file can be stored, including text, music, video, and images, which is especially useful for non-fungible tokens (NFTs). In contrast to HTTP, data in IPFS is identified by content rather than location. When we upload a file to IPFS, a hash of the content is generated. This hash identifies the content uniquely and can be used to retrieve the file. If we upload a different file, the hash will be completely different, but we can always recompute the file’s hash locally to ensure it matches the original IPFS hash. We selected the IPFS protocol in our proposed scheme because it is a well-known and working decentralized file storage protocol. _3.2. Ethereum_ Ethereum [19] is, at its core, a decentralized global software platform that utilizes blockchain technology. It is most well-known for its native cryptocurrency, ether, abbreviated as ETH. Anyone can use Ethereum to start creating any protected digital technology. It has a token intended to be utilized by the blockchain network, but it may also be employed to pay participants for blockchain work. It is a platform for various DApps that can be deployed through smart contracts. An Ethereum Private Network is a blockchain that is completely separate from the main Ethereum network. The Ethereum Private Network is primarily used by organizations to limit blockchain read permissions. _3.3. Web3.js_ Web3.js [20] is a set of libraries that allows developers to communicate with a remote or local Ethereum node via HTTP, IPC, or WebSocket. You can use this library to create websites or clients that communicate with the blockchain. _3.4. Ethers.js_ Ethers.js [21] connects to Ethereum nodes using Alchemy, JSON-RPC, Etherscan, Infura, Metamask, or Cloudflare. Developers can use ethers. js to take advantage of full functionality for their various Ethereum needs. ----- _Electronics 2023, 12, 1545_ 6 of 23 _3.5. Smart Contract_ Smart contracts are programs that are implemented and stored on a blockchain when certain requirements are fulfilled. They are frequently used to automate agreement execution so that all groups have instant surety of the results even without the involvement of an additional party. They also can automate a workflow by automatically performing the next action if certain requirements are fulfilled. _3.6. Smart Contract Events_ When a transaction is mined, smart contracts could also emit events and logs to the blockchain, which the front end can then process. Events are essential on any blockchain because they make connections between smart contracts, which are self-executing software programs that have the terms of the buyer’s and seller’s agreement straight integrated into lines of code for response with user interfaces. To use a smart contract, a user must first manually sign a transaction and interact with the blockchain. This is where automation can help users by simplifying things. Event-driven automation initiates processes without requiring human intervention. An automation tool can start a predefined process or workflow of smart contracts after detecting an event. _3.7. Decentralized Applications (DApp)_ A decentralized application [22] is an application that can run autonomously, typically using smart contracts and running on a decentralized computing, blockchain, or other distributed ledger system. DApps, like traditional applications, provide some function or utility to their users. _3.8. React.js_ React.js [23], also known as simply React, is a free and open-source JavaScript library. It is best to create user interfaces by combining code sections (components) into complete websites. We can use React as much or as little as we want. React enables developers to use separate software components across the client and server sides, which also speeds up development. _3.9. Dependencies_ 3.9.1. Node Package Manager (NPM) The node package manager (NPM) is a command-line tool for installing, updating, and removing Node.js packages from our application. It also serves as a repository for open-source Node.js packages. A package manager is essentially a set of software tools that can be used by a developer to automate and standardize package management. 3.9.2. Node.js Node.js is a simple programming language that can be used for prototyping and agile development, as well as to create extremely fast and scalable services. 3.9.3. MetaMask MetaMask is a non-custodial Ethereum-based decentralized wallet that also lets users save, buy, send, transform, and swap crypto tokens, as well as sign transactions. Using Metamask in conjunction with Web3.js in a web interface simplifies communication with the Ethereum network. 3.9.4. Truffle Framework Truffle is a set of tools that allows us to create smart contracts, write tests against them, and deploy them to blockchains. It also provides a development console and allows us to create client-side applications within our project. Truffle is the most widely used framework for creating smart contracts. It supports Solidity and Viper as smart contract languages. Truffle has three main functions: it compiles, deploys, and tests smart contracts. ----- _Electronics 2023, 12, 1545_ 7 of 23 **4. Proposed Data Storing Scheme** Our proposed scheme divides data storage, retrieval, and searching into four steps. The system uploads a file, file hash is stored on the blockchain, monitors smart contract events, and searches for relevant information. _4.1. File Uploading_ The main concept of the file uploading process is depicted in Figure 2. The file is selected from the DApp (browser) (1), and when the DApp form’s submit button is clicked, the uploaded file is stored on IPFS (2). The hash of the file uploaded is returned to the DApp (3); this hash is the file’s location. The file’s hash is saved to a smart contract (4), which is subsequently kept on the blockchain (5), and the hash and other information of the uploaded file were also listed on the DApp (6), from which we can obtain all of the files we have uploaded to IPFS. **Figure 2. File Upload.** To connect to an Ethereum wallet Metamask, we used a web browser as a front end which will communicate with the blockchain and store the smart contract on it. We will upload the file directly to an IPFS, and then IPFS will return to us a hash. We will then store this hash on the smart contract, and it will store that hash on the blockchain, allowing us to access all of the files we have created when we list them on the DApp. A smart contract stores the hash value on the blockchain, and another smart contract lists the uploaded files on the DApp. The smart contract handles file uploading, file storage, and file listing. Figures 3 and 4 show our smart contract. Our project’s smart contract is responsible for four tasks. Define a data structure for file management, upload the files, store the file hash in the blockchain, and display the uploaded files on the DApp. We use a struct to manage the files inside Solidity. Solidity structs allow us to create more complex data types with multiple properties. By creating a struct, we can define our own type. They are useful for organizing related data. Structures can be declared outside of one contract and imported into another. ----- _Electronics 2023, 12, 1545_ 8 of 23 **Figure 3. Solidity code for creation of a blockchain register and events to facilitate interoperability (1/2).** **Figure 4. Solidity code for creation of a blockchain register and events to facilitate interoperability (2/2).** The following steps show the tasks of a smart contract: (i) Define data structure for the management of files: Figure 3 shows step one in modeling the file (6). We created a file object, and inside we defined a unit id, which will be the unique identifier for the file inside our smart contract. The string will be the hash of the file, and this will be its location on IPFS, and a description of the file, which contains the location of the file and events related to the uploaded file. The address-payable uploader is the person who uploads the file, and it is the Ethereum address of that person’s wallet address as they are connected to the blockchain; it is like their username on the blockchain. (ii) Store and list the files: Step two is to store the file on IPFS, and step three is to list the event logs on the DApp. We used mapping inside of Solidity to store the files, as shown in Figure 3. Mapping is another data structure. It can be utilized to store data as key-value pairs, with the key being any of the built-in data types but just not reference types, as well as the value being any type. We created mapping (5) as shown in Figure 3. A mapping inside of Solidity is ----- _Electronics 2023, 12, 1545_ 9 of 23 just a key-value store. We can give it a key and a value. The data type of the key in our smart contract is an unsigned integer, and the return value is file struct (6), as shown in Figure 2. When we place a file with an id within this mapping, it will write and store it on the blockchain. Mapping is also going to give us the ability to list the files because mapping is public, and thus it gives us a function called “files” (5) that we can call, pass in the id, and fetch out each individual file. We can get back a file with all the data, such as the id, hash, file name, description, and uploader. (iii) Upload File: The solidity code has a function called fileUpload (8). “fileUpload” takes the following arguments: fileHash, fileSize, fileType, fileName, fileDescription. Whenever we upload a new file, we will just add a new file to the mapping. We created a new file (6) and put it inside the file’s mapping (5). We are going to store the file based on the id inside the mapping, as shown in (5). We stored the file onto the blockchain as shown in (11). Inside the smart contract, Solidity has a global variable called “msg” or “message” that has many different attributes, one of which is the person calling the function, “message sender” is the Ethereum address of the person uploading the file. We created a video struct and saved it inside the “files mapping”, which we simply say “files”, pass in the id, and it will be equal to a new file (11). **fileCount (4) is a variable that stores the number of files that have been created.** Whenever we create the smart contract, the counter value will be zero, but we can change this value inside the function (11) as fileCount anytime the function is called. We could write fileCount ++ (10) and then pass in fileCount in (11). fileCount keeps track of all the files; it is basically our ID management system, and we save it inside the file mapping, which acts like our database. (iv) Creating an Event: The event allows us to know when the file was uploaded. We can create events from the Solidity code. We define an event called “fileUpload” and we pass in the same arguments as the struct (7); this is going to allow us to subscribe to the event whenever it is triggered from our application. We can trigger the upload event (12). We use the emit keyword, then FileUploaded which has the same name as the event (7) and we pass in the arguments file count, fileHash, fileSize, fileType, filename, file description, and now, msg.sender. Next, we added some requirements to the function to make it robust. We can use Solidity’s **require function (9). The require function checks that a set of parameters is true before the rest** of the function executes. Table 2 shows the list of variables used in our smart contract. **Table 2. Smart Contract variables.** **Variables** **Why It Is Used** Keeps track of how many files have been fileCount added to the current smart contract. mapping File key value store and lists the files struct Manage the files event FileUploaded Allows us to know when the file was uploaded function fileUpload Uploads new file emit FileUploaded Trigger an event Recently, diverse types of formal methods are investigated to enhance the security of smart contracts, since the compromise of smart contracts can lead to a catastrophic monetary loss [24]. However, our smart contract codes have not been analyzed using those formal methods yet, and we will verify our codes in our future work. Our first project element is a private Ethereum blockchain that will act as the back end for our DApp. Ethereum nodes maintain an archive of the blockchain’s code. The information is dispersed throughout the network. The Geth is utilized to run an Ethereum node. ----- _Electronics 2023, 12, 1545_ 10 of 23 By running a node on the Ethereum network, we could also perform transactions as well as communicate with smart contracts. The uploaded file’s hash is saved in a smart contract, and then immutably stored also on the Ethereum blockchain. The next component is IPFS, which enables us to keep files in a distributed fashion. Because files are large, storing megabytes and gigabytes of files on the blockchain may not be feasible. This is where IPFS comes into play. It has nodes, just like Ethereum, and we distribute files that cannot be tampered with across the network. IPFS uses hashes. When you upload a file to IPFS, it will be stored somewhere and identified by its hash. We run our own IPFS node, which supports an IPFS gateway for file retrieval and storage and runs the IPFS Daemon server. We cannot store or retrieve data unless the Daemon server is up and running, or unless we link to public gateways such as Infura [25]. When a user uploads CCTV footage to our DApp, they can specify the location as well as event details such as whether it was an accident or a traffic violation. This information is fed into the DApp as a file description. This information is critical when uploading a video to the DApp because users can quickly search for location and event information using the DApp’s keyword search function. We first must import and link our Ethereum blockchain account to Metamask before we can use the DApp. Our web browser now supports blockchain networks, and we can upload files to IPFS using our custom-designed DApp user interface (UI). First, we must select the file, enter its description (such as file event and location), and then click the submit button. When we click the submit button, the file is sent to IPFS and we receive the IPFS result, which contains the hash value and path of the file. Metamask directs us to accept the transaction, save the hash in a smart contract, and store the smart contract on the blockchain via a confirmation pop-up. To store the hash on the blockchain, we should pay some gas in the manner of ethers. When we confirm the Metamask transaction, the hash of the uploaded file is preserved on the Ethereum blockchain. The DApp monitors the “file upload” event and updates the DApp’s User interface automatically. The event log of the smart contract is generated by retrieving and displaying all events from the smart contract within our DApp. The smart contract event log includes the file no, file description, type of file, file size, timestamp, Ethereum information of the uploaded person, and the hash value of the file after it has been stored in IPFS. By clicking on the file description, individuals may view the uploaded files in their web browser. The hash value does not need to be remembered or stored separately by the user. _4.2. Keyword Searching_ Users of the blockchain network can view transaction details but cannot identify the individuals who made the transactions. On our DApp, we can see the transactions and use the data for keyword searching. (i) Read information from the blockchain: When events occur in the smart contract, the smart contract emits events in order to communicate with DApps and other smart contracts. When we invoke a smart contract function, it has the ability to generate an event. It is critical for us to be able to listen to these events in real time when developing DApps. To listen for smart contract events, we used Ethers.js smart contract event listener. To communicate with a smart contract using Ethers.js, we must first create a new contract object with Ethers as shown in step (1) of Figure 5. ----- _Electronics 2023, 12, 1545_ 11 of 23 **Figure 5. Ethers.js filter to read events from blockchain.** As shown in steps (2), (3), and (4) of Figure 5, we need the blockchain address for the smart contract, the ABI of the smart contract, and the signer or provider (4). The ABI is a JSON object that describes how the smart contract works; it describes the interface, which essentially means what functions the smart contract has, what function arguments it accepts, and what it responds to when we try to read data from it. Ether.js allows us to store ABIs as an array and only pull in the parts we want when we are setting up a smart contract object. We require file upload information for our project, so we included ABI, which is related to the file upload event. Then, we need a provider or a signer; in our project, we have a provider. A provider is an abstraction of an Ethereum network connection that provides a concise, consistent interface to standard Ethereum node functionality. We take our smart contract ABI and create a new contract address ABI, and then we provide all of the required information as shown in step (5) of Figure 5. We used contract.queryFilter to filter the information, as shown in step (6) of Figure 5. Using this command, we will examine every single FileUploaded event that has ever occurred on our blockchain. We include this filter to reduce the search space inside the Ethereum blockchain. Ethers.js allows us to examine the FileUploaded events and specify which blocks we want to examine as shown in step (7) of Figure 5. (ii) Keyword search text file creation: We can create a text file for keyword searches once the events are retrieved from the blockchain. The smart contract events are written into a text file. We store only necessary information in the text file, e.g., information such as file name, event type, location, Ethereum account number, and smart contract. Figure 6 shows how to retrieve data from the blockchain and conduct keyword searches. To listen to smart contract events, we used a command prompt to send requests to the blockchain (1). Blockchain responded with a filtered smart contract event log containing all of the information about the uploaded file, including the smart contract address, file name, file hash and description, Ethereum address of the uploaded file, and so on (2). When we received a smart contract event log, we saved some of the event logs in a text file (3). We wrote code in react.js to filter the results and search for keywords on the DApp. When a user searches for a keyword on the DApp, the request is sent to a text file containing smart contract events, which is then filtered, and the result is returned to the DApp (4). Users can look up a word or an alphabet. ----- _Electronics 2023, 12, 1545_ 12 of 23 **Figure 6. Keyword Search Function of the proposed DApp.** **5. Implementation** On the Windows 10 operating system, we used a private Ethereum blockchain to implement a proposed scheme. The Ethereum core network is not connected to a private Ethereum network. Organizations primarily use it to limit blockchain read permissions. Installing geth/parity allows the current node to join the Ethereum network and download the blockchain to local storage. We used Go Ethereum to create our Ethereum blockchain (Geth). _5.1. Steps to Create Private Ethereum Network_ The following steps show how we built our private Ethereum network: 5.1.1. Download “Geth” Go Ethereum (Geth) can be directly downloaded and installed from geth.ethereum. **org, accessed on 16 February 2023. Because Geth is a command line interface, we execute** all commands from the command line. After installing Geth on our system, we typed geth and pressed enter in a command prompt and obtained the output as shown in Figure 7. **Figure 7. Geth command.** We used the geth command to connect to a blockchain, and the geth command will run in fast sync mode. Fast sync is Geth’s current default sync mode. Fast Sync nodes download the headers of each block and retrieve all the nodes beneath them until they ----- _Electronics 2023, 12, 1545_ 13 of 23 reach the leaves. Instead of reprocessing all transactions that have ever taken place, fast sync downloads the blocks but only validates the affiliated proof-of-works (which could take weeks). When we stop and restart the geth, it will operate in full sync mode. Full sync needs to download all blocks and incrementally generate the blockchain state by running each block since genesis. The data size of the Ethereum blockchain is currently around 800–1000 gigabytes, and we do not need to download the entire Ethereum blockchain on our system. 5.1.2. Make a Folder for Our Private Ethereum Network For the private Ethereum network, we created a separate folder called “Private Ethereum”. This folder separates the Ethereum private network files from the public files. 5.1.3. Construct a Genesis Block In blockchain, all transactions are recorded in the form of blocks in sequential order. There are an infinite number of blocks, but there is always one distinct block that gives rise to the entire chain, known as the genesis. The genesis block, also known as Block 0 or Block 1, is the first block ever recorded on its respective blockchain network. There are no transactions. The genesis block is used to initialize the blockchain, as shown in Figure 8. A genesis block is required to create a private blockchain. The genesis block can be created with any text editor and saved with the JSON extension in the Private Ethereum folder. Figure 9 shows the genesis block file. **Figure 8. Genesis block in a blockchain.** **Figure 9. Genesis block file.** 5.1.4. Run the Genesis File To extract the genesis file, we open the Private Ethereum folder in Visual Studio Code and run the command geth init ./genesis.json -datadir eth. Eth is the name of a folder. Geth is connected to the genesis file after running the above command. ----- _Electronics 2023, 12, 1545_ 14 of 23 5.1.5. Set Up the Private Network We created a private network in which multiple nodes can add new blocks. We must use the command geth –datadir ./eth/ –nodiscover to accomplish this. When–nodiscover is used to start a geth node, it prevents the node from being discovered by the network’s bootnode. Every time the private network chain is needed, commands in the console must be executed to connect to the genesis file and the private network. A private Ethereum network and a personal blockchain are now available. Figure 10 shows the running status of a private Ethereum network. **Figure 10. Private Ethereum network.** 5.1.6. Make Externally Owned Account (EOA) EOAs are controlled by users who have access to the account’s private keys. These accounts, which can both send transactions and trigger contract accounts, are typically used in conjunction with a wallet. To manage the blockchain network, EOA is required. To make it, we launched Geth in two windows. One terminal to run Geth as shown in Figure 10 and another terminal to create EOA. We entered the command geth attach \\.\pipe\geth.ipc in the second terminal (console window). This will connect the second terminal to the private Ethereum network established in Figure 10. We used the command personal.newAccount() to create a new account. After executing this command, we entered our password to obtain our account number and saved it for future use as shown in Figure 11. **Figure 11. Externally owned account, Mining Start and Stop.** 5.1.7. Ethereum Mining on Our Private Chain If we mine on the Ethereum main chain, we will need expensive equipment with powerful graphics processors. ASICs are typically used for this but high performance is not ----- _Electronics 2023, 12, 1545_ 15 of 23 required in our private network, and we can begin mining with the command miner.start () as shown in Figure 11. After a few seconds, some ether was found in the default account if the balance status is checked as shown in Figure 11. To check the balance, we used the command **eth.getBalance(eth.accounts[0]). Figure 12 shows the mining process. We used the com-** mand miner.stop() to stop mining as shown in Figure 11. **Figure 12. Mining Process.** 5.1.8. Connecting the Private Ethereum Network to Metamask We closed the terminal in which our private network was running and opened a new terminal and typed the command geth –datadir ./eth/ –nodiscover –http –http.addr “local**host” –http.port “8545” –http.corsdomain=“*” –http.api web3,eth,debug,personal,net –** **ws.api web3,eth,debug,personal,net –networkid 7777 –allow-insecure-unlock, as shown** in the Figure 13 and now our private Ethereum is connected to Metamask. Explanation of the used commands as follows: – http.addr value: Listening interface for HTTP-RPC servers (default: “localhost”). – http.port value: Listening port for HTTP-RPC server (default: 8545). – http.corsdomain value: A list of domains separated by commas that will accept cross-origin queries (browser enforced). Because the HTTP server can be accessed from any local application, the server includes additional safeguards to prevent API abuse from web pages. The server must be configured to accept Cross-Origin requests in order to allow API access from a web page. The —http.corsdomain flag is used to accomplish this. The —http.corsdomain command accepts wildcards, allowing access to the RPC from any location: —corsdomain ’*’. – http.api value: APIs accessible via the HTTP-RPC protocol. – ws.api value: APIs accessible via the WS-RPC interface. – nodiscover: The peer discovery mechanism is disabled. – networkid value: Sets network id explicitly. – allow-insecure-unlock: When account-related RPCs are exposed via http, this allows for insecure account unlocking. ----- _Electronics 2023, 12, 1545_ 16 of 23 **Figure 13. Importing Ethereum account in Metamask.** We launched Metamask and added the Network “Local Host 8545” with the Chain ID “2022”. It is the chain ID we specified in our private Ethereum network’s genesis block. By importing a JSON file from our private Ethereum folder, we imported a private Ethereum account. The JSON file can be found in the keystore’s Private Network folder. Figure 13 depicts how to add a Private Ethereum account to Metamask. _5.2. Running Our Own IPFS Node_ To store information on IPFS, we must run an IPFS Daemon server on our own IPFS node. To use IPFS, we must first download and install the Go language from the golang website, then go to the IPFS command line install page and download “install go-ipfs”. We navigate to the download path, extract the files to C drive, and then run ipfs.exe to start the Daemon server, as shown in Figure 14. **Figure 14. IPFS Execution and Daemon server.** _5.3. Deploying Smart Contract_ A smart contract stores the hash of the uploaded file. To make smart contracts in the Solidity programming language, the Truffle framework is used. The Truffle Suite is a collection of tools specifically designed for Ethereum blockchain development. The suite includes three pieces of software. Truffle is capable of helping compile and deploy smart contracts in addition to injecting them into web apps and building DApp front ends. Truffle is now a popular Ethereum Blockchain IDE. ----- _Electronics 2023, 12, 1545_ 17 of 23 _5.4. File Uploading and Retrieving_ After writing the smart contract, deploying, and publishing it to our Ethereum blockchain, we then utilize Metamask to connect our DApp to the Ethereum blockchain. A Metamask is required to communicate with the blockchain. The client-side application, which is also going to communicate with IPFS, was built with React. Figures 15 and 16 show how we initially deployed the smart contract to Ethereum, then launched the DApp with the command npm run start, imported an Ethereum account into Metamask, and linked Metamask to our DApp. Figures 17 and 18 show how to submit a file to IPFS, deposit the file’s hash in a smart contract, record the smart contract on the Ethereum blockchain, and successfully retrieve the file using our DApp. **Figure 15. Smart contract deploy.** **Figure 16. Connecting Metamask to the DApp.** We chose the file and entered the location as well as the location of the file in the user interface of DApp after logging into Metamask, then clicked the submit button also confirmed the transaction of Metamask, as shown in Figure 17. To deploy the smart contract, upload files, and store hash values on the blockchain, we start and maintain mining. ----- _Electronics 2023, 12, 1545_ 18 of 23 **Figure 17. Choosing a file and confirming Metamask transaction.** As the transaction is confirmed, the DApp listens for the event “File Upload” and updates the DApp’s user interface automatically. Whenever a transaction has been mined, smart contracts generate events and logs to the blockchain, which can then be processed by the front end. Our DApp retrieves and displays all smart contract events. It is referred to as a “smart contract event log”. The event log of the smart contract contains the file number, file description (which includes an event and location of the file), type of the file, file size, date and time, the uploader’s Ethereum account details, and the hash of the file. By having to click on the file’s file details, users are able to view the uploaded files through their web browser. Figure 18 depicts a smart contract event log and various file types retrieved. **Figure 18. Event log and file retrieve.** _5.5. Keyword Searching_ Our DApp supports the keyword search method. In order to conduct keyword searches, we obtain event information from Blockchain. Smart contracts could even emit logs as well as events to the blockchain whenever an Ethereum transaction is mined, which the front end can then process. An event broadcasts information about a file upload, and we could have access to all of the events so that we could listen to them in real time, or we could just use them to obtain all of the most recent file uploads on the blockchain. We can read smart contract events outside of the DApp’s user interface by using Web3.js or Ethers.js. In our implementation, Ethers.js is used to read smart contract events. We only have one event in our smart contract, so we use a filter to retrieve information from that event, which is File Upload. A smart contract event log is shown in Figure 19. ----- _Electronics 2023, 12, 1545_ 19 of 23 **Figure 19. Smart contract events.** The smart contract events are then written to a text file, allowing our DApp to conduct keyword searches. We store only necessary information in the text file, e.g., information such as file name, event type, location, Ethereum account number, and smart contract. When looking for sensitive information on the DApp, keyword searching is essential. Entering an alphabet or a keyword into keyword searching will filter the results to show only the keyword we entered. In the case of an alphabet search, the DApp will display all events that include the letter we typed into the search box. This method makes navigating an event easier and more efficient. Keyword searching is shown in Figure 20. **Figure 20. Keyword Searching.** **6. Performance Evaluation** The majority of applications we use today are centralized, which means they are managed by a single authority. Google [26] and Facebook [27], for example, retain complete ownership of their respective products, running their apps and storing user data on private servers and databases. While this gives Google and Facebook control over their applications and user experiences, it can also be discouraging to users. Users of centralized apps have little control over their data or experience within the app. They must have faith in the app’s developer to listen to their feedback, provide product services, and treat them and their data with dignity. However, with other centralized applications facing backlash over privacy and the monetization of user data, many users are wary of relying on them. Centralized applications run programs and store critical user information on centralized servers. The entire application may fail if a single, central server is compromised. DApps enable users to complete transactions, verify claims, and collaborate in real time without relying on a centralized intermediary. Our DApp operates on a peer-to-peer network, similar to a distributed ledger, with each network member contributing to the program. Each of the roles that a central server would normally provide, from computing power to storage, is distributed across the network. We do not need to keep and secure a central server, and users can directly participate in the app’s operation. Our system is robust to system failure. There is no single point of failure in our DApp and is distributed across a network of public nodes, with copies of critical information distributed ----- _Electronics 2023, 12, 1545_ 20 of 23 among them. The application is unaffected if one or more IPFS nodes are compromised. Even if there is a virus attack, a hardware failure, or the system is turned off, the user can still retrieve the uploaded files and perform keyword searches. When a user uploads data to IPFS, it is chopped into smaller chunks, hashed, and assigned a unique content identifier (CID), which serves as a fingerprint. This makes it faster and easier to store small amounts of data on the network. A cryptographic hash (CID) is generated for each piece of data, making each upload to the network unique and resistant to security breaches or tampering. The experiment we conducted demonstrates that our DApp is resistant to system failure, robust, and transparent. The experiments we carried out are listed below. **Scenario 1:** In Scenario 1, the system unexpectedly shuts down, and when it is restarted, the DApp’s event log vanishes, as illustrated in Figure 21. We can retrieve the event log outside of the DApp using smart contract event listeners. In Figure 19, we used Ethers.js to retrieve the event log. The data associated with the uploaded file is included in the event log. As a result, system failure has no effect on the uploaded data. **Figure 21. No event log listed on the DApp.** **Scenario 2:** The information in the keyword search text file was accidentally deleted in Scenario 2 as shown in Figure 22, and we were unable to perform the keyword search on the DApp. As demonstrated in Scenario 1, we recreated the keyword search text file using information retrieved from the smart contract event log and performed a keyword search as illustrated in Figure 23. Table 3 summarizes the scenarios of performance evaluation. **Table 3. Performance Evaluation Scenario Summarization.** **Scenario #** **Description** The system unexpectedly shuts down. When the system restarted, the DApp’s 1 event log vanished. We used ether.js to retrieve the event log. The information in the keyword search text file was accidentally deleted. We 2 recreated the keyword search text file by using information retrieved from the smart contract event log and performed a keyword search. ----- _Electronics 2023, 12, 1545_ 21 of 23 **Figure 22. Text file with no data.** **Figure 23. Keyword Search.** If a malicious actor manages to compromise the blockchain network, any changes are visible on a public network, allowing both users and developers to respond quickly. Our DApp operates on a public ledger, which means that anyone with internet access can participate in the application and network. As a result, anyone can view the transaction record and any changes made to those records. Therefore, this system can provide better transparency than centralized applications can provide. On a publicly distributed ledger, no central entity can revoke transparency, limit viewership, or censor participation. **7. Conclusions** In this paper, we present the design and implementation of a decentralized application that uses Ethereum blockchain and IPFS to store CCTV and black box footage securely and efficiently. The DApp allows users to easily manage their storage. For scalability, only hashes of the files are stored on the blockchain via smart contracts. Our proposed scheme works in a decentralized manner. When a file is uploaded, the DApp listens for the event File Upload and automatically updates the DApp’s user interface. All smart contract events are fetched and displayed on our DApp. The extracted information is called a smart contract event log, and it includes information about the file, timestamp, the uploader’s account information, and the hash of the IPFS file returned. By clicking on the file’s description, users can gain access to it. The selected file is then displayed in the web browser. DApp also includes a keyword search ----- _Electronics 2023, 12, 1545_ 22 of 23 feature to help us find any information quickly. To filter and read data from the blockchain, we used ether.js’ smart contract event listener and contract.queryFilter. We used the smart contract address as well as the smart contract’s ABI. The smart contract events are then written into a text file. The text file only contained necessary information, such as the file name, event type, location, Ethereum account number, and smart contract. Our experiment shows that our DApp is not affected by system failure. We can secure an application by managing the data in a decentralized manner. Because our DApp runs on a public ledger, anyone with internet access can participate in the application and network. As a result, anyone can view and modify the transaction record. As a result, unlike centrally managed applications, this system provides greater transparency. We anticipate that our DApp can be used in a variety of fields, such as for keeping records of student research securely at universities, the medical information of patients at hospitals, and customer information at banks due to its ability to store various file types. In our current system, the access control function is not included in the smart contract yet, and thus, the hash values of one’s files can be exposed to anyone who knows his or her smart contract address. We will investigate the access control scheme for the smart contract to resolve this issue in our future work. In addition, we will also verify the source code of our smart contract using well-known formal methods. Recently, Ethereum has been upgraded by changing its consensus mechanism from proof-of-work (PoW) to proof of stake (PoW), and this new version is also known as Ethereum 2.0. However, this new consensus mechanism has not been verified intensively compared to the PoW mechanism, and thus, we used an old version of Ethereum and its corresponding Ethereum Virtual Machine (EVM) environment in this paper. We will implement and investigate our proposed system on the new version of Ethereum in our future work. **Author Contributions: Conceptualization, N.S. and S.Y.N.; data curation, N.S.; formal analysis, N.S.;** methodology, N.S.; project administration, N.S. and S.Y.N.; resources, N.S.; software, N.S.; supervision, S.Y.N.; validation, N.S.; visualization, N.S.; writing—original draft, N.S.; writing—editing and review, N.S. and S.Y.N. All authors have read and agreed to the published version of the manuscript. **Funding: This research was supported in part by the National Research Foundation of Korea (NRF),** with a grant funded by the Korean government (MSIT) (2020R1A2C1010366). This research was supported in part by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2021R1A6A1A03039493). **Data Availability Statement: No new data were created or analyzed in this study. Data sharing is** not applicable to this article. **Conflicts of Interest: The authors declare no conflict of interest.** **References** 1. Mateen, A.; Khalid, A.; Nam, S.Y. Management of Traffic Accident Videos using IPFS and Blockchain Technology. KICS Summer _Conf. 2022, 1, 1366–1368._ 2. [Singh, A.; Chatterjee, K. Cloud security issues and challenges: A survey. J. Netw. Comput. Appl. 2017, 79, 88–115. [CrossRef]](http://doi.org/10.1016/j.jnca.2016.11.027) 3. Shin, Y.; Koo, D.; Hur, J. A survey of secure data deduplication schemes for cloud storage systems. ACM Comput. Surv. 2017, 49, 1–38. [[CrossRef]](http://dx.doi.org/10.1145/3017428) 4. Yinghui, Z.; Dong, Z.; Deng, R.H. Security and privacy in smart health: Efficient policy-hiding attribute-based access control. _IEEE Internet Things J. 2018, 5, 2130–2145._ 5. Zhang, Y.; Chen, X.; Li, J.; Wong, D.S.; Li, H.; You, I. Ensuring attribute privacy protection and fast decryption for outsourced data [security in mobile cloud computing. Inf. Sci. 2017, 379, 42–61. [CrossRef]](http://dx.doi.org/10.1016/j.ins.2016.04.015) 6. [Dropbox. Available online: https://www.theguardian.com/technology/2016/aug/31/dropbox-hack-passwords-68m-data-](https://www.theguardian.com/technology/2016/aug/31/dropbox-hack-passwords-68m-data-breach) [breach (accessed on 17 February 2023).](https://www.theguardian.com/technology/2016/aug/31/dropbox-hack-passwords-68m-data-breach) 7. [Arrington, M. Gmail Disaster: Reports of Mass Email Deletions. December 2006. Available online: https://techcrunch.com/2006](https://techcrunch.com/2006/12/28/gmail-disaster-reportsof-mass-email-deletions/) [/12/28/gmail-disaster-reportsof-mass-email-deletions/ (accessed on 17 February 2023).](https://techcrunch.com/2006/12/28/gmail-disaster-reportsof-mass-email-deletions/) 8. [Amazon. Amazon s3 Availability Event: 20 July 2008. Available online: https://simonwillison.net/2008/Jul/27/aws/ (accessed](https://simonwillison.net/2008/Jul/27/aws/) on 17 February 2023). 9. [Krigsman, M. Apple’s MobileMe Experiences Post-Launch Pain. July 2008. Available online: https://www.zdnet.com/article/](https://www.zdnet.com/article/apples-mobileme-experiences-post-launch-pain/) [apples-mobileme-experiences-post-launch-pain/ (accessed on 17 February 2023).](https://www.zdnet.com/article/apples-mobileme-experiences-post-launch-pain/) ----- _Electronics 2023, 12, 1545_ 23 of 23 10. [Benet, J. Ipfs-Content Addressed, Versioned, p2p File System. 2014. Available online: https://arxiv.org/abs/1407.3561 (accessed](https://arxiv.org/abs/1407.3561) on 17 February 2023). 11. [Geth. Available online: https://geth.ethereum.org/ (accessed on 17 February 2023).](https://geth.ethereum.org/) 12. [Metamask. Available online: https://metamask.io/ (accessed on 17 February 2023).](https://metamask.io/) 13. Hao, J.; Sun, Y.; Luo, H. A Safe and Efficient Storage Scheme Based on BlockChain and IPFS for Agricultural Products Tracking. _J. Comput. 2018, 29, 158–167._ 14. Rajalakshmi, A.; Lakshmy, K.V.; Sindhu, M.; Amritha, P. A blockchain and IPFS based framework for secure Research record keeping. Int. J. Pure Appl. Math. 2018, 119, 1437–1442. 15. Vimal, S.; Srivatsa, S.K. A new cluster P2P file sharing system based on IPFS and blockchain technology. J. Ambient. Intell Hum. _[Comput. 2019, 1–8. [CrossRef]](http://dx.doi.org/10.1007/s12652-019-01453-5)_ 16. Chen, Y.; Li, H.; Li, K.; Zhang, J. An improved P2P file system scheme based on IPFS and Blockchain. In Proceedings of the 2017 [IEEE International Conference on Big Data (Big Data), Boston, MA, USA, 11–14 December 2017; pp. 2652–2657. [CrossRef]](http://dx.doi.org/10.1109/BigData.2017.8258226) 17. Wang, R.; Tsai, W.-T.; He, J.; Liu, C.; Li, Q.; Deng, E. A Video Surveillance System Based on Permissioned Blockchains and Edge Computing. In Proceedings of the 2019 IEEE International Conference on Big Data and Smart Computing (BigComp), Kyoto, [Japan, 27 February–2 March 2019; pp. 1–6. [CrossRef]](http://dx.doi.org/10.1109/BIGCOMP.2019.8679354) 18. Sun, J.; Yao, X.; Wang, S.; Wu, Y. Blockchain-Based Secure Storage and Access Scheme For Electronic Medical Records in IPFS. _[IEEE Access 2020, 8, 59389–59401. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.2982964)_ 19. [Ethereum. Available online: https://ethereum.org/ (accessed on 17 February 2023).](https://ethereum.org/) 20. [Web3. Available online: https://web3js.readthedocs.io/en/v1.8.0/ (accessed on 17 February 2023).](https://web3js.readthedocs.io/en/v1.8.0/) 21. [Ethers. Available online: https://docs.ethers.io/v5/ (accessed on 17 February 2023).](https://docs.ethers.io/v5/) 22. Cai, W.; Wang, Z.; Ernst, J.B.; Hong, Z.; Feng, C.; Leung, V.C.M. Decentralized Applications: The Blockchain-Empowered Software [System. IEEE Access 2018, 6, 53019–53033. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2018.2870644) 23. [React. Available online: https://reactjs.org/ (accessed on 17 February 2023).](https://reactjs.org/) 24. Krichen, M.; Lahami, M.; Al-Haija, Q.A. Formal Methods for the Verification of Smart Contracts: A Review. In Proceedings of the 15th International Conference on Security of Information and Networks (SIN), Sousse, Tunisia, 11–13 November 2022; pp. 1–8. 25. [Infura. Available online: https://infura.io/ (accessed on 17 February 2023).](https://infura.io/) 26. [Google. Available online: https://www.google.com/ (accessed on 17 February 2023).](https://www.google.com/) 27. [Facebook. Available online: https://www.facebook.com/ (accessed on 17 February 2023).](https://www.facebook.com/) **Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual** author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/electronics12071545?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/electronics12071545, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/2079-9292/12/7/1545/pdf?version=1679978401" }
2,023
[]
true
2023-03-24T00:00:00
[ { "paperId": "0f880f7152719d5f34a1f2dd0472dc92bc52b8ce", "title": "Blockchain-Based Secure Storage and Access Scheme For Electronic Medical Records in IPFS" }, { "paperId": "a53f2a6bdd0bde4def8447c562d2265774c40fea", "title": "A new cluster P2P file sharing system based on IPFS and blockchain technology" }, { "paperId": "759117d27b07ace9d2fa58fa4b7ff6f08bc7a326", "title": "A Safe and Efficient Storage Scheme Based on BlockChain and IPFS for Agricultural Products Tracking" }, { "paperId": "acea0aada3b7c58b3711507f0935cfb4606eab72", "title": "Decentralized Applications: The Blockchain-Empowered Software System" }, { "paperId": "13cb37c279dee9e674bc151395c048b1bd60354e", "title": "Security and Privacy in Smart Health: Efficient Policy-Hiding Attribute-Based Access Control" }, { "paperId": "b2f207a7819c71fcec0de9eb57e5d7822291f674", "title": "Ensuring attribute privacy protection and fast decryption for outsourced data security in mobile cloud computing" }, { "paperId": "cd470eaaf7b1117c6c637408ef3a104636d1ac03", "title": "Cloud security issues and challenges: A survey" }, { "paperId": "b9c50d2fedecf434e477788e46cc730d9e2a0b34", "title": "A Survey of Secure Data Deduplication Schemes for Cloud Storage Systems" }, { "paperId": "714311007bc41cffeb11ca2968cec4f89128ce95", "title": "A Blockchain and IPFS based framework for secure Research record keeping" } ]
15,280
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Economics", "source": "external" }, { "category": "Economics", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/012136c4ef2730ee762ae3110798b0648dfade38
[ "Computer Science", "Medicine", "Economics" ]
0.886622
The Bitcoin as a Virtual Commodity: Empirical Evidence and Implications
012136c4ef2730ee762ae3110798b0648dfade38
Frontiers in Artificial Intelligence
[ { "authorId": "98204516", "name": "Cinzia Baldan" }, { "authorId": "2073965620", "name": "Francesco Zen" } ]
{ "alternate_issns": null, "alternate_names": [ "Front Artif Intell" ], "alternate_urls": null, "id": "6a8c0041-d0b7-4e32-b52c-33adef005c7e", "issn": "2624-8212", "name": "Frontiers in Artificial Intelligence", "type": null, "url": "https://www.frontiersin.org/journals/artificial-intelligence#" }
The present work investigates the impact on financial intermediation of distributed ledger technology (DLT), which is usually associated with the blockchain technology and is at the base of the cryptocurrencies' development. “Bitcoin” is the expression of its main application since it was the first new currency that gained popularity some years after its release date and it is still the major cryptocurrency in the market. For this reason, the present analysis is focused on studying its price determination, which seems to be still almost unpredictable. We carry out an empirical analysis based on a cost of production model, trying to detect whether the Bitcoin price could be justified by and connected to the profits and costs associated with the mining effort. We construct a sample model, composed of the hardware devices employed in the mining process. After collecting the technical information required and computing a cost and a profit function for each period, an implied price for the Bitcoin value is derived. The interconnection between this price and the historical one is analyzed, adopting a Vector Autoregression (VAR) model. Our main results put on evidence that there aren't ultimate drivers for Bitcoin price; probably many factors should be expressed and studied at the same time, taking into account their variability and different relevance over time. It seems that the historical price fluctuated around the model (or implied) price until 2017, when the Bitcoin price significantly increased. During the last months of 2018, the prices seem to converge again, following a common path. In detail, we focus on the time window in which Bitcoin experienced its higher price volatility; the results suggest that it is disconnected from the one predicted by the model. These findings may depend on the particular features of the new cryptocurrencies, which have not been completely understood yet. In our opinion, there is not enough knowledge on cryptocurrencies to assert that Bitcoin price is (or is not) based on the profit and cost derived by the mining process, but these intrinsic characteristics must be considered, including other possible Bitcoin price drivers.
Edited by: Alessandra Tanda, University of Pavia, Italy Reviewed by: Jürgen Hakala, Leonteq Securities AG, Switzerland Marika Vezzoli, University of Brescia, Italy *Correspondence: Cinzia Baldan [cinzia.baldan@unipd.it](mailto:cinzia.baldan@unipd.it) Specialty section: This article was submitted to Artificial Intelligence in Finance, a section of the journal Frontiers in Artificial Intelligence Received: 19 November 2019 Accepted: 20 March 2020 Published: 30 April 2020 Citation: Baldan C and Zen F (2020) The Bitcoin as a Virtual Commodity: Empirical Evidence and Implications. Front. Artif. Intell. 3:21. [doi: 10.3389/frai.2020.00021](https://doi.org/10.3389/frai.2020.00021) [ORIGINAL RESEARCH](https://www.frontiersin.org/journals/artificial-intelligence#editorial-board) [published: 30 April 2020](https://www.frontiersin.org/journals/artificial-intelligence#editorial-board) [doi: 10.3389/frai.2020.00021](https://doi.org/10.3389/frai.2020.00021) # The Bitcoin as a Virtual Commodity: Empirical Evidence and Implications [Cinzia Baldan* and Francesco Zen](http://loop.frontiersin.org/people/770598/overview) Department of Economics and Management, University of Padova, Padua, Italy ### The present work investigates the impact on financial intermediation of distributed ledger technology (DLT), which is usually associated with the blockchain technology and is at the base of the cryptocurrencies’ development. “Bitcoin” is the expression of its main application since it was the first new currency that gained popularity some years after its release date and it is still the major cryptocurrency in the market. For this reason, the present analysis is focused on studying its price determination, which seems to be still almost unpredictable. We carry out an empirical analysis based on a cost of production model, trying to detect whether the Bitcoin price could be justified by and connected to the profits and costs associated with the mining effort. We construct a sample model, composed of the hardware devices employed in the mining process. After collecting the technical information required and computing a cost and a profit function for each period, an implied price for the Bitcoin value is derived. The interconnection between this price and the historical one is analyzed, adopting a Vector Autoregression (VAR) model. Our main results put on evidence that there aren’t ultimate drivers for Bitcoin price; probably many factors should be expressed and studied at the same time, taking into account their variability and different relevance over time. It seems that the historical price fluctuated around the model (or implied) price until 2017, when the Bitcoin price significantly increased. During the last months of 2018, the prices seem to converge again, following a common path. In detail, we focus on the time window in which Bitcoin experienced its higher price volatility; the results suggest that it is disconnected from the one predicted by the model. These findings may depend on the particular features of the new cryptocurrencies, which have not been completely understood yet. In our opinion, there is not enough knowledge on cryptocurrencies to assert that Bitcoin price is (or is not) based on the profit and cost derived by the mining process, but these intrinsic characteristics must be considered, including other possible Bitcoin price drivers. Keywords: Bitcoin, FinTech, Vector Autoregression model, distributed ledger technology, cryptocurrencies price determination JEL Codes: G12, C52, D40 ## INTRODUCTION A strict definition of FinTech seems to be missing since it embraces different companies and technologies, but a wider one could assert that FinTech includes those companies that are developing new business models, applications, products, or process based on digital technologies applied in finance. [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 1 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity Financial Stability Board (FSB) (2017) defines FinTech as “technology-enabled innovation in financial services that could result in new business models, applications, processes, or products with an associated material effect on the provision of financial services.” OECD (2018) analyzes instead various definitions from different sources, concluding that none of them is complete since “FinTech involves not only the application of new digital technologies to financial services but also the development of business models and products which rely on these technologies and more generally on digital platform and processes.” The services offered by these companies are indeed various: some are providing financial intermediation services (FinTech companies), while others offer ancillary services relating to the financial intermediation activity (TechFin companies). Technology is, for FinTech firms, an instrument, a productive factor, an input, while for TechFin firms, it is the final product, the output. The latter are already familiar with different technologies and innovation; hence, they could easily diversify their production by adding some digital and financial services to the products they already offer. They enjoy a situation of privileged competition because they are already known in the market due to their previous non-financial services and thus could take advantage of their customers’ information to enlarge their supply of financial services. TechFin firms are the main competitors for FinTech companies (Schena et al., 2018). Indeed FinTech, or financial technology, is changing the way in which financial operations are carried out by introducing new ways to save, borrow, and invest, without dealing with traditional banks. FinTech platforms, firms, and startups rose after the global financial crisis in 2008 as a consequence of the loss of trust in the traditional financial sector. In addition, digital natives (or millennials, born between 1980 and 2000) seemed interested in this new approach proposed by FinTech entrepreneurs. Millennials were old enough to be potential customers, who feel much more related to these new, fresh mobile services offered through mobile platforms and apps, rather than bankers. The strength of these new technologies lies in their transparent and easy-to-use interfaces that was seen as an answer to the trust crisis toward banks (Chishti and Barberis, 2016). After the first Bitcoin (Nakamoto, 2008) has been sent in January 2009, hundreds of new cryptocurrencies started being traded in the market, whose common element is to rely on a public ledger (or blockchain technology; Hileman and Rauchs, 2017). In fact, in addition to Bitcoin, other cryptocurrencies gained popularity, such as: Ethereum (ETH), Dash, Monero (XMR), Ripple (XRP), and Litecoin (LTC). Ethereum (ETH) was officially launched in 2015 and is a decentralized computing platform characterized by its own programming language. Dash was introduced in 2014 but its market value was rising in 2016. The peculiarity of this digital coin is that, in contrast with other cryptocurrencies, block rewards are equally shared among community participants and a revenue percentage (equal to 10%) is stored in the “treasury” to fund further improvements, marketing, and network operations. Monero (XMR), launched in 2014, is a system that guarantees anonymous digital cash by hiding the features of the transacted coins. Its market value raised in 2016. Ripple (XRP) has the unique feature to be based on a “global consensus ledger” rather than on blockchain technology. Its protocol is adopted by large institutions like banks and money service businesses. Litecoin (LTC) appeared for the first time in 2011 and is characterized by a large supply of 84 million LTC. Its functioning is based on that of Bitcoin, but some parameters were altered (the mining algorithm is based on Scrypt rather than Bitcoin’s SHA-265). Despite the creation of these new cryptocurrencies, Bitcoin remains the main coin in terms of turnover. The main advantage of this new digital currency seems to be the low cost of transaction (even if this is actually a myth, since BTC transactions topped out at 50 USD per transaction in 2017–2018, while private banks charge less these days) and, contrary on what many people think, anonymity was not one of its main features when this network was designed. An individual could attempt to make his identity less obvious but the evidences available by now do not support the claim that it could be hidden easily; it may be probably impossible. To this purpose, fiat physical currencies remain the best option. Hayes (2015, 2017, 2019) analyzes the Bitcoin price formation. In particular, he assumes the cryptocurrency as a virtual commodity, starting from the different ways by which an individual could obtain it. A person could buy Bitcoins directly in an online marketplace by giving in exchange fiat currencies or other types of cryptocurrencies. Alternatively, he can accept them as payment and finally an individual can decide to “mine” Bitcoins, which consists in producing new units, by using computer hardware designed for this purpose. This latter case involves an electrical consumption and a rational agent would not be involved in the mining process if the marginal costs of this operation exceed its marginal profits. The relation between these values determines price based on the cost of production that is the theoretical value underlying the market price, around which it is supposed to gravitate. Abbatemarco et al. (2018) resume Hayes’ studies introducing further elements missed in the previous formulation. The final result confirms Hayes’ findings: the marginal cost model provides a good proxy for Bitcoin market price, but the development of a speculative bubble is not ruled out. We study the evolution of Bitcoin price by considering a cost of production model introduced by Hayes (2015, 2017, 2019). Adding to his analysis some adjustment proposed by Abbatemarco et al. (2018), we recover a series for the hypothetical underlying price; then, we study the relationship between this price and the historical one using a Vector Autoregression (VAR) model. The remainder of the paper proceeds as follows: in section Literature Review, we expose a literature overview, presenting those papers that investigate other drivers for Bitcoin price formation, developing alternative approaches. In section Materials and Methods, we exploit the research question, describing the methodology behind the implemented cost of production model, the sources accessed to collect data, the hardware sample composition, and the formula derivations. In section Main Outcomes, we analyze and comment on the main findings of the analysis; section Conclusions concludes the work with our comments on main findings and their implications. [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 2 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity ## LITERATURE REVIEW Researchers detect a number of economic determinants for Bitcoin price; it seems that given the new and particular features of this cryptocurrency, price drivers will change over time. For this reason, several authors analyze various potential factors, which encompass technical aspects (such as the hashrate and output volume), user-based growth, Internet components (as Google Trends, Wikipedia queries, and Tweets), market supply and demand, financial indexes (like S&P500, Dow Jones, FTSE100, Nikkei225), gold and oil prices, monetary velocity, and exchange rate of Bitcoin expressed in US dollar, euro, and yen. Among others, Kristoufek (2015) focuses on different sources of price movements by examining their interconnection during time. He considers different categories: economic drivers, as potential fundamental influences, followed by transaction and technical drivers, as influences on the interest in the Bitcoin. The results show how Bitcoin’s fundamental factors, such as usage, money supply and price level, drive its price over the long term. With regard to the technical drivers, a rising price encourages individuals to become miners but this effect eclipses over time, since always more specialized mining hardware have increased the difficulty. Evidences show that price is even driven by investors’ interest. According to previous studies (Kristoufek, 2013; Garcia et al., 2014), the relationship appears as most evident in the long run, but during episodes of explosive prices, this interest drives prices further up, while during rapid declines, it pushes them further down. He then concludes that Bitcoin is a unique asset with properties of both a speculativefinancial asset, and a standard one and because of his dynamic nature and volatility, it is obvious to expect that its price drivers will change over time. The interest element seems to be particularly relevant when analyzing the behavior of Bitcoin price, leading many researchers to study its interconnection with Internet components, such as Google Trends, Wikipedia queries, and Tweets. Even Matta et al. (2015) investigate whether information searches and social media activities could predict Bitcoin price comparing its historical price to Google Trends data and volume of tweets. They used a dataset based only on 60 days, but, in addition to the other papers regarding this topic, they implement an automated sentiment analysis technique that allows one to automatically identify users’ opinions, evaluations, sentiments, and attitudes on a particular topic. They use a tool called “SentiStrength,” which is based on a dictionary only made by sentiment words, where each of them is linked to a weight representing a sentiment strength. Its aim is to evaluate the strength of sentiments in short messages that are analyzed separately, and the result is summed up in a single value: a positive, negative, or neutral sentiment. The study reveals a significant relationship between Bitcoin price and volumes of both tweets and Google queries. Garcia et al. (2014) study the evolution of Bitcoin price based on the interplay between different elements: historical price, volume of word-of-mouth communication in on-line social media (information sharing, measured by tweets, and posts on Facebook), volume of information search (Google searches and Wikipedia queries), and user base growth. The results identify an interdependence between Bitcoin price and two signals that could form potential price bubbles: the first concerns the word-ofmouth effect, while the other is based on the number of adopters. The first feedback loop is a reinforcement cycle: Bitcoin interest increases, leading to a higher search volume and social media activity. This new popularity encourages users to purchase the cryptocurrency driving the price further up. Again, this effect would raise the search volume. The second loop is the user adoption cycle: after acquiring information, new users join the network, growing the user base. Demand rises but since supply cannot adjust immediately but changes linearly with time, Bitcoin price would increase. Ciaian et al. (2016) adopt a different approach to identify the factors behind the Bitcoin price formation by studying both the digital and traditional ones. The authors point out the relevance of analyzing these factors simultaneously; otherwise, the econometric outputs could be biased. To do so, they specify three categories of determinants: market forces of supply and demand; attractiveness indicators (views on Wikipedia and number of new members and posts on a dedicated blog), and global macro-financial development. The results show that the relevant impact on price is driven by the first category and it tends to increase over time. About the second category, they assert that the short-run changes on price following the first period after Bitcoin introduction are imputable to investors’ interest, which is measured by online information search. Its impact eases off during time, having no impact in the long run and may be due to an increased trust among users who become more willing to adopt the digital currency. On the other hand, the results suggest that investor speculations can also affect Bitcoin price, leading to a higher volatility that may cause price bubbles. To conclude, the study does not detect any correspondences between Bitcoin price and macroeconomics and financial factors. Kjærland et al. (2018) try to identify the factors that have an impact on Bitcoin price formation. They argue that the hashrate, CBOE volatility index (VIX), oil, gold, and Bitcoin transaction volume do not affect Bitcoin price. The study shows that price depends on the returns on the S&P500, past price performance, optimism, and Google searches. Bouoiyour and Selmi (2015) examine the links between Bitcoin price and its potential drivers by considering investors’ attractiveness (measured by Google search queries); exchange– trade ratio; monetary velocity; estimated output volume; hashrate; gold price; and Shanghai market index. The latter value is due to the fact that the Shanghai market is seen as the biggest player in Bitcoin economy, which could also drive its volatility. The evaluation period is the one from 5th December 2010 to 14th July 2014 and it is investigated through the adoption of an ARDL Bounds Testing method and a VEC Grander causality test. The results highlight the speculative nature of this cryptocurrency stating that there are poor chances that it becomes internationally recognized. Giudici and Abu-Hashish (2019) propose a model to explain the dynamics of bitcoin prices, based on a correlation network VAR process that models the interconnections between different crypto and classic asset price. In particular, they try to assess [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 3 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity whether bitcoin prices in different exchange markets are correlated with each other, thus exhibiting “endogenous” price variations. They select eight exchange markets, representative of different geographic locations, which represent about 60% of the total daily volume trades. For each exchange market, they collect daily data for the time period May 18th, 2016 to April 30th, 2018. The authors also try to understand whether bitcoin price variations can also be explained by exogenous classical market prices. Hence, they use daily data (market closing price) on some of the most important asset prices: gold, oil, and SP500, as well as on the exchange rates USD/Yuan and USD/Eur. Their main empirical findings show that bitcoin prices from different exchanges are highly interrelated, as in an efficiently integrated market, with prices from larger and/or more connected trading exchanges driving the others. The results also seem to confirm that bitcoin prices are typically unrelated with classical market prices, thus bringing further support to the “diversification benefit” property of crypto assets. Katsiampa (2017) uses an Autoregressive model for the conditional mean and a first-order GARCH-type model for the conditional variance in order to analyze the Bitcoin price volatility. The study collects daily closing prices for the Bitcoin Coindesk Index from 18th July 2010 to 1st October 2016 (2,267 observations); the returns are then calculated by taking the natural logarithm of the ratio of two consecutive prices. The main findings put on evidence that the optimal model in terms of goodness of fit to the data is the AR-CGARCH, a result that suggests the importance of having both a short-run and a long-run component of conditional variance. Chevallier et al. (2019) investigate the Bitcoin price fluctuations by combining Markov-switching models with Lévy jump-diffusion to match the empirical characteristics of financial and commodity markets. In detail, they try to capture the different sub-periods of crises over the business cycle, which are captured by jumps, whereas the trend is simply modeled under a Gaussian process. They introduce a Markov chain with the existence of a Lévy jump in order to disentangle potentially normal economic regimes (e.g., with a Gaussian distribution) vs. agitated economic regimes (e.g., crises periods with stochastic jumps). By combining these two features, they offer a model that captures the various crashes and rallies over the business cycle, which are captured by jumps, whereas the trend is simply modeled under a Gaussian framework. The regime-switching Lévy model allows identifying the presence of discontinuities for each market regime, and this feature constitutes the objective of the proposed model. ## MATERIALS AND METHODS We study the evolution of Bitcoin price by considering a cost of production model introduced by Hayes (2015, 2017). Adding to his analysis some adjustment proposed by Abbatemarco et al. (2018), we recover a series for the hypothetical underlying price, and we study the relationship between this price and the historical one using a VAR model. In detail, Hayes backtests the pricing model against the historical market price to consolidate the validity of his theory. The findings show how Bitcoin price is significantly described by the cryptocurrency’s marginal cost of production and suggest that it does not depend on other exogenous factors. The conclusion is that during periods in which price bubbles happen, there will be a convergence between the market price and the model price to shrink the discrepancy. Abbatemarco et al. (2018) resume Hayes’ studies introducing further elements missed in the previous formulation. The final result confirms Hayes’ findings: the marginal cost model provides a good proxy for Bitcoin market price, but the development of a speculative bubble is not ruled out. Since these studies were published before Bitcoin price raise reached its peak on 19th December 2017 (the value was $19,270), the aim of our work is to extend the analysis considering a larger time frame and verify if, even in this case, the results are unchanged. In particular, we consider the period from 9th April 2014 to 31st December 2018. We start with some unit root tests to verify if the series are stationary in level or need to be integrated and then we identify the proper number of lags to be included in the model. We then check for the presence of a cointegrating relationship to verify whether we should adopt a Vector Error Correction Model (VECM) or a VAR model; the results suggest that a VAR model is the best suited for our data [1] . We thus collect the final results of the analysis and we improve them by correcting the heteroscedasticity in the regressions. The marginal cost function, which estimates the electrical costs of the devices used in the mining process, is presented as Equation (1): COST $ [hash] ∗ Eff J $ (1) day [=][ H] s hash [∗] [CE] kWh [∗] [24] [ h] day Where: H hash/s is the hashrate (measured by hash/second); EFF J/hash is the energy efficiency of the devices involved in the process and it is measured by Joule/hash; CE $/kWh is the electricity cost expressed in US dollar per kilowatt/hour; 24 is the number of hours in a day; A marginal profit function, which estimates the reward of the mining activity, is instead depicted as Equation (2): � PROFIT BTC day [=][ BR] [BTC] [ ∗] 3, 600 s h [∗] [24] [ h] day BTs � (2) Where: 1 According to Abbatemarco et al. (2018), the nature of the variables considered suggests that they probably are mutually interdependent. Lütkepohl and Krätzig (2004) state that the analysis of interdependencies between time series is subject to the endogenous problem; part of the literature proposes to specify a Vector Auto Regressive model (VAR) that analyzes the causality between the two series estimated by the model. Engle and Granger (1987), instead, demonstrated that the estimate of such a model in the presence of non-stationary variables (i.e., with mean and variance non-constant over time) can lead to erroneous model specification and hence to unconditional regressions (spurious regressions). Scholars’ intuition suggests that the price trend of a cryptocurrency and that of its estimated equilibrium prices are non-stationary time series, as there is a constant increase in their values over time. [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 4 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity BR BTC is the block reward that refers to new Bitcoins distributed to miners who successfully solved a block (hence it is measured by BTC) and it is given by a geometric progression (Equation 3): n−1 BR BTC = BR 1 ∗ [1] (3) 2 n increases by 1 every 210,000 blocks. At the beginning, it was BR 1 = 50, but during the course of time, it halved twice: on 29th November 2012 and on 10th July 2016. 3,600 is the number of seconds in an hour; 24 is again the number of hours in a day; BT s is the block time, which is expressed as the seconds needed to generate a block (around 600 s = 10 min), and it is computed as Equation (4): BT s = [D][ ∗] [2] [32] (4) H Where H = hashrate and D = difficulty. The latter variable specifies how hard it is to generate a new block in terms of computational power given a specific hashrate. This is the value that changes frequently to ensure a BT s close to 10 min [2] . In addition to the variables already considered, we introduce some adjustments proposed by Abbatemarco et al. (2018), who thought there were two elements missing in Hayes’ formulations. They add, on the cost side, the one required to maintain and update miners’ hardware (MAN, expressed in US dollar), and on the profit side, the fees (FEES) received by miners who place transactions in a block [3] . Maintenance costs are computed as a ratio between the weighted devices’ price and their weighted lifespan (5), while fees, expressed in BTC, are measured as a ratio between the daily total transaction fees and the number of daily transactions [4] (6). MAN $ = [Wei] [g] [hted Devices Price] [$] (5) Weighted Lifespan FEES BTC = [Total Transaction Fees][ (][BTC][)] (6) Daily Transaction Fees The new equations become: COST $ [hash] ∗ Eff J $ (7) day [=][ H] s hash [∗] [CE] kWh [∗] [24] [ h] day [+][ MAN] [$] PROFIT BTC 3, 600 s h [∗] [24] [ h] day + FEES BTC (8) day [=][ BR] [BTC] [ ∗] BTs � � Moreover, due to the equality 1 joule = 1 watt [∗] second, Equation (7) could be expressed as follows: COST $/day = H hash/s ∗ Eff W hash ∗ s [∗] [CE] kWh $ [∗] [24] [h][/][day] [ +][ MAN] [$] [ (9)] 2 Results are shown in Table A.1 ( **Supplementary Material** ). In order to simplify the presentation, we display only the values for the last day of each month. 3 Bitcoin could be obtained through both the mining process and the registration of transactions but, since Bitcoin supply is limited to 21 million, once it is reached, fees become the only remuneration source in the future. 4 Fees computation results are displayed in Table A.1 ( **Supplementary Material** ). TABLE 1 | Sources. Variables Sources P hist$ Historical price in US [https://Bitcoinvisuals.com](https://Bitcoinvisuals.com) dollar H hash/s Hashrate BR BTC Block reward D Difficulty BT s Block time Computed using D and H hash/s FEES BTC Transaction fees [https://charts.Bitcoin.com/bch/](https://charts.Bitcoin.com/bch/) CE $/kWh Cost of energy Computed using data from: en.Bitcoin.it/wiki/Mining_ hardware_comparison [https://archive.org/web/](https://archive.org/web/) MAN $ Hardware maintaining cost EFF J/hash Hardware energy efficiency Source: Authors’ elaboration. By converting watt in kilowatt/hour, it can be written as: Eff W ∗ s hash COST $/day = H hash/s ∗ 1000 ∗ CE kWh $ [∗] [24] [h][/][day] [ +][ MAN] [$] (10) COST $/day = H hash/s ∗ Eff kWh hash ∗ s ∗ CE kWh $ [∗] [24] [h][/][day] [ +][ MAN] [$] (11) According to the competitive market economic theories, the ratio between the cost and profit functions must lead to the price under equilibrium condition (Equation 12): COST $ day P $/BTC = (12) PROFIT BTC day A historical price below the one predicted by the model would force a miner out of the market, since he is operating in loss, but at the same time, the removal of its devices from the network increases others’ marginal profits (competition decreases), and at the end, the system would return to equilibrium. On the other hand, a historical price higher than what predicted by the model attracts more miners, thus increasing the number of devices operating in the network and decreasing others’ marginal profits (competition increases). Again, the system would return in balance (Hayes, 2015). We must remark that the assumption of an energy price per hemisphere is not very realistic. In fact, for large consumers, energy price is contractually set differently for peak times and less busy times. There is a lot of variation in the energy price of mines in different countries and circumstances (see, for example, Iceland with its geothermal cheap energy as a cheap energy example; Soltani et al., 2019). Taking more variation around energy prices into account would probably add a wider range of BTC prices (de Vries, 2016); due to the difficulties on collecting comparable data, we adopted a simplified proxy of the cost of energy. [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 5 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity **Table 1** presents the sources used to collect and compute the required information. We start the analysis by constructing a hardware sample that evolves during a chosen time window (2010–2018), which is divided in semesters associated with the introduction of a specific device ( **Table 2** ). Since the first Bitcoin was traded, there has been an evolution of the devices used by miners. The first ones adopted were GPU (Graphical Processing Unit) and later FPGA (Field-Programmable Gate Array), but these days, only ASIC (Application-Specific Integrated Circuit) is suitable for mining purposes. For each device model, we collect the efficiency, expressed in Mhash/J, and the dollar price at the release day. Technical data were collected from the Wikipedia pages [https://en.Bitcoin.it/wiki/Mining__hardware__comparison](https://en.Bitcoin.it/wiki/Mining__hardware__comparison) and [https://en.Bitcoin.it/wiki/Non-specialized__hardware__](https://en.Bitcoin.it/wiki/Non-specialized__hardware__comparison) [comparison by using in addition the online archive https://](https://en.Bitcoin.it/wiki/Non-specialized__hardware__comparison) [archive.org/web/,](https://archive.org/web/) which allows the recovery of different webpages at the date in which they were modified, enabling the comparison before and after reviews [5] . Since only ASIC devices were created with specifications to mining purpose, there is homogeneity among FPGA and especially among GPU hardware. Due to this fact and considering the difficulty to recover the release prices, we make some simplified assumptions about them based on the information available online. This means that given the same computational power, we assume price homogeneity among devices when they were not available for specific models [6] . Given the hardware sample, we construct a weights distribution matrix (Table A.3 in **Supplementary Material** ) that represents the evolution of the devices used during each semester of the time window selected, which are replaced following a substitution rate that increases over time. In fact, until 2012, before FPGA took roots, it is equal to 0.05; until 2016, we set it equal to 0.1, and in the last 2 years of the analysis, it is equal to 0.15 [7] . All computations are based on this matrix; indeed, we multiplied it by a specific column of the hardware sample table to obtain the biannual Efficiency (Table A.4 in **Supplementary Material** ) (J/Hash), Weighted Devices’ Prices ($) (Table A.5 in **Supplementary Material** ), and Weighted Lifespans (Table A.6 in **Supplementary Material** ). Regarding this latter matrix, we made further assumptions on the device lifespans by implementing Abbatemarco et al. (2018) assumptions. Hence, we set a lifespan equal to 2,880 days for 5 When possible, we double check Wikipedia prices with those on the websites of the companies producing mining hardware, and if they are not identical, we choose the latter. 6 In detail, we approximate the prices of ATI FirePro M5800, Sapphire Radeon 5750 Vapor-X, GTX460, FireProV5800, Avnet Spartan-6 LX150T, and AMD Radeon 7900. 7 Despite that ASIC devices have been released for the first time in 2013, they became the main devices used in the mining process only in 2015–2016. In the last 2 years of the analysis, we increase the substitution rate up to 0.15 because the competition among miners has been driven up as more sophisticated hardware was developed with a larger frequency. FIGURE 1 | Historical market price vs. implied model price (July 2010–December 2018). Source: Authors’ elaboration. GPU, 1,010 days for FPGA, and 540 days for ASIC, but after 2017, due to a supposed market growth phase, we halved these numbers ( **Table 2** ). To evaluate the cost of energy, we follow the assumptions suggested by the cited researchers and we divide the world into two parts relative to Europe: East and West, each one with a fix electricity price equal to 0.04 and 0.175 $/kWh, respectively. The weights’ evolution of the mining pool is set up in 2010 equal to 0.7 for the West part and 0.3 for the East part and it changes progressively until reaching in 2018 a 0.2 for the West and 0.8 for the East. We obtained a biannual cost of energy evolution measured by $/kWh by multiplying the biannual weights to the electricity costs and summing up the value for the West and the East (Table A.7 in **Supplementary Material** ). At this point, to smooth the values across the time window, we take the differences between biannualMAN $, biannualEFF J/Hash, and biannualCE $/kWh at time t and t – 1 and we divide these values by the number of days in each semester, obtaining DeltaMAN, DeltaEFF, and DeltaCE (Table A.8 in **Supplementary Material** ). Starting the first day of the analysis with the first value of the biannual matrixes, we compute the final variables as follows: MAN $ (t) = MAN $ (t − 1) + DeltaMAN (13) EFF hash J [(][t][)][ =][ EFF] [J][/][hash] [ (][t][ −] [1][)][ +][ DeltaEFF] (14) CE kWh $ [(][t][)][ =][ CE] kWh $ [(][t][ −] [1][)][ +][ DeltaCE] (15) ## MAIN OUTCOMES By applying Equations (8), (11), and (12), we obtain the model price [8] and compare its evolution to the historical one ( **Figure 1** ). The evolution of the model (or implied) price shows a spike during the second semester of 2016, probably because on 10th July 2016, the Block Reward halved from 25 to 12.5, leading to a reduction on the profit side and a consequent price increase. Despite this episode, the historical price seems to fluctuate around the implied one until the beginning of 2017, the period 8 Table A.2 ( **Supplementary Material** ) displays all the variables required to compute the model price and compares it with the historical price. Since our time window involves 3,107 observation days, for the sake of simplicity, we present only the results for the last day of each month. [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 6 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity TABLE 2 | Hardware sample. TYPE MODEL TIME EFF. (Mhash/J) PRICE (USD) LIFESPAN Before [′] 17 After [′] 17 GPU ATI FirePro M5800 2 s. 2010 1.45 175 2,880 1,440 GPU Sapphire Radeon 5750 Vapor-X 2 s. 2010 1.35 160 2,880 1,440 GPU GTX460 2 s. 2010 1.73 200 2,880 1,440 GPU FirePro V5800 1 s. 2011 2.08 469 2880 1,440 FPGA Avnet Spartan-6 LX150T 2 s. 2011 6.25 995 1,010 505 FPGA AMD Radeon 7900 1 s. 2012 10.40 680 1,010 505 FPGA Bitcoin Dominator X5000 2 s. 2012 14.70 750 1,010 505 FPGA X6500 1 s. 2013 23.25 989 1,010 505 ASIC Avalon 1 2 s. 2013 107.00 1,299 540 270 ASIC Bitmain AntMiner S1 1 s. 2014 500.00 1,685 540 270 ASIC Bitmain AntMiner S2 2 s. 2014 900.00 2,259 540 270 ASIC Bitmain AntMiner S3 1 s. 2015 1,300.00 1,350 540 270 ASIC Bitmain AntMiner S4 2 s. 2015 1,429.00 1,400 540 270 ASIC Bitmain AntMiner S5 1 s. 2016 1,957.00 1,350 540 270 ASIC Bitmain AntMiner S5+ 2 s. 2016 2,257.00 2,307 540 270 ASIC Bitmain AntMiner S7 1 s. 2017 4,000.00 1,832 540 270 ASIC Bitmain AntMiner S9 2 s. 2017 10,182.00 2,400 540 270 ASIC Ebit E9++ 1 s. 2018 10,500.00 3,880 540 270 ASIC Ebit E10 2 s. 2018 11,100.00 5,230 540 270 Source: Authors’ elaboration. in which Bitcoin price started raising exponentially, reaching its peak with a value equal to $19,270 on 19th December 2017. It declined during 2018, converging again to the model price. Another divergence was detected at the end of 2013, but it was of a lower amount and resolved quickly. Given the historical and implied price series, we make a further step than what Hayes (2019) and Abbatemarco et al. (2018) did, by including in the analysis a time frame even in the divergence phase. Therefore, we consider the period from 9th April 2014 to 31st December 2018. We select this time window also to base the analysis on solid data. Because of the difficulty to obtain reliable information on the hardware used in the mining process, we make some simplified assumptions on their features. By choosing this time window, we include the hardware sample whose data are more precise. ## Unit Root Tests We first try to determine with different unit root tests whether the time series is stationary or not. The presence of a unit root indicates that a process is characterized by time-dependent variance and violates the weak stationarity condition [9] . We test the presence of a unit root with three procedures: the augmented Dickey–Fuller test (Dickey and Fuller, 1979), the Phillips–Perron test (Phillips and Perron, 1988), and the Zivot–Andrews test (Zivot and Andrews, 1992). Given a time series {y t }, both the augmented Dickey–Fuller test (Dickey and Fuller, 1979) and the Phillips–Perron test are 9 The condition of weak stationarity asserts that Var(r t ) = γ o, which means that the variance of the process is time invariant and equal to a finite constant. based on the general regression (Equation 16): �y t = α + βt + θ y t−1 + δ 1 �y t−1 + . . . + δ p−1 �y t−p+1 + ε t (16) Where �y t indicates changes in time series, α is the constant, t is the time trend, p is the order of the autoregressive process, and ε is the error term (Boffelli and Urga, 2016). For both tests, the null hypothesis is that the time series contains a unit root; thus, it is not stationary (H 0 : θ = 0), while the alternative hypothesis asserts stationarity (H 0 : θ < 0). Considering only the augmented Dickey–Fuller test, its basic idea is that if a series {y t } is stationary, then {�y t } can be explained only by the information included in its lagged values (�y t−1 . . . �y t−p+1 ) and not from those in y t− 1 . For each variable, we conduct this test firstly with a constant term and later by including also a trend [10] . **Table 3** presents the main findings of the test. The Phillips–Perron test points out that the process generating y t might have a higher order of autocorrelation than the one admitted in the test equation. This test corrects the issue, and it is 10 In order to select the proper number of lags to include in this test, we used, only for this part of the analysis, the open-source software Gretl. Its advantage is to apply clearly the Schwert criterion for the maximum lag (p max ) estimation, which is given by: p max = integer part of �12 ∗ � 100 T � 1/4 [�], where T is the number of observations. The test is conducted firstly with the suggested value of p max, but if the absolute value of the t statistic for testing the significance of the last lagged value is below the threshold 1.6, p max is reduced by 1 and the analysis is recomputed. The process stops at the first maximum lag that returns a value >1.6. When this value is found, the augmented Dickey–Fuller test is estimated. [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 7 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity robust in case of unspecified autocorrelation or heteroscedasticity in the disturbance term of the equation. **Table 4** displays the test results. The main difference between these tests is that the latter applies Newey–West standard errors to consider serial correlation, while the augmented Dickey–Fuller test introduces additional lags of the first difference. Since the previous tests do not allow for the possibility of a structural break in the series, Zivot and Andrews (1992) propose to examine the presence of a unit root including the chance of an unknown date of a break-point in the series. They elaborate three models to test for the presence of a unit root considering a one-time structural break: a) permits a one-time change in the intercept of the series: �y t = α + βt + θ y t−1 + γDU t + δ 1 �y t−1 + . . . + δ p−1 y t−p+1 + ε t (17) b) permits a one-time change in the slope of the trend function: �y t = α + βt + θ y t−1 + ϑDT t + δ 1 �y t−1 + . . . + δ p−1 �y t−p+1 + ε t (18) c) combines the previous models: �y t = α + βt + θ y t−1 + γDU t + ϑDT t + δ 1 �y t−1 + . . . + δ p−1 �y t−p+1 + ε t (19) Where DU t is a dummy variable that relates to a mean shift at a given break-date, while DT t is a trend shift variable. The null hypothesis, which is the same for all three models, states that the series contains a unit root (H 0 : θ = 0), while the alternative hypothesis asserts that the series is a stationary process with a one-time break occurring at an unknown point in time (H 0 : θ < 0) (Waheed et al., 2006). The results in **Table 5** confirm what the other tests predict: both series are integrated of order 1. Since this last test identifies for �lnPrice the presence of a structural break on 18th December 2017 and after this date the Bitcoin price reaches its higher value to start declining later, we add to the analysis a dummy variable related to this observation, in order to take into account a broken linear trend in a series. ## Identifying the Number of Lags The preferred lag length is the one that generates the lowest value of the information statistic considered. We follow Lütkepohl’s intuition that “the SBIC and HQIC provide consistent estimates of the true lag order, while the FPE and AIC overestimate the lag order with positive probability” (Becketti, 2013). Therefore, for our analysis, we select 1 lag ( **Table 6** ) [11] . 11 To identify the proper lag length to be included in the VAR model, we use the “varsoc” command in Stata that displays a table of test statistics, which reports for each lag length, the log of the likelihood functions (LL), a likelihood-ratio test statistic with the related degrees of freedom and p value (LR, df, and p), and also four information criteria: Akaike’s final prediction error (FPE); Akaike’s information criterion (AIC), Hannan and Quinn’s information criterion (HQIC), ## Identifying the Number of Cointegrating Relationships A cointegrating relationship is a relationship that describes the long-term link among the levels of a number of the nonstationary variables. Given K non-stationary variables, they can have at most K – 1 cointegrating relationships. Since we have only two non-stationary variables (lnPrice and lnModelPrice), we could obtain, at most, only one cointegrating relationship. If series show cointegration, a VAR model is no more the best suited one for the analysis, but it is better to implement a Vector Error-Correction Model (VECM), which can be written as (20): ′ Where the first part α β u t−1 + ν + ρt represents � � the cointegrating equations, while the second p−1 � i=1 [Ŵ] [i] [�][y] [t][−][i] [ +][ γ][ +][ τ] [t][ +][ ε] [t] [ refers to the variables in levels.] This representation allows specifying five cases that Stata tests: 1) Unrestricted trend: allows for quadratic trend in the level of y t (τ t appears in the equation) and states that the cointegrating equations are trend stationary, which means they are stationary around time trends. 2) Restricted trend (τ = 0): excludes quadratic trends but includes linear trends (ρt). As in the previous case, it allows the cointegrating equations to be trend stationary. 3) Unrestricted constant (τ = 0, ρ = 0): lets linear trends in y t to present a linear trend (γ ) but the cointegrating equations are stationary around a constant means (ν). 4) Restricted constant (τ = 0, ρ = 0, γ = 0): rules out any trends in the levels of the data but the cointegrating relationships are stationary around a constant mean (ν). and Schwarz’s Bayesian information criterion (SBIC). Every information criteria provide a trade-off between the complexity (e.g., the number of parameters) and the goodness of fit (based on the likelihood function) of a model. Since the output is sensitive to the maximum lag considered, we try different options by changing the one included in the command computation. We tried with 4, 8, 12, 16, 20, and 24 lags. After selecting a maximum lag length equal to 16, the optimal number of lags suggested changes: while the previous results agree recommending 1 lag with each information criteria, now the FPE and AIC diverge and propose 13 lags. �y t = µ + δt + αβ [′] u t−1 + p−1 � Ŵ i �y t−i + ε t (20) i=1 Where the deterministic components µ + δt are, respectively, the linear and the quadratic trend in y t that can be separated into the proper trends in y t and those of the cointegrating relationship. This depends on the fact that in a first-difference equation: a constant term is a linear trend in the level of the variables (y t = κ + λt → �y t = λ), while a linear trend derives from the quadratic one in the regression in levels (y t = κ + λt + ωt [2] → �y t = λ + 2ωt − ω). Therefore, µ ≡ αν + γ, and δt = αρt + τ t. By substituting in the previous expression, the VECM can be expressed as Equation (21): �y t = α �β [′] y t−1 + ν + ρt� + p−1 � Ŵ i �y t−i + γ + τ t + ε t (21) i=1 [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 8 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity TABLE 3 | Augmented Dickey–Fuller test. Augmented Dickey–Fuller test Constant Constant + trend Result t stat p-value t-stat p-value lnPrice −0.606 0.8696 −1.839 0.6856 NO stationary lnModelPrice −0.467 0.8982 −1.669 0.7644 NO stationary �lnPrice −7.694 0.0000 −7.697 0.0000 Stationary �lnModelPrice −8.041 0.0000 −8.038 0.0000 Stationary Critical values Constant Constant + trend 1% 5% 10% 1% 5% 10% −3.430 −2.860 −2.570 −3.960 −3.410 −3.120 Source: Authors’ elaboration. TABLE 4 | Phillips–Perron test. Phillips–Perron test Constant Constant + trend Result t stat p-value t stat p-value lnPrice −0.437 0.9037 −1.546 0.8130 NO stationary lnModelPrice −0.637 0.8624 −1.805 0.7021 NO stationary �lnPrice −34.394 0.0000 −34.385 0.0000 Stationary �lnModelPrice −42.972 0.0000 −42.959 0.0000 Stationary Critical values Constant Constant + trend 1% 5% 10% 1% 5% 10% −3.430 −2.860 −2.570 −3.960 −3.410 −3.120 Source: Authors’ elaboration. 5) No trend (τ = 0, ρ = 0, γ = 0, ν = 0): considers no non-zero means or trends. Starting from these different specifications, the Johansen test can detect the presence of a cointegrating relationship in the analysis. The null hypothesis states, again, that there are no cointegrating relationships against the alternative that the null is not true. H 0 is rejected if the trace statistic is higher than the 5% critical value. We run the test with each case specification and the results agree to detect zero cointegrating equations (a maximum rank of zero). Only the unrestricted trend does not display any conclusion from the test but, since the other results matched, we consider rank = 0 the right solution. This implies that the two time series could be fitted into a VAR model. ## VAR Model The VAR model allows investigating the interaction of several endogenous time series that mutually influence each other. We do not only want to detect if Bitcoin price could be determined by the one suggested by the cost of production model; we also want to check if the price has an influence on the model price. This latter relation can occur if, for example, a price increase leads to a higher cost for the mining hardware. In fact, a raise in the price represents also a higher reward if the mining process is successfully conducted, with the risk to push hardware price atop, which in turn could boost the model price up. To explain how a VAR model is constructed, we present a simple univariate AR(p) model, disregarding any possible exogenous variables, which can be written as (22): y t = µ + φ 1 y t−1 + . . . + φ p y t−p + ε t (22) Or, in a concise form (23): φ(L)y t = µ + ε t (23) where y t depends on its p prior values, a constant (µ) and a random disturbance (ε t ). [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 9 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity TABLE 5 | Zivot–Andrews test. Zivot–Andrews test Intercept Trend Intercept + trend Result t stat Break Date t stat Break Date t stat Break Date lnPrice −2.964 1,083 26/03/2017 −2.049 261 25/12/2014 −2.562 1,196 17/07/2017 NON-stationary lnModelPrice −3.221 281 14/01/2015 −3.357 408 21/05/2015 −3.914 620 19/12/2015 NON-stationary �lnPrice −34.905 1,350 18/12/2017 −34.626 1,285 14/10/2017 −34.895 1,350 18/12/2017 Stationary �lnModelPrice −42.848 582 11/11/2015 −42.781 1,469 16/04/2018 −42.858 582 11/11/2015 Stationary Critical values Intercept Trend Intercept + trend 1% 5% 10% 1% 5% 10% 1% 5% 10% −5.34 −4.8 −4.58 −4.93 −4.42 −4.11 −5.57 −5.08 −4.82 Source: Authors’ elaboration. TABLE 6 | Proper number of lags. Lag LL LR df P FPE AIC HQIC SBIC 0 7160.95 8.0e−07 −8.36581 −8.3611 −8.35308 1 7190.57 59.237 4 0.000 7.7e−07 −8.39575 −8.38633* −8.37029* 2 7192.42 3.7134 4 0.446 7.8e−07 −8.39325 −8.37911 −8.35506 3 7194.48 4.1059 4 0.392 7.8e−07 −8.39097 −8.37231 −8.34005 4 7195.74 2.5346 4 0.638 7.8e−07 −8.38778 −8.36422 −8.32413 5 7197.81 4.1319 4 0.388 7.8e−07 −8.38552 −8.35725 −8.30914 6 7199.73 3.8486 4 0.427 7.8e−07 −8.38309 −8.35011 −8.29399 7 7201.63 3.8014 4 0.434 7.9e−07 −8.38064 −8.34295 −8.2788 8 7204.56 5.8468 4 0.211 7.9e−07 −8.37938 −8.33698 −8.26482 9 7208.36 7.6003 4 0.107 7.9e−07 −8.37914 −8.33204 −8.25185 10 7212.23 7.7429 4 0.101 7.9e−07 −8.37899 −8.32717 −8.23897 11 7213.48 2.5086 4 0.643 7.9e−07 −8.37578 −8.31925 −8.22304 12 7225.63 24.303 4 0.000 7.8e−07 −8.38531 −8.32407 −8.21983 13 7243.57 35.872* 4 0.000 7.7e−07* −8.4016* −8.33565 −8.2234 14 7244.29 1.4495 4 0.836 7.7e−07 −8.39777 −8.32711 −8.20684 15 7246.50 4.4025 4 0.354 7.7e−07 −8.39567 −8.3203 −8.19201 16 7248.86 4.7357 4 0.316 7.8e−07 −8.39376 −8.31368 −8.17737 Source: Authors’ elaboration. A vector of n jointly endogenous variables is express as (24): Where µ is a vector (Equation 26) of the n-constants: µ 1 µ 2 ... µ p  (26)   (24)    y t =   y 1,t y 2,t ... y n,t µ = the matrix of coefficients � i is Equation (27): This n-element vector can be rearranged as a function (Equation 25) of n constants, p prior values of Y t, and a vector of n random disturbances, ǫ t : y t = µ + φ 1 y t−1 + . . . + φ p y t−p + ǫ t (25)  (27)  � 1 = φ i,11 φ i,12 - · · φ i,1n φ i,21 φ i,22 - · · φ i,2n ... ... ... ... φ i,n1 φ i,n2 . . . φ i,nn [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 10 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity TABLE 7 | Regressions of the Vector Autoregression model. Variables (1) (2) dlnPrice dlnModelPrice L.dlnPrice 0.18330223*** 0.00799770 (0.02359822) (0.02055802) L.dlnModelPrice −0.00655017 −0.02899205 (0.02762476) (0.02406582) Dummy −0.00588960*** 0.00027999 (0.00185465) (0.00161571) Constant 0.00236755*** 0.00149779** (0.00086910) (0.00075713) Observations 1,726 1,726 R [2] 0.04178812 0.00092579 Standard errors in parentheses. ***p < 0.01, **p < 0.05, and *p < 0.1. Source: Authors’ elaboration. and ǫ t consists in Equation (28): ε 1 ε 2 ... ε p TABLE 8 | Regressions with robust standard errors. Variables (1) (2) dlnPrice dlnModelPrice L.dlnPrice 0.18330223*** 0.00799770 (0.04306718) (0.01592745) L.dlnModelPrice −0.00655017 −0.02899205*** (0.02681078) (0.00979148) Dummy −0.00588960*** 0.00027999 (0.00225058) (0.00142356) Constant 0.00236755*** 0.00149779* (0.00078480) (0.00078942) Observations 1,726 1,726 R [2] 0.04178812 0.00092579 Robust standard errors in parentheses. ***p < 0.01, **p < 0.05, and *p < 0.1. Source: Authors’ elaboration. This feature does not compromise the unbiasedness or the consistency of the OLS coefficients but invalidates the usual standard errors. In time series analysis, heteroscedasticity is usually neglected, as the autocorrelation of the error terms is seen as the main problem due to its ability to invalidate the analysis. Since it is not possible to check and correct heteroscedasticity while performing the VAR model, we run each VAR regression separately and check the presence of heteroscedasticity by running the Breusch-Pagan test, whose null hypothesis states that the error variance are all equal (homoscedasticity) against the alternative hypothesis that the error variances change over time (heteroscedasticity). H 0 : σ 1 [2] [=][ σ] [ 2] 2 [=][ . . .][ =][ σ] [ 2] (31) The null hypothesis is rejected if the probability value of the chisquare statistic (Prob < chi2) is <0.05. The results of the test for both regressions show that the null hypothesis is always rejected, implying the presence of heteroscedasticity in the residuals (Table A.10 in **Supplementary Material** ). We try to correct the issue using heteroscedasticity-robust standard errors. The results are displayed in **Table 8** . These new robust standard errors are different from the standard errors estimated with the VAR model, while the coefficients are unchanged. The first difference of lnPrice depends even in this case on its lag, but, contrary from the VAR, now the first difference of lnModelPrice is not independent from its previous values. This new specification confirms the previous finding that each variable does not depend on the lagged value of the other one. Therefore, it seems that during the time window considered, the Bitcoin historical price is not connected with the price derived by Hayes’ formulation, and vice versa. Recalling **Figure 1**, it seems that the historical price fluctuated around the model (or implied) price until 2017, the year in which Bitcoin price significantly increased. During the last months of 2018, the prices seem to converge again, following a common path. In our analysis, we focus on the time window in which Bitcoin experienced its higher price volatility (Figure  (28)  ǫ t =   With Eǫ t = 0 and Eǫ t ǫ [′] s = ��0,, t t ̸= = s s the elements of ǫ t can be contemporaneously correlated. Given these specifications, a pth-order VAR can be presented as Equation (29): �(L) u t = µ + ǫ t (29) To clarify this expression, the ith endogenous time series can be extracted from these basic VAR and be represented as (30): y i,t = µ i + φ 1,i1 y 1,t−1 + . . . + φ 1,in y n,t−1 + φ 2,i1 y 1,t−2 + . . . + φ 2,in y n,t−2 + . . . + φ p,i1 y 1,t−p + . . . + +φ p,in y n,t−p + ε i,t (30) The result of the VAR model considering the dummy variable is presented in **Table 7** : As expected, the dummy is significant in the dlnPrice function but not in dlnModelPrice. Looking at the significance of the parameters, we can see how dlnPrice depends on its lagged value, on the dummy and on the constant term, but it seems not to be linked with the lagged value of dlnModelPrice. The regression of dlnModelPrice appears not to be explained by any variable considered in the model. We then check the stationarity of the model. The results confirm that the model is stable and there is no residual autocorrelation (Table A.9 in **Supplementary Material** ). ## Heteroscedasticity Correction Given the series’ path and the daily frequency of the data, the variables included in the model are probably heteroskedastic. [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 11 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity A.1 in **Supplementary Material** ) and the results suggest that it is disconnected from the one predicted by the model. These findings may depend on the features of the new cryptocurrencies, which have not been completely understood yet. The previous analyses, conducted on different time periods, by Hayes (2019) and Abbatemarco et al. (2018) assert that Bitcoin price could be justified by the costs and revenues of its blockchain network, leading to an opposite result from ours. We suggest that the difference could be based on the time window analyzed since we make a further step evaluating also the months in which Bitcoin price was pushed atop and did not follow a stable path. We think that there is not enough knowledge on cryptocurrencies to assert that Bitcoin price is (or is not) based on the profit and cost derived by the mining process, but these intrinsic characteristics must be considered and checked also in further analysis that include other possible Bitcoin price drivers suggested by the literature. ## CONCLUSIONS The main findings of the analysis presented show how, in the considered time frame, the Bitcoin historical prices are not connected with the price derived from the model, and vice versa. This result is different from the one obtained by Hayes (2019) and Abbatemarco et al. (2018), who conclude that the Bitcoin price could be explained by the cost of production model. The reason behind these opposite outcomes could be the considered time window. In fact, our analysis includes also those months where Bitcoin price surges up, reaching a peak of $19,270 on 19th December 2017, without following a seasonal path (Figure A.1 in **Supplementary Material** ). This has a relevant impact on the results even if the historical price started declining in 2018, converging again to the model one. Looking at the overall time frame, it seems that the increasing value of the historical price from the beginning of 2017 to the end of 2018 is a unique episode that required some months to get back to more standard behavior (Caporale et al., 2019). It seems now possible to assert that Bitcoin could not be seen as a virtual commodity, or better not only. According to Abbatemarco et al. (2018), the implemented approach does not rule out the possibility of a bubble development and, given the actual time frame, this is the reason why it would be more precise to explain Bitcoin price not only with the one implied by the model, but also with other explanatory variables that the literature seems to identify as meaningful. Therefore, to avoid misleading results, Bitcoin intrinsic characteristics must be considered and checked by adding to the profit and cost functions also these suggested parameters that range from technical aspects and Internet components to financial indexes, commodity prices, and exchange rate. This could open new horizons for research, which, despite the traditional drivers, should consider also new factors such as Google Trends, Wikipedia queries, and Tweets. These elements are related to the Internet component and appear to be particularly relevant given the social and digital Bitcoin’s nature. Kristoufek’s (2013) intuition, which considers Bitcoin as a unique asset that presents properties of both a speculative financial asset and a standard one, whose price drivers will change over time considering its dynamic nature and volatility, seems to be confirmed. The explanatory power of the VAR specification we implemented to inspect fundamental vs. market price dynamics could be quite low, which is to ascribe to missing factors and volatility. Further researches could include more tests on the VAR specification also including other controls/factors to check whether, for example, the VIX is another and important explanatory factor. More involved analyses should also explore for latent factors and/or time-varying relationships with stochastic and jump components. Although there are highlighted elements of uncertainty, Bitcoin has undoubtedly introduced to the market a new way to think about money transfers and exchanges. The distributed ledger technology could be a disruptive innovation for the financial sector, since it can ease communication without the need of a central authority. Moreover, the spread of private cryptocurrencies, which enter into competition with the public forms of money, could affect the monetary policy and the financial stability pursued by official institutions. For these reasons, central banks all over the world are seeking to understand if it is possible to adopt this technology in their daily operations, with the aim of including it in the financial system and controlling its implementations, enhancing its benefits, and reducing its risks (Gouveia et al., 2017; Bank for International Settlements, 2018). ## DATA AVAILABILITY STATEMENT All datasets generated for this study are included in the article/ **Supplementary Material** . ## AUTHOR CONTRIBUTIONS FZ: Introduction, Literature Review, and Conclusions. CB: Materials and Methods, Main Outcomes, and Conclusions. ## ACKNOWLEDGMENTS We acknowledge useful comments and suggestions from two anonym ous referees that have helped to substantially improve the paper. We are also grateful to Alessia Rossi, who has helped us in collecting and processing data. ## SUPPLEMENTARY MATERIAL The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frai.2020. 00021/full#supplementary-material [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 12 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) ----- Baldan and Zen Bitcoin as a Virtual Commodity ## REFERENCES Abbatemarco, N., De Rossi, L., and Salviotti, G. (2018). An econometric model to estimate the value of a cryptocurrency network. The Bitcoin case. Association for Information Systems. AIS Electronic Library (AISeL) ECIS. [Available online at: https://aisel.aisnet.org/ecis2018_rp/164?utm_source=aisel.](https://aisel.aisnet.org/ecis2018_rp/164?utm_source=aisel.aisnet.org%2Fecis2018_rp%2F164&utm_medium=PDF&utm_campaign=PDFCoverPages@@uline@) [aisnet.org%2Fecis2018_rp%2F164&utm_medium=PDF&utm_campaign=](https://aisel.aisnet.org/ecis2018_rp/164?utm_source=aisel.aisnet.org%2Fecis2018_rp%2F164&utm_medium=PDF&utm_campaign=PDFCoverPages@@uline@) [PDFCoverPages@@uline@ (accessed February 20, 2020).](https://aisel.aisnet.org/ecis2018_rp/164?utm_source=aisel.aisnet.org%2Fecis2018_rp%2F164&utm_medium=PDF&utm_campaign=PDFCoverPages@@uline@) Bank for International Settlements (2018). Central bank digital currencies. Committee on Payments and market Infrastructures. Markets Committee. Becketti, S. (2013). Introduction to Time Series Using Stata. College Station, TX: Stata Press. Boffelli, S., and Urga, G. (2016). Financial Econometrics Using Stata. College Station, TX: Stata Press, 17–30. Bouoiyour, J., and Selmi, R. (2015). What does bitcoin look like? Ann. Econ. [Financ. 16, 449–492. Available online at: http://aeconf.com/Articles/Nov2015/](http://aeconf.com/Articles/Nov2015/aef160211.pdf) [aef160211.pdf](http://aeconf.com/Articles/Nov2015/aef160211.pdf) Caporale, G. M., Plastun, A., and Oliinyk, V. (2019). Bitcoin fluctuations and the frequency of price overreactions, Financ. Mark. Portf. Manag. 33, 109–131. [doi: 10.1007/s11408-019-00332-5](https://doi.org/10.1007/s11408-019-00332-5) Chevallier, J., Goutte, S., Guesmi, K., and Saadi, S. (2019). Study of the Dynamic [of Bitcoin’s Price, HAL. Available online at: https://halshs.archives-ouvertes.fr/](https://halshs.archives-ouvertes.fr/halshs-02175669) [halshs-02175669 (accessed February 20, 2020).](https://halshs.archives-ouvertes.fr/halshs-02175669) Chishti, S., and Barberis, J. (2016). The FinTech Book: The Financial Technology Handbook for Investors, Entrepreneurs and Visionaries. Chichester: Wiley. Ciaian, P., Rajcaniova, M., and Kancs, D. A. (2016). The economics of Bitcoin price [formation, Appl. Econ. 48, 1799–1815. doi: 10.1080/00036846.2015.1109038](https://doi.org/10.1080/00036846.2015.1109038) de Vries, A. (2016). Bitcoin’s growing energy problem, Joule 2, 801–809. [doi: 10.1016/j.joule.2018.04.016](https://doi.org/10.1016/j.joule.2018.04.016) Dickey, D. A., and Fuller, W. A. (1979). Distribution of the estimators for autoregressive time series with a unit root, J. Am. Stat. Assoc. 74, 427–431. Engle, R. F., and Granger, C. W. (1987). Co-integration and error correction: representation, estimation, and testing, Econometrica 55, 251–276. Financial Stability Board (FSB) (2017). Financial Stability Implications from FinTech. Supervisory and Regulatory Issues that Merit Authorities’ Attention. Available online at: [http://www.fsb.org/wp-content/uploads/R270617.pdf](http://www.fsb.org/wp-content/uploads/R270617.pdf) (accessed February 20, 2020). Garcia, D., Tassone, C. J., Mavrodiev, P., and Perony, N. (2014). The digital traces of bubbles: feedback cycles between socio-economic signals in the Bitcoin [economy. J. R. Soc. Interface 11, 1–9. doi: 10.1098/rsif.2014.0623](https://doi.org/10.1098/rsif.2014.0623) Giudici, P., and Abu-Hashish, I. (2019). What determines bitcoin exchange prices? A network VAR approach. Financ. Res. Lett. 28, 309–318. [doi: 10.1016/j.frl.2018.05.013](https://doi.org/10.1016/j.frl.2018.05.013) Gouveia, O. C., Dos Santos, E., de Lis, S. F., Neut, A., and Sebastián, J. (2017). Central Bank Digital Currencies: Assessing Implementation Possibilities and Impacts. BBVA Working Paper, No. 17/04. Hayes, A. (2015). A Cost of Production Model for Bitcoin. Department of Economics. The New School for Social Research, Working Paper No. 5. Hayes, A. (2017). Cryptocurrency value formation: an empirical analysis leading to a cost of production model for valuing bitcoin, Telemat. Inform. 34, 1308–1321. [doi: 10.1016/j.tele.2016.05.005](https://doi.org/10.1016/j.tele.2016.05.005) Hayes, A. (2019). Bitcoin price and its marginal cost of production: support for a fundamental value, Appl. Econ. Lett. 26, 554–560. [doi: 10.1080/13504851.2018.1488040](https://doi.org/10.1080/13504851.2018.1488040) Hileman, G., and Rauchs, M. (2017). Global Cryptocurrency Benchmarking Study. Cambridge Centre for Alternative Finance, University of Cambridge, Judge Business School. Katsiampa, P. (2017). Volatility estimation for Bitcoin: a comparison of GARCH [models, Econ. Lett. 158, 3–6. doi: 10.1016/j.econlet.2017.06.023](https://doi.org/10.1016/j.econlet.2017.06.023) Kjærland, F., Khazal, A., Krogstad, E. A., Nordstrøm, F. G. B., and Oust, A. (2018). An analysis of Bitcoin’s price dynamics. J. Risk Financ Manag. 11:63. [doi: 10.3390/jrfm11040063](https://doi.org/10.3390/jrfm11040063) Kristoufek, L. (2013). Bitcoin meets Google Trends and Wikipedia: quantifying the relationship between phenomena of the Internet era. Sci. Rep. 3:3415. [doi: 10.1038/srep03415](https://doi.org/10.1038/srep03415) Kristoufek, L. (2015). What are the main drivers of the bitcoin price? evidence from wavelet coherence analysis. PLoS ONE 10:e0123923. [doi: 10.1371/journal.pone.0123923](https://doi.org/10.1371/journal.pone.0123923) Lütkepohl, H., and Krätzig, M. (2004). (Eds.). Applied Time Series Econometrics. Cambridge University Press. Matta, M., Marchesi, M., and Lunesu, M. I. (2015). “Bitcoin spread prediction using social and web search media,” in Proceedings of Conference on Workshop Deep Content Analytics Techniques for Personalized & Intelligent Services. (Dublin). Nakamoto, S. (2008). Bitcoin: A Peer-to-Peer Electronic Cash System. Available [online at: https://Bitcoin.org/Bitcoin.pdf (accessed February 20, 2020).](https://Bitcoin.org/Bitcoin.pdf) OECD (2018). Financial Markets, Insurance and Pensions. Digitalisation and Finance. Phillips, P. C. B., and Perron, P. (1988). Testing for a unit root in time series [regression. Biometrika 75, 335–346. doi: 10.1093/biomet/75.2.335](https://doi.org/10.1093/biomet/75.2.335) Schena, C., Tanda, A., Arlotta, C., and Potenza, G. (2018). Lo sviluppo del FinTech. CONSOB. Quaderni Fintech. Soltani, M., Kashkooli, F. M., Dehghani-Sanij, A. R., Kazemia, A. R., Bordbar, N., Farshchi, M. J., et al. (2019). A comprehensive study of geothermal heating and cooling systems. Sustain. Cities Soc. 44, 793–818. doi: 10.1016/j.scs.2018. 09.036 Waheed, M., Alam, T., and Ghauri, S. P. (2006). Structural Breaks and Unit Root: Evidence from Pakistani Macroeconomic Time Series. Available Online [at: https://ssrn.com/abstract=963958 (accessed February 20, 2020).](https://ssrn.com/abstract=963958) Zivot, E., and Andrews, D. (1992). Further evidence of great crash, the oil price shock and unit root hypothesis. J. Bus. Econ. Stat. 10, 251–270. [doi: 10.1080/07350015.1992.10509904](https://doi.org/10.1080/07350015.1992.10509904) **Conflict of Interest:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Copyright © 2020 Baldan and Zen. This is an open-access article distributed [under the terms of the Creative Commons Attribution License (CC BY). The use,](http://creativecommons.org/licenses/by/4.0/) distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. [Frontiers in Artificial Intelligence | www.frontiersin.org](https://www.frontiersin.org/journals/artificial-intelligence) 13 [April 2020 | Volume 3 | Article 21](https://www.frontiersin.org/journals/artificial-intelligence#articles) -----
{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC7861307, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.frontiersin.org/articles/10.3389/frai.2020.00021/pdf" }
2,020
[ "JournalArticle" ]
true
2020-04-30T00:00:00
[ { "paperId": "d7e8216c2ccf352092faee4c8efc37ac75bd8748", "title": "Study of the dynamic of Bitcoin's price" }, { "paperId": "0681f2b62a76d0deeff58d8927df30f8f3fc8e09", "title": "Bitcoin fluctuations and the frequency of price overreactions" }, { "paperId": "c44a39fcacc18debbdc02a94a1deefe1a222726a", "title": "Central Bank digital currencies" }, { "paperId": "835cd1d2eb825837466de11bd52a51ee192caa92", "title": "What determines bitcoin exchange prices? A network VAR approach" }, { "paperId": "6acd750d0beb00a8e204e30d5538c4323c764da3", "title": "A comprehensive study of geothermal heating and cooling systems" }, { "paperId": "5bcbe099925ebec9e9ff66818dfad08d2562b733", "title": "An Analysis of Bitcoin’s Price Dynamics" }, { "paperId": "d4a4785d6f75f0b83a5b3105e49eb605285d9d80", "title": "Bitcoin price and its marginal cost of production: support for a fundamental value" }, { "paperId": "51648f57cd9d445519ebba8ff583a7150d9bd0e7", "title": "Bitcoin's Growing Energy Problem" }, { "paperId": "886e6a9047853392a0e695b1f7f466a0de1fc158", "title": "Volatility estimation for Bitcoin: A comparison of GARCH models" }, { "paperId": "570d73250449ba3b9f0c4d43912e7bc2e1f09b51", "title": "Central Bank Digital Currencies: assessing implementation possibilities and impacts" }, { "paperId": "dc4a51341b37fc376af3958e1a4b2f705335cd3b", "title": "Financial Econometrics Using Stata" }, { "paperId": "aeccfade09d20de2b0ad8071aec6be9c2a445d37", "title": "The FINTECH Book: The Financial Technology Handbook for Investors, Entrepreneurs and Visionaries" }, { "paperId": "9ab35bdcec458f0c4ad818c48f158b8117ff719b", "title": "THE FINTECH BOOK: THE FINANCIAL TECHNOLOGY HANDBOOK FOR INVESTORS, ENTREPRENEURS AND VISIONARIES" }, { "paperId": "58252961cea2898d6fe39a1529ff11595b867193", "title": "Cryptocurrency Value Formation: An Empirical Analysis Leading to a Cost of Production Model for Valuing Bitcoin" }, { "paperId": "c1757f6ff1faac403294eebf809df251944b7c9b", "title": "Bitcoin Fluctuations and the Frequency of Price Overreactions" }, { "paperId": "4d4d41bd4821db58d5d813dfa659ac949555a953", "title": "A Cost of Production Model for Bitcoin" }, { "paperId": "f4391720375a2ad269a95826a2eb15a7ddb48ba8", "title": "What makes an accurate and reliable subject-specific finite element model? A case study of an elephant femur" }, { "paperId": "7328d31f9b5078658718ada22e94088c5a094291", "title": "The digital traces of bubbles: feedback cycles between socio-economic signals in the Bitcoin economy" }, { "paperId": "b2c445eb9b5b917f51ac14e91a7603e58553d29a", "title": "What Are the Main Drivers of the Bitcoin Price? Evidence from Wavelet Coherence Analysis" }, { "paperId": "a4c31b5d6cc6ca750a3bc982101ff12354e07805", "title": "The economics of BitCoin price formation" }, { "paperId": "f544743ba489bce65e1a4df0d80d48754f81bcd1", "title": "BitCoin meets Google Trends and Wikipedia: Quantifying the relationship between phenomena of the Internet era" }, { "paperId": "4f6c218687cd155d00ef17eb1fe3793429ca3a9f", "title": "Introduction to Time Series Using Stata" }, { "paperId": "caee2834e07a27c255d3cb45645a1bb2ec77ef41", "title": "Structural Breaks and Unit Root: Evidence from Pakistani Macroeconomic Time Series" }, { "paperId": "33c4502a781403a69a3bc2f57b1f7ef7abbe2994", "title": "Further Evidence on the Great Crash, the Oil-Price Shock, and the Unit-Root Hypothesis" }, { "paperId": "fafcf2e5f7878549d6537aa8ff5582b8fdcc3a34", "title": "Testing for a Unit Root in Time Series Regression" }, { "paperId": "9885b611792aedb706ba905e610b4a0a409d5984", "title": "Co-integration and error correction: representation, estimation and testing" }, { "paperId": "5cbbb5deb4d92dc0504fb7f2af0f6fe7da355d98", "title": "Distribution of the Estimators for Autoregressive Time Series with a Unit Root" }, { "paperId": null, "title": "Conflict of Interest" }, { "paperId": "d6bebe45c128e4d0bca46c2c4c2b11b44de373ab", "title": "An econometric model to estimate the value of a cryptocurrency network. The Bitcoin case" }, { "paperId": null, "title": "Lo sviluppo del FinTech" }, { "paperId": null, "title": "Financial Markets, Insurance and Pensions" }, { "paperId": "f4bee5e8eeacfd6abb059c901c19155da385861b", "title": "Financial Stability Implications from FinTech: Supervisory and Regulatory Issues that Merit Authorities' Attention" }, { "paperId": "94c347a7c426011d68da497e761efdc472852bfb", "title": "Global Cryptocurrency Benchmarking Study" }, { "paperId": "b7410adc691b805bab0129850b021d5285193f59", "title": "What Does Bitcoin Look Like" }, { "paperId": "1345a50edee28418900e2c1a4292ccc51138e1eb", "title": "Bitcoin Spread Prediction Using Social and Web Search Media" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": "0eb41ce10d763ce2cbe1006fae83d911b89d23a4", "title": "Co-Integration and Error Correction : Representation , Estimation , and Testing" }, { "paperId": "f75230a9f5a0a5a39d53099f7815234abeeb38ad", "title": "Testing for a Unit Root in Time Series with TARCH-Skew-t Errors" }, { "paperId": "247ce4c5e0cd2935bfe5a5e58c2828ca51986444", "title": "Applied Time Series Econometrics" }, { "paperId": null, "title": "Unrestricted constant ( τ = 0, ρ = 0): lets linear trends in y t to present a linear trend ( γ ) but the cointegrating equations are stationary around a constant means ( ν )" }, { "paperId": null, "title": "Bitcoin as a Virtual Commodity" }, { "paperId": null, "title": "Unrestricted trend: allows for quadratic trend in the level of y t ( τ t appears in the equation) and states that the cointegrating equations are trend stationary" }, { "paperId": null, "title": "Restricted trend ( τ = 0): excludes quadratic trends but includes linear trends ( ρ t )" } ]
19,897
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0127a973ed71969408342e74ef1b27ebbb67a76d
[]
0.903294
FOG Computing: The new Paradigm
0127a973ed71969408342e74ef1b27ebbb67a76d
[ { "authorId": "11013929", "name": "H. Alwageed" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
_Foundation of Computer Science FCS, New York, USA_ _Volume 3– No.5, November 2015 – www.caeaccess.org_ # FOG Computing: The new Paradigm ## Hathal Salamah A. Alwageed #### Department of Computer Engineering and Network Aljouf University, Saudi Arabia ## ABSTRACT As the Internet of Everything (IoE) heats up, Cisco engineers put forward a new networking, compute, and storage paradigm that extends to the edge of the network [http://newsroom.cisco]. Fog Computing is a paradigm that stretches out or extends Cloud Computing and services to the systems or network edge. Like Cloud, Fog gives information/data, process or compute, storage, and application services to end-clients. The recognizing Fog attributes are its closeness to end-clients, its tightly packed geographical conveyance or distribution, and its backing for mobility. Services are facilitated at the network edge or even end devices, for example, set-top-boxes or end points. Thusly, Fog diminishes services latency, and enhances QoS, bringing about prevalent client experience. Fog Computing holds up up-and-coming Internet of Everything (IoE) applications that request real timing/unsurprising latency (Industrial computerization/automation, transportation, sensors networks and actuators). On account of its geographical distribution the Fog paradigm is very much situated for real-time huge information or big data and analytics. Fog bolsters compactly distributed data collection points, subsequently adding a fourth pivot to the frequently specified Big Data measurements such as volume, variety, and velocity. ## Keywords Cloud Computing, Distributed Computing, Networking, IoT ## 1. INTRODUCTION The Internet of Everything is changing how we interface with this present reality," Milito says. "Things that were totally isolated from the Internet some time as of late, for instance, cars, are at present continuing onto it. Regardless, as we go from one billion endpoints to one trillion endpoints around the globe, that makes an authentic adaptability/scalability issue and the defy of overseeing complex gatherings or cluster of endpoints – what we call 'rich systems' – as opposed to overseeing individual endpoints. Fog's hardware infrastructure and software platform handle that [http://newsroom.cisco]. The information and communication technology (ICT) gather routinely puts aside time to yield to the authentic meaning, extension and setting of the new terms that show up associated to new development examples and their related hype. Web services, big data, cloud computing are a few instances of developed terms that were puzzling when at first founded. The term Fog Computing is resulting in these present circumstances starting wreckage now. Not in the slightest degree like the representations over, 'the fog' is not obliged to a particular inventive reach. In this manner, we can expect the beginning perplexity about 'what the fog is?' to reach uncommon levels. As it consistently happens with new developments, an understanding definition ought to be surrendered to by the community to tone down hype and chaos. The central definitions tend to focus on just two or three perspectives, like flexibility in the cloud or interoperability in web services. The way that the Fog stuck together various uniting imaginative examples makes this issue fundamentally more genuine. To be sure, looking at any of the progressions related to the fog from a singular point may offer the false view that there is insignificant new to it. For example, late definition attempts have shown it as just advancement to our present cloud model. It's out and out selfevident, for event, Cisco's point of view of the fog [Flavio Bonom et al]. Fog is an expansion of the Cloud Paradigm," says Technical Leader Rodolfo Milito, one of Cisco's thought pioneers in fog computing, "It's similar to cloud yet closer to the ground. Fog computing architecture enhances the cloud out into this present reality, the physical world of things." Fog supplements the Cloud, tending to creating IoT applications that are geo-scattered or geographical distributed oblige low latency, or snappy flexibility support or mobility. Fog computing would prop up sensors (which ordinarily measure, recognize, and accumulate data) and actuators – which are devices that can perform a physical movement, for instance, closing a valve, moving the arms of a robot, or rehearsing the brakes in an auto [http://newsroom.cisco]. Not at all like ordinary server homesteads or data centers, Fog devices are geologically passed on over heterogeneous platforms, spreading over diverse management territories. Cisco is involved with creative proposals that energize service flexibility across over stages or platforms, and headways that defend end-customer and content security and confidentiality transversely over territories. Fog gives exceptional perks over a couple of verticals, for instance, IT, incitement, advancing, personal computing et cetera. Cisco is uncommonly motivated by proposals that accentuation on Fog Computing circumstances related to Internet of Everything (IoE), Sensor Networks, Data Analytics and other data concentrated services to show the upsides of such another perspective, to survey the trade offs in both exploratory and fabrication deployments and to address potential examination issues for those course of action ## 2. THE FOG DESCRIPTION Fog takes the data and workload technology to another level. We are currently discussing edge computing – the home of Fog. While Fog insightfully expands Cloud computing and impacts Cloud's essential progressions, Fog, by definition, compasses more broad geographic territories than Cloud, and in a denser way. Similarly, Fog devices are significantly more heterogeneous in nature, running from end-customer devices, access points, to edge routers and switches. To oblige this heterogeneity, Fog services are engrossed inside a holder for effortlessness of association. Holder or container technologies are Linux containers and Java Virtual Machine (JVM). Implications that look into service versatility transversely over Fog platform are astoundingly convincing. Specifically, - Technologies that support workload adaptability amidst Cloud and Fog platform in light of methodologies and the essential's infrastructure. ----- _Foundation of Computer Science FCS, New York, USA_ _Volume 3– No.5, November 2015 – www.caeaccess.org_ - Technologies that enhance various parts of service mobility. - Fog services will be orchestrated transversely over management domains; services will be provisioned, looked at and took after over these zones. Suggestion looking at security and insurance in the association of Fog Computing are engaged. Specifically, - Privacy, security peril examination for distinctive Fog players (ex: service supplier, end-customer, content supplier) in the association of particular Fog service verticals (ex: IoE, Sensor Networks, Data Analytics, IT, redirection, Personal Computing). - Technologies that ensure security and insurance of customer/substance across over zones. - Technologies that reliably fuse and widen existing Cloud security/insurance courses of action in the association of Fog. - While Fog gives astounding central focuses to advantages over a couple of verticals, for instance, IT, incitement, publicizing, personal computing so as to register et cetera., Cisco is outstandingly fascinated Fog ideal circumstances for Big Data services in a couple of verticals including IoE. Specifically, improvements in compute, storage offerings for data intensive services, for instance, the going with: - Interaction between the Fog and the Cloud. Typically, the Fog platform supports real time, critical examination, processes, and channels the data, and pushes to the Cloud data that is worldwide in time and geographical scope. - Collection of data and analytics (pulled from access devices, pushed to Cloud) - Data storage for redistribution (pushed from Cloud, pulled by downstream devices) - Technologies that empower data fusion in the above settings. - Analytics noteworthy for neighborhood communities transversely over distinctive verticals (ex: advancements video examination, social protection, sensing and performance observation et cetera.) - Methodologies, Models and Algorithms to streamline the cost and execution through workload flexibility amidst Fog and Cloud. - PC frameworks or networks can be portrayed into differing sorts in perspective of their size of operation. They include: - **LAN: Local Area Networks spread or cover a bit** physical area, like a home, office, or a small group of buildings, for instance, a school or university etc. - **WLAN: Wireless Local Area Networks engage** customers to move around within a greater degree domain, yet be remotely connected with the framework/network. - **WAN: Wide Area Networks spread a far reaching** district, like communication links that cross metropolitan, neighborhood, or national points of confinement. The Internet is the best outline of a WAN. - **MAN:** Metropolitan Area Networks are unfathomable frameworks that cover an entire city. **SAN** Storage Area Networks facilitate associate remote PC storage devices, for instance, disk arrays, optical jukeboxes and tape libraries to servers in such a way that they reserves of being secretly joined to the O.S. Considering this information, we put forward the going with importance of the Fog: Fog computing is a circumstance where a monster number of heterogeneous (remote/wireless and autonomous) widespread and decentralized devices bestow and conceivably cooperate among them and with the framework/network to perform storage and processing assignments without the intervention of third-parties. These errands can be for supporting major framework/network limits or new services and applications that continue running in a sandboxed circumstance. Customers leasing bit of their devices to have these services get encouragements for doing in that capacity. This definition incorporates the parts which we consider will be key components of the fog: all inclusiveness, improved framework capacities as an encouraging circumstance, and better sponsorship for support among devices. In the event that in light of the fact that the deficient front of the terms, the complexities amidst fog and cloud computing could be hard to handle for a couple of customers. Some could consider the fog just an "extension" of the cloud. ## 3. APPLICATIONS & USAGE CASES [datacenterknowledge.com] The expression "Fog computing" has been clutched by Cisco Systems as another perspective to reinforce remote data trade to sponsorship distributed devices in the "Web of Things." different passed on handling and storage new services are in like manner getting the expression. It develops earlier thoughts in distributed computing, for instance, content transport frameworks/networks, however allows the movement of more personality boggling services using cloud propels. Before you get confused for yet another development term, it's discriminating to grasp where Fog Computing expect a section. Regardless of the way that it is another wording, this development starting now has a spot within the present day's universe server ranch and the cloud. Passing on data close to the customer- The volume of data being passed on by method for the cloud makes a quick need to store data or diverse services. These services would be discovered closest to the end-customer to upgrade stillness concerns and data access. As opposed to cabin information at server ranch regions far from the end-point, the Fog expects to put the data close to the end-customer. Making dense geographical allotment- Fog making in order to process widens direct cloud services an edge framework/network which sits at different core interests. This, thick or dense, topographically scattered system helps from various perspectives. As an issue of first significance, gigantic data and examination ought to be conceivable speedier with better results. By then, administrators have the ability to ----- _Foundation of Computer Science FCS, New York, USA_ _Volume 3– No.5, November 2015 – www.caeaccess.org_ support range based adaptability demands and not have to explore the entire WAN. Finally, these edge (Fog) structures would be made in a way that continuous data examination transforms into a reality on a truly gigantic scale. Authentic sponsorship for adaptability and the IoE- As said some time recently, there is a quick augmentation in the measure of devices and data that we bring into play. Executives have the ability to impact the Fog and control where customers are coming in and how they get to this information or data. Not simply this upgrade customer execution, it similarly helps with security and insurance issues. By controlling data at diverse edge centers, Fog preparing consolidates focus cloud services with those of a truly distributed server datacenter platform. As more services are made to advantage the end-customer, edge and Fog frameworks will end up being more pervasive. Reliable joining with the cloud and diverse services- The musing isn't to supplant the cloud. With Fog services, we are prepared to enhance the cloud experience by disengaging customer data that needs to live on the edge. Starting there, heads have the ability to tie-in examination, security, or distinctive services clearly into their cloud model. This base still keeps up the cloud's thought while uniting the power of Fog Computing at the edge. ## 4 TECHNOLOGIES 4.1 The ubiquity of devices There is a tremendous augmentation in the amount of devices getting connected with the framework/network. This augmentation is driven by two sources: customer devices and sensors/actuators. Cisco unadventurously assesses that there will be 50 billion joined devices by 2020 [D.Evans]. This impact in the amount of devices per individual is illuminated by the increase of mobile phones e.g. cell phones and tablets, remarkably in developing countries. Yet, these imperative numbers will soon be overpasses by the group of distinguishing/acting devices put in every way that really matters everywhere on the assumed Internet of Things, IoT, and pervasive sensor networks. Wearable computing devices (smart watches, glasses, et cetera.), sharp urban zones [Taewoo Nam et al], smart metering devices sent by energy suppliers to explore usage at the home level [Beth Plale et al], self-driving vehicles, sensor networks et cetera will be genuine drivers to the all inclusiveness of related devices. Each one of these applications are developing the region of devices everywhere around us. Along these lines ubiquity has incited intensive investigation, provoking another kind of particular achievement that hopes to handle today's repressions in device size and battery lifespan. This may itself encourage the association of more devices, making a calm circle. ## 4.1.1. Battery Size and lifetime: Cost is an essential issue driving devices to be as meager as would be reasonable. This also grows device portability and lessens power use, which may be noteworthy in some association e.g. advantageous phones or sturdy fire sensors in remote boondocks. Packaging and power management headways hope to make smaller and more independent devices that can run way more in any event cost. System on Chip (SoC) headways addition fragments, for instance, CPU, memory e.g. HP's memristor [Duncan R et al], checks and outside interfaces in a singular chip. They oblige less room and exhaust less power than common multi-chip systems. System in Package (SiP) is an answer some spot amidst SoCs and multi chip structures: it outfits circuits in a single unit or 'package', and is used today for little devices, for instance, propelled cellular telephones or smart phones. Despite when better packaging may improve power consumption, this alone may not be adequate for it to last more. The IoT is calling for long life sensors which here and there won't have the ability to join with any power supply. Today's lithium-molecule batteries (LiB) are brought into play for flexible devices of different sorts; solid state LiB plans are obliged to supplant them in the medium term, extending up to three times today's energy thickness. Still, batteries in perspective of chemical power sources can transform into a compelling component in future upgrades: higher power requirements in an unobtrusive piece of the degree of current batteries. Research efforts are revolved around 3D micro-batteries. "3D" is a term that incorporates the efforts to sort out the anode and cathode of batteries in 3D plans (past the typical 2D courses of action), to enhance density of both its power and energy. Using those 3D structures at minute scale is realizing batteries of humble size and tremendous power. Moreover, we have to watch the advancement of RF-powered computing [Shyamnath Gollakota et al], which speaks to that energy can be harvested from encompassing radiofrequency signs, (for instance, TV, cell) to power low-end devices that sense, compute and communicate. Also renewable energy empowered devices are presently available. ## 4.2 Network Management or Administration Having various devices can be especially helpful to improve our systems at all levels from our home to the planet all things considered and help us with understanding them better. These devices ought to be masterminded and kept up once they get passed on e.g. a future phone encouraging a service sold to an outcast customer or third party or a remote sensor at the sea's base. Administering frameworks or networks of billions of heterogeneous devices that run one or more services is boundlessly trying and complex. A couple Fog advances have been creating to help disciplined this versatile quality: "softwareisation/Programming" of framework and service management for better flexibility; conclusive techniques for scaling management; "little or small" edge clouds to host services close to the endpoints or at the endpoints themselves; and circulated (P2P)- and sensor framework/network like approaches for application auto-coordination. ## 4.2.1 NetworkManagementSoftwareisation or Programming Organizing and keeping upgraded and secure fog networks, services and devices is done autonomously e.g. switches, servers, services and devices are freely administered by unusual inhabitants. These assignments are work raised and slip by slanted. For example, definitely comprehended Internet associations ensure a single chairman handles large number machines running a lone service sort. Planning and keeping up various diverse sorts of services running on billions of heterogeneous devices will simply fuel our present management issues. The Fog needs heterogeneous devices and their running services to be dealt with in a more homogeneous manner; ideally totally automated by programming. Network Function Virtualization (NFV) is obviously the most bewildering advancement in such way. NFV is the re-movement of telco overseers to their ----- _Foundation of Computer Science FCS, New York, USA_ _Volume 3– No.5, November 2015 – www.caeaccess.org_ nonattendance of ability and unvarying prerequisite for tried and true or reliable systems. NFV tries to give the limit of intensely passing on-interest network services e.g. a firewall, a switch or a WAN enlivening specialist, another LAN or a VPN or customer services e.g. a database where and when desirable. Software Defined Networks (SDN’s) are one of the sections needed for NFV, since some network services like making new "virtual" networks on top of the physical system ought to be conceivable by programming just. For instance, a couple of entries can be sent as virtual machines and their traffic can be solidly controlled because of SDN capacities in an area edge cloud. The programming of a generally hardware driven business amassed around switches and servers where services got passed on will achieve not so much lavish but rather more deft operations. A corresponding close estimation is proposed by Cisco with its first programming simply type of the IOS wrapped in with a Linux transport (IOx). The switch itself becomes a SDN-enabled virtualization establishment where NFV and application services are sent close to the spot where they are truly going to be used. On the other hand, IOx's computing capacities will even now be limited. [Arati Baliga et al], [Arijit Banerjee et al] put forwards, however NFV limits don't accomplish end customer devices or sensors yet. In like manner, NFV and IOx simply consider requirements of vendors, telco overseer's or operators. Network gear equipment is only a little division of the Fog' devices. Billions of customer handheld devices and conceivably trillions of sensors need to have a near automation set up that can adjust to the obliged scale. ## 4.2.2. Arbitrary or Asymptotic methods: At fog scale, simply definitive and asymptotic methods have all the earmarks of being achievable [G. Pollock et al]. These procedures join with parts in their own specific management endeavors so that: a) the manager just shows the last desired state (life-changing) rather than individual charges; and b) she/she acknowledge the setup may never happen in light of the way that when it is set out the system may have changed e.g. fog nodes are gone or fresh nodes show up. As a delineation of these methodologies, see exertion on definitive/declarative and asymptotic management ended by HP Labs in the past [G. Pollock et al]. Diverse vendors are similarly starting to bring into play dramatic structures to reasonable scale and multifaceted nature, for instance see Cisco's technique at managing OpFlex (a kind of Cisco's OpenFlow reinforced by IBM and Midokura) SDNs. ## 4.2.3. Clouds at the Edge Littler than anticipated or Mini-clouds are getting sent closer to the edge to the customer by means of private clouds. Telcos and gear venders are moving on that course also. Long Term Evolution (LTE's) Enhanced Packet Core (EPC) can without a doubt be stretched out to take account of their own specific mini-clouds. Having a modest cloud at the EPC can lend a hand to pass on services close customers (at the edge) and confine traffic there while diminishing trombone routes with the help of SDNs. In like manner, IOx is just a progression of the present cloud model in which routers can transform into the virtualization infrastructure given that their pervasiveness and hierarchical position help to fulfill domain. The fog engages customer devices to wind up or become the virtualization platform themselves. In this manner, they can lease some computing and storage aptitude of confinement for applications to continue running on them. In the Fog, both the framework/network and the services running on top of it can be passed on enthusiasm for a fog of edge devices. Service delivery to specific regions in the framework or network is remarkably streamlined. For example [S. Sae Lor et al] gives a specimen of storage functions being dynamically passed on in diff erent mini-fogs in picked framework territories so that lumbering data trades are quickened. ## 4.2.4. Scattered or Distributed Management: The management practices discussed so far relies on upon a supplier e.g. the telco administrator as the sole aware of framework/network and service operation. In any case, there are in like manner P2P and sensor framework/network like procedures that allow endpoints to team up in order to perform equivalent results, yet can scale better. P2P advances have been around for quite a while and they are growing enough to help pass on the fog's vision. They can abuse neighborhood while emptying the prerequisite for a central management or administration point. Applications like Popcorn Time have shown the benefits of a P2P model to pass on overall services at scale. Various the musings of P2P content distribution networks (CDNs) are pertinent to the fog also; a fog application could be seen as a content distribution network where some sort of data is exchanged between peers. Thusly, in the fog a subset of framework/network and customer device/sensor segments can go ahead as a minicloud or in other words a littler than ordinary fogs. Hence the fog becomes an area where applications and data are not any more expected to stay in united server ranches. This perks up versatility and draws in customers to hold control and obligation regarding own data/applications. Applications will then be completed by bringing into play droplets or little bits of code that can securely continue running in devices at the edge with bare minimum communication with central parts, reducing undesired exchanges of data to central servers in corporate server ranches (data centre’s). ## 4.3 The connectivity at a fog scale The region of perhaps unassuming devices all around is one and just of the fog's components. As indicated over, each one of these devices ought to be joined. The sheer volume of devices 50 billion handheld customer devices in 2020 together with various moreover identifying/acting devices of the IoT working throughout the day, consistently will likely minute individual present bandwidth and connectivity issues. A remarkable report in The EconoFog titled (Augmented Business) depicted how cows will be checked to ensure healthier, more sufficient supply of meat for people to eat up. In light of current circumstances, every year each cow produces around 200 Mega Byte of information. ## 4.3.1 The Physical Connectivity A result of having numerous billions of devices using and conveying data at the framework's or networks edge is that these networks transform into an enormous bottleneck [Metro network traffic growth]. Network managers have been genuinely placing assets into a blended sack of new remote access advances to adjust to the sudden augmentation in devices per customer; however these LAN and Personal Network, WAN and MAN hypotheses may come up short in an IoT world. Most efforts in WAN/MAN are revolved around LTE; LTEv12 will be the first feature that fulfills each one of the essentials of the International Telecommunications Union to be labeled 4G. 4G LTE/EPC ought to be totally taken off by 2017 [Metro network traffic growth] and it will augment the available information exchange limit or ----- _Foundation of Computer Science FCS, New York, USA_ _Volume 3– No.5, November 2015 – www.caeaccess.org_ bandwidth of edge frameworks/networks [D. Astely et al]. LAN development has improved to lessen congestion and boost the on-hand bandwidth at lower power utilization, see for instance the latest Wi-Fi determination, 802.11ac. Finally, there have been monstrous improvements in PNs. These short range advances oblige center points to deal with themselves, as no central access point may be open. Bluetooth Low Energy, ANT+, ZigBee and RF4CE are the most striking. ### 4.3.2 Network Connectivity Past upgrades on remote frameworks, other expansions are relied upon to engage correspondence in circumstances where having all endpoints joined with some LAN & WAN is not possible as a result of costs, nonappearance of enough links centers, for instance, base stations, etc. In the fog, each center must have the ability to go about as a router for its neighbors and must be adaptable to nodes entering and leaving the network and compactness. Mobile Ad-hoc Networks (MANET), which have been a discriminating research subject for a long time now [S. K. Sarkar et al], could be the reason for future fog frameworks as they will engage the course of action of thickly populated frameworks without obliging adjusted and costly establishments to be open to this point. Frankly, Bluetooth LE, ANT+, ZigBee and RFC4CE all allow the advancement of MANETs at any rate up to adjacent reach. How-ever, most capacity is still to be done to engage MANET in MAN and WAN frameworks. Remote Mesh Networks are answers close to MANETs. A WMN can bring into play system routers at its core, which have no transportability or connectivity. Nodes bring into play those routers to get accessibility, or diverse nodes if no quick association with the routers can be set up. Routers facilitate access to distinctive frameworks, for instance, cell, Wi-Fi, etc. There is still a raised examination development on WMNs and MANET. On top of WMNs and MANET or right on top of the wireless framework if achievable we come across the protocols that have been delivered for the IoT, as MQTT [MQTT Protocol] and CoAP [CoAP Protocol]. All are sketched out in perspective of two targets: low resource consumption and adaptability to dissatisfaction; they tend to take after a publish/subscribe (pub/sub) communication model. Both IoT protocols and network can benefit by data region: they not any more need to send all the data around the world continually. Just aggregates may be sent or a pub/sub model can be approved that can colossally facilitate our accessibility needs, tying potential blockage tribulations at the framework's or network’s edge more so with the happening to edge switch/handheld/sensor enabled littler than ordinary fogs. In addition to confining traffic at the edge, this has an extraordinarily constructive outcome or optimistic on confidentiality. ## 4.4 Confidentiality or Privacy Today, we ceaselessly discharge personal information by employing unusual things, services and platforms. Albrecht et al. picture a blunt, however reasonable, reality: we may think we are in charge of our client cards and our mobile applications and our smart fridges, yet we should not deceive ourselves. The information is not our own. It has a spot with Google, and IBM, and Cisco Systems and the overall MegaCorp that has your adjacent store. If you don't believe us, essentially make a go at evacuating your data from their databases [K. Albrecht et al]. Customers are ending up being logically stressed over the risk of having their private data revealed. As needs be, other than the specific troubles exhibited by the inescapability of devices, another example will push for a fog circumstance where data is not sent to a couple united services, but instead it is fairly kept in the framework/network for better assurance. Data proprietorship will be a discriminating establishment of the fog, where some applications will have the ability to bring into play the framework/network to run applications and administer data without relying upon united services. Securing mixed sensitive data in standard fogs is a particular alternative for keep security. Nevertheless, this makes it genuinely hard to perform any taking care of over such data. There is fundamental examination wear down this subject, for case using crypto-processors or applying excellent encryption lives up to expectations that figure while keeping some of its interesting properties, thusly allowing performing certain obliged endeavors on it [Raluca Ada Popa et al]. Still, such decisions have obliged suitability. In this way, customers will ask for inventive ways to deal with shield their assurance from any potential colossal kin like component. This will be an extraordinary impulse to get fog developments, as they will engage the framework/network to supplant centralized or united services. ## 5 PROPSECTS DEFIES Regardless of the way that the investigation pains and customer examples depicted in past sections are pushing to bring the Fog, the way is far from cleared. There are various open issues that will must be tended to make the fog a reality. It is essential to unforgivably recognize these issues so prospect investigation works can focus on them. The game plan of open defies for the fog to wind up the fact of the matter is: - _Restriction of Compute/Storage: Current examples_ are improving this with smaller, more energy proficient and all the more exceptional devices e.g. one of today's phones is more prevailing than various top notch desktops from fifteen years back. Still new-fangled changes are yielded for non buyer devices. - _Administration or Management: despite setting up_ the communication routes transversely over end center points or nodes, IoT/general handling nodes and applications running on top ought to be genuinely setup and intended to fill in as needed. Having potentially billions of little or small devices to be orchestrated, the fog will strongly rely on upon decentralized (adaptable) management methods that are yet to be attempted at this unprecedented scale. One thing that can be foreseen with certain level of buoyancy is that there will be no jam-packed control of the complete fog and asymptotic dramatic setup methods will turn out to be more indispensable. - Sync or Discovery: Applications running on devices may have need of either some agreed united or centralized points e.g. set up an upstream fortification if there are unreasonably few peers in our storage application. - Standardization: At the moment no systematized instruments are available so every individual from the framework/network can proclaim its accessibility to host others software components, and for others to sent it their software to be run. ----- _Foundation of Computer Science FCS, New York, USA_ _Volume 3– No.5, November 2015 – www.caeaccess.org_ - Accountability: Enabling customers to share their extra assets for host applications is discriminating to engage new plans of activity around the fog's thought. An honest to goodness course of action of rousing strengths ought to be made. The spurring powers can be funds related or for the most part e.g. unrestricted free data rates. On the other hand the nonappearance of central controlling entity in the fog makes it difficult to confirm if a given device is to make sure encouraging a section droplet or not. - Programmability: Controlling application lifecycle is by now a test in cloud circumstances [19]. The existence of minimal utilitarian units ―droplets‖ in more territories (devices) obliges the right reflections to be set up so programming designers don't need to deal with these difficult issues [12]. Easy to use APIs for programming designers will overwhelmingly rely on upon fundamental Management segments that outfit them with the right reflections to disguise the tremendous manysided nature of the Fog. A couple of vendors like Microsoft have successfully ventured in arranging themselves in this space. ##  Protection/Security: The similar security stresses that apply to contemporary virtualized circumstances can be expected to affect Fog devices encouraging applications. The region of secure sand-boxes for the execution of droplets applications acts up-to-the-minute out of the ordinary predicaments (Privacy & Trust). Prior to using distinctive devices or downsized fogs in the framework/network to run some software’s, withdrawal and sandboxing parts must be set up to ensure bidirectional trust among cooperating parties. The fog will allow applications to transform customer’s data in third parties hardware and software. This clearly displays strong stresses over data security and its detectable quality to those third parties. ## 6 CONCLUSION Fog Computing [dataversity.net] identifies with a to an extraordinary degree essential progression in Cloud Computing and in handling when all is said in done. Its improvement emphasizes the ascendance of a decentralized model for computing that is more versatile and facilitated than the routine centralized paradigm. Such deftness and versatility are fundamental with Big Data applications taking the kind of the IoT and its low or no inertia necessities. Fog Computing may not exhibit a panacea for the exceptional solicitations of the IoT and the rigid advancement towards mobile computing. In any case, it at any rate sees and attempts to address an extensive parcel of the circumscriptions of bound together models which simply attract more movement—with less and less transmission limit and frameworks organization/management capacities—as Big Data continues creating. It gives a sensible building response for these stresses which may even perk up in the near prospect. ## 7 REFERENCES [1] http://newsroom.cisco.com/featurecontent?type=webcont ent&articleId=1365576) [2] http://www.datacenterknowledge.com/archives/2013/08/ 23/welcome-to-the-fog-a-new-type-of-distributedcomputing/) [3] CoAP Protocol - IETF Draft‖ [http://tools.ietf.org/html/draft-ietf-core-coap-18.](http://tools.ietf.org/html/draft-ietf-core-coap-18) [4] MQTT Protocol - OASIS Specification. ―http://www.oasisopen.org/committees/mqtt/ [5] Metro network traffic growth: An architecture impact study. Technical report, Bell Labs Alcatel-Lucent, December 2013. [6] ―K. Albrecht et al‖, Connected: To everyone and everything [guest editorial: Special section on sensors]. _Technology and Society Magazine, IEEE, 32(4):31–34,_ winter 2013. [7] ―D. Astely et al‖: the evolution of mobile broadband. _Communications Magazine, IEEE, 47(4):44–51, April_ 2009. [8] ―Arati Baliga et al‖, Virtual private mobile network towards mobility-as-a-service. In _Proceedings_ _of the_ _Second_ _International_ _Workshop_ _on_ _MobileCloud_ _Computing and Services, MCS ’11, pages 7–12, New_ York, NY, USA, 2011. ACM. [9] ―Arijit Banerjee, et al ―, A lightweight mobile cloud offloading architecture. In _Proceedings of the Eighth_ _ACM International Workshop on Mobility in the_ _Evolving Internet Architecture, MobiArch ’13, pages 11–_ 16, New York, NY, USA, 2013. ACM. [10] ―Flavio Bonom et al‖, Fog computing and its role in the internet of things. In Proceedings of the First Edition _of_ _the MCC Workshop on Mobile Cloud Computing, MCC_ ’12, pages 13–16, New York, NY, USA, 2012. ACM. [11] ―Duncan R et al‖, The missing memristor found. Nature, (7191):8083, 2008. [12] ―D.Evans‖, ―The internet of things how the next evolution of the internet is changing everything‖. Technical report, CISCO IBSG, April 2011. [13] ―Shyamnath Gollakota et al‖, The emergence of rf powered computing. Computer, 99(PrePrints):1, 2013. [14] Kirak Hong, David Lillethun, Umakishore Ramachandran, Beate Ottenw¨alder, and Boris Koldehofe. Mobile fog: A programming model for largescale applications on the internet of things. In _Proceedings of the Second ACM SIGCOMM Workshop_ _on Mobile Cloud Computing, MCC ’13, pages 15–20,_ New York, NY, USA, 2013. ACM. [15] ―Taewoo Nam et al‖, Smart city as urban innovation: Focusing on management, policy, and context. In _Proceedings of the 5th International_ _Conference on_ _Theory_ _and_ _Practice_ _of Electronic_ _Governance,_ ICEGOV ’11, pages 185–194, New York, NY, USA, 2011. ----- _Foundation of Computer Science FCS, New York, USA_ _Volume 3– No.5, November 2015 – www.caeaccess.org_ [16] ―Beth Plale et al‖, Adaptive cyberinfrastructure for real time multiscale weather forecasting. _Computer,_ 39(11):56–64, November 2006. [17] ―G. Pollock et al‖, The asymptotic configuration of application components in a distributed system. Technical report, University of Glasgow, Glasgow, UK. [18] ―Raluca Ada Popa et al‖, Processing queries on an encrypted database. [19] _Communications of the ACM, 55(9):103–111, September_ 2012.―S. Sae Lor et al‖, The cloud in core networks. _Communications Letters,_ _IEEE,_ 16(10):1703–1706, October 2012. [20] ―S. K. Sarkar et al‖, Ad Hoc Mobile Wireless Networks: _Principles, Protocols, and Applications. CRC Press,_ 2007. [21] ―Luis M. Vaquero et al‖, Towards runtime reconfiguration of application control policies in the cloud. J. Netw. Syst. Manage., 20(4):489–512, December 2012. [22] http://www.dataversity.net/the-future-of-cloud computing-fog-computing-and-the-internet-of-things/) -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.5120/cae2015651946?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5120/cae2015651946, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://doi.org/10.5120/cae2015651946" }
2,015
[]
true
null
[]
9,553
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01290c447f2304ff3c347c3ac0dfa22f053e281b
[ "Computer Science" ]
0.910503
A Balanced Trust-Based Method to Counter Sybil and Spartacus Attacks in Chord
01290c447f2304ff3c347c3ac0dfa22f053e281b
Secur. Commun. Networks
[ { "authorId": "2998809", "name": "R. Pecori" }, { "authorId": "1685596", "name": "L. Veltri" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
A Sybil attack is one of the main challenges to be addressed when securing peer-to-peer networks, especially those based on Distributed Hash Tables (DHTs). Tampering routing tables by means of multiple fake identities can make routing, storing, and retrieving operations significantly more difficult and time-consuming. Countermeasures based on trust and reputation have already proven to be effective in some contexts, but one variant of the Sybil attack, the Spartacus attack, is emerging as a new threat and its effects are even riskier and more difficult to stymie. In this paper, we first improve a well-known and deployed DHT (Chord) through a solution mixing trust with standard operations, for facing a Sybil attack affecting either routing or storage and retrieval operations. This is done by maintaining the least possible overhead for peers. Moreover, we extend the solution we propose in order for it to be resilient also against a Spartacus attack, both for an iterative and for a recursive lookup procedure. Finally, we validate our findings by showing that the proposed techniques outperform other trust-based solutions already known in the literature as well.
Hindawi Security and Communication Networks Volume 2018, Article ID 4963932, 16 pages [https://doi.org/10.1155/2018/4963932](https://doi.org/10.1155/2018/4963932) #### Research Article A Balanced Trust-Based Method to Counter Sybil and Spartacus Attacks in Chord ###### Riccardo Pecori 1 and Luca Veltri 2 _1SMARTEST Research Centre, eCAMPUS University, Novedrate, CO 22060, Italy_ _2Department of Engineering and Architecture, University of Parma, Parma, PR 43124, Italy_ Correspondence should be addressed to Riccardo Pecori; riccardo.pecori@uniecampus.it Received 23 February 2018; Revised 27 September 2018; Accepted 14 October 2018; Published 12 November 2018 Academic Editor: Carmen Fernandez-Gago [Copyright © 2018 Riccardo Pecori and Luca Veltri. This is an open access article distributed under the Creative Commons](https://creativecommons.org/licenses/by/4.0/) [Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is](https://creativecommons.org/licenses/by/4.0/) properly cited. A Sybil attack is one of the main challenges to be addressed when securing peer-to-peer networks, especially those based on Distributed Hash Tables (DHTs). Tampering routing tables by means of multiple fake identities can make routing, storing, and retrieving operations significantly more difficult and time-consuming. Countermeasures based on trust and reputation have already proven to be effective in some contexts, but one variant of the Sybil attack, the Spartacus attack, is emerging as a new threat and its effects are even riskier and more difficult to stymie. In this paper, we first improve a well-known and deployed DHT (Chord) through a solution mixing trust with standard operations, for facing a Sybil attack affecting either routing or storage and retrieval operations. This is done by maintaining the least possible overhead for peers. Moreover, we extend the solution we propose in order for it to be resilient also against a Spartacus attack, both for an iterative and for a recursive lookup procedure. Finally, we validate our findings by showing that the proposed techniques outperform other trust-based solutions already known in the literature as well. ###### 1. Introduction By now, peer-to-peer (P2P) networks, be they structured or not, concern a significant and mature slice of the overall Internet traffic. Their usage ranges from file-sharing to VoIP applications [1], from on-line alerting systems to intrusion detection, and so forth. Moreover, their deployment is experiencing a new flourishing in all those scenarios involving the Internet of Things (IoT) [2]. The most important type of current structured P2P networks is the one relying on Distributed Hash Tables (DHT) algorithms. Among these, we can mention the well-known Kademlia [3], Chord [4], CAN [5], and Pastry [6], but other ones are emerging. A DHT is a distributed structure that maps identifiers to values, similar to a hash table. The lookup is performed through an efficient routing mechanism leading to the peer that actually maintains the mapping. In a DHT-based P2P network, the identifiers generally refer to both the peers (peer IDs) and the resources (resource IDs). The logical overlay and the routing tables of DHTs feature a solid structure allowing them to perform in a quick and simple way, but, at the same time, make them subject to various malicious attacks. The so-called Sybil attack [7] is a prominent example of these types of attacks. This attack usually involves an attacker managing a great amount of multiple false IDs, called sybils, able to taint the routing tables and, as a consequence, capable to disrupt or degrade the main operations of the DHT. _1.1. Novelty, Contribution and Motivation. In this work we_ consider a scenario that is implicitly infected by sybils, but we also address Spartacus-behaving nodes (spartaci), i.e., nodes that steal the IDs of other nodes. This is one of the main novelties of this work, since a Spartacus attack, i.e., a variant of the Sybil attack, is not very studied in the relevant literature yet. Differently from [8], which provides an admission control system, we do not focus on limiting the access of malicious nodes to the P2P network, but, similarly to the contribution in [9], involving a clean routing process, we tried to devise a trusted lookup and storage mechanism, involving only those nodes turning out to be the most trustworthy. As a consequence, the querying node will be able to take a decision by itself about which nodes to trust, by ----- 2 Security and Communication Networks considering, in a dynamic and evolving way, the interactions it experienced. In the aforementioned framework, we investigate and compare some reasonable trust-based techniques to avoid, or better still, to moderate the misconduct of malicious peers, be they sybils or spartaci. In particular, we focus on the wellknown Chord as DHT, trying to modify its lookup as well as its storage and retrieval procedures in a way similar to what has already been done for Kademlia in [10]. Since Chord has a different way of computing the distance between pairs of nodes and its lookup and storage and retrieval procedures are different from those of Kademlia, we needed to devise proper modification with respect to the solution presented in [10]. The contribution of this work is multifaceted and recapitulated in the following points: (1) We readapted a trust-based mixed strategy, which already proved effective in Kademlia both for routing [11] and for storage and retrieval procedures, to Chord, verifying its effectiveness during lookup as well as storage and retrieval procedures in a network infected by sybils. (2) Then we tested the proposed solution, called SChord, in a Spartacus attack scenario, considering both iterative and recursive lookup procedures, and we propose some improvements in order to make SChord resilient also to a Spartacus attack. (3) Finally, we compare the proposed strategy with other trust-based solutions published in the literature along the years. _1.2. Structure of the paper. The remainder of this article_ is structured in this way: in Section 2 we summarize the main concepts about the Sybil and Spartacus attacks and we describe some possible solutions already analyzed in the literature, and in Section 3 we investigate the effects of a Sybil attack in Chord, while we accurately present the extension of the technique in [10] to Chord in Section 4. In Section 5 we study the proposed strategy in a Spartacus attack scenario showing how the aforementioned procedure may result effective also in presence of spartaci through some degrees of fine-tuning. Finally Section 6 seals up the work and provides some suggestions on possible future followups. ###### 2. Background and Related Works The Sybil attack, firstly described in 2002 by Douceur in [7], exploits the redundancy of DHT networks, and it is usually launched through a malicious physical entity owning multiple virtual and logical identities. The trick is to introduce into the P2P network many fake identities (the so-called _sybils), which are controlled by one single attacker in the_ physical layer. This allows attackers to monitor the traffic, to partition the network, e.g., through an eclipse attack, or to misuse the DHT in different ways, e.g., performing a Distributed Denial of Service (DDoS) attack [12]. Different types of threats can be generated by the sybils and they can be classified into the following categories: (i) routing table invasions, (ii) storage and retrieval malfunctions, and (iii) miscellaneous attacks [13]. The first types encompass incorrect lookups and wrong routing table updates, and the second ones regard denying to store resources or simulating to have resources the nodes actually do not own, while the third types may concern inconstant behaviors, overloading of specific nodes, quick joining and leaving the network, and so on. Many solutions have been devised during the years for solving or mitigating the effects of a Sybil attack: they range from the introduction of trusted certification and computational puzzle approaches to costs and fees and trusted devices [10]. Some renowned solutions are Whanau [14], X-vine [15], and Persea [16]. However, these are protocols that leverage on a further social network of trusted relationships between the users of the P2P network and they need proper datasets to be evaluated. These are important limitations and overheads that, if not available, do not allow one to enact any countermeasure against the sybils. Moreover, these solutions take for granted that friends in a social network are trusted users in a P2P network, something not always true. Furthermore, they restrict the access of the sybils to the P2P system by constraining the number of possible paths among good and malicious peers. Conversely, we resort to a solution that applies to an open scenario where the sybils may freely join the network: we do not pose limitations to the possibilities of an attacker. The Sybil attack is still investigated nowadays, as demonstrated by some recent works such as [17, 18]. In [17] the authors focus on solving a limitation of Persea through lookup inspection for detecting the sybils, while the latter presents VoteTrust, a system to detect sybils through the friend invitation graph of social networks. However, both of these systems focus on sybils detection. The scenario of our work envisages a P2P network inherently infected by malicious nodes instead. Moreover, both works provide an active collaboration among good nodes for sybils detection and this may introduce a further overhead that our solution carefully avoids through an automatic trust computation. In this work, we analyze routing as well as storage and retrieval attacks in a Chord DHT and provide a solution, already proven effective in Kademlia [10], using trust metrics in a balanced way. We chose to focus on Chord, as it is one of the most well-known, relevant, and representative DHTs, and even if introduced in the early 2000s, it is still currently studied and considered a reference DHT for its simplicity and efficiency. This is, for example, shown by the following: a recent work that uses Chord in large scaled P2P networks in conjunction with a dynamic trust computing model [19], as well as recent studies on the further improvement of Chord inherent efficiency [20] and correctness [21], or on the joining time of Chord nodes through the usage of anchor peers in an educational scenario [22]. Further recent contributions have concerned the usage of Chord in mobile networks [23] and for location-based services [24]; therefore, Chord’s security and reliability are still important aspects to be carefully assessed and studied. ----- Security and Communication Networks 3 Some solutions for improving Chord’s security are already present in the relevant literature, such as [25] where the authors deploy certificates and signatures. Nevertheless, an increased number of messages, compared with standard Chord, are required; moreover, the authors try to remove malicious nodes from the network, something leading to a scenario that differs from the one considered in this work where we preferred another approach, i.e., accepting sybils in the network and trying to overcome their malicious activity, through direct trust, in such a way an attacker cannot recognize a countermeasure has been enacted. This is performed by following some successful solutions mainly studied in [10, 26]. Concerning trust and reputation in general, in [27] a method for comparing trust models based on a hierarchical fuzzy inferring model is proposed. However, the contribution focuses only on a file downloading scenario, like the mechanism proposed in [28], which is moreover applied only to nonstructured P2P networks. Conversely, the mechanism proposed in [29] can be applied in P2P networks based on DHTs, but it is difficult to implement in a real world scenario. Concerning the application of trust and reputation directly to Chord, some ideas can be found in [30–33]. In [31] the authors apply some of the strategies proposed by Koyanagi in [30], extending them to security purposes and not limiting them to maintenance aims. More precisely, the authors of [31] exploit Bayesian networks in order to enrich Chord finger tables with a further column, called “trust,” whose value will be used in order to finalize trustworthy lookups. They took advantage of a sort of variable central entity. In contrast, in our work, we prefer to keep a kind of decentralized web-of-trust, as it is the case with [34], and we consider a more general trust score, not limited to some features like age or downloads and uploads that could be specific to some P2P network usages. In [32], Koyanagi et al. propose a solution similar to ours where a general trust score computation and its weighing are used. However, our work differs in the fact that we consider a different balanced strategy, a mix between standard and trusted Chord using a tunable parameter. We also take into account the so-called environmental risk, analyzing the outcomes of a growing number of sybils, not only monitoring a constant percentage of them. Moreover, unlike [32], (i) we consider churn a normal action of peers, (ii) our solution does not require a look-ahead phase, and (iii) the computation of trust is simpler as we do not consider trust propagation but only direct experiences. Finally, the work in [33] is the most recent one attempting to apply a trust model to Chord by leveraging onto guarantee_ing and archive peers, whose reputation is evaluated together_ with the one of service peers. The proposal provides also an incentive system and an anonymous reputation management strategy; nevertheless, the authors consider at least 4 different roles for a peer and they envision a complex system for establishing guarantee relationships (at least 10 messages) and a lookup transaction (at least 14 messages). This is in contrast with our idea of keeping the system as simple as possible, featuring a minimum number of changes and a minimum overhead with respect to standard Chord; moreover, we do not set up different roles for the peers, which would lead to extra overhead, but we prefer a distributed approach, where all peers are at the same level. The Spartacus attack, a variation of the Sybil attack, has not yet been thoroughly examined within the relevant literature. In this attack, itself especially dangerous in an environment focusing on trust and reputation, the malicious physical entity does not generate a great amount of pseudonymous logical IDs but simply steals them from other real nodes, inheriting their trust value. Such an action takes place in the bootstrap phase, when a spartacus, intending to enter the network, looks for a possible bootstrap node, and later for nodes with high trust score. The spartacus pings these nodes and, when they appear to be disconnected, it replaces them by copying the hash of their IDs. If no way exists to bind a node to its virtual ID, and such is the scenario we consider, and the node replaced by the spartacus needs to choose another virtual ID and must start again to create its own trust score from the beginning. This malicious activity allows spartaci to acquire a higher trust score than sybils, at least initially, and if churn is enacted, this attack is much more feasible and dangerous. An example of possibly disrupting consequences can be found in a Fog computing IoT scenario, where smart gateways, belonging to the fog layer, may be part of a virtual peer-to-peer overlay. If one of these fog gateways is attacked by a spartacus node, all the information coming from the sensors managed by the attacked node may be easily lost or counterfeited. Some solutions against the Spartacus attack exist, but they are based on public cryptography, as it is the case with the “kad-spartacus” extension of the “kadtool” package [(http://www.npmjs.com/package/kad-spartacus), on a cen-](http://www.npmjs.com/package/kad-spartacus) tral server, on the binding between IP address and virtual ID or on an active verification of the node about to join the network, carried out by those nodes having already joined the network. However, these are not fully decentralized solutions. They require certification authorities and a public key infrastructure, leading to possible bottlenecks due to the presence of central servers or further operations on the joining procedure, such as the binding with network or physical addresses or other verification activities, which limit the entering of the peer-to-peer network and can cause additional overheads. Conversely, in our scenario we consider the joining procedure free from further constraints or limitations, and we do not use central authorities or servers. In the last part of this work we propose a new adaptation of the solution firstly proposed for a Sybil attack, called SChord, to work effectively also against a Spartacus attack. Furthermore, we provide also a comparison between the mixed trust-based technique and those described in [32, 33] and demonstrate they reach worse results when considering successful lookups and the average number of hops. ###### 3. Sybil Attack in Chord In this part of the paper, after briefly recalling Chord [4], we study the degradation of the performances of the DHT in presence of a Sybil attack by means of simulations, ----- 4 Security and Communication Networks considering both lookups and storage and retrieval operations. Some distinctions are made also between iterative and recursive lookup procedures. _3.1. A Brief Overview of Chord. Peers and resources are_ associated with a unique 𝑚-bit identifier (𝑚 is the ID length), calculated by using consistent hashing of the peer addresses. In the following, we indifferently name Chord entities as “peers” or as “nodes.” Chord logical IDs are ordered following a numerical circle called Chord ring, where distances are computed modulo 2[𝑚]. The owner of a resource key 𝑘 is defined as the first node with an ID matching or following 𝑘 in the key-ring modulo 2[𝑚]. This node is named 𝑠𝑢𝑐𝑐𝑒𝑠𝑠𝑜𝑟(𝑘). The definition of distance between a node 𝑛1 and a node 𝑛2 is (𝑛2 −𝑛1) 𝑚𝑜𝑑2[𝑚]; therefore it is asymmetric. Often, due to reliability reasons, redundancy is added and more than one node may be in charge of a certain key. Hereafter we separately describe the procedures for (i) lookup and (ii) join. _3.1.1. Chord Lookup. A lookup operation is used to find_ out what node is in charge of a specific identifier. Chord features two lookup procedures, called, respectively, 𝑏𝑎𝑠𝑖𝑐 and 𝑎𝑐𝑐𝑒𝑙𝑒𝑟𝑎𝑡𝑒𝑑. In the basic lookup, each node simply contacts its current successor, since the requests are transferred around the ring through successor pointers until they find a pair of node identifiers spanning the requested key identifier; the second in the pair is the target of the lookup query. The accelerated lookup takes advantage of further routing information: a so-called finger table made up of m entries. The 𝑖[𝑡ℎ] entry in the table of node n is the first node 𝑠𝑖 following n by a distance of minimum 𝑑𝑖 = 2[𝑖−1]. Essentially, 𝑠𝑖 is defined as 𝑠𝑢𝑐𝑐𝑒𝑠𝑠𝑜𝑟(𝑛+2[𝑖−1]) 𝑚𝑜𝑑2[𝑚], with 1 ≤𝑖≤𝑚. 𝑠𝑖 is termed as the 𝑖[𝑡ℎ] finger of peer n, and it can be denoted by 𝑓𝑖𝑛𝑔𝑒𝑟(𝑛, 𝑖). An entry on any finger table includes both the identifier and the peer address. It is possible to summarize the accelerated lookup procedure for any given key through the following steps: (i) A node is requested for a certain key. (ii) Should the node be in charge of the key being requested, the lookup ends. (iii) Otherwise, should the node’s successor be in charge for that key, the requested node responds with the successor node. (iv) Else the requested node returns the highest entry in the finger table that comes before the key. This version of the lookup algorithm assures that lookups, in a 𝑁-node network, require at most 𝑂(log2𝑁) hops in order to succeed. This is because the topological distance between the current requested node and the target node is divided by two at each repetition of the algorithm. Both the basic and the accelerated lookup may be performed in an iterative or recursive way [35]; during the former the querying peer has full control of the lookup process, while in the latter case at each step a different peer is in charge to forward the request. _3.1.2. Chord Join. A new node joins a Chord network by_ computing its ID and contacting a bootstrap node. This is performed by executing a find successor procedure, using the joining node ID as an argument. Once the joining node has discovered its successor, it establishes its successor node to be the admitting peer and sends to it a 𝑛𝑜𝑡𝑖𝑓𝑦 message that informs the admitting peer that a new predecessor has joined the network. _3.2. Simulation Conditions. In order to perform numerical_ simulations, we employed DEUS [36] together with its inherent Chord implementation. This simulator is based on discrete events and on virtual seconds (V𝑠) as regards simulation time. For the sake of simplicity, we only considered stationary conditions so that, regardless of peers churning, a fixed number of them are constantly active in the considered scenario. The number of peers that behave well, i.e., according to the standard Chord algorithm, is constant and set to 10,000, while malicious peers, i.e., the sybils, vary across the different simulations, and they can acquire the following figures: 0, 2,000, 4,000, 6,000, 8,000, 10,000, 15,000, and 20,000. Those values reaching more than 10,000 are considered as cases of overwhelming presence of malicious nodes. The simulation time is set to 30,000 V𝑠. Good nodes are subject to a churning Poisson process of parameter 𝜆 (equaling 10 joins/leaves every 1 V𝑠); malicious nodes, instead, join and leave following a random period process. The ID length 𝑚 is 128 bits, while the observation cycle amounts to 1,000 V𝑠, allowing for 30 observations for each simulation. The DEUS implementation of Chord does also feature a parameter, accounting for a maximum waiting time, named 𝑀𝑊, which is set to 1,000 V𝑠. This parameter is used to simulate those nodes responding too slowly to a request. Indeed, we only concentrate on lookup procedures that correctly terminated, that begin only when the network is in a stationary condition, and that do not cause 𝑀𝑊 to expire. Should 𝑀𝑊 expire, we labeled these lookups as “unended.” As a final remark, for all parameter sets, we considered the averages of the results computed over the 30 datasets coming from each observation cycle and over various simulation seeds. This was done in order to achieve a confidence interval amounting at least to 95% for the obtained results. _3.3. How the Sybils Affect Standard Chord. We study both_ lookup attacks as well as storage and retrieval attacks and both iterative and recursive accelerated lookups. In particular two malicious basic behaviors have been considered for lookups. _Sybils asked for a next hop could either_ (1) refuse to answer or (2) respond with a randomly chosen peer ID. Concerning storage and retrieval, we analyze different behaviors: (3) refusing to store a resource, (4) accepting a resource and then refusing the relative retrieval queries, ----- Security and Communication Networks 5 (5) returning a null resource in response to a retrieval request, (6) blocking storage and retrieval operations for certain specific resources chosen randomly. 100 90 80 70 60 50 40 30 20 10 In particular, since just the first and third behaviors are able to effectively disrupt standard Chord operations, the second and the fourth behaviors are studied only against our trustbased proposal in the following sections, in order to stress its effectiveness. Although various metrics to study a Sybil attack are present in the literature ([37, 38]), we mainly focused on the following: (i) for lookups, the average amount of successful requests and the mean hops per lookup; (ii) for storage and retrieval operations, the mean of successful store (PUT) and retrieval (GET) procedures. 0 0 5000 10000 15000 20000 Sybils number In Figure 1 we show lookups in a scenario where sybils refuse to provide an answer to the querying peer (case (1)) of the aforementioned list). We refer to successful operations when lookups reached the peer correctly in charge of the searched key, regardless of whether the lookup procedure has involved sybils or not. The outcomes demonstrate an intense decrease in successful lookup procedures while malicious nodes grow. Unsuccessful operations, which overtake successful lookups if the sybils overtake the 8,000 threshold, refer to lookups that have not yet finished once the 𝑀𝑊 timeout has expired; therefore they are labeled as “unended.” Furthermore, in the same figure, we show that a little fraction (2,000 or 4,000) of malicious nodes does not influence too much the overall performances, so there should be a minimum amount of malicious nodes that makes a Sybil attack successful. Conversely, those lookup requests being unsuccessful are somehow bounded when malicious peers are less than 6,000, while they increase very much and exceed successful lookups from 8,000 on. It is also relevant to the saturation behavior experienced by the curve when the sybils are more than 10,000. This could be due to the fact that the sybils have reached and overtaken good nodes and their negative effect tends to stabilize. In Figure 2 we consider the other malicious behavior: a random response (case (2)). In this case, the sybils respond to lookup queries with the ID of a random peer, regardless of whether it is malicious or not. In these simulation conditions, we point out a different assessment for genuine and nongenuine lookups, as this could be useful to analyze the behavior of iterative and recursive lookups. As a matter of fact, when we consider a no response attack, the iterative and recursive lookups are similarly affected; that is, the procedure is unended, whereas when we consider a random response, this could lead to different outcomes according to the number of sybils encountered during the procedure. More specifically, there are three cases: (i) successful genuine lookups, (ii) successful non-genuine lookups, and (iii) unended lookups. Genuine lookups refer to those lookups not affected by sybils at all, whereas non-genuine lookups are the lookups in which at least a sybil has been contacted during the procedure. It succ. unended Figure 1: Successful and unended lookups versus the number of _sybils, whenever sybils do not respond._ is evident from these definitions that in case of no response from the sybils this different behavior could not be detected. Returning to Figure 2, while the number of genuine lookups decreases gently in a way similar to the one depicted in Figure 1, the number of non-genuine ended lookups experiences a peak in correspondence to a number of sybils equaling 8,000 and then it sharply decreases reaching the performances of genuine successful operations. The opposite behavior can be observed for unended lookups that, when the _sybils number more than 8,000, undergo a sharp increase._ While non-genuine lookups have the chance of considering good nodes leading to the correct target, when sybils are not overwhelming, this cannot take place anymore when the quantity of sybils surpasses a given threshold, which, looking 100 90 80 70 60 50 40 30 20 10 0 0 5000 10000 15000 20000 Sybils number genuine not genuine unended Figure 2: Successful genuine, successful non-genuine,and unended lookups versus the number of sybils, whenever sybils respond with random IDs. ----- 6 Security and Communication Networks at the trends in the graph, can be set to 8,000. On the other hand, when this threshold is passed, the probability for the lookups to encompass a great majority of sybils increases and, as a consequence, the lookup operations may be stalled into never-ending cycles. Figure 2 shows aggregated data for both iterative and recursive lookup; however, we noticed that successful lookups belong more to the iterative procedures, rather than to the recursive ones. This is because the Chord requesting node can check whether a lookup is moving towards the queried key and, in the case of the iterative procedure, can decide not to follow the wrong suggestion of a sybil and to query again the previous peer. This is not possible in a recursive lookup procedure. We will return on the difference between recursive and iterative lookups later, as a difference will emerge when we apply our trust-based extension to Chord. In Figure 3 we compare the two malicious behaviors considering the average of hops per lookup, regardless of the fact that the lookup is genuine or not, and not considering if it succeeds or it is unended after the 𝑀𝑊 timeout has been reached. As it may be noticed, the curve concerning the _sybils not responding, experiences a first increase and then a_ decrease, whereas the curve of the second malicious behavior undergoes a constant growth of the hops per lookup. This could be due to the fact that in the second case an increasing number of sybils may lead to a never-ending lookup process, while in the first case, above a certain threshold of sybils in the network, the chance to reach good behaving nodes decreases and so lookups experience a longer waiting time without increasing the number of hops that tends to decrease on average. It must be highlighted that in the first case the mean quantity of hops does not surpass the typical Chord threshold (log2𝑁, with 𝑁 the overall nodes in the network: good nodes plus sybils), ranging from 13.28 to 14.87 depending on the variable amount of sybils in the considered scenario. In the second case this happens whenever the sybils equal or go over 10,000. In Figure 4 we show the trend of successful storage and retrieval operations when the amount of sybils in the network increases. We encompass both PUT and GET procedures together and the basic malicious behaviors related to these two operations: refusing to accept a legitimate PUT and responding with a null value when receiving a GET request. However, it is unuseful to consider whether the procedure is iterative or recursive or whether it is genuine or not, because we do not regard maliciousness in the lookups: the nature of the last peer reached in the lookup process is the only thing that counts in this case. In the figure, we report also the curve of successful lookups in presence of no response from the sybils for the sake of comparison, since all these attacks are quite similar, regarding the refusal of performing the requested operation. As can be inferred from the figure, the decrease of the successful storage or retrieval operations is more evident than that of lookups (already shown in Figure 1), especially when the sybils number more than 4,000. We can explain this trend considering that, in this case, we consider two malicious behaviors at the same time, i.e., both for PUTs and for GETs, while the lookups were affected by only one In this section, we investigate some possible countermeasures to limit the impact of sybils on lookups, as well as on 18 17 16 15 14 13 12 11 10 0 5000 10000 15000 20000 Sybils number no response random response Figure 3: Average number of hops versus the number of sybils, when comparing two different malicious behaviors. 100 90 80 70 60 50 40 30 20 10 0 0 5000 10000 15000 20000 Sybils number succ. lookups succ. storage and retrieval unended Figure 4: Comparison between successful lookups and successful storage and retrieval operations versus an increasing number of _sybils in standard Chord._ form of attack, i.e., no response. This trend is not so marked in those cases where there are few sybils in the network, confirming again that a small number of malicious nodes do not influence the performances very much. Moreover, we can notice that the number of unended storage and retrieval procedures is almost constant and very limited, and this is correct since this percentage is only dependent on the 𝑀𝑊 parameter rather than on malicious behaviors in the storage and retrieval procedures. ###### 4. S-Chord ----- Security and Communication Networks 7 storage and retrieval procedures, by integrating a simple but effective trust-based mechanism into the standard Chord environment and procedures, called S-Chord. The proposed approach is similar but yet different, to the one proposed for Kademlia in [10]. We propose improving the resilience of Chord to a Sybil attack, using trust information in the lookup as well as in storage and retrieval operations, but with a differentiated trust management for the two types of operations. The proposed solution considers a trust metric when sorting local finger tables. We suppose, contrarily from Koyanagi [32], that sorting the entries of such tables regarding only trust would not be effective in some cases, since the best solution is obviously the one of standard Chord. We introduce a novel metric for computing the distance of peers and call it “new distance” (𝑛𝑑𝑖). This new distance takes into account also trust, and it is computed, for each peer 𝑖 in the finger table, according to 𝑛𝑑𝑖 = 𝑏⋅2[𝑖−1] + (1 −𝑏) 𝑇𝑖 ⋅2[𝑖−1] (1) where 𝑛𝑑𝑖 is the new distance, 𝑏 is a balancing term, whose values may range from 0 to 1, while 𝑇𝑖 is the trust factor, with values from 0 to 1, of the 𝑖[𝑡ℎ] entry of the finger table. The formula in (1) guarantees that the actual successor is not skipped, as the base 2 exponentiations are always multiplied by factors less than 1. The farthest node according to the new distance is chosen as the candidate for the next step of the iterative or recursive lookup process. This is mandatory also in case a recursive lookup is considered and the peer currently in charge for the lookup is a sybil. _4.1. Trust Score. The trust score is calculated following a_ definition coming from the PET model [26], and not from the model in [32] or in [33] as regards Chord, or from the model in [9, 39] for Kademlia. In this way, we try to maintain the procedures as simple as possible and, at the same time, to insert the least overhead as possible. In case there are no previous interactions with other peers, we account for a global risk factor, which allows us to escape the grace phase used in [9] and the indirect trust employed in [32]. Conversely, we propose a proper mixture of direct trust and risk by means of two numerical weights. The trust score (𝑇) is defined as follows: 𝑇= 𝑊𝑅𝑒 ⋅𝑅𝑒 + 𝑊𝑅𝑟 ⋅(1 −𝑅𝑟) (2) where 𝑅𝑒 and 𝑅𝑟 are the direct trust and the risk, respectively, trust and to the global risk of the network (values rangingwhereas 𝑊𝑅𝑒 and 𝑊𝑅𝑟 are the weights assigned to the direct between 0 and 1, and one the complement to 1 of the other one), respectively. We simplify further the model with the introduction of a unique negative level of interaction (L = low grade). This implies that the risk 𝑅𝑟 is calculated with the subsequent equation: 𝑅𝑟 = [𝑁][𝐿] . (3) 𝑁𝑇 𝑁𝐿 accounts for all interactions, with a low grade of service, provided by a certain peer in the whole network. It may concern various behaviors, depending on the particular scenario: (i) considering only the trust in lookups, 𝑁𝐿 may encompass timeouts of lookups, no responses, and the like; (ii) considering only the trust in PUT and GET procedures, 𝑁𝐿 encompasses storage refusals, retrieval of null resources, and so on. (iii) considering trust in all possible operations, 𝑁𝐿 is computed considering all the previously mentioned malicious responses. 𝑁𝑇 represents all the requests generated by the nodes of the considered scenario towards the peer under evaluation, be they lookups or PUT or GET operations. The risk of a certain peer is a global measure that in a realistic implementation could be maintained through mechanisms similar to those of blockchain [40]. Conversely, the value of the direct trust depends on the particular trustor node, which stores various values of trust, one for each trustee peer, in a local table. If no communications with a particular peer have taken place yet (e.g., new joining node),𝑅𝑟); in the opposite situation it comprises the direct trust 𝑡 is defined only through 𝑊𝑅𝑟 ⋅(1 − (𝑅𝑒) contribution as well, which is computed and updated according to the following steps: (i) when only lookups are examined: +1/𝑀 to all peers involved in a lookup that leads to a correct resource, or −1.5/𝑀 to the peers leading to dead-points or wrong targets; (ii) when only PUT and GET procedures are of interest: +1/𝑀 to all nodes that accept to store a legitimate PUT or that return a correct resource, or −2/𝑀 to all nodes refusing a storage request or providing a null value as a response to a legitimate GET; (iii) in case both lookups and storage and retrieval operations are considered: the update of the single direct trusts is included in considering the before mentioned procedures. 𝑀 represents the total number of queries, launched by the trustor node, in which the trustees have been involved; they can be lookups, PUTs, or GETs. This asymmetry in rewarding or punishing the involved peers is devised to foster the correct behavior of nodes and, at the same time, to discourage malicious behaviors by punishing a bogus node more than it could recover with a single good interaction. The greater punishment in storage and retrieval operations (−2/𝑀 in place of −1.5/𝑀) is given by the fact that we consider jointly a malicious behavior in both PUTs and GETs and this leads to worse performances than in the lookup case as shown in Figure 4. Similarly to what was done in [9], trust values expire after a certain time (we set this time to 24 virtual hours in our simulations) in order to avoid temporal attacks. This is to prevent that a bogus peer, by colluding with other malicious nodes, may obtain from them a high trust score before starting behaving badly. The negative direct trust update is ----- 8 Security and Communication Networks a significant difference with [32] and it could be tuned in a more fine-grained way to involve only the last in charge peer, or a certain subgroup of peers, for a single lookup path. This is done with the aim of considering the chance that not every peer of a wrong path acted in a bad way. _4.2. Evaluation of S-Chord_ _4.2.1. Parameter Tuning. In order to evaluate the effectiveness_ of S-Chord, we consider the same simulation conditions as those in Section 3.2, and we analyze its performances in a network scenario with a growing number of sybils. The tuning of parameter 𝑏 is very important since it may determine outcomes that may be worse than those obtained with traditional Chord. Therefore, a great simulation campaign was carried out for optimizing parameter 𝑏, obtaining a final value of 0.68 when considering only lookups and of 0.60 when considering only storage and retrieval operations. This may be explained by the fact that more importance should be given to the standard procedure for an actual convergence of the lookups rather than for the successfulness of storage and retrieval operations, something depending only on the nature (malicious or not) of the last queried peer. for a node joining for the first time andConcerning the other parameters of the model, 0.3 for those nodes 𝑊𝑅𝑟 is 1 having already joined the network, and0 for new joining nodes (this accounts for the lack of direct 𝑊𝑅𝑒 is, by definition, trust) and 0.7 for already joined ones (this allows having a greater impact of direct trust over the environmental risk, whenever such data are available). In Figure 5 we compare the performances of different values of the 𝑏 parameter in case of no response from the sybils during a lookup procedure: it is evident that if the balance factor 𝑏 is not properly set, it can lead to performances worse than in the case of standard Chord. This happens when 𝑏 approaches zero and trust becomes preponderant, to the detriment of the optimum strategy of standard Chord. Standard Chord strategy is obviously the best one and would be based only on a distance given by 2[𝑖−1], where the base 2 exponentiation is not affected by the trust factor. On the other hand, whenever 𝑏 gets close to 1, the outcomes appear to be similar to standard Chord, and this is correct, as the formula in (1) reduces to 2[𝑖−1], the standard form of the distance in classic Chord. From the same figure, it may also be inferred that S-Chord, whenever 𝑏 is properly set, can limit or mitigate, at least partly, the negative effects of a Sybil attack. This is especially marked whenever malicious nodes are more than 6,000, since the percentage increase in the performances, compared to standard Chord, ranges from 28% (6,000 sybils) to 184% (10,000 sybils). In Figure 6 we compare standard Chord and S-Chord in case of sybils responding with a random node in a lookup procedure and variable values for 𝑏. In this case, we do not consider genuine and non-genuine lookups separately, but they equally contribute to successfully ended lookups as the objective is to tune parameter 𝑏. As one can see, whenever the balancing factor is correctly set (𝑏 equal to 0.68) the performances of S-Chord are far better than standard Chord, with an improvement ranging from +4.27% (4,000 𝑠𝑦𝑏𝑖𝑙𝑠) 100 90 80 70 60 50 40 30 20 10 0 0 5000 10000 15000 20000 Sybils number standard Chord S-Chord b=0.9 S-Chord b=0.68 S-Chord b=0.2 Figure 5: A comparison between standard Chord and S-Chord for different values of 𝑏 considering the percentage of successful lookups versus the number of sybils in the network; “no response” is considered as malicious behavior. 100 90 80 70 60 50 40 30 20 10 0 0 5000 10000 15000 20000 Sybils number standard Chord S-Chord b=0.68 S-Chord b=0.2 S-Chord b=0.9 Figure 6: Comparison of successful lookups versus sybils number for standard Chord and S-Chord for different values of 𝑏. The considered malicious behavior is random response. to +94.84% (20,000 𝑠𝑦𝑏𝑖𝑙𝑠). On the contrary, when 𝑏 is not correctly set to its optimum value, two main cases are possible and they are described in the following. The first one of such cases is when 𝑏 is close to zero: in this situation, represented in Figure 6 by the case with 𝑏 equal to 0.2, the performances are far worse than standard Chord, even worse than in the case of no response attack shown in Figure 5. This could be explained considering that, in this situation, the new distance is mainly driven by the trust score and the distance metric reduces, more or less, to 𝑇𝑖 ⋅2[𝑖−1]. This has as an ultimate effect on the reinforcement of the randomness in the responses given by the sybils themselves. ----- Security and Communication Networks 9 The second one is when 𝑏 is close to 1 and this is depicted in Figure 6 by the case of 𝑏 equal to 0.9. In this case, whenever the number of sybils is small, the performances seem to follow the ones of standard Chord, like in Figure 5. This is because the distance metric reduces to the classic one: 2[𝑖−1]. However, this is not true anymore when malicious peers are more than or equal to 10,000. As a matter of fact, in these conditions the curve undergoes a saturation behavior across the 20% value. This is a very interesting attitude and a tentative explanation is that there could be a threshold of the overall network nodes, be they sybils or not, after which the random response behavior should lead always to the same results. This threshold should depend on the balancing factor, as it is not present for lower values of 𝑏, the 𝑀𝑊 parameter, since it determines the timeout and the classification of a lookup as “unended,” and the relative ratio of malicious nodes versus good nodes. _4.2.2. Effectiveness in the Routing Process. In this subsec-_ tion, we carry out a comparison between S-Chord, standard Chord, Koyanagi’s solution [32], Kohnen’s solution [9] readapted to Chord, and GeTrust [33] both in terms of successfulness of the lookups and considering overhead complexity. In these simulations, we set the parameter 𝑏 of S-Chord to its optimum value. In Figure 7 we make a comparison in terms of successful lookups with an increasing number of sybils. The outcomes prove that a “balanced” strategy is better than a solution based only on trust, particularly whenever malicious nodes are not preponderant. It must be stressed that when the sybils number significantly more than or equal to 10,000, the number of good nodes, the chance to obtain a successful lookup is always less than 50%. From the same figure, it can be inferred that our solution reaches similar performances as GeTrust; however, S-Chord is less computationally intensive in terms of simulation time, in comparison both to Koyanagi’s solution and especially to GeTrust. This is shown in Figure 8, where we make a graph of the average time spent by each trust model after 30 simulation cycles. The best performances, in terms of temporal overhead, reached by our solution are due to the fact that Koyanagi’s method requires trust propagation, while in GeTrust there is an overhead of messages both to establish warranted relationships and in the lookup transactions themselves. The solution of Kohnen is the worst in terms of temporal overhead and this is probably due to the usage of certificates and public key cryptography. Because of this temporal overhead, and since its performances are only slightly better than Koyanagi’s ones (see Figure 7), we will not take it into account anymore in the rest of the paper. In Figure 9 we compare standard Chord, S-Chord, and GeTrust solutions, in case of sybils responding with a random ID and performing different analyses for genuine and non-genuine lookups as well as for iterative and recursive lookups. Considering genuine lookups are of no interest, as these decrease with the same trend for both our solution and GeTrust, therefore, they are not shown here. What is interesting, instead, is the analysis of the behavior of successful non-genuine lookups and of iterative and recursive procedures separately. As per the figure, our solution is better than both standard Chord and GeTrust, especially when we look at the recursive lookup procedure. While the iterative lookup procedure undergoes a similar improvement as GeTrust compared with standard Chord (more or less the same curve, and this is why they are not shown), the recursive implementation of S-Chord outperforms them both, particularly whenever the _sybils number more or less like the good nodes (8,000-_ 10,000). This can be explained thanks to a better spreading of the trust information, since, differently from the iterative procedure, more than a good node is involved in the recursive 100 90 80 70 60 50 40 30 20 10 0 0 5000 10000 15000 20000 Sybils number standard Chord S-Chord b=0.68 Koyanagi GeTrust Kohnen Figure 7: Comparison between standard Chord, the method employed by Koyanagi [32], the method employed by Kohnen [9], GeTrust [33], and S-Chord considering successful lookups versus _sybils number. The considered malicious behavior is no response._ 100 80 60 40 20 0 S-Chord Koyanagi GeTrust Kohnen Strategy Figure 8: Comparing different strategies using trust in Chord, in terms of computational time of 30 cycles in a simulation. ----- 10 Security and Communication Networks 90 80 70 60 50 40 30 20 10 0 0 5000 10000 15000 20000 Sybils number Chord iterative S-Chord/GeTrust recursive S-Chord Figure 9: Comparison of standard Chord, GeTrust [33], and SChord, considering successful non-genuine lookups versus sybils number. The considered malicious behavior is random response. grows constantly from a certain point on, even in those cases when the sybils are more than 8,000. Such a behavior is quite different from what happens for standard Chord. However, the threshold of standard Chord performances (log2𝑁) is reached and surpassed only from 15,000 sybils on, confirming again the goodness of the proposed strategy in presence of a preponderant number of malicious nodes. Obviously, if our solution is compared with standard Chord algorithm, it leads to worse performances (more number of hops) in some circumstances. This can be mainly inferred from the first parts of the curves, when the number of _sybils is limited, and thus the performances can be assimilated_ with an adequate degree of precision to those of standard Chord. As one can notice, in this situation, when bogus peers are fewer than or equal to 4,000, the lines of S-Chord are always above the corresponding ones of standard Chord, meaning that it requires more hops. This is correct, since SChord is based on a distance metric that is not optimum but encompasses also the fading value of the trust score. _4.2.3. Effectiveness in the Storage and Retrieval Processes._ In this subsection, we perform an evaluation of S-Chord considering storage and retrieval procedures. We do not report, for the sake of brevity, the same comparisons, reported previously for lookups, on the fine-tuning of parameter 𝑏, but we take for granted it has been optimized for these procedures to the above-stated value of 0.60. Moreover, this value does not depend on the type of lookup (iterative or recursive), since it obviously depends only on the choice of the last peer in the query. Considering successful PUT and GET operations in presence of basic malicious behaviors leads to graphs similar to the ones provided before for the lookup process, thus in this part we concentrate only on the study of the sophisticated behaviors described in Section 3.3, which consider a good behavior immediately followed by a bad behavior or a selective bad behavior. They are summarized in the following: 18 17 16 15 14 13 12 11 10 0 5000 10000 15000 20000 Sybils number no response Chord random response Chord no response S-Chord random response S-Chord Figure 10: Average number of hops per lookup with standard Chord and S-Chord algorithm when considering two different malicious behaviors. lookup and thus they can be aware of the right or wrong responses received by other peers contacted in the procedure. Finally, in Figure 10 we analyze S-Chord concerning the hops mean value per lookup. Under these conditions, when the sybils perform a random response behavior our solution seems to lead to better performances (fewer hops), provided that the number of malicious nodes is greater than 4,000. In this case, the increase in the performances provided by SChord ranges from +2.35% (6,000 𝑠𝑦𝑏𝑖𝑙𝑠) to +14.62% (8,000 𝑠𝑦𝑏𝑖𝑙𝑠). On the other hand, a particular situation takes place in the case of no response from the sybils: by using our solution the curve does not experience a maximum anymore; rather it (i) accepting a resource and then refusing the relative retrieval queries; (ii) blocking storage and retrieval operations for certain specific resources chosen randomly. The first of these malicious behaviors does not influence the effectiveness of storing operations; therefore we focus only on retrieval procedures. In such a situation, the outcomes of the previously proposed solution can decline, as Figure 11 demonstrates; however, this is linked to the quantity of GETs compared to the PUTs, to the number of sybils and to the lookup type (iterative or recursive) used to reach the queried node. Since also in this case we experienced the fact that the GET operations preceded by a recursive lookup reach the same performances like the ones of the basic malicious behaviors, we decided to employ the relative curve as a benchmark for the following considerations. Defining 𝑟 as the number of PUTs divided by the number of GETs, within a certain period of time (30,000 V𝑠 in the considered scenario), one observes, according to the following evaluations, that its fluctuation can worsen the ----- Security and Communication Networks 11 100 90 80 70 60 50 40 30 20 10 100 90 80 70 60 50 40 30 20 10 0 0 5000 10000 15000 20000 Sybils number 0 0 5000 10000 15000 20000 Sybils number standard Chord recursive S-Chord iterative S-Chord s=0.3 iterative S-Chord s=0.8 standard Chord recursive S-Chord iterative S-Chord r=2 iterative S-Chord r=0.9 Figure 11: Successful GETs versus sybils number with standard Chord and S-Chord considering two different ways of performing the lookup and variable ratio (𝑟) between PUTs and GETs. The maliciousbehavior consistsintoacceptingallPUTsandintorefusing to answer to a retrieval request. outcomes of our proposal, even if they are always better than standard Chord’s. This happens especially if 𝑟 is higher than 1, and malicious nodes pass 6,000 and with an iterative lookup. Conversely, when 𝑟 is less than 1, the performances of the proposed solution degrade later in case of GETs preceded by an iterative lookup: when malicious peers become more than 8,000. On the contrary, as already stated, when the GET is preceded by a recursive lookup, the performances are not influenced by 𝑟 and they mirror the ones experienced when the basic malicious behaviors are enacted. Conclusively, one can notice that when the sybils are not overwhelming in number, the role of 𝑟 is no more crucial and the outcomes are more or less the same as they are in those situations without these more advanced attacks. The second sophisticated malicious behavior, concerning, like the first one, storages and retrievals, could influence both PUT and GET procedures. Such an attack is to be evaluated concerning the spreading of a content or resource. In case the resource is very well-liked, S-Chord (be it iterative or recursive) can be employed with no changes, while if the searched contents are not so popular, the outcomes of SChord fluctuate according to a spreading factor 𝑠, which measures the popularity of the searched resource and can be determined as the number of storages for a certain content divided by all PUT procedures. This may be easily inferred from Figure 12, where successful storages are analyzed for standard Chord, iterative SChord with a varying 𝑠 parameter and recursive S-Chord. Also, in this case, the curve of recursive S-Chord is used as a benchmark since it is very similar to the one obtained under basic malicious behaviors for PUTs and GETs using S-Chord. As it may be inferred, if 𝑠 is higher than 0.5 (widespread content) the trend of iterative S-Chord is very similar to the one of recursive S-Chord, whereas when 𝑠 is lower Figure 12: Successful PUTs versus sybils number with standard Chord and S-Chord considering two different ways of performing the lookup and variable spreading parameter (𝑠). The malicious behavior consists into blocking storage and retrieval operations for certain specific resources chosen randomly. than 0.5 the trend becomes irregular, resulting sometimes below and sometimes above standard Chord, according to the number of malicious nodes. In this attack scenario, the trust management could be readapted to execute storages or retrievals in a correct way. The adaptation may influence both the evaluation of the risk, introducing another parameter in addition to 𝑁𝐿, and the update of the trust score. However, such a focused attack could influence the final outcome of the whole attack strategy, since not all storage procedures are compromised and executing various attacks for every existing content could be computationally very heavy. As we have shown, recursive S-Chord usually obtains better results and features fewer problems compared to the iterative one. This is due to the fact that, in the considered implementation of recursive S-Chord, the last searched peer transmits back, to every previously contacted peer, some data about the outcome of storages or retrievals, while, in the considered implementation of iterative S-Chord, these data are available only to the peer that originated a storage or retrieval query. As a consequence, such an implementation could foster the diffusion of trust data across nodes. ###### 5. Spartacus Attack in S-Chord In this section, we study S-Chord in further detail, trying to understand its effectiveness and resilience under more complex attack scenarios. For this purpose, we consider only the recursive way of performing lookups, as it proved to be the best solution (see Section 4.2). Furthermore, we assess the effects of a network infected by spartaci against SChord and we try to provide some slight modifications in the proposed algorithms, to soothe the negative consequences of a Spartacus attacker. We focus solely on routing attacks coming from spartaci and particularly the already studied ----- 12 Security and Communication Networks cases of (i) no response or of (ii) random response, since the results for storage and retrieval operations are very similar. Moreover, the balancing factor corresponds with the optimum value we found, i.e., 0.68 for S-Chord routing, unless otherwise stated. The two main differences in the following analysis, compared to the previous ones, regard the presence of the temporal dimension on the abscissas of the graphs and the constancy of the number of peers in the system; i.e., we do not consider a variable amount of bogus peers. The reasons for these changes are the following: (i) first of all, because the effectiveness of a Spartacus attack varies in time, reasonably maximum in the first instants and then decreasing and, (ii) secondly, because the Spartacus attack provides the replacement of already valid IDs rather than the addition of new nodes. In order to properly assess the temporal dimension, in the simulation sets of this section the churn period of malicious nodes is no more random but it is established to 150 V𝑠, and the observation interval is set to 50 V𝑠, in order to achieve more granular observations. Obviously, in a real network malicious nodes could have different times in which they enter the network; however, the simplified assumption of a constant and common churn period is reasonable for an attacker characterized by bounded computational resources. Despite the churning, the average amount of bogus peers is constant during a simulation cycle. _5.1. The Effects of a Spartacus Attack onto S-Chord. In this_ subsection, we analyze the effects of a Spartacus attack directly onto the proposed S-Chord. We focus our attention on the performance degradation concerning both the number of successful lookups and the average hops per lookup. As we have already explained in advance, the analyses are performed in time, limiting to the first 2,000 V𝑠, but the results are still an average over various simulation seeds to get a confidence of 90%. The limitation to the first 2,000 V𝑠 is due to the inherent nature of a Spartacus attack: this is more effective in its beginning and its effects fade afterward. In particular, we report only this time interval as in the following the curves do not experience other significant trends. Concerning the first metric, we focus only on the case of no response, since this was the case of better improvement of S-Chord compared to standard Chord, for both 8,000 and 10,000 malicious nodes (see Figure 7). As it may be inferred from Figure 13, when we consider a Spartacus attack the performances decrease over time, something happening, by a lesser degree, also on the standard Chord algorithm. This may be caused by the fact that Chord, unlike other DHT, e.g., Kademlia, has no inherent proximity routing procedures that reinforce the whole algorithm during the joining phase (see Section 3.1.2). The performances of S-Chord dramatically drop, especially in the case of 10,000 spartaci and this is quite obvious because the good nodes have to face the same number of malicious counterparts. Particularly, the performances in case of 8,000 spartaci decrease of about 26%, while in the case of 10,000 spartaci they fall of about 36%. The worsening of the performances of S-Chord under a Spartacus attack may be motivated by the combined effect, caused by malicious nodes, of both inheriting high trust scores and replacing of good behaving nodes. That is, not only are malicious nodes trusted by good nodes in the beginning, but their presence also decreases the number of good behaving peers and, therefore, the overall performances drop accordingly. What is to be highlighted is also the presence of a periodic trend in the curves, with period circa equivalent to the churning period of malicious nodes, i.e., 150 V𝑠. In Figure 14 we assess the effects of spartaci on Chord and S-Chord whenever the mean number of hops per lookup is concerned. In this circumstance, we consider only the case of 8,000 malicious nodes, since this is one of the two points of the graph in Figure 10 where our trust-based solution obtains better results than the classic algorithm. As the performances are very similar for both the no response and 70 60 50 40 30 20 10 0 200 400 600 800 1000 1200 1400 1600 1800 2000 Virtual time (vs) Chord 8,000 spartaci Chord 10,000 spartaci S-Chord 8,000 spartaci S-Chord 10,000 spartaci Figure 13: Successful lookups versus sybils number in standard Chord and S-Chord with a different quantity of spartaci. 18 17 16 15 14 13 12 11 10 0 500 1000 1500 2000 Virtual time (vs) Chord S-Chord standard Chord threshold Figure 14: Average hops per lookup versus time in standard Chord and S-Chord with 8,000 spartaci in the network. ----- Security and Communication Networks 13 0 500 1000 1500 2000 Virtual time (vs) the random response cases, we consider an average value and, as a consequence, only one combined malicious behavior. The graphs depicted in Figure 14 show an almost constant behavior for classic Chord, whereas the performances of SChord quickly degrade, overtaking the standard algorithm and even the log2𝑁 threshold (log2(18, 000) = 14.13) by 1,500 V𝑠. Furthermore, in this situation, the periodic tendency, following the churn of malicious nodes, appears again, especially in the curve of S-Chord. _5.2. Improving S-Chord. In this subsection, we propose some_ improvements to S-Chord to combat a Spartacus attack. We focus the improvement efforts mainly as regards the percentage of successful lookups, since, also in presence of _sybils, the advantage of our balanced trust-based method,_ considering average hops, is visible only in some limited circumstances, namely, when malicious peers are limited to 6,000 or 8,000 (see Figure 10). As a consequence, we vary S-Chord mainly using the following two countermeasures: 70 60 50 40 30 20 10 0 (i) the first one regards the opportune tuning of the weight 𝑊𝑅𝑟 of the risk environment. (ii) the second one concerns the utilization of an opportune time decay function to be applied to the trust score. S-Chord 8,000 spartaci S-Chord 10,000 spartaci improved S-Chord 8,000 spartaci improved S-Chord 10,000 spartaci Figure 15: Successful lookups versus sybils number in S-Chord and enhanced S-Chord featuring different spartaci in the network. What can be inferred, from Figure 15, is a sort of improvement in the performances of improved S-Chord: the curves, both for 8,000 and for 10,000 spartaci, maintain an overall constant trend. This does not take place with standard S-Chord. As a consequence, we could manage to get a stable behavior in time and thus soothe a little the drop in the performances. This is more evident in the case of 10,000 _spartaci than in that of 8,000 malicious nodes; as a matter of_ fact, in the first case, the decrease of improved S-Chord is only about 1.89%, greatly better if compared with 33% of S-Chord, while in the second case the drop is more than three times, i.e., 4.35%, however, better than 15% of S-Chord. Moreover, the periodic trend, following the churning period of malicious nodes, tends to disappear both in the case of 8,000 and in the case of 10,000 spartaci. _5.3. Comparisons with Other Techniques. Finally, in this_ section, we assess the effectiveness of our proposal in comparison with other solutions, already known in the literature, under a Spartacus attack. We address the comparison between improved S-Chord, Koyanagi’s solution [32], and GeTrust both in terms of successful lookups and in terms of average hops. The simulation conditions, as well as the utilized metrics, are the same as the ones of the previous subsections. In Figure 16 we analyze the performances of S-Chord, GeTrust, and Koyanagi’s solution with the ones of the improved version of S-Chord, in terms of successful lookups with a growing simulation time and a combination of malicious behaviors regarding the routing process (no response and random responses). The number of malicious nodes is 8,000. We can see that GeTrust performs more or less like SChord, especially for a growing simulation time, according to the analyses showed before in the paper, but its curve is always under the one of improved S-Chord, which tends to The first improvement we devised involves augmenting the weight of the network scenario. This seems a straightforward consequence since the Spartacus attack appears more dangerous than the Sybil-based one. However, we do not simply increase the environment weight to a higher value for those peers having already joined the network, since, following our first tests, this static countermeasure would have affected only the first moments of the simulation. Therefore, we make 𝑊𝑅𝑟 vary according to a period equal to the one of the churn of the spartaci, increasing till 0.5 and decreasing to 0.3. This obviously implies the knowledge of such a period; however, it could be easily detected through a cooperation between those peers that are going to admit possible malicious nodes. The second aforementioned countermeasure encompasses the introduction of a time decay function. More in detail, we multiply the trust score by a negative exponential function of time, with mean value corresponding to the churning period of the spartaci, i.e., 150 V𝑠. This is better explained in (4), where 𝑇[̃] stands for new trust, 𝑇 for old trust, and 𝐷𝑓 for decay function. ̃𝑇= 𝑇⋅𝐷𝑓 (4) The decay function is, instead, expressed by the following formula: 1 𝐷𝑓 (𝑡) = 150V𝑠 [⋅𝑒][−(1/150][V][𝑠)𝑡] (5) Therefore, also 𝑇[̃] in (4) becomes a function of time. 𝑇[̃] is going to replace 𝑇 in the formulas of (2) and (1). ----- 14 Security and Communication Networks Table 1: Comparison between different trust-based solutions in terms of decrease of successful lookups with a variable amount of spartaci in the network. **S-Chord** **improved S-Chord** **GeTrust** **Koyanagi’s** **8,000 spartaci** -14.93% -4.49% -15.77% -26.96% **10,000 spartaci** -33.07% -2.00% -32.81% -58.14% 70 65 60 55 30 25 50 45 20 15 40 0 500 1000 1500 2000 Virtual time (vs) 10 0 500 1000 1500 2000 Virtual time (vs) S-Chord improved S-Chord GeTrust Koyanagi S-Chord standard Chord threshold improved S-Chord GeTrust Koyanagi Figure 16: Successful lookups versus time in S-Chord, enhanced SChord, Getrust, and Koyanagi’s solution with 8,000 spartaci in the network. 50 45 40 35 30 25 20 15 10 0 500 1000 1500 2000 Virtual time (vs) S-Chord improved S-Chord GeTrust Koyanagi Figure 17: Successful lookups versus time in S-Chord, enhanced SChord, Getrust, and Koyanagi’s solution with 10,000 spartaci in the network. be almost constant. The decreasing performances of Getrust may be due to the chance that spartaci could assume the role of guarantors: this does not lead to a complete worsening of the overall performances, since the direct trust towards the guarantors decrease, but they, however, tend to follow the decreasing trend of S-Chord, where no guarantors are Figure 18: Average number of hops versus time in S-Chord, enhanced S-Chord, Getrust, and Koyanagi’s solution with 8,000 _spartaci in the network._ present. The worst performances are those of Koyanagi’s solution. This may be due to its reliance on trust aggregation and propagation that may boost the collusion activity of _spartaci._ In Figure 17 we show the performances of S-Chord, GeTrust, and Koyanagi’s solution with the ones of the improved version of S-Chord, when bogus peers are 10,000 and considering successful lookups. The same conditions used for simulations of Figure 16 apply and we can draw more or less the same considerations. The curves of S-Chord and GeTrust seem to overlap much more and the decrease in the performances is more marked as reported in Table 1. From that table, it can be inferred that the improved S-Chord solution is better than S-Chord and than GeTrust and much better than Koyanagi’s strategy. In Figure 18 we show a comparison between the aforementioned solutions in terms of average hops per lookup and considering 8,000 spartaci, the most critical situation as seen previously. As we can see, Koyanagi’s solution is the worst since it is based solely on trust and its propagation and aggregation and thus much more vulnerable to a Spartacus attack, even if its hop number is on average already the double of standard Chord by default [32]. In order to compare GeTrust and the other solutions effectively, we do not consider the messages exchanged with guarantor nodes and archive nodes, but only those used to actually perform the lookup. GeTrust performs similarly to ----- Security and Communication Networks 15 S-Chord, with a trivial growth for the hops according to the simulation time, while the improved version of S-Chord succeeds in maintaining an almost constant trend, similarly to standard Chord, and to remain under the 𝑂(log2𝑁) threshold. ###### 6. Conclusions In this article a deep analysis of some trust-based countermeasures for Chord, under a Sybil or a Spartacus attack, has been presented. We have numerically studied the consequences of a Sybil attack in routing as well as in storage and retrieval operations, and we introduced a solution (namely, S-Chord), based on direct trust, to make Chord procedures more resilient. Moreover, we evaluated the still not deeply investigated Spartacus attack, both in Chord and in S-Chord, proposing some effective improvements to S-Chord itself. The results of our simulations are encouraging compared to standard Chord and to existing methods using exclusively trust, or other complex trust management systems. In conclusion, our approach may be regarded as a good candidate for a security solution applied to P2P networks. Using simple trust metrics is, as a matter of fact, far less power consuming than other cryptography-based solutions, opening its applicability to new emerging scenarios featuring low power devices, like the Internet of Things one. This is a matter for possible future research work, along with considering other DHTs, focusing on the application of trust to lookups and PUT and GET operations jointly, analyzing the spreading of trust information across the peers, or varying the optimum value of the balancing factor according to various scenarios. ###### Data Availability The data used to support the findings of this study are available from the corresponding author upon request. ###### Conflicts of Interest Dr. Riccardo Pecori and Dr. Luca Veltri declare that no conflicts of interest, regarding the publication of this paper, are present at the moment of submission. ###### Acknowledgments Dr. Riccardo Pecori would like to thank Mr. Antonio Enrico Buonocore for carefully proof-reading the paper and polishing English and Stefano Marmani for making him discover the Spartacus attack. ###### References [1] R. Pecori and L. Veltri, “3AKEP: Triple-authenticated key exchange protocol for peer-to-peer VoIP applications,” Com_puter Communications, vol. 85, pp. 28–40, 2016._ [2] C. Liao, S. Cheng, and M. Domb, “On Designing Energy Efficient Wi-Fi P2P Connections for Internet of Things,” in _Proceedings of the IEEE 85th Vehicular Technology Conference_ _(VTC Spring ’17), pp. 1–5, Sydney, NSW, June 2017._ [3] P. Maymounkov and D. Mazi`eres, “Kademlia: a peer-to-peer information system based on the XOR metric,” in Peer-to-Peer _Systems, vol. 2429 of Lecture Notes in Computer Science, pp. 53–_ 65, Springer-Verlag, 2002. [4] I. Stoica, R. Morris, D. Karger, M. F. Kaashoek, and H. Balakrishnan, “Chord: a scalable peer-to-peer lookup service for internet applications,” in Proceedings of the Conference _on Applications, Technologies, Architectures, and Protocols for_ _Computer Communications (SIGCOMM ’01), pp. 149–160, San_ Diego, Calif, USA, August 2001. [5] S. Ratnasamy, P. Francis, M. Handley, R. Karp, and S. Schenker, “A scalable content-addressable network,” in Proceedings of _the Conference on Applications, Technologies, Architectures, and_ _Protocols for Computer Communications, ser. SIGCOMM ’01, pp._ 161–172, San Diego, California, USA, 2001. [6] A. Rowstron and P. Druschel, “Pastry: Scalable, Decentralized Object Location, and Routing for Large-Scale Peer-to-Peer Systems,” in Middleware 2001, vol. 2218 of Lecture Notes in _Computer Science, pp. 329–350, Springer Berlin Heidelberg,_ Berlin, Heidelberg, 2001. [7] J. R. Douceur, “The sybil attack,” in Peer-to-Peer Systems, P. Druschel, F. Kaashoek, and A. Rowstron, Eds., vol. 2429 of ser. _Lecture Notes in Computer Science, pp. 251–260, Springer Berlin_ Heidelberg, 2002. [8] H. Rowaihy, W. Enck, P. McDaniel, and T. La Porta, “Limiting sybil attacks in structured P2P networks,” in Proceedings of the _26th IEEE International Conference on Computer Communica-_ _tions (INFOCOM ’07), pp. 2596–2600, May 2007._ [9] M. Kohnen, “Analysis and optimization of routing trust values in a Kademlia-based distributed hash table in a malicious environment,” in Proceedings of the 2nd Baltic Congress on Future _Internet Communications, BCFIC ’12, pp. 252–259, Lithuania,_ April 2012. [10] R. Pecori, “S-Kademlia: A trust and reputation method to mitigate a Sybil attack in Kademlia,” Computer Networks, vol. 94, pp. 205–218, 2016. [11] R. Pecori and L. Veltri, “Trust-based routing for Kademlia in a sybil scenario,” in Proceedings of the 22nd International _Conference on Software, Telecommunications and Computer_ _Networks, SoftCOM ’14, pp. 279–283, Croatia, September 2014._ [12] M. De Donno, N. Dragoni, A. Giaretta, and A. Spognardi, “DDoS-Capable IoT Malwares: Comparative Analysis and Mirai Investigation,” Security and Communication Networks, vol. 2018, pp. 1–30, 2018. [13] S. Delgado-Segura, C. P´erez-Sol`a, J. Herrera-Joancomart´ı, G. Navarro-Arribas, and J. Borrell, “Cryptocurrency Networks: A New P2P Paradigm,” Mobile Information Systems, vol. 2018, Article ID 2159082, 16 pages, 2018. [14] C. Lesniewski-Laas and M. F. Kaashoek, “Whanau: A sybilproof distributed hash table,” in Proceedings of the 7th USENIX _Conference on Networked Systems Design and Implementation,_ _ser. NSDI’10, 2010._ [15] P. Mittal, M. Caesar, and N. Borisov, “X-vine: Secure and pseudonymous routing in dhts using social networks,” in _Proceedings of the in 19th Annual Network and Distributed_ _System Security Symposium, (NDSS ’12), San Diego, California,_ USA, 2012. [16] M.N.Al-AmeenandM.Wright,“Persea:A sybil-resistant social dht,” in Proceedings of the Third ACM Conference on Data and _Application Security and Privacy, ACM, p. 169, New York, NY,_ USA, Feburary 2013. ----- 16 Security and Communication Networks [17] M. N. Al-Ameen and M. Wright, “IPersea: Towards improving the Sybil-resilience of social DHT,” Journal of Network and _Computer Applications, vol. 71, pp. 1–10, 2016._ [18] Z. Yang, J. Xue, X. Yang, X. Wang, and Y. Dai, “VoteTrust: Leveraging friend invitation graph to defend against social network sybils,” IEEE Transactions on Dependable and Secure _Computing, vol. 13, no. 4, pp. 488–501, 2016._ [19] Z. Tan, X. Wang, and X. Wang, “A Novel Iterative and Dynamic Trust Computing Model for Large Scaled P2P Networks,” _Mobile Information Systems, vol. 2016, Article ID 3610157, 12_ pages, 2016. [20] L. Shi, J. Zhou, Q. Huang, and W. Yan, “A modification on the Chord finger table for improving search efficiency,” in _Proceedings of the 13th IEEE/ACIS International Conference on_ _Computer and Information Science, ICIS ’14, pp. 395–398, China,_ June 2014. [21] P. Zave, “Reasoning About Identifier Spaces: How to Make Chord Correct,” IEEE Transactions on Software Engineering, vol. 43, no. 12, pp. 1144–1156, 2017. [22] C. T. Min and L. T. Ming, “Investigate SPRON Convergence Time Using Aggressive Chord and Aggressive AP-Chord,” in _Proceedings of the 12th International Conference on Information_ _Technology: New Generations, ITNG ’15, pp. 61–66, USA, April_ 2015. [23] I. Woungang, F.-H. Tseng, Y.-H. Lin, L.-D. Chou, H.-C. Chao, and M. S. Obaidat, “MR-Chord: Improved Chord Lookup Performance in Structured Mobile P2P Networks,” IEEE Systems _Journal, vol. 9, no. 3, pp. 743–751, 2015._ [24] T. Amft and K. Graffi, “Moving peers in distributed, locationbased peer-to-peer overlays,” in Proceedings of the International _Conference on Computing, Networking and Communications,_ _ICNC ’17, pp. 906–911, USA, January 2017._ [25] W. Zhang, B. Sun, and Y. Sun, “Trustchord: chord protocol based on the trust management mechanism,” in Proceedings _of the International Conference on Advanced Intelligence and_ _Awareness Internet (AIAI ’10), pp. 64–67, Beijing, China._ [26] Z. Liang and W. Shi, “PET: A PErsonalized Trust Model with Reputation and Risk Evaluation for P2P Resource Sharing,” in _Proceedings of the 38th Annual Hawaii International Conference_ _on System Sciences, HICSS ’05, pp. 201b–201b, Big Island, HI,_ USA. [27] J. Wang and J. Liu, “The comparison of distributed P2P trust models based on quantitative parameters in the file downloading scenarios,” Journal of Electrical and Computer Engineering, vol. 2016, Article ID 4361719, pp. 1–10, 2016. [28] L. Mekouar, Y. Iraqi, and R. Boutaba, “Reputation-based trust management in peer-to-peer systems: Taxonomy and anatomy,” in Handbook of Peer-to-Peer Networking, X. Shen, H. Yu, J. Buford, and M. Akon, Eds., pp. 689–732, Springer US, 2010. [29] X. L. Xie, “Creditability assessment of dealers in P2P ecommerce,” in Proceedings of the 2016 IEEE Advanced Informa_tion Management, Communicates, Electronic and Automation_ _Control Conference, IMCEC ’16, pp. 1326–1333, China, October_ 2016. [30] X. Ding and K. Koyanagi, “Study on trust-based maintenance of overlays in structured P2P systems,” in Proceedings of the _International Conference on Computational Problem-Solving,_ _ICCP ’11, pp. 598–603, China, October 2011._ [31] R. R. Rout and D. Talreja, “Trust-based decentralized service discovery in structured Peer-to-Peer networks,” in Proceedings _of the 11th IEEE India Conference, INDICON ’14, India, Decem-_ ber 2014. [32] Y. Han, K. Koyanagi, T. Tsuchiya, T. Miyosawa, and H. Hirose, “A trust-based routing strategy in structured P2P overlay networks,” in Proceedings of the 27th International Conference _on Information Networking, ICOIN ’13, pp. 77–82, Thailand,_ January 2013. [33] X. Meng and D. Liu, “GeTrust: A Guarantee-Based Trust Model in Chord-Based P2P Networks,” IEEE Transactions on _Dependable and Secure Computing, vol. 15, no. 1, pp. 54–68, 2018._ [34] R. Pecori, “A comparison analysis of trust-adaptive approaches to deliver signed public keys in P2P systems,” in Proceedings of _the 7th International Conference on New Technologies, Mobility_ _and Security, NTMS ’15, pp. 1–5, France, July 2015._ [35] F. Tomonori, N. Yoshitaka, Y. Shiraishi, and T. Osamu, “An effective lookup strategy for recursive and iterative lookup on hierarchical dht,” International Journal of Informatics Society _(IJIS), vol. 4, no. 3, pp. 143–152, 2012._ [36] M. Amoretti, M. Picone, F. Zanichelli, and G. Ferrari, “Simulating mobile and distributed systems with DEUS and ns3,” in Proceedings of the 11th International Conference on High _Performance Computing and Simulation, HPCS ’13, pp. 107–114,_ Finland, July 2013. [37] J. Zhang, R. Zhang, J. Sun, Y. Zhang, and C. Zhang, “TrueTop: A Sybil-Resilient System for User Influence Measurement on Twitter,” IEEE/ACM Transactions on Networking, vol. 24, no. 5, pp. 2834–2846, 2016. [38] M. S. Khan and N. M. Khan, “Low Complexity Signed Response Based Sybil Attack Detection Mechanism in Wireless Sensor Networks,” Journal of Sensors, vol. 2016, pp. 1–9, 2016. [39] M. Kohnen, “Applying trust and reputation mechanisms to a Kademlia-based Distributed Hash Table,” in Proceedings of the _IEEE International Conference on Communications, ICC ’12, pp._ 1036–1041, Canada, June 2012. [40] A. Anjum, M. Sporny, and A. Sill, “Blockchain Standards for Compliance and Trust,” IEEE Cloud Computing, vol. 4, no. 4, pp. 84–90, 2017. ----- |Col1|VLSI Design Hwiwndwa.whiindawi.com Volume 2018| |---|---| **Modelling &** International Journal of **Simulation** Navigation and **in Engineering** Observation Hindawiwww.hindawi.com Volume 2018 Hindawiwww.hindawi.com Volume 2018 Hindawiwww.hindawi.com Volume 2018 Hindawiwww.hindawi.com Advances in ###### Shock and Vibration Volume 2018 Hindawiwww.hindawi.com Volume 2018 Hindawiwww.hindawi.com International Journal of Antennas and Propagation Hindawiwww.hindawi.com Volume 2018 Journal of ###### Sensors Hindawiwww.hindawi.com Volume 2018 _Advances in_ ### Multimedia _Hindawiwww.hindawi.com_ _Volume 2018_ Journal of Electrical and Computer Engineering Hindawiwww.hindawi.com Volume 2018 # AerospaceInternational Journal of Engineering Hindawiwww.hindawi.com Volume 2018 ###### The Scientific World JournalHindawi Publishing Corporation Hindawihttp://www.hindawi.comwww.hindawi.com Volume 2018Volume 2013 _Advances in_ _OptoElectronics_ _Hindawiwww.hindawi.com_ _Volume 2018_ Volume 2018 -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1155/2018/4963932?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1155/2018/4963932, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "http://downloads.hindawi.com/journals/scn/2018/4963932.pdf" }
2,018
[ "JournalArticle" ]
true
2018-11-12T00:00:00
[ { "paperId": "660bed056b20b9bc54093d78b7b61b0947cf4739", "title": "Cryptocurrency Networks: A New P2P Paradigm" }, { "paperId": "ece4cc22924a45619416d0ef3e7fb20d8ea4ece7", "title": "Vulnerability Analysis of Network Scanning on SCADA Systems" }, { "paperId": "658bd6ad713bd86e5a2add7dcab4847c0eb0a353", "title": "DDoS-Capable IoT Malwares: Comparative Analysis and Mirai Investigation" }, { "paperId": "17dafb105068b17b3eb4970fd80dd7e6b05931cf", "title": "Blockchain Standards for Compliance and Trust" }, { "paperId": "cc0c287a05adf5cfed3c65982e68c69800d23319", "title": "On Designing Energy Efficient Wi-Fi P2P Connections for Internet of Things" }, { "paperId": "e5c541ac892b182a9597a06f0b6e1e0e6b8a637e", "title": "Reasoning About Identifier Spaces: How to Make Chord Correct" }, { "paperId": "3555b40711a2b0d5790430f8f848977a0cea4f1d", "title": "Creditability assessment of dealers in P2P e-commerce" }, { "paperId": "f6796627ee20187d71dc2d97f8ba71cedcb43b6c", "title": "iPersea: Towards improving the Sybil-resilience of social DHT" }, { "paperId": "395c08b916fab4d39f376c1cbcedc1ed087e2d34", "title": "3AKEP: Triple-authenticated key exchange protocol for peer-to-peer VoIP applications" }, { "paperId": "60b2708c90f9ea0a86d6970336fad1c7e220d405", "title": "The Comparison of Distributed P2P Trust Models Based on Quantitative Parameters in the File Downloading Scenarios" }, { "paperId": "d26c0ffb387829262196980b4ed6550bac001f55", "title": "A Novel Iterative and Dynamic Trust Computing Model for Large Scaled P2P Networks" }, { "paperId": "5bfce77ea7799d826e6faca26426f601def52b7f", "title": "S-Kademlia: A trust and reputation method to mitigate a Sybil attack in Kademlia" }, { "paperId": "b2e66961cab5f00e2873b291f9c31382d29e64ea", "title": "Autonomous Gait Event Detection with Portable Single-Camera Gait Kinematics Analysis System" }, { "paperId": "e15e8f17300255c8ce95258f4fb2df58365ce82f", "title": "MR-Chord: Improved Chord Lookup Performance in Structured Mobile P2P Networks" }, { "paperId": "336629800964489fa2cf1d78a0180696a557e93f", "title": "A comparison analysis of trust-adaptive approaches to deliver signed public keys in P2P systems" }, { "paperId": "5a715e2fe02bddd83c7be8e91ca40a6546a5203f", "title": "TrueTop: A Sybil-Resilient System for User Influence Measurement on Twitter" }, { "paperId": "cfff11f0c072ebac5d506254b52b8228d475d1b5", "title": "Investigate SPRON Convergence Time Using Aggressive Chord and Aggressive AP-Chord" }, { "paperId": "690c74bf1448fe9faf1e3246be69b3581068acf0", "title": "Trust-based decentralized service discovery in structured Peer-to-Peer networks" }, { "paperId": "0f98863c8ba3069863aaea5da7839c4caaae1614", "title": "Trust-based routing for Kademlia in a sybil scenario" }, { "paperId": "a9632676d019a8f00b6b8994472561d882b1d086", "title": "A modification on the Chord finger table for improving search efficiency" }, { "paperId": "47faf047a4157073d42e0988e4ebc1a660b7e4e4", "title": "Simulating mobile and distributed systems with DEUS and ns-3" }, { "paperId": "56adc6a44b1d7397c80d219321b055b2ccea6c22", "title": "VoteTrust: Leveraging friend invitation graph to defend against social network Sybils" }, { "paperId": "7499e266887d1b5b070e486e694591697929fe7f", "title": "Persea: a sybil-resistant social DHT" }, { "paperId": "3d9434f1da75b88a1a950443768e1bdc600b857d", "title": "A trust-based routing strategy in structured P2P overlay networks" }, { "paperId": "fc159c799be60d6bcafa5a4744e089828c9af9ca", "title": "An Effective Lookup Strategy for Recursive and Iterative Lookup on Hierarchical DHT" }, { "paperId": "81ae41d0caa85b0c2d2e423fa8e7f41b94c86145", "title": "Applying trust and reputation mechanisms to a Kademlia-based Distributed Hash Table" }, { "paperId": "ee0d55b58d006d783437c458704bbd4c4dadfb18", "title": "Analysis and optimization of routing trust values in a Kademlia-based Distributed Hash Table in a malicious environment" }, { "paperId": "04434837cb5487cf4b487ef328c450c40c375a73", "title": "Study on trust-based maintenance of overlays in structured P2P systems" }, { "paperId": "d072cf541b29e9aad70ebf9340bfe0171ff9fc02", "title": "Whanau: A Sybil-proof Distributed Hash Table" }, { "paperId": "d0dd4fee9d433d16428e8530da5742da6543e50f", "title": "Handbook of Peer-to-Peer Networking" }, { "paperId": "067b124d5635b12bdd868d71fcbb27e4f53aa470", "title": "Limiting Sybil Attacks in Structured P2P Networks" }, { "paperId": "a0811379b8d2ea7d17c9a4aed00f76d20d855cf9", "title": "PET: A PErsonalized Trust Model with Reputation and Risk Evaluation for P2P Resource Sharing" }, { "paperId": "eb51cb223fb17995085af86ac70f765077720504", "title": "Kademlia: A Peer-to-Peer Information System Based on the XOR Metric" }, { "paperId": "35516916cd8840566acc05d0226f711bee1b563b", "title": "The Sybil Attack" }, { "paperId": "cf025469b2d7e4b37c7f2d2bf0d46c6776f48fd4", "title": "Pastry: Scalable, Decentralized Object Location, and Routing for Large-Scale Peer-to-Peer Systems" }, { "paperId": "680ba806b8a651e8cb2e2d64a9d6bc2325a8eea1", "title": "A scalable content-addressable network" }, { "paperId": "f03db79dc2922af3ec712592c8a3f69182ec5d65", "title": "Chord: A scalable peer-to-peer lookup service for internet applications" }, { "paperId": "eb452b225ab5a047e7d1ede136a3ca8a39c31dc8", "title": "GeTrust: A Guarantee-Based Trust Model in Chord-Based P2P Networks" }, { "paperId": "b812ff1ff57b0beededb72bf8b9e0470b64cdc49", "title": "Moving peers in distributed, location-based peer-to-peer overlays" }, { "paperId": "66be23d2d7e120e4d727fbfb7d188b4a68e010e0", "title": "X-Vine: Secure and Pseudonymous Routing in DHTs Using Social Networks" }, { "paperId": "a1e9cf0348b873370461cf6f02772792a70898aa", "title": "Reputation-Based Trust Management in Peer-to-Peer Systems: Taxonomy and Anatomy" }, { "paperId": "e0a9226104a8469c9b077033e3ff8f0da50cb0d6", "title": "Trustchord: chord protocol based on the trust management mechanism" }, { "paperId": null, "title": "LowComplexity Signed Response Based Sybil Attack Detection Mechanism in Wireless Sensor Networks" } ]
22,017
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/012b5aa479f62062944ab4190a60d6f940f10c37
[]
0.919339
A Sponge-Based Key Expansion Scheme for Modern Block Ciphers
012b5aa479f62062944ab4190a60d6f940f10c37
Energies
[ { "authorId": "2185802696", "name": "Maciej Sawka" }, { "authorId": "144009429", "name": "Marcin Niemiec" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-155563", "https://www.mdpi.com/journal/energies", "http://www.mdpi.com/journal/energies" ], "id": "1cd505d9-195d-4f99-b91c-169e872644d4", "issn": "1996-1073", "name": "Energies", "type": "journal", "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-155563" }
Many systems in use today require strong cryptographic primitives to ensure confidentiality and integrity of data. This is especially true for energy systems, such as smart grids, as their proper operation is crucial for the existence of a functioning society. Because of this, we observe new developments in the field of cryptography every year. Among the developed primitives, one of the most important and widely used are iterated block ciphers. From AES (Advanced Encryption Standard) to LEA (Lightweight Encryption Algorithm), these ciphers are omnipresent in our world. While security of the encryption process of these ciphers is often meticulously tested and verified, an important part of them is neglected—the key expansion. Many modern ciphers use key expansion algorithms which produce reversible sub-key sequences. This means that, if the attacker finds out a large-enough part of this sequence, he/she will be able to either calculate the rest of the sequence, or even the original key. This could completely compromise the cipher. This is especially concerning due to research done into side-channel attacks, which attempt to leak secret information from memory. In this paper, we propose a novel scheme which can be used to create key expansion algorithms for modern ciphers. We define two important properties that a sequence produced by such algorithm should have and ensure that our construction fulfills them, based on the research on hashing functions. In order to explain the scheme, we describe an example algorithm constructed this way, as well as a cipher called IJON which utilizes it. In addition to this, we provide results of statistical tests which show the unpredictability of the sub-key sequence produced this way. The tests were performed using a test suite standardized by NIST (National Institute for Standards and Technology). The methodology of our tests is also explained. Finally, the reference implementation of the IJON cipher is published, ready to be used in software. Based on the results of tests, we conclude that, while more research and more testing of the algorithm is advised, the proposed key expansion scheme provides a very good generation of unpredictable bits and could possibly be used in practice.
# energies _Article_ ## A Sponge-Based Key Expansion Scheme for Modern Block Ciphers **Maciej Sawka *[,†]** **and Marcin Niemiec** **[†]** Department of Telecommunications, AGH University of Science and Technology, Mickiewicza 30, 30-059 Krakow, Poland *** Correspondence: maciejsawka@gmail.com** † These authors contributed equally to this work. **Citation: Sawka, M.; Niemiec, M.** A Sponge-Based Key Expansion Scheme for Modern Block Ciphers. _[Energies 2022, 15, 6864. https://](https://doi.org/10.3390/en15196864)_ [doi.org/10.3390/en15196864](https://doi.org/10.3390/en15196864) Academic Editor: Wei-Hsin Chen Received: 8 August 2022 Accepted: 15 September 2022 Published: 20 September 2022 **Publisher’s Note: MDPI stays neutral** with regard to jurisdictional claims in published maps and institutional affil iations. **Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **Abstract: Many systems in use today require strong cryptographic primitives to ensure confidentiality** and integrity of data. This is especially true for energy systems, such as smart grids, as their proper operation is crucial for the existence of a functioning society. Because of this, we observe new developments in the field of cryptography every year. Among the developed primitives, one of the most important and widely used are iterated block ciphers. From AES (Advanced Encryption Standard) to LEA (Lightweight Encryption Algorithm), these ciphers are omnipresent in our world. While security of the encryption process of these ciphers is often meticulously tested and verified, an important part of them is neglected—the key expansion. Many modern ciphers use key expansion algorithms which produce reversible sub-key sequences. This means that, if the attacker finds out a large-enough part of this sequence, he/she will be able to either calculate the rest of the sequence, or even the original key. This could completely compromise the cipher. This is especially concerning due to research done into side-channel attacks, which attempt to leak secret information from memory. In this paper, we propose a novel scheme which can be used to create key expansion algorithms for modern ciphers. We define two important properties that a sequence produced by such algorithm should have and ensure that our construction fulfills them, based on the research on hashing functions. In order to explain the scheme, we describe an example algorithm constructed this way, as well as a cipher called IJON which utilizes it. In addition to this, we provide results of statistical tests which show the unpredictability of the sub-key sequence produced this way. The tests were performed using a test suite standardized by NIST (National Institute for Standards and Technology). The methodology of our tests is also explained. Finally, the reference implementation of the IJON cipher is published, ready to be used in software. Based on the results of tests, we conclude that, while more research and more testing of the algorithm is advised, the proposed key expansion scheme provides a very good generation of unpredictable bits and could possibly be used in practice. **Keywords: cybersecurity; cryptography; block ciphers; symmetric key; iterated ciphers; smart grids** **1. Introduction** Data integrity and confidentiality are the crucial security requirements of information systems and communication networks, including smart grids [1,2]. Deployment of protection methods allows for secure data transmission in cyberspace. However, the security services need efficient cryptography algorithms, such as symmetric block ciphers [3]. These kinds of algorithms are building blocks of modern security services—from privacy protection to authentication of smart meters [4]. Although nowadays most discussion on security focuses on modern algorithms and protocols, cryptography itself is a very old art. One of the oldest mentions of encryption is the Caesar’s cipher, allegedly used by the ruler of ancient Rome. It was a simple cipher operating on text written in Latin alphabet. As centuries passed, other ciphers have improved on its design, with a notable example being the Vigènere cipher. It was also based on Latin alphabet but used a secret key in a way similar to modern constructions. Much later, in the 20th century, the Enigma machine was used by German army during World War II to ----- _Energies 2022, 15, 6864_ 2 of 18 encrypt classified military information. The next step in evolution of cryptography came with the invention of a computer which surpassed humans’ computing ability. This made it necessary to create more complex encryption algorithms that would be able to withstand the newfound computing power without breaking. Works in the field of telecommunication and information theory, based on research of Claude Shannon (among others) served an important role in this development. This led to invention of Data Encryption Standard (DES). This could be considered the beginning of era of modern cryptography, based on ciphers designed for digital computers. Following into the 1990s and early 2000s, improvements made in the area of processing units and other integrated circuits caused DES to become obsolete. Its key sizes were deemed too small and attempts to improve its security through multiple iterations (Triple DES) were thwarted by new attacks. A new standard for cryptography became a necessity. Out of this necessity, Advanced Encryption Standard (AES) was introduced. Despite the time passed since its inception, AES remains in widespread use and is still considered a secure choice for data confidentiality. Nonetheless, new algorithms are still being created in order to improve—if not in terms of security then in terms of performance and ease of implementation [5,6]. Both DES and AES, as well as many modern ciphers, follow a structure called an iterated cipher in which the encryption process is split into a number of rounds. Every round requires a separate secret sub-key. To fulfill this requirement, each cipher of this type defines its own key expansion algorithm. The role of the algorithm is to derive a sequence of sub-keys from the main key. _1.1. Motivation_ A lot of effort has been put over the years into making modern symmetric ciphers secure. A lot of this effort was focused on the encryption process itself, and less on the process of key expansion. This may be observed in the fact that many currently used and upcoming ciphers have fully reversible key expansion algorithms. This means that, given a long enough sequence of sub-keys, the attacker is able to decipher not only the rest of the sequence, but often the main key itself. Example ciphers which exhibit this behaviour include AES and LEA. AES has been the standard for symmetric encryption for around 20 years, and no practical attacks against it have been found. Specifically, no attacks were found that would target its key expansion. Despite this, we cannot be certain that a new attack is not developed in 1, 5 or perhaps 10 years. Modern constructions should not ignore this possibility. This is especially true when one takes into account the research into side-channel attacks, which do not attack the cipher directly. Instead, they attack the environment the algorithm is performed in. The goal of such attacks is to leak secret values from memory. If a sufficiently large part of the sub-key sequence was to be leaked this way, any cipher with reversible key expansion would be instantly compromised. To prevent this from happening, the sub-key sequence should have two important properties: - The main key should not be directly used as part of the sub-key sequence; - Every sub-key should be sufficiently difficult to derive from any other sub-key, including the one happening before and after it in the sequence. The word “sufficiently” in this context means a varying degree of security, but, in general, it should be practically impossible to guess one sub-key based on the knowledge of another. This would mean that, in order to break the encryption, an attacker would have to leak every sub-key in the sequence. Given the cost, complexity and low reliability of side-channel attacks should render this attack vector impractical. _1.2. Contribution_ The authors of this paper propose a key expansion scheme based on the sponge construction. This solution can be used to create key expansion algorithms for modern block ciphers. The scheme produces a sequence of sub-keys which is difficult to reverse thanks to the properties of the sponge construction. This makes it difficult to retrieve the original key or other sub-keys from part of the sequence. This is achieved by using the excess bits ----- _Energies 2022, 15, 6864_ 3 of 18 of the state of the sponge as a variable unknown to the attacker, as well as increasing the work performed between absorbing the input and squeezing the output. This is done to protect ciphers from attacks on block ciphers which aim to retrieve original key material from singular sub-keys, e.g., slide attacks or attacks based on side-channel data extraction. In addition to the scheme itself, a cipher called IJON (pronounced “e-yon”) is proposed. It serves as an example application of the scheme. It is a block cipher with 128 bits of block and key size optimized for processing units capable of operating on 32-bit words. The sequence of sub-keys used in the cipher is generated using a key expansion algorithm based on the sponge construction with 96 bits of sponge state and 32 bitrate. Finally, the test results are described. The sequence of sub-keys produced by the key expansion algorithm of IJON was tested using a suite of tests for cryptographically secure pseudo-random number generators (CSPRNGs) standardized by National Institute of Standards and Technology (NIST) [7]. The suite checks whether a sequence behaves like a truly random, unpredictable stream of bits. Specific methodology which was assumed during tests is described as well. The remainder of the paper proceeds as follows: Section 2 provides an introduction to cryptography techniques applied in modern block ciphers. Sponge construction is explained in Section 3. In Section 4, a new sponge-based key expansion scheme is proposed. The IJON cipher is explained in Section 5 in detail, including both the key expansion as well as the encryption processes. Section 6 describes security considerations, testing methodology and results of statistical tests of the cipher. Finally, Section 7 concludes the paper. The paper is intended for cryptographers working on new symmetric block ciphers. The authors hope it provides them with tools necessary to create secure key expansion algorithms for their constructions. Additionally, any readers interested in developments of cryptography should find the paper interesting. _1.3. Related Works_ The problems arising from usage of reversible sub-key sequences have been noticed before. Latest developments meant to create a secure key expansion scheme have been mostly focused on advanced mathematics, specifically chaos maps [8–10]. While these approaches are most likely to result in a solution, they are also difficult to follow for readers unfamiliar with the topic. When new ciphers are created, it is not only important that they are safe, but also that it is relatively easy to prove that they are safe, or to approximate their level of security. That is why we propose a solution that is based on less complex concepts and constructions— specifically, the sponge construction. Instead of placing the trust in mathematics, we place it in the research previously performed in the field of hashing functions. This way, the resulting solution is much easier to understand for people not acquainted with advanced mathematics—for example software developers, project managers, and smart grid engineers. We believe that, since those are the people who will benefit from results of our work, it is important that they are able to comprehend it. At the same time, we strongly believe that the increase in simplicity will not negatively impact the practical application of our solution. In fact, we provide a reference implementation of the proposed cipher which is ready to be used in software which needs symmetric ciphers with strong key expansion algorithms. _1.4. Acronyms_ The acronyms used in the paper are listed and expanded in Table 1 below. The name “IJON” is not an acronym. ----- _Energies 2022, 15, 6864_ 4 of 18 **Table 1. Acronyms used in the paper.** **Acronym** **Meaning** AES Advanced Encryption Standard ARX Add-Rotate-XOR ASCII American Standard Code for Information Interchange CPU Central Processing Unit CSPRNG Cryptographically Secure Pseudo-Random Number Generator DES Data Encryption Standard LEA Lightweight Encryption Algorithm LTS Long Trail Strategy NIST National Institute for Standards and Technology P-BOX Permutation box—the permutation layer of an SPN PRNG Pseudo-Random Number Generator S-BOX Substitution box—the substitution layer of an SPN SHA-3 Secure Hashing Algorithm 3 SPN Substitution-Permutation Network WTS Wide Trail Strategy **2. Symmetric Block Ciphers** An algorithm is a set of instructions meant to be performed in specific order with a certain purpose behind it. Therefore, a cipher can be viewed as a set of algorithms. Every cipher defines at least two algorithms: encryption and decryption. Purpose of encryption is to transform secret data (called plaintext) in a way that prevents anyone without the knowledge of the special secret key from recovering it. At the same time, anyone who knows the key can easily recover plaintext from the encrypted data (called ciphertext) using the decryption algorithm. Additional algorithms may also be a part of the cipher if they are necessary. Ciphers are divided into two categories as seen in Figure 1: symmetric ciphers and asymmetric ciphers. Symmetric ciphers use the same key during encryption and decryption, while asymmetric ciphers use two distinct, albeit related keys for each operation. The difference does not have any security implications—neither type of cipher is inherently “more secure”. Instead, it necessitates different assumptions about the privacy of the key. This in turn causes each type of ciphers to have different use cases. In practice, both types often are used together. This way they are able to complement each other. Symmetric ciphers can further be divided into stream and block ciphers. Stream ciphers encrypt and decrypt one bit at a time. Block ciphers operate on blocks of data, which have fixed length usually defined in bits or bytes. Encryption and decryption algorithms of such ciphers take a block of data of specified length as input and produce a different block of data of the same length. **Figure 1. Types of ciphers.** A popular design choice for block ciphers is a construction called an iterated cipher, shown in Figure 2. Instead of creating one large algorithm for encryption, a round function is defined. It modifies internal state of the cipher during execution. It is applied to the plaintext a certain number of times called number of rounds. Output of the last round is the ----- _Energies 2022, 15, 6864_ 5 of 18 ciphertext. Each iteration of the round function usually uses a sub-key. It is a smaller piece of data generated from the original key. An inverse round function must also be defined. It is used during decryption, to undo the work performed during encryption. Decryption usually applies sub-keys in reverse order. This approach not only makes creating a cipher simpler, but also minimizes code size. It makes it necessary to define additional algorithm within the cipher, called the key expansion algorithm. Its role is to generate a sequence of sub-keys from the main key. **Figure 2. Encryption in an iterated cipher with r number of rounds.** Iterated ciphers may be implemented in a multitude of ways. One of them is a substitution–permutation network (SPN), as seen in Figure 3. This type of construction is divided into two layers: the substitution layer and permutation layer. The role of the first one is to achieve nonlinearity. This means that the cipher is more difficult to approximate with linear functions. This mitigates attacks based on linear cryptanalysis [11]. Nonlinearity is achieved by using a function usually called an S-BOX. Substitution is the act of replacing one value with another. It is often implemented using a lookup-table (LUT) to allow any possible mapping from input to output. The other part of an SPN is a permutation layer. Its role is to mix the bits of the state together. This layer might only swap bits around, or also mix them using XOR, matrix multiplication or other operations. In the end, the purpose of the P-BOX is to increase diffusion. Making bits of state change positions with each round causes each bit of the output to depend on multiple input bits. **Figure 3. Full round of SPN.** An alternative to SPN is a construction called Feistel network. Assume that desired block size of the cipher is N bits. To create a Feistel network, an F function needs to be defined. The function has to accept two inputs: half of the block (N/2 bits long) and a sub-key. The Feistel network begins by splitting the input block into two halves. During ----- _Energies 2022, 15, 6864_ 6 of 18 each round, the left half is combined with a sub-key by the F function. The output of the function is then combined with the right half using the XOR operation. As the last step in a round, the halves are reordered. The left half becomes the right and vice versa. This may continue for any number of rounds. After the last round, the two halves are concatenated into an N bit block of ciphertext. Feistel network is a simple construction with great potential. It was proven to be secure even with a small number of rounds as long as the input of the F function is sufficiently hard to predict based on its output [12]. Additionally, the function F does not need to be reversible as the decryption algorithm uses the function itself rather than its inverse. The only requirement is that the sub-keys have to be supplied to the round function in reverse order during decryption. In contrast to a Feistel network or an SPN, Add-Rotate-XOR (ARX) does not directly refer to the construction of a block cipher. Instead, ARX can be thought of as a special category of ciphers. Ciphers of this category are built entirely out of three operations: addition modulo 2[n], XOR of n-bit words and n-bit rotations by a constant amount. These operations are very simple and easy to approximate in various ways by potential attackers. Because of this, creation of a secure cipher based solely on them is not a trivial task. However, if used properly, the ARX operations provide the resulting cipher with important advantages, listed below. - All three operations are very fast, usually taking small number of cycles on various CPU architectures. This causes software implementations of such ciphers to be very efficient. - Not only is the time of their execution low but also constant. This means that ciphers built out of them are naturally immune to side-channel attacks based on the time of execution of certain parts of code [13]. - Since the ciphers use only basic operations, they are often very easy to implement and analyze. **3. Sponge Construction** Sponge construction [14] is a scheme most often considered as the core of hashing algorithms rather than block ciphers. It was introduced as a part of the Keccak hash algorithm, the winner of the SHA-3 (Secure Hashing Algorithm 3) competition [15]. Since it is intended to be used in hashing functions, a sequence of bits generated using a sponge construction is usually difficult to reverse. By “reversing” a sequence, we mean calculating one part of the sequence knowing another part of it, or finding the “seed” it was based on. This property also makes it useful for creation of key expansion algorithms. Sponge construction requires three elements. The first is a function f which takes b-bit data block as input and outputs another b-bit data block (b is the size of the internal state of the sponge construction). The second is the bitrate r, which defines the size of chunks in which sponge consumes input and returns output (r should be always smaller than b). Finally, the third element is a pad function, which makes sure that input to the sponge is always a multiple of r. In this paper, we can ignore the pad function and assume that input always has correct length. The sponge construction allows creation of a function which generates output of any size (limited to multiples of bitrate) from input of any size. Sponge functions work in the way visualized in Figure 4. In the figure, the vertical dashed line divides the absorbing and squeezing stages. The horizontal lines mark the part of the state directly modified by the input and directly copied to output. Other bits of the state have to be populated by the f function. The steps of the function are presented below. 1. Set all b bits of internal state to 0. 2. Divide input data into chunks of r bits: I0, I1, up to Ik for selected k. 3. For each chunk of input data perform the absorbing procedure: (a) Apply the input chunk to the first r bits of internal state through the XOR operation, (b) Apply the f function to the internal state. ----- _Energies 2022, 15, 6864_ 7 of 18 4. After all input has been absorbed by the sponge, start squeezing out the output: (a) Append first r bits of the state to the output, (b) Apply the f function to the internal state. 5. Stop after all necessary output has been squeezed out. **Figure 4. Sponge function.** **4. Sponge-Based Key Expansion** The role of the key expansion algorithm in an iterated cipher is to generate sub-keys for all the rounds of encryption and decryption. The sub-keys have to depend on the value of the main key. The bits of the main key are also collectively known as key material. At the same time, we suggest that it should be difficult to guess one part of the sequence of sub-keys from another. It should also be difficult to guess the seed that the sequence was generated from. To make it possible, we propose usage of the sponge construction as the framework for key expansion algorithms. In order to explain how one would use the sponge this way, we propose a novel key expansion algorithm to serve as an example. Our key expansion algorithm is based entirely on 32-bit ARX operations. This way it should be easy to implement and be optimized for 32-bit processing units. The size of its internal state is 96 bits, which can be easily implemented as three 32-bit words. The input to the key expansion algorithm is 128 bits (16 bytes) of key material, divided into four input words. In terms of the sponge construction, the bitrate parameter is equal to 32 bits [14]. Output is also split into 32-bit words, and each output word is a single sub-key. Key expansion is split into four stages, as seen in Figure 5. The stages, in order, are: initialization, absorbing, mixing and squeezing. Initialization, absorbing and squeezing are all standard phases of the sponge construction.The mixing stage might be thought of as part of the squeezing stage, with output discarded. It is added to increase the amount of work performed on the internal state before output words are collected. The function f is defined later in this section. **Figure 5. Sponge-based key expansion algorithm.** ----- _Energies 2022, 15, 6864_ 8 of 18 The algorithm is described in steps below. 1. Initialization: Set 96 bits of internal state to 0. 2. Absorbing: (a) Absorb a word of input K[i] through a XOR operation with the first 32 bits of the state (b) Apply 4 iterations of the f function to the internal state; Repeat until all of the key material has been absorbed (4 times). 3. Mixing: Apply 24 iterations of the f function 4. Squeezing: (a) Apply 12 iterations of the f function to the internal state; (b) Squeeze an output word Sk[j], by saving first 32 bits of the internal state. Repeat until all of the sub-keys have been squeezed out. The number of iterations of each stage of the algorithm as well as the number of applications of the f function is based on the length of the input and required length of the output. They also take into account the results of security analysis, to find a good trade-off between security and performance of the algorithm. _The f Function_ The f function used in the key expansion algorithm transforms 96-bit input into 96-bit output. Its purpose is to fill the bits of state that are not modified by input data and to mix all the bits of the state. The security of the key expansion algorithm relies heavily on the excess bits of the state. Because of this, the design of the f function is very important. The function is designed as three applications of a function fot (“f one-third”), which is presented in Figure 6. It splits input state into three 32-bit words a, b and c. Then, it applies constants C1, C2 and C3 to the state using XOR. This is followed by a series of ARX operations between the words. In the end, the state is rotated one position to the right to produce the output words, a[′], b[′] and c[′]. **Figure 6. The fot function (one third of the entire f function).** The application of constants as the first step of the function ensures that any bits are set to 1 prior to the additions, XORs and rotations being performed. This mitigates a fundamental weakness of ARX operations. Every ARX operation will produce an all-zero word as output if given an all-zero input. Thanks to the constants, if the initial state of a, b and c is set to 0, some of the bits will change value to 1 before any other operations take place. In case the input state happens to be equal to the constants, applying the constants will actually have the opposite effect and clear all bits. However, those will be again set back to 1 in the next iteration of the fot function. Usage of constants at the beginning removes an entire class of trivial weak keys with all bits set to 0 or with a very small number of bits set to 1. ----- _Energies 2022, 15, 6864_ 9 of 18 The values of the constants proposed by the authors are given below: ``` C1 = 0x1763af12, C2 = 0xd1bb5770, C3 = 0x2b3a55bb ``` These are so called nothing-up-my-sleeve numbers. A nothing-up-my-sleeve number is a type of a constant generated in a complicated manner, based on values which are hard to control. Oftentimes, fractional parts of mathematical constants are used or values derived from the name of the algorithm. This is done to eliminate any suspicion—a skilled cryptanalyst could theoretically develop a cipher with meticulously chosen constants that allow a backdoor into the algorithm. By using nothing-up-my-sleeve numbers, this is made significantly harder. This, in turn, makes the algorithm more trustworthy. The fot function constants have been generated from the name “IJON” (The algorithm was developed in the year 2021, which also marked the hundredth anniversary of the birth of Stanisław Lem—a Polish writer of science fiction and futurologist. Ijon Tichy is the main character in many of his novels and the cipher was named after him.) in the way described in steps below: 1. Four bytes which form the string IJON (in ASCII encoding) were interpreted as a 32-bit floating point number. In addition to the number itself, its square root and second power were calculated. This resulted in a total of three floating point values. 2. All three values from previous step were reinterpreted as unsigned integers. The following procedure was performed on each of them; (a) The integer was multiplied by itself, generating a 64-bit value; (b) The upper and lower halves of the result from previous step were XORed together to make a 32-bit number; (c) This number then served as input to the next iteration of the procedure, for a total of 128 iterations; 3. The result of the last iteration of the procedure became the output of the entire algorithm. Since procedure was performed on three integers, it resulted in three constants: _C1, C2 and C3._ The ARX operations performed after the application of constants are the core of the function. The order of operations was decided by attempting to obtain the highest possible diffusion between all three words of the state to utilize the state to its full potential. Both order and rotation amounts were determined by trial and error, with resulting sub-key sequences rated by statistical tests which measure randomness [7]. Another consideration was performance on various CPUs—some architectures support shifts by any amount, but some 8-bit architectures might only support rotations through arithmetic shifts, which are limited to 8 bits at most. In those cases, amounts chosen are a multiple of 8, with a potential additional shift by 1 (for example 17 = 8 + 8 + 1, which results in only three rotations). The last part of the fot function is 32-bit words swap. During each of three iterations, every word enters the fot function at a different position to perform different operations. This allows for chaining of three fot iterations into one full f iteration and results in a larger and more complex procedure being constructed out of smaller steps. The full f function is presented in Figure 7. The first word of the state and its path through the function are highlighted. This construction both increases the diffusion and allows for memory/time trade-off during implementation. If performance is more important than space, loop unrolling might be performed as is usually done. However, if the implementation targets the embedded environment, space might be more important than speed of execution. In such case, the key expansion algorithm may be implemented as a loop which repeatedly executes the fot function. Since fot is relatively small—consisting of only 5 XORs, 3 additions and 4 rotations—it would result in very small code size, at the expense of performance. ----- _Energies 2022, 15, 6864_ 10 of 18 **Figure 7. The f function made of three iterations of fot.** **5. IJON Cipher** The proposed sponge-based key expansion algorithm was used in a selected iterated cipher. The authors proposed such cipher and decided to call it IJON. The IJON cipher encrypts data over the span of 10 rounds. Each round consumes eight sub-keys. This results in a total of 80 sub-keys in a sequence generated by the key expansion algorithm. The design of IJON was based on the research into the Long Trail Strategy (LTS) [16]. LTS is a cipher design strategy applicable to ARX ciphers. It was inspired by previous work on AES and the Wide Trail Strategy (WTS) [17]. Its aim is to allow bounding of the possible probabilities of differential trails within the ciphers. To achieve that, the ciphers are constructed in a way similar to traditional substitution–permutation networks, except with S-BOXes implemented as series of ARX operations (called ARX-BOX). LTS focuses on the substitution layer, by introducing multiple applications of the S-BOX intertwined with sub-key applications before the permutation layer. This is done to place the primary burden of achieving diffusion on the substitution layer. This allows differential probabilities to be bounded, approximating complexity of an attack. All of this is reflected in the architecture of IJON encryption algorithm. To match the nomenclature of the LTS research, in this section, applications of the S-BOX will be referred to as rounds, while what is usually called a round in iterated ciphers will be referred to as a step. _5.1. Encryption Algorithm_ The encryption algorithm is visualized in Figure 8 and described below. 1. Plaintext Pt contains 128 bits of data and serves as an input to the algorithm. 2. Split Pt into 4 words of 32 bits each. 3. Perform ten steps on the words of the state. Each step has 8 sub-keys K assigned from the sequence generated by the key expansion. (a) Combine four first sub-keys with the words of the state using XOR. (b) Apply the S-BOX S twice in parallel to the state. (c) Combine four last sub-keys. ----- _Energies 2022, 15, 6864_ 11 of 18 (d) Perform the second application of S-BOXes. (e) Apply the P-BOX P. 4. The output of the last step is the resulting ciphertext Ct. As the core of the substitution layer IJON can use a selected S-BOX. The authors decided to use a 64-bit ARX-BOX called Alzette [18]. It offers great statistical properties even at just two applications while being very efficient in both software and hardware. Due to the IJON block size being 128 bits long, two parallel applications of this S-BOX are performed on the input during each application. **Figure 8. The encryption algorithm of IJON.** The permutation layer, while deemphasized in the LTS, still remains a vital part of the SPN construction. It ensures mixing of all the bits of the input together. A construction similar to a Feistel network is used in IJON as the permutation layer. The F function is inspired by the one used in the SPARX family of ciphers, introduced as part of the LTS research [16]. The entire Feistel-like P-BOX is shown in Figure 9, while the F function itself is shown in Figure 10. **Figure 9. Feistel-like P-BOX of IJON.** **Figure 10. F function used in P-BOX.** ----- _Energies 2022, 15, 6864_ 12 of 18 _5.2. Decryption Algorithm_ The role of the decryption algorithm is to reverse the work performed by encryption— a given ciphertext and sub-key sequence should return the original plaintext. Because of this, its design is closely related to the encryption function through inverse operations. Luckily, all ARX operations have trivial inverses: inverse of XOR is the XOR itself, inverse of rotation right by N bits is either rotation left by the same amount, or rotation right by (32 _N) bits (when working with 32-bit words) and inverse of addition modulo 2[32]_ is _−_ subtraction. Thanks to this, many complex functions built out of those three operations are easily invertible. When defining the decryption algorithm, the goal is to find those inversions and perform them in reverse order. The decryption algorithm is listed in steps below and is presented in Figure 11. 1. Ciphertext Ct contains 128 bits of encrypted data and serves as the input to the algorithm. 2. Split Ct into 4 words of 32 bits each. 3. Perform ten reverse steps on the words of the state. Each step has 8 sub-keys K assigned from the sequence generated by the key expansion. (a) Apply the inverse P-BOX P[−][1] to reverse the mixing of bits. (b) Apply the inverse S-BOX S[−][1] twice in parallel to the state. (c) Combine four last sub-keys with the words of the state using XOR. (d) Apply the inverse S-BOX again. (e) Combine the four first sub-keys with the state. 4. The output of the last decryption step is the resulting plaintext Pt. **Figure 11. Decryption algorithm.** It is worth mentioning that the inverse of the P-BOX must be found. This structure is presented in Figure 12. It is used to reverse the order of operations: the application of the F function and the reordering of halves. What is important is that the F function itself stays the same. It is one of the advantages of the Feistel-like structure on which the P-BOX is based. An inverse of the S-BOX is harder to find, although still trivial. This is due to the fact that ARX operations are easily invertible. In fact, most of the operations in the S-BOX do not even have to be replaced—they are their own inverse. The only operation that has to be replaced by its inverse is addition modulo 2[32] which becomes subtraction. ----- _Energies 2022, 15, 6864_ 13 of 18 **Figure 12. Inverse P-BOX.** **6. Security Considerations** The functionality of the IJON cipher which contains a sponge-based key expansion algorithm was verified. Additionally, basic diffusion and confusion properties have been tested. This means that a single bit changed in key or plaintext results in massive changes in the ciphertext, with around 50% of bits changing value. This is important to verify, but it is not enough to fully evaluate the security of the cipher. IJON, like many iterated ciphers, has parameters—for example, the number of steps, the number of rounds within a step (S-BOX applications), block size, key size, etc. All of them should have values that are justified either by tests and experiments or by common practice and logic. This section contains security considerations and explains the choices made in design of the cipher. _6.1. Key Size and Block Size_ Minimal values for these two parameters are slowly growing as hardware evolves. Ciphers with 64-bit block or key size, albeit sometimes still used in memory-constrained environments, are usually deemed not secure enough. This is mostly due to the key space being too small and prone to a brute force attack. The block size and shortest key of AES cipher is 128 bits, which still holds up very well today after more than 20 years in use. Therefore, 128 bits seems to be the perfect middle-ground between ’too small to be secure’ and ’too big to be practical’. Due to those reasons, this key size was chosen. The introduced key expansion algorithm may theoretically use keys of any length, as long as the length is a multiple of 32 bits. This can be done by performing the absorbing stage of the algorithm different amount of times. In the future, if 128 bits of key material proves to be insufficient, the algorithm could easily be expanded for longer keys. _6.2. Side-Channel Attacks_ The cipher was consciously designed with resistance against time-based side-channel attacks [13] in mind. Thanks to the exclusive usage of constant-time ARX operations, it is easier to implement the cipher in a way that is immune to this type of attack. This includes the reference implementation created to verify functionality of the proposed algorithm [19]. _6.3. Slide Attack_ Slide attack is a technique which targets iterated ciphers that base their security on a large number of rounds and re-use a single sub-key between multiple rounds [20]. Since IJON is an iterated cipher, it may be susceptible to a slide attack. To counter that, a lot of effort was directed into creating a strong, secure key expansion algorithm. Every sub-key is used only once, in order to make potential attack as complicated as possible. Additionally, even if an attacker comes in possession of a single sub-key, it should be difficult to retrieve the main key or other sub-keys from that information. This is due to the non-invertible sponge construction used as the framework for the key expansion algorithm. ----- _Energies 2022, 15, 6864_ 14 of 18 _6.4. Construction of Encryption_ The important component of the encryption algorithm is the selected S-BOX (in IJON, the Alzette S-BOX was chosen) due to Long Trail Strategy [16] guiding the design of the cipher. Because of that, security approximation was based entirely on its properties. Authors of the Alzette S-BOX [18] claim that two iterations of the ARX-BOX have the Maximum Expected Differential Characteristic Probability (MEDCP) bounded at around 2[−][32]. This was decided to be reasonably low for a single step of encryption, and so the number of rounds within a single step was set to 2. The number of steps was set to 10, which allows approximation of maximum differential trail probability at around 2[−][320] under the assumption that no other characteristics arise from the repeated use of the S-BOX or from connection with P-BOX within the encryption algorithm. This probability is very small, which means that the number of steps could have been lowered to possibly 8 or 7. However, a security margin was assumed and number of steps during encryption was chosen to be 10. _6.5. Randomness of Key Expansion_ Key expansion of IJON has been tested as a Cryptographically Secure Pseudo-Random Number Generator (CSPRNG). The tests were performed according to a specification published by the National Institute of Standards and Technology (NIST) [7]. 6.5.1. Methodology The suite defines a number of tests meant to measure the unpredictability of a random sequence of bits. Each test generates a p-value as an output which characterizes randomness in quantitative way. The sequence should result in a p-value bigger than 0.01 to pass a test. The tests are listed below: 1. Frequency (monobit) test 2. Frequency test within a block 3. Runs test 4. Test for the longest run of ones in a block 5. Binary matrix rank test 6. Discrete Fourier transform (spectral) test 7. Non-overlapping template matching test 8. Overlapping template matching test 9. Maurer’s “Universal Statistical” test 10. Linear complexity test 11. Serial test 12. Approximate entropy test 13. Cumulative sums (cusum) test 14. Random excursions test 15. Random excursions variant test All tests put constraints on the input given to them, with one of the constraints being the minimal length of the input sequence. When testing IJON, tests with numbers 5, 8, 9, 10, 14 and 15 have been omitted due to their requirements on length. Additionally, in tests where an arbitrary length of a block was to be chosen, a value of 256 bits was used—the total length of 8 sub-keys used during a single round of encryption. Sequences generated from expansion of a set of keys have been tested using the test suite. Out of all the keys, most attention was given to the theoretical weak keys. This is a group of keys with only 0, 1 or 2 bits set to 1. Those keys could become trivial weak keys if the algorithm was not properly designed and tested due to the nature of ARX operations used. Additionally, two variants of pseudorandom keys have been tested. In the first variant, 128 bits of the main key were generated, and then expanded using the IJON key expansion algorithm. In the second variant, the entire sequence of 320 bytes was generated, and no expansion was used. All pseudorandom data in both cases have been generated using/dev/urandom—Pseudo-Random Number Generator (PRNG) present in ----- _Energies 2022, 15, 6864_ 15 of 18 Linux environment. The second variant has been performed to compare the results of IJON key expansion algorithm to a well-known PRNG. 6.5.2. Test Results Test results for the three sets of keys are presented in Tables 2–4. Meaning behind values in columns is described below: - Max Diff—applicable only to the monobit test, it is the maximal absolute difference between expected and actual number of ones in the sequence. The percentage in the brackets is given in relation to the expected value (100% = 1280 bits). - Success rate—number of samples from the given set that successfully passed a given test. Percentage in brackets is in relation to the number of samples in the set. - Min/Max/Avg P—respectively minimal, maximal and average value of P among all samples, both successful and failing. The p-value has to be greater than 0.01 to pass a test. The results for the second pseudorandom variant should be treated as a benchmark. They were generated using an industry-standard PRNG and are independent from the IJON cipher structure. Overall, results for all three groups are very similar. Success rates in all cases are above 98%, and average p-values are relatively high: 0.40 and more. In some rare cases, the IJON-based sequences achieve better results than those generated through Linux PRNG (an example would be the Min P for the longest run of ones test). The differences are so small that they can be attributed to the randomness during bit generation. Therefore, IJON key expansion algorithm could be considered comparable to the Linux/dev/urandom PRNG in terms of generating an unpredictable sequence of bits. What is worth emphasizing is that the difference between results for potential weak keys and first variant random keys is also very small. This means that the potential weak keys are not weaker than random keys generated using a PRNG. In turn, they should not be considered weak at all. **Table 2. Test results for potential weak keys (8257 samples total).** **Test Number/Name** **Success Rate** **Min P** **Max P** **Avg P** 1. Monobit—Max Diff 109 (8.52%) 8182 (99.09%) 0.000016 1.000000 0.500998 2. Frequency within block 8172 (98.97%) 0.000023 0.999526 0.499081 3. Runs 8162 (98.85%) 0.000000 1.000000 0.501611 4. Longest run of ones 8178 (99.04%) 0.000100 0.993439 0.496900 6. DFT 8166 (98.90%) 0.000066 1.000000 0.494753 7. Non-overlapping template match 8242 (99.82%) 0.000188 1.000000 0.923600 11. Serial 8135 (98.52%) 0.000182 0.997537 0.414052 12. Approximate entropy 8200 (99.31%) 0.000075 0.999842 0.502494 13. Cusum 8157 (98.79%) 0.000016 0.999526 0.422093 **Table 3. Test results for the first variant random (10,000 samples total).** **Test Number/Name** **Success Rate** **Min P** **Max P** **Avg P** 1. Monobit—Max Diff 97 (7.58%) 9907 (99.07%) 0.000126 1.000000 0.499881 2. Frequency within block 9902 (99.02%) 0.000082 0.999928 0.499588 3. Runs 9895 (98.95%) 0.000124 1.000000 0.497403 4. Longest run of ones 9920 (99.20%) 0.000004 0.993439 0.498367 6. DFT 9901 (99.01%) 0.000006 1.000000 0.490965 7. Non-overlapping template match 9981 (99.81%) 0.000236 1.000000 0.920993 11. Serial 9824 (98.24%) 0.000022 0.998638 0.403821 12. Approximate entropy 9893 (98.93%) 0.000106 0.999955 0.493109 13. Cusum 9884 (98.84%) 0.000111 0.999798 0.423216 ----- _Energies 2022, 15, 6864_ 16 of 18 **Table 4. Test results for the second variant random (10,000 samples total).** **Test Number/Name** **Success Rate** **Min P** **Max P** **Avg P** 1. Monobit—Max Diff 98 (7.66%) 9918 (99.18%) 0.000065 1.000000 0.502647 2. Frequency within block 9910 (99.10%) 0.000021 0.999950 0.503439 3. Runs 9894 (98.94%) 0.000122 1.000000 0.501850 4. Longest run of ones 9907 (99.07%) 0.000000 0.993439 0.497037 6. DFT 9912 (99.12%) 0.000066 1.000000 0.489584 7. Non-overlapping template match 9985 (99.85%) 0.000094 1.000000 0.922027 11. Serial 9832 (98.32%) 0.000012 0.999070 0.409102 12. Approximate entropy 9882 (98.82%) 0.000026 0.999896 0.497468 13. Cusum 9883 (98.83%) 0.000014 0.999526 0.426680 **7. Conclusions** Development of modern ciphers leads us towards higher security of smart grids. This paper describes a novel key expansion scheme based on the sponge construction. This construction is able to produce a sequence of subkeys which is difficult to reverse and can be used in iterated modern ciphers. The authors also introduced a new block cipher called IJON. The design of this solution was described in details and a reference implementation was developed. The cipher is able to encrypt and decrypt data as long as the same key is used in both operations, which is a standard way of operation of symmetric ciphers. The new construction is well suited to execute on 32-bit CPUs. The algorithm is based on fast operations on 32-bit words that perfectly fit into registers of such processing units. At the same time, it does not require hardware implementation to perform well, due to low clock cycles required for used operations. Tests and approximations of security bounds were performed. The results indicate that the algorithm is safe to use. The sequence generated by the key expansion algorithm is very unpredictable and hard to reverse. It is worth mentioning that the security margin left by the number of rounds is very high. It was calculated based on differential probabilities of the used Alzette S-BOX. However, this does not give a guarantee about the cipher’s security. Further tests and analysis are still required in this area. On the other hand, research might also assess that the number of rounds is actually too big. In such case, it may be reduced in future versions of the algorithm to increase efficiency. This could help to increase performance and resolve potential memory requirement problem. However, results of security analysis indicate that the cipher seems to have great potential. IJON with the sponge-based key expansion algorithm is a fully functional cipher; however, there are still a lot of possibilities for improvements. First of all, further tests are needed to fully evaluate the encryption process in terms of security. In addition, further performance tests and benchmarks are needed. Speed of execution should be tested and compared to other modern ciphers. In addition to that, further optimizations in assembly may be possible using vector processing instructions on various architectures—for instance, SIMD extensions to the x86 instruction set. Furthermore, while the cipher should execute very well on 32-bit microcontrollers, its memory requirements may be a problem in embedded environments. The length of the buffer necessary to hold all sub-keys after key expansion is equal to 320 bytes. This is by no means a small amount, even by today’s standards. This problem could be mitigated by generating sub-keys as needed during encryption, but it is a solution which wastes a lot of cycles by duplicating work. It also does not work with decryption, as sub-keys are used in reverse order. In summary, further research on cipher’s security as well as performance optimizations are required before the algorithm could be widely used. However, the cipher is pretty much ready to be incorporated into software thanks to the reference implementation. ----- _Energies 2022, 15, 6864_ 17 of 18 **Author Contributions: Conceptualization, M.S. and M.N.; methodology, M.S. and M.N.; software,** M.S.; validation, M.S.; formal analysis, M.S. and M.N.; investigation, M.S. and M.N.; writing—original draft preparation, M.S. and M.N.; writing—review and editing, M.S. and M.N.; visualization, M.S.; supervision, M.N.; project administration, M.N.; funding acquisition, M.N. All authors have read and agreed to the published version of the manuscript. **Funding: This work has been funded by the European Union’s Horizon 2020 Research and In-** novation Programme, under Grant Agreement No. 830943, project ECHO (European network of Cybersecurity centres and competence Hub for innovation and Operations). The research was also partially supported by the National Centre for Research and Development, Grant No. CYBERSECIDENT/381319/II/NCBR/2018 on “The federal cyberspace threat detection and response system” (acronym DET-RES) as part of the second competition of the CyberSecIdent Research and Development Program–Cybersecurity and e-Identity. **Institutional Review Board Statement: Not applicable.** **Informed Consent Statement: Not applicable.** **Data Availability Statement: The data presented in this study—reference implementation of the IJON** [block cipher—are available online in: https://github.com/msaw328/ijon (accessed on 8 September 2022).](https://github.com/msaw328/ijon) **Conflicts of Interest: The authors declare no conflict of interest.** **References** 1. Tufail, S.; Parvez, I.; Batool, S.; Sarwat, A. A Survey on Cybersecurity Challenges, Detection, and Mitigation Techniques for the [Smart Grid. Energies 2021, 14, 5894. [CrossRef]](http://doi.org/10.3390/en14185894) 2. Alghassab, M. Analyzing the Impact of Cybersecurity on Monitoring and Control Systems in the Energy Sector. Energies 2022, _[15, 218. [CrossRef]](http://dx.doi.org/10.3390/en15010218)_ 3. Jain, N.; Chauhan, S.S. Novel Approach Transforming Stream Cipher to Block Cipher. In Proceedings of the 2021 International Conference on Technological Advancements and Innovations (ICTAI), Tashkent, Uzbekistan, 10–12 November 2021; pp. 182–187. 4. Di Matteo, S.; Baldanzi, L.; Crocetti, L.; Nannipieri, P.; Fanucci, L.; Saponara, S. Secure Elliptic Curve Crypto-Processor for Real-Time [IoT Applications. Energies 2021, 14, 4676. [CrossRef]](http://dx.doi.org/10.3390/en14154676) 5. Rodinko, M.; Oliynykov, R. Comparing Performances of Cypress Block Cipher and Modern Lighweight Block Ciphers on Different Platforms. In Proceedings of the 2019 IEEE International Scientific-Practical Conference Problems of Infocommunications, Science and Technology (PIC S&T), Kyiv, Ukraine, 8–11 October 2019; pp. 113–116. 6. Alasaad, A.; Alghafis, A. Key-Dependent S-box Scheme for Enhancing the Security of Block Ciphers. In Proceedings of the 2019 2nd International Conference on Signal Processing and Information Security (ICSPIS), Dubai, United Arab Emirates, 30–31 October 2019; pp. 1–4. 7. Rukhin, A.; Soto, J.; Nechvatal, J.; Smid, M.; Barker, E.; Leigh, S.; Levenson, M.; Vangel, M.; Banks, D.; Heckert, A.; et al. A Statistical _Test Suite for Random and Pseudorandom Number Generators for Cryptographic Applications; National Institute of Standards & Technology:_ Gaithersburg, MD, USA, 2010. 8. Xu, Y.; Zhao, M.; Liu, H. Design an irreversible key expansion algorithm based on 4D memristor chaotic system. Eur. Phys. J. Spec. _[Top. 2022. [CrossRef]](http://dx.doi.org/10.1140/epjs/s11734-022-00561-2)_ 9. Liu, H.; Wang, X.; Li, Y. Cryptanalyze and design strong S-Box using 2D chaotic map and apply to irreversible key expansion. _arXiv 2021, arXiv:2111.05015 ._ 10. Zhao, M.; Liu, H. Construction of a Nondegenerate 2D Chaotic Map with Application to Irreversible Parallel Key Expansion [Algorithm. Int. J. Bifurc. Chaos 2022, 32, 2250081. [CrossRef]](http://dx.doi.org/10.1142/S021812742250081X) 11. Matsui, M. Linear Cryptanalysis Method for DES Cipher. In Proceedings of the Advances in Cryptology— EUROCRYPT’93; Helleseth, T., Ed.; Springer: Berlin/Heidelberg, Germany, 1994. 12. Luby, M.; Rackoff, C. How to Construct Pseudorandom Permutations from Pseudorandom Functions. SIAM J. Comput. 1988, _[17, 373–386. [CrossRef]](http://dx.doi.org/10.1137/0217022)_ 13. Kocher, P.C. Timing Attacks on Implementations of Diffie-Hellman, RSA, DSS, and Other Systems. In Proceedings of the Advances _in Cryptology—CRYPTO’96; Koblitz, N., Ed.; Springer: Berlin/Heidelberg, Germany, 1996; pp. 104–113._ 14. [Bertoni, G.; Daemen, J.; Peeters, M.; Van Assche, G. Cryptographic Sponge Functions. 2011. Available online: https://keccak.](https://keccak.team/files/CSF-0.1.pdf) [team/files/CSF-0.1.pdf (accessed on 8 September 2022).](https://keccak.team/files/CSF-0.1.pdf) 15. Dworkin, M. SHA-3 Standard: Permutation-Based Hash and Extendable-Output Functions; Federal Inf. Process. Stds. (NIST FIPS); National Institute of Standards and Technology: Gaithersburg, MD, USA, 2015. 16. Dinu, D.; Perrin, L.; Udovenko, A.; Velichkov, V.; Großschädl, J.; Biryukov, A. Design Strategies for ARX with Provable Bounds: Sparx and LAX. In Proceedings of the Advances in Cryptology—ASIACRYPT 2016; Cheon, J.H., Takagi, T., Eds.; Springer: Berlin/Heidelberg, Germany, 2016; pp. 484–513. 17. Daemen, J.; Rijmen, V. The Wide Trail Design Strategy. In Proceedings of the Cryptography and Coding; Honary, B., Ed.; Springer: Berlin/Heidelberg, Germany, 2001; pp. 222–238. ----- _Energies 2022, 15, 6864_ 18 of 18 18. Beierle, C.; Biryukov, A.; Cardoso dos Santos, L.; Großschädl, J.; Perrin, L.; Udovenko, A.; Velichkov, V.; Wang, Q. Alzette: A 64-Bit ARX-box. In Proceedings of the Advances in Cryptology—CRYPTO 2020; Micciancio, D., Ristenpart, T., Eds.; Springer International Publishing: Cham, Switzerland, 2020; pp. 419–448. 19. [Sawka, M. Reference Implementation of the IJON Block Cipher. 2021. Available online: https://github.com/msaw328/ijon](https://github.com/msaw328/ijon) (accessed on 8 September 2022). 20. Biryukov, A.; Wagner, D. Slide Attacks. In Proceedings of the Fast Software Encryption; Knudsen, L., Ed.; Springer: Berlin/Heidelberg, Germany, 1999; pp. 245–259. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/en15196864?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/en15196864, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/1996-1073/15/19/6864/pdf?version=1663744451" }
2,022
[]
true
2022-09-20T00:00:00
[ { "paperId": "56244a2dea676d3ff025a6a604cb418b4e4c2b7c", "title": "Design an irreversible key expansion algorithm based on 4D memristor chaotic system" }, { "paperId": "157284c5da3baadf122db30b06f98a996c7825d4", "title": "Analyzing the Impact of Cybersecurity on Monitoring and Control Systems in the Energy Sector" }, { "paperId": "8bfb36b4ae66b3794946b6dc2f789f7a56cb950d", "title": "Novel Approach Transforming Stream Cipher to Block Cipher" }, { "paperId": "4fd2be3f0549fbb3b3464b297bab433de82a9e66", "title": "Cryptanalyze and design strong S-Box using 2D chaotic map and apply to irreversible key expansion" }, { "paperId": "cdead98d0e59c098ec2b225225e51cbf97abc4c3", "title": "A Survey on Cybersecurity Challenges, Detection, and Mitigation Techniques for the Smart Grid" }, { "paperId": "bda236ee1940bc2078e74f9b276a522a8c4c860c", "title": "Secure Elliptic Curve Crypto-Processor for Real-Time IoT Applications" }, { "paperId": "0d7822069f69fa33477307e5e106a9f6a4b71b21", "title": "Alzette: A 64-bit ARX-box" }, { "paperId": "dd0987a4327aed336e70d219639b3a9ce7313a7e", "title": "Key-Dependent S-box Scheme for Enhancing the Security of Block Ciphers" }, { "paperId": "55ba14cd22ec8ca8b52d85313bd127fb37626967", "title": "Comparing Performances of Cypress Block Cipher and Modern Lighweight Block Ciphers on Different Platforms" }, { "paperId": "84343afddb829367110b053edb8424f9f8db1b19", "title": "Design Strategies for ARX with Provable Bounds: Sparx and LAX" }, { "paperId": "0a1a4a9a4e19c58445a89339719b6d23c427fac7", "title": "The Wide Trail Design Strategy" }, { "paperId": "4c7b001199c2e95d4fcdf25b3e1592f4cbd9bfce", "title": "Linear Cryptanalysis Method for DES Cipher" }, { "paperId": "172cb8e1f02f57f1893edab976be9fed99a4d2b3", "title": "How to Construct Pseudo-Random Permutations from Pseudo-Random Functions (Abstract)" }, { "paperId": "5e2595b9a40bae56c789e82b2ed735c49146d618", "title": "First Things First" }, { "paperId": "c8f6a8f081f49325eb97600eca05620887092d2c", "title": "Timing Attacks on Implementations of Di(cid:14)e-Hellman, RSA, DSS, and Other Systems" } ]
13,991
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/012bcc9a33bf8686741b315847b6c45f3b23fbb9
[ "Computer Science" ]
0.886741
Safe fusion compared to established distributed\ fusion methods
012bcc9a33bf8686741b315847b6c45f3b23fbb9
International Conference on Multisensor Fusion and Integration for Intelligent Systems
[ { "authorId": "2715984", "name": "J. Nygårds" }, { "authorId": "2947805", "name": "Viktor Deleskog" }, { "authorId": "2486923", "name": "Gustaf Hendeby" } ]
{ "alternate_issns": null, "alternate_names": [ "MFI", "Int Conf Multisens Fusion Integr Intell Syst" ], "alternate_urls": null, "id": "c0c48886-e760-4ad7-b692-040bb5c2a621", "issn": null, "name": "International Conference on Multisensor Fusion and Integration for Intelligent Systems", "type": "conference", "url": null }
null
## Safe Fusion Compared to Established Distributed Fusion Methods #### Jonas Nygårds, Viktor Deleskog and Gustaf Hendeby ### Conference Publication #### N.B.: When citing this work, cite the original article. Original Publication: Jonas Nygårds, Viktor Deleskog and Gustaf Hendeby, Safe Fusion Compared to Established Distributed Fusion Methods, Proceedings of IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems, 2016. Copyright: http://ieee.org Postprint available at: Linköping University Electronic Press http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-131425 ----- # Safe Fusion Compared to Established Distributed Fusion Methods ##### Jonas Nygårds[∗], Viktor Deleskog[∗], and Gustaf Hendeby[†] _∗_ Div. of C4ISR, Swedish Defence Research Agency (FOI), Linköping, Sweden e-mail: firstname.lastname@foi.se _† Dept. of Electrical Engineering, Linköping University, Linköping, Sweden_ e-mail: hendeby@isy.liu.se **_Abstract—The safe fusion algorithm is benchmarked against_** **three other methods in distributed target tracking scenarios. Safe** **fusion is a fairly unknown method similarly to, e.g., covariance** **intersection, that can be used to fuse potentially dependent** **estimates without double counting data. This makes it suitable** **for distributed target tracking, where dependencies are often** **unknown or difficult to derive. The results show that safe fusion** **is a very competitive alternative in five evaluated scenarios, while** **at the same time easy to implement and compute compared to** **the other evaluated methods. Hence, safe fusion is an attractive** **alternative in track to track fusion systems.** I. INTRODUCTION Methods for track-to-track fusion (T2TF) are important in distributed tracking systems. T2TF enables tracks from multiple sources to be fused in a close to optimal way. Nowadays it is also common that sensors contains some sort of tracking functionality which limits access to the “raw” sensor output. To integrate such a sensor in a tracking network of different sensors T2TF is necessary. In the literature, different methods for T2TF have been presented and analyzed. In this paper we implement and compare four different methods, with focus on robustness and tracking accuracy and how they perform against centralized measurement fusion (CMF), which is the optimal choice if possible. The methods studied in this paper are: (i) naïve _information matrix fusion (naïve IMF); (ii) covariance inter-_ _section (CI) fusion [5, 11]; (iii) generalized information matrix_ _filter (GIMF) fusion [19], and (iv) safe fusion (SF) [9]. The_ authors have not manged to find any publication where SF (or equivalent methods), contrary to the other three methods, have been applied to the T2TF problems. However, the ellipsoidal _intersection (IE, [15, 17]) method, as mentioned below to_ produce the same result as SF, has been evaluated in other contexts, e.g., vehicle platooning in [16]. One of the contributions of this paper is hence to compare SF to other methods in the context of T2TF. To get optimal T2TF performance the cross-correlations between tracks must be taken into account [18], one unavoidable source of cross-correlations is the shared process noise in tracks that describe the same target. The naïve fusion method assumes that tracks are uncorrelated which leads to overconfident error covariance matrices and overuse of data. Both CI (and variations thereof, [3, 14]) and SF assumes that there exist an unknown cross-correlation between tracks and provides a conservative solution, which is a more sound approach than naïve fusion. Exact methods as those in [18], on the other hand, calculates the cross-correlations between tracks and attains optimal performance, with the drawback that it requires a lot of information to be transferred between sensors. The ellipsoidal intersection method combines the two approaches, approximating the cross-correlation with a worst case scenario and then compensate for it. Though, derived based on different fundamental principles, it turns out that SF and EI produces identical estimates. The GIMF method also uses an information theoretic approach to handle the crosscorrelation. It is known that information is additive, hence you can subtract information to avoid double counting data. In the real world, communication between nodes in a network is not ideal, especially in wireless network configurations with limited communication rate and possible communication delays. Such issues are a current focus in T2TF community. Here, we consider communication to be mostly synchronous, in the sense that data is current when fusing, but not necessarily at full rate. In some cases delayed measurements are also considered. In recent research issues regarding communication rate and transmission load are highlighted to pursue CMF performance. For example the distributed version of the augmented state _density (ASD) filter called DASD [13] and the recent devel-_ oped distributed Kalman filter (DKF) [7]. In [6], the two previous methods are compared in terms of fusion performance, process noise sensitivity, and level of global knowledge about sensor parameters. The main contribution of this paper is to compare the SF algorithm to other T2TF methods in a security setting, this way bringing the attention to the SF method. To evaluate the chosen T2TF methods we have picked three different datasets to accentuate the differences between the fusion methods. The first dataset is based on data used in previous T2TF evaluations [19]; the second dataset consists of recorded target trajectories [20]; and the third dataset is data from a real world field trial suitable for target tracking in surveillance applications highlighting security problems. The paper is organized as follows. In Sec. II the general notation and description of methods are presented. How the evaluation of the methods was performed is presented in Sec. III and the results and discussion in Sec. IV. Sec. V ----- _x1(tf |t2)_ _x2(tf |t1)_ _xf (t0|t0)_ _xf (tf_ 1|t1) _xf (tf |tf )_ _x2(tf_ 1|t1) _x2(tf |t2)_ _x2(t0|t0)_ _z2(t1)_ _z2(t2)_ Fig. 1. Sampling sequence for asynchronous fusion of two sensors. concludes the paper. II. METHOD DESCRIPTION It is assumed that bandwidth restrictions in combination with communication delays leads to an architecture where not every sample is communicated between the sensor locations. Consider a generic asynchronous case of decentralized fusion, as depicted in Fig. 1. We have two sensors and a fusion center that could be co-located with one of the sensors (or both if fully decentralized). At time t2 we have new information from the first sensor z1(t2) which is brought forward (under communication delays) to the fusion center as the prediction _x1(tf_ _|t2) and since we assume asynchronous reports we have_ also the current and predicted values for the the second sensor; _i.e. information at the time of the most current sample and_ at the previous communication time x2(tf _|t2) and x2(tf_ _|t1),_ respectively. To keep the notation minimal we will in the following drop the argument for the most recent estimates from the two sensors and only keep it for the previous information for the second sensor. We assume that the estimates can be represented by their first two moments thus {xˆ1, P1}, _{xˆ2, P2}, {xˆf_ _, Pf_ _} and {xˆ2(tf_ _|t1), P2(tf_ _|t1)} are the notation_ for the variables considered in the track fusion steps. In the following sections brief descriptions of the studied methods are given, for more information the reader is referred to the provided main references of the methods. All methods are assumed to have some sort of fusion memory, i.e. a state that keeps the predicted first two moments from the last fusion at the fusion center. For GIMF this is already included, but also for CIF and SF this memory is used to enable the use of asynchronous sensor information. _A. Local and Centralized Filters_ As reference a centralized extended Kalman filter (EKF) [12] is run on all measurements from the sensors. The EKF is also used as the local filter in the sensor nodes for most scenarios. For one scenario an interacting multiple models (IMM, [4]) filter is used with a bank of two EKFs tuned with a low and a high process noise level. The model transition probability matrix in the IMM is tuned for sojourn times of 30 s for the low noise and 8 s for the high noise case following the model in [1] giving the transition probability matrix: �0.9983 0.0017� Π = _._ (1) 0.0062 0.9938 _B. Generalized Information Matrix Filter_ The GIMF is a generalization of the information filter for asynchronous tracklets following [19]. At the time of fusion the previous information is redacted giving the resulting equations: _Pf[−][1]_ = P1−1 + [P2−1 − _P2(tf_ _|t1)−1]_ (2a) _Pf[−][1]xˆf = P1−1 ˆx1 + [P2−1 ˆx2 −_ _P2(tf_ _|t1)−1 ˆx2(tf_ _|t1)]. (2b)_ The decorrelation by removing the previous information can also be performed at the local estimates as in the channel filter [8]. _C. Covariance Intersection_ The CI fusion rule [11] was explicitly developed to handle the problem of fusion of two (Gaussian) estimates with unknown correlations. The fusion equations are _Pf[−][1]_ = ωP1−1 + (1 − _ω)P2−1_ (3a) _Pf[−][1]xˆf = ωP1−1 ˆx1 + (1 −_ _ω)P2−1 ˆx2,_ (3b) where _ω = arg min_ det(Pf ). (4) _ω_ The choice to use the determinant in the criteria can be seen as minimizing the Shannon information [10]. _D. Naïve Independence Assumption_ For comparison the result of applying naïve fusion, i.e., fusion of the track reports ignoring the correlation introduced by the common process noise in the target trajectory, are also presented. The corresponding fusion equations for the naïve information matrix filter (naïve IMF) are: _Pf[−][1]_ = P1−1 + P2−1 (5a) _Pf[−][1]xˆf = P1−1 ˆx1 + P2−1 ˆx2_ (5b) This filter is thus the information form of the Kalman filter [12] under the (for track to track fusion naïve) assumption of uncorrelated errors between the local tracks. _E. Safe Fusion_ Similar to the covariance intersection method, safe fusion (SF, [9]) avoids double counting information from two possibly dependent estimates by decoupling the components in the estimates and using the most informative one from each estimate. This is achieved by repeatedly applying singular _value decomposition (SVD). The largest ellipsoid method [3]_ utilizes an eigen vector basis factorization to obtain the same covariance matrix as SF, but differs in how the mean of the estimate is computed. Contrary to CI, SF has not been shown to provide consistent estimates. ----- _xˆ2_ _xˆ2_ _xˆ1_ _xˆ2_ _xˆ1_ _xˆ1_ |Col1|ˆx| |---|---| ||| ||| Fig. 2. Illustration of the important steps of safe fusion between the two possibly correlated estimates ˆx1 and ˆx2. This is illustrated in Fig. 2. To the very right in the figure, the two estimates, ˆx1 and ˆx2, to be fused are illustrated by their covariance ellipses. In step 1 of the algorithm, a linear transformation is applied (obtained from an SVD) to transform the covariance matrix of ˆx1 into a unit matrix (the middle of the illustration). The components in ˆx1 are now independent. In step 2, a rotation is applied (obtained by an SVD) to make the components in ˆx2 independent (the right part of the illustration). Note that the components in **_xˆ1 remain independent under the rotation as they are equally_** uncertain in all directions. Next the most informative estimate is used in each direction, resulting in the grayed ellipsoid to the right. It is allowed to treat the components independently as the different directions have been decoupled by the two transformations. Double counting information is hence avoided by using only information from one of the two estimates to be fused in each direction. Finally, the fused estimate is obtained by simply applying the inverse of the two transformations (not illustrated). For completeness, the algorithm is provided in Algorithm 1, explicitly stating how to derive the necessary transformations. It should be noted that SF can be implemented using standard linear algebra functions, without the requirement of the optimization step found in CI. Hence, as SF also does not need to store or compute correlations it can be efficiently implemented with predictable execution time. The interested reader is referred to [9] for details and a motivation. III. METHOD EVALUATION The methods will be evaluated through three different datasets. The first two datasets are chosen from literature to allow for comparisons. The first set is inspired by [19] to provide direct comparisons for the GIMF. The second set is an adaption of one of the air scenarios in [20] (for brevity called Blair in the figures) used as a ground person tracking scenario. Finally results from field trials of a security scenario are presented. For the second dataset permutations in parametrizations of the process models or the introduction of IMM models gives two additional scenarios for a total of 5 scenarios described below. **Algorithm 1 Safe Fusion [9]** Given two possibly correlated estimates of x, ˆx1 and ˆx2 such that P1 = cov(ˆx1), and P2 = cov(ˆx2): 1) Compute U1 and D1, using an SVD of the positive definite matrix P1, such that _P1 = U1D1U1[T]_ _[.]_ (6) 2) Similarly, derive U2 and D2 using an SVD, such that _D1[−][1][/][2]U1[T]_ _[P][2][U][1][D]1[−][1][/][2]_ = U2D2U2[T] _[.]_ (7) 3) Let _T = U2[T]_ _[D]1[−][1][/][2]U1[T]_ (8a) **_xˆ¯_** 1 = T ˆx1 **_xˆ¯_** 2 = T ˆx2, (8b) where by construction cov(x¯[ˆ] 1) = I and cov(x¯[ˆ] 2) = D2. 4) Select the most informative source for each component _i = 1, 2, . . ., dim(x), let_ [x¯[ˆ]]i = [x¯[ˆ] 1]i, [D]ii = 1 if [D2]ii 1, (9a) _≥_ [x¯[ˆ]]i = [x¯[ˆ] 2]i, [D]ii = D2[ii] if [D2]ii < 1. (9b) 5) The final estimate given by **_xˆf = T_** _[−][1]x¯[ˆ]_ (10a) _Pf = T_ _[−][1]D[−][1]T_ _[−][T]_ _._ (10b) _A. Scenario 1_ The first setup is basically the same as Scenario 3 in [19], but with the same sample time for both sensors and without initial delay. Thus the scenario consist of two sensors one at origin and one at (5000 m, 0 m). They sample the position with 2 s interval ([19] uses 2 and 2.5 s). The range-bearing uncertainty is modeled as white noise with standard deviation _σr = 10 m in range and σθ = 1[◦]_ in bearing. The fusion is only performed with a rate of 8 s and a delay of 8 s. The motion model used is a continuous white noise acceleration model (CWNA) with the process noise σw = 0.1 ms[−][3][/][2]. In the first scenario the process model used for the motion is the same as the one used in the filter, hence the centralized Kalman filter is optimal. However, for a typical security scenario the human motion model is more like the aircraft flight of [20]. _B. Scenario 2_ For the second scenario we modify flight Scenario 6 of [20] to a ground scenario by scaling both positions and velocities with 1/1000 giving a small velocity but reasonable motions for a ground scenario. (See Fig. 3.) The sensors are placed at (0 m, 5 m) and (0 m, 45 m), respectively, and sampled at 5 Hz. It runs without any delay but with fusion in a subsampled rate of 2.5 s. The motion model is still a CWNA with the process noise σw = 0.1 ms[−][3][/][2]. ----- 50 40 20 0 30 20 10 0 -20 -40 -20 0 20 40 60 S2 S1 -10 -20 X [m] -30 -10 0 10 20 30 40 50 X [m] Fig. 3. Trajectory of Scenario 6 in [20] adopted to a ground scenario. _C. Scenario 3_ The third scenario is identical to Scenario 2, except for an additional delay of 2.5 s. _D. Scenario 4_ In Scenario 4 we return to Scenario 2 but introduces the IMM filter for the local sensors where the high/low (H/L) process models are chosen as σwH = 4 × 0.1 ms[−][3][/][2] and _σwL = 0.1/4 ms[−][3][/][2], respectively._ _E. Scenario 5_ The final scenario illustrates a real world example of tracking a person in surveillance cameras. There are two cameras pointing at the same area from different angles where one person walks through as illustrated in Fig. 4. The person is assumed to move according to a constant velocity model with process noise σw = 0.2 ms[−][3][/][2]. The fusion runs at a rate of 0.5 s. _F. Evaluation_ Fig. 4. Map overview of Scenario 5 where the optimal track is marked as red. The two sensors, S1 and S2, are illustrated as black boxes. The field of view of each sensor is illustrated as lines for S1 and dashed lines for S2. The area considered for evaluation is the common area seen by both sensors. The target moves from left to right. 60 Sensor 1 EKF Sensor 2 EKF 50 GIMF CIF Safe fusion CKF 40 30 20 10 0 0 50 100 150 Time [s] Fig. 5. 100 Monte Carlo simulations on a continuous white noise acceleration scenario similar to [19]. The results are evaluated against ground truth when available. Monte Carlo simulations are performed with 100 samples for Scenario 1 and 40 samples for Scenarios 2–4. To allow easy comparison with [19] the solutions of Scenario 1 are tested for consistency by normalized state estimation error squared (NEES, [2]) NEES(t) = �x(t)−xˆf (t|t)�T P −f 1(t|t)�x(t)−xˆf (t|t)�, (11) from the optimal track, i.e. the CKF. The optimal track and each sensor track is generated using a multi-sensor-multi-target tracker which associates visual detections from each sensor to tracks in a world coordinate frame. In the project both performance measures were studied for all scenarios but for brevity only the most interesting plots are reproduced here. IV. RESULTS AND DISCUSSION evaluated as a mean over Monte Carlo evaluations. The performance is evaluated as root mean square error (RMSE) averaged over Monte Carlo evaluations in Scenarios 1–4. In Scenario 5 no ground truth is available so the results will be evaluated as the root mean square deviation (RMSD) Scenario 1 was chosen to relate the obtained results to the results in [19]. The result in Fig. 5 compare favorably with the results in [19], but note that the initialization here was not as advanced as in the initial reference, causing larger errors for the first two updates in our implementation. The SF filter gives comparable results with the GIMF while the CIF performance varies along the trajectory. In Fig. 6 the consistency of the fusion is tested by the NEES. Apart from the inconsistent initialization, both the GIMF and SF filter ----- Time [s] Fig. 6. Mean NEES of 100 Monte Carlo simulations on a continuous white noise acceleration scenario similar to [19]. perform consistently while as expected the CI show an overly conservative covariance. The grouping of estimates four by four is due to the reduced rate of information used for the fusion. New information only arrives every 8 s, i.e., every fourth sample. For the second dataset in Scenario 2–4 with the trajectory illustrated in Fig. 3 the true motion model is nonlinear and thus the baseline CKF is no longer the optimal filter but can only aspire to be the best linear filter. In this case the local filters will not necessarily provide consistent estimates at all times either. In Fig. 7 the performance for Scenario 2, without delay is illustrated. Here the SF performs better than the CKF that actually cannot be optimal for this scenario. Since the path is based on basically linear motion with maneuvers the use of a CWNA has to be a compromise between good tracking in the corners or along the straight lines. It is interesting to note that for the more linear parts of the paths the SF is actually better than the CKF. However in the corners when the local filters lag the SF has worse performance but still at the same level as the GIMF. When delays are introduced as in Fig. 8 in Scenario 3 the SF can no longer beat the CKF but actually reach the level of CKF on the straight lines. In the corners the decentralized filters give worse performance than the individual local filters this is probably due to the predictions of the local states being inconsistent due to small process variance in the linear model during corners. The naïve IMF actually has performance worse than the local filters in the corners. The results suggest that an IMM would actually perform better, hence in Fig. 9 IMM filters have been applied both on the local and central level. On the central level the IMM does not improve the situation as much as expected on the straight lines, probably due to relatively uncertain sensors. For the local filters, especially of sensor 1, an improvement can be seen which also translates to an improvement for the fused filters. Now at the update times the filters are better than the local estimates. Again the SF filter is better than the central IMM filter on the straight line parts. Using local IMM filters complicated the picture and no straightforward way to implement exact methods [18] was seen, but even the GIMF poses problems in which model to use for the predictions. In the results presented, the model with the large process noise was used to predict the previous information forward. Trials using smaller variance made the filter unstable during shifts in the modes between large and small process noise. Here the SF and CI filters were the only ones that were straight-forward to implement and since the SF filter is more consistent than the CI with a less conservative covariance it would be the better choice. In Fig. 10 the performance for Scenario 5 is illustrated with a fusion rate set to 2 Hz. Here the GIMF filter performs best, especially at the fusion points, with the smallest deviation from CKF. This was the expected result since GIMF is optimal at full rate. Here, both SF, CIF, and naïve IMF show almost equal performance. The naïve assumption shows almost no degradation in performance as in previous scenarios. This could be an effect of the simple target movements in the scenario. When the estimated error covariance is inspected, naïve IMF instead turns out to be overconfident contrary to the other methods, which is due to the information double counting. The scenario does not cover advanced target trajectories as in the previous scenarios, but it shows once again that SF is an applicable method for T2TF. V. CONCLUSION In this paper a less known alternative to the covariance _intersection (CI) method for fusion of correlated estimates,_ _safe fusion (SF), was evaluated using simulated and a real_ world experimental data. It provides a less conservative covariance than the CI method, hence it provides estimates with better consistency. In the scenarios of interest, motivated by camera surveillance scenarios, the SF performed well, on a level comparable to established methods such as generalized _information matrix fusion (GIMF). The algorithm works well_ in conjunction with local interacting multiple models (IMM) filters on level with the GIMF and where the exact methods [18] are intractable. SF can be implemented using standard linear algebra methods and has predictable computation time, making it an attractive alternative to the other described methods. When local IMM filters are used, careful considerations need to be taken in the choice of process model for the prediction used for the GIMF making the SF the more robust alternative. ACKNOWLEDGMENTS This paper was supported by research projects at the Swedish Defence Research Agency (FOI) funded by the Swedish Armed Forces. G. Hendeby was supported by The Swedish Research Council through their grant Scalable Kalman Filters. The authors would also like to thank the anonymous reviewers for insightful comments and for pointing ----- Fig. 7. Root mean square error for the trajectory of Fig. 3 to the ellipsoid intersection method and its relations to safe fusion. REFERENCES [1] Y. Bar-Shalom and H. Chen. Covariance reconstruction for track fusion with legacy track sources. Journal of Advances in Information Fusion, 2008. [2] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan. Estimation with applica_tions to tracking and navigation: theory algorithms and software. John_ Wiley & Sons, 2004. [3] A. R. Benaskeur. Consistent fusion of correlated data sources. In 28th _Annual Conference of the Industrial Electronics Society, volume 4, pages_ 2652–2656, Nov. 2002. [4] S. Blackman and R. Popoli. Modern tracking systems. Artech House, 1, 1999. 10[1] 10[0] Sensor 1 EKF 10[-1] Sensor 2 EKF GIMF CIF Safe fusion NaiveIMF CKF 10[-2] 0 50 100 150 200 Time [s] Fig. 8. Mean square error for delayed local EKF filters for the trajectory of Fig. 3. The legend is the same as in Fig. 7 [5] L. Chen, P. O. Arambel, and R. K. Mehra. Fusion under unknown correlation — covariance intersection as a special case. In Proceedings _of 7th IEEE International Conference on Information Fusion, pages_ 905–912, Annapolis, MD, USA, July 2002. [6] C.-Y. Chong, W. Koch, and F. Govaers. Comparsion of tracklet fusion and distributed kalman filter for track fusion. In Proceedings of 17th _IEEE International Conference on Information Fusion, 2014._ [7] F. Govaers and W. Koch. Distributed kalman filter fusion at arbitrary instants of time. In Proceedings of 13th IEEE International Conference _on Information Fusion, 2010._ [8] S. Grime and H. F. Durrant-Whyte. Data fusion in decentralized sensor networks. Control engineering practice, 2(5):849–863, 1994. [9] F. Gustafsson. Statistical Sensorfusion. Studentlitteratur, 2010. [10] M. B. Hurley. An information theoretic justification for covariance intersection and its generalization. In Proceedings of 7th IEEE International _Conference on Information Fusion, volume 1, Annapolis, MD, USA,_ ----- July 2002. [11] S. J. Julier and J. K. Uhlmann. A non-divergent estimation algorithm in the presence of unknown correlations. In Proceedings of American _Control Conference, pages 2369–2373, Albuquerque, NM, USA, June_ 1997. [12] T. Kailath, A. H. Sayed, and B. Hassibi. Linear Estimation. PrenticeHall, Inc, 2000. ISBN 0-13-022464-2. [13] W. Koch and F. Govaers. On decorrelated track-to-track fusion based on accumulated state densities. In Proceedings of 17th IEEE International _Conference on Information Fusion, 2014._ [14] W. Niehsen. Information fusion based on fast covariance intersection filtering. In Proceedings of 7th IEEE International Conference on _Information Fusion, volume 2, pages 901–904, Annapolis, MD, USA,_ July 2002. [15] B. Noack, J. Sijs, M. Reinhardt, and U. D. Hanebeck. Treatment of dependent information in multisensor Kalman filtering and data fusion. In H. Fourtati, editor, Multisensor Data Fusion: From Algorithms and 10[1] 10[0] Sensor 1 IMM Sensor 2 IMM 10[-1] GIMF CIF Safe fusion NaiveIMF CKF Central IMM 10[-2] 0 50 100 150 200 Time [s] Fig. 9. The results of applying IMM filters on the trajectory of Fig. 3 _Architectural Design to Applications, page 169–192. CRC Press, 2015._ [16] J. Sijs and M. Lazar. Emperical case-study of state fusion via ellipsoidal intersection. In Proceedings of 14th IEEE International Conference on _Information Fusion, Chicago, IL, July 2011._ [17] J. Sijs and M. Lazar. State fusion with unknown correlation: Ellipsoidal intersection. Automatica, 48:1847–1878, Aug. 2012. [18] X. Tian and Y. Bar-Shalom. Exact algorithms for four track-to-track fusion configurations: All you wanted to know but were afraid to ask. In Proceedings of 12th IEEE International Conference on Information _Fusion, pages 537–544, 2009._ [19] X. Tian and Y. Bar-Shalom. On algorithms for asynchronous track-totrack fusion. In Proceedings of 13th IEEE International Conference on _Information Fusion, pages 1–8, 2010._ [20] G. Watson and W. Blair. Benchmark problem for radar resource allocation and tracking maneuvering targets in the presence of ECM. Technical report, Technical Report NSWCDD/TR-96/10, 1996. Fig. 10. RMSD from the optimal generated CKF track for the trajectory in Fig. 4. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/MFI.2016.7849499?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/MFI.2016.7849499, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "http://liu.diva-portal.org/smash/get/diva2:972030/FULLTEXT01" }
2,016
[ "JournalArticle", "Conference" ]
true
2016-09-01T00:00:00
[]
7,584
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/012cfea30b807059314c2e98d96bdbd502bdf409
[]
0.879638
Self-Sovereign Identity Solution for Blockchain-Based Land Registry System: A Comparison
012cfea30b807059314c2e98d96bdbd502bdf409
Mobile Information Systems
[ { "authorId": "2065563657", "name": "Mohammed Shuaib" }, { "authorId": "2177762", "name": "N. H. Hassan" }, { "authorId": "30958007", "name": "S. Usman" }, { "authorId": "2068848160", "name": "Shadab Alam" }, { "authorId": "2147349647", "name": "Surbhi Bhatia" }, { "authorId": "73488595", "name": "Arwa A. Mashat" }, { "authorId": "1712222777", "name": "Adarsh Kumar" }, { "authorId": "1780130", "name": "M. Kumar" } ]
{ "alternate_issns": null, "alternate_names": [ "Mob Inf Syst" ], "alternate_urls": [ "https://www.iospress.nl/journal/mobile-information-systems/", "https://www.iospress.nl/html/1574017x.php", "http://content.iospress.com/journals/mobile-information-systems" ], "id": "6b6df2de-21bc-4137-9859-3fcef46f6a21", "issn": "1574-017X", "name": "Mobile Information Systems", "type": "journal", "url": "https://www.hindawi.com/journals/misy/" }
Providing an identity solution is essential for a reliable blockchain-based land registry system. A secure, privacy-preserving, and efficient identity solution is essential but challenging. This paper examines the current literature and provides a systematic literature review in three stages based on the three research questions (RQ) that show the assessment and interpretation process step by step. Based on the parameters and RQ specified in the research methodology section, a total of 43 primary articles have been selected from the 251 articles extracted from various scientific databases. The majority of these articles are concerned with evaluating the existing self-sovereign identity (SSI) solutions and their role in the blockchain-based land registry system to address the compliance issues in the existing SSI solutions with SSI principles and find the best possible SSI solution to address the identity problems in the land registry. The existing digital identity solutions cannot handle the requirements of the identity principle and are prone to various limitations like centralization and dependency on third parties that further augment the chance of security threats. SSI has been designed to overcome these limitations and provide a secure, reliable, and efficient identity solution that gives complete control to the users over their personal identity information (PII). This paper reviews the existing SSI solutions, evaluates them based on the SSI principles, and comes up with the best possible SSI solution for a blockchain-based land registry system. It further provides a detailed investigation of each SSI solution to present its functionalities and limitations for further improvement.
Hindawi Mobile Information Systems Volume 2022, Article ID 8930472, 17 pages [https://doi.org/10.1155/2022/8930472](https://doi.org/10.1155/2022/8930472) # Review Article Self-Sovereign Identity Solution for Blockchain-Based Land Registry System: A Comparison ## Mohammed Shuaib,[1,2] Noor Hafizah Hassan,[1] Sahnius Usman,[1] Shadab Alam,[2] Surbhi Bhatia,[3] Arwa Mashat,[4] Adarsh Kumar,[5] and Manoj Kumar 5 1Razak Faculty of Technology and Informatics (RFTI), Universiti Teknologi Malaysia (UTM), Kuala Lumpur, Malaysia 2College of Computer Science & IT, Jazan University, Saudi Arabia 3Department of Information Systems, College of Computer Science and Information, Technology, King Faisal University, Al Hasa, 36362, Saudi Arabia 4Faculty of Computing & Information Technology, King Abdulaziz University, P.O. Box 344, Rabigh 21911, Saudi Arabia 5Department of Systemics, School of Computer Science, University of Petroleum and Energy Studies, Dehradun, India Correspondence should be addressed to Adarsh Kumar; adarsh.kumar@ddn.upes.ac.in and Manoj Kumar; wss.manojkumar@gmail.com Received 20 January 2022; Accepted 17 March 2022; Published 4 April 2022 Academic Editor: Sebastian Podda [Copyright © 2022 Mohammed Shuaib et al. This is an open access article distributed under the Creative Commons Attribution](https://creativecommons.org/licenses/by/4.0/) [License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is](https://creativecommons.org/licenses/by/4.0/) properly cited. Providing an identity solution is essential for a reliable blockchain-based land registry system. A secure, privacy-preserving, and efficient identity solution is essential but challenging. This paper examines the current literature and provides a systematic literature review in three stages based on the three research questions (RQ) that show the assessment and interpretation process step by step. Based on the parameters and RQ specified in the research methodology section, a total of 43 primary articles have been selected from the 251 articles extracted from various scientific databases. The majority of these articles are concerned with evaluating the existing self-sovereign identity (SSI) solutions and their role in the blockchain-based land registry system to address the compliance issues in the existing SSI solutions with SSI principles and find the best possible SSI solution to address the identity problems in the land registry. The existing digital identity solutions cannot handle the requirements of the identity principle and are prone to various limitations like centralization and dependency on third parties that further augment the chance of security threats. SSI has been designed to overcome these limitations and provide a secure, reliable, and efficient identity solution that gives complete control to the users over their personal identity information (PII). This paper reviews the existing SSI solutions, evaluates them based on the SSI principles, and comes up with the best possible SSI solution for a blockchain-based land registry system. It further provides a detailed investigation of each SSI solution to present its functionalities and limitations for further improvement. ## 1. Introduction The land registry is an important economic pillar for any country in nation-building. Blockchain technology can improve the security and transparency in the land registry by recording land-related details on the blockchain. Blockchain technology also hastens property identification and enhances trust and accuracy in transactions by enabling digital monitoring by stakeholders. Through an increasingly digital world, robust, useful, and flexible digital identity management systems are critical to electronically identifying and authenticating ourselves and to know who we communicate. As per McKinsey, “Good Digital ID” contains a high level of digital channel protection, verification, and authenticated identity, specially created with the user consent [1]. In 2005, Cameron wrote “The Law of Identity as an Identity and Access Architect” at Microsoft Corporation [2]. This law consists of 7 principles that translate several guidelines on managing and disclosing a user’s identity and identifying various entities with different types of identification. These ----- 2 Mobile Information Systems principles describe digital identity systems’ success and failure. So digital identity solutions are needed to facilitate the users of the land registry system to initiate a transaction [3]. However, many researchers [4, 5] working in the field of applying digital identity solutions for blockchain-based land registry systems confirmed the issue of noncompliance with digital identity principles given by Cameron [2]. So while developing an identity solution for a blockchainbased land registry system, these issues need attention [6, 7]. A digital identity is a collection of credentials and identifiers expressed in an appropriate context, for instance, the name, ID, and other relevant attributes [8, 9]. Digital identity describes the attribute of an entity digitally in providing access to systems and application of identity management process [10]. Traditionally, digital identities are mediums to validate users at the workplace. Existing digital identities are controlled by identity providers, not by the users themselves. Identity providers have complete ownership over an individual’s identity, making it vulnerable to identity misuse. Identity owners often share their credentials for registering or accessing a service with no standard or guidelines on what data they need to share and store on the Internet. In addition, oversharing of data contributes to privacy issues for the identity owner [11]. Since the challenges of current digital identity are severe and damaging, a new concept of digital identity is required. That can offer users complete control over their identities, reduce management costs, increase efficiency, and improve overall online identity [12]. In [13], the author presented the privacy-preserving blockchain-based identity management system for remote healthcare. The author evaluated the proposed system on the parameters like transaction gas cost, transaction per second, number of blocks lost, and block propagation time. The developed identity system can be applied to cancer patients and can be further extended by integrating the blockchain with IPFS. Additionally, in [14], purpose the scheme of digital coupon and explained the desired properties and features in the couponing system, which can be utilized to identify the nonrepudiation property using malicious issuers. Further in [15], the author presented a privacypreserving blockchain architecture for IoT using Hierarchical Identity Based Encryption (HIBE) suitable for IoT devices and mobile edge and cloudlet environments. The presented architecture is evaluated in a simulation environment named Contiki OS. The presented architecture provides the confidentiality, integrity, and availability of the data for the mobile edge nodes. SSI provides a decentralized identity and fully controls their identity and personal data. It only shares the necessary information with a third party, known as selective disclosure [16]. Issuing identity credential built on the trusted network among two parties is the main objective of self-sovereign identity. Blockchain technology utilizes a distributed ledger to achieve consensus using a cryptographic protocol, fulfilling the requirement of providing a decentralized system in self-sovereign identity [17, 18]. While several blockchainbased SSI frameworks are available, no SSI model is available specifically for the land registry systems. The SSI used in the land registry will provide individuals with identities that can be used for communication with land management services. SSI can also allow individuals to create evidence of their property, such as a certified survey plan or a notarized declaration. SSI offers an opportunity to design a gradually more secure and trustworthy identity in lieu of a government-approved identity document by collecting certificates issued by reliable third parties, such as a land registry and financial institutions [19]. SSI can provide a framework for data transformation into credentials to use their verified location history from a mobile provider and land registry certificates to provide proof of ownership claim [20]. SSI may directly connect individuals to land plots and provide a mechanism for recording land claims and related data. An SSI holder can use a verifiable claim issued for land ownership to access other services such as banking, loans, and government benefits. Individuals could submit a digital title to obtain financial assistance or agricultural subsidies. A verifiable claim will be a permanent record by government authority acknowledging the rights of a property owner at a certain stage. If property certificates are lost, or the owners were relocated, the verifiable claim will remain [21]. SSI development is still at the initial stage. Many governments and enterprises are currently involved in developing SSI solutions that are mainly based on blockchain technology. Some of the prominent SSI solutions are Sovrin [22], UPort [23], Civic [24], Blockstack [25], Selfkey [26], and ShoCard [27]. These SSI solutions are being used in different domains. These SSI solutions should satisfy the principles of digital identity solutions given by Cameron [2]. In [2], Cameron looked at SSI solutions to figure out the cause of their failure and market adaptability. He also came up with a requirement to comply with the SSI principles for building a successful SSI solution [27]. So every SSI solution should comply with the SSI principles [28]. This study is aimed at identifying how the self-sovereign identity solves the issues of compliance with the digital identity principle in a blockchain-based land registry system. This paper tries to identify the role of SSI in a blockchainbased land registry system. It further aims to review the various SSI principles by different researchers and come up with an evaluation criterion to evaluate the existing SSI solutions. Finally, it evaluates the existing SSI solutions to identify the most suitable one for applying in the blockchain-based land registry system. Various classification of the principle of SSI is given by [29] [11]. However, none of these classifications is complete since several properties are still missing. However, it appears that some principles under one group can be irrelevant, as described in [29]. We identified the criteria based on the classifications given by [11, 30] to compare the SSI solutions in SSI compliance principles, which should be taken care of while designing an SSI-based identity model for a blockchain-based land registry. A systematic analytical study of existing SSI solutions has been conducted based on the defined SSI principles and finalized evaluation criteria. This article is divided into five sections. Section 2 provides a detailed background study. Section 3 presents the research methodology that includes identified research questions (RQ), data sources used, search mechanism, and ----- Mobile Information Systems 3 inclusion/exclusion criteria to shortlist the study sources. Section 4 presents a detailed analysis of the outcomes extracted from the literature based on each research question. Finally, Section 5 concludes the findings and reviews. ## 2. Background and Literature Review This section provides a detailed study of the background literature required for this study that includes concepts of selfsovereign identity (SSI), the role of SSI in information flow, blockchain technology, and its application in SSI and applications of SSI in the land registry system. 2.1. Concept of Self-Sovereign Identity. Self-sovereign identity (SSI) is a revolutionary way to address identity. In the early days, centralized organizations controlled digital identities, while in the real world, people stored their issued identity information in a decentralized manner using a physical wallet. SSI’s objective is to connect online identity systems to the actual world and give users control over their identities. In the actual world, after the birth of a child, identity credentials like birth certificates, identification numbers, etc., are provided by the government authorities [16]. The person utilizes these credentials on several occasions to identify themselves or establish a relationship throughout life. The self-sovereign identity is a well-developed concept in the academic and industry fields. However, there is still no consensus on its exact definition. Generally, the SSI is defined by considering the principles of self-sovereign by de Marneffe [31] and descriptions of identity by [32]. Self-sovereign identity is a digitalized form of personal features, details, and attributes. No entity can breach the right to choose a level of privacy or reputation of identity attribute. While working as an identity and access architect in Microsoft Corporation, Cameron wrote identity laws in 2005. The identity law [2] follows a distributed ledger [33], which first explains the concept of SSI [34]. Although Cameron was unaware of the advancement of distributed ledgers in the upcoming years, proposing the Microsoft Passport is an unnecessary reliance on a single organization without user control and can lead to identity failures. The necessity of user access, minimal disclosure, and a portable, interoperable structure is required. The first occurrence of sovereign identity happened in 2019 [35]. In 2016, Allen presented ten principles of the selfsovereign identity (SSI) [34], focusing upon identity laws by describing how identity could work, why systems and algorithms need to be transparent, and how is it permanent despite being portable and interoperable. The details required for the concept of self-sovereign identity were proposed by [36]. The definition provided by Abraham is congruent with the ten principles provided by Allen [34]. Abraham extends the control concept and adds, “All user identity information will be recorded for further authentication.” It is trade-of-security and privacy, which should be based on the chosen user. SSI is considered as a long-lasting identity possessed and controlled by the individual without any external authority sans the possibility of identity removal. It requires user consent for interoperability of user identity across several locations and ownership over the identity to provide user autonomy. SSI may prove to be the new normal in the evaluation of identity management. 2.2. Roles of Self-Sovereign Identity. The self-sovereign identity (SSI) environment structure is defined as a peer-to-peer model where the independent identity works as a peer and communicate with each other. Communication is done so that people and organizations can affirm the information from individuals by assigning claims or credentials [12, 16]. The significant elements in SSI are identity verifier, identity issuer, and credential issuer. The functions of each entity of SSI are represented in Figure 1. Figure 1 explains the roles of SSI in the credential flow order. The issuer provides the credential for making the statement as it is often given through off-chain. The credentials and self-attached data of the identity owner are available in the wallet. Issuers may withdraw credentials if requirements are not fulfilled. The identity owner stores the credentials provided in a digital wallet, which function as an agent in the SSI environment. The entire identity credential is held in a digital wallet as proof of verification displayed in a disclosed manner. The identity owner has complete control over data sharing and usage. The consent of the identity owner is required to access information for verifier services. Accessible records in public registries such as the issuer’s identification key, DID, are confirmed to ensure the actual issuer issues these credentials. When the identity owner’s information meets the criteria, access is given, where the presented credentials are checked without contacting an issuer. Similarly, offering alternative credentials like a student ID does not require the university’s permission in the actual world. The blockchain uses a distributed ledger technology which allows the creation of identity without a central authority where the ledger acts as a basis of trust. An essential feature of SSI is the backend data storage in off-ledger. Most DID methods use a public or private repository, such as a private database or IPFS (Interplanetary File System), to collect offledger information. IPFS generates content-based hashes using particular IPFS data. Wallet files are stored as a backup in the backend off-ledger, making it easy to recover if lost. 2.3. Blockchain Technology and Self-Sovereign Identity. Selfsovereign identity systems are based on blockchain technology. The blockchain is an evolving technology that uses cryptocurrency to provide a decentralized, open shared ledger [37] that can be used for electronic voting [38] land registration [39, 40]. It is evident that cryptocurrency is not the only feasible use case for blockchain [41, 42]. Blockchain technology is well placed due to its technical features in facilitating a notable change in digital identity [43]. The self-sovereign identity is based on the sharing and storing verifiable claims held in off-ledger [44]. The authenticity of these signed data objects is assured by storing a hash of the thing on a blockchain. Once subjects submit a verifiable claim to a relying party, the hash of the claim with available blockchain record can be compared and verified through an integrated signature where the relying party can quickly and precisely ascertain the claim’s validity. A blockchain provides a way to revoke or store an auditable record of consent behavior and maintain the security of data ----- 4 Mobile Information Systems Issuer Identify owner 3.Presents credential or creates proof 2. Stores credential in wallet Verifier SSI-Ledger Figure 1: Roles of SSI with information flow [12]. objects to assure the integrity of the data object. Blockchain is built on a decentralized public-key infrastructure and provides robust methods that can be used for encryption and authentication, apart from self-sovereign identity [45, 46]. Additionally, [47], blockchain offers several key features that have ample opportunities for identity systems, including immutability, usability, and low transaction cost [48, 49]. 2.4. Self-Sovereign Identity and Land Registry. The selfsovereign identity (SSI) fundamental application for the land registry is to provide individuals with identities so that they can be used for communication with land management services. There is no identification record for one billion people across the world. SSI offers an opportunity to design a gradually more secure and trustworthy identity in lieu of a government-approved identity document by collecting certificates issued by reliable third parties, such as a land registry and financial institutions [19]. In the absence of legal documentation, SSI can also allow individuals to create evidence of their property, such as a certified survey plan or a notarized declaration. SSI credentials are robust and should not be limited to the digital version of the traditional paper [50]. SSI can provide a framework for data transformation into credentials so that administrative agencies trust it. For example, a person can use their verified location history from a mobile provider and land registry certificates to provide proof of ownership claim [20]. In the absence of land registries, the self-sovereign identity may directly connect individuals to land plots and provide a mechanism for recording land claims and related data to access other services such as banking, loans, and government benefits. An SSI holder can use a verifiable claim issued for land ownership. Individuals could submit a digital title to obtain financial assistance or agricultural subsidies. A verifiable claim will be a permanent record by government authority acknowledging the rights of a property owner at a particular stage. In the case of property, if the certificate is lost or the owner relocated, and the verifiable claim will remain [51]. (i) User Control. Self-sovereign identity solutions using a cryptographic signature, pairwise connection, and digital identities provide the user with complete control over his identity information. The user or the groups will be attached to the assets through self-sovereign identity, which improves the functions and scope of the land registry. Moreover, verifying and exchanging identity information will evolve to provide validated credentials and manage the remaining registry components that do not benefit through Self-sovereign identity (ii) Facilitate Access to Finance. Self-sovereign identitybased land registers can also provide more detailed and trusted information about potential borrowers in developing countries. The financial-market specialists at the Inter-American Development Bank, Juan Antonio Ketterer, and Gabriela Andrade, acknowledged that transparent and more accurate asset registers as collateral could mitigate knowledge-related asymmetry constraints and provide financial access [52]. As shown in recent initiatives in the United States of America, The expansion of mobile assets can have a major impact on economic growth for small and medium scale enterprises [53] (iii) Efficiency in Real Estate Markets. To reduce the possibility of fraud in the real estate markets, a high degree of due diligence is required for the identity of the involved parties, leading to inefficiency and more transaction fees. A self-sovereign identity solution will securely associate the owner with its properties and legally bind the digital signature to provide trusted and transparent online working (iv) Land Ownership in Postconflict Situations. Legal reestablishment of land for refugees and internally displaced persons (IDPs) helps postconflict restoration. However, the restoration process is complicated as many refugees do not have any essential land records or fear consequences [54]. An SSI secures land ownership records and receives verifiable credentials from an NGO to help record a claim in lieu of a proper land registry [1] (v) Natural Disaster Resilience. Land ownership is important for preparing for disasters and can improve the restoration process. New programs for disaster preparedness use innovative ----- Mobile Information Systems 5 technologies. Nevertheless, a solution to SSI will give users a safer and more accessible tool to show their land ownership and submit a request for assistance and restoration grants. Decentralized record management will guarantee the preservation of land ownership records. The use of biometrics in SSI allows people to prove their identities and authorized services, even though documents are deleted or lost ## 3. Research Methodology This paper performs a systematic literature review to explore the latest state-of-the-art academic research on self-sovereign identities and blockchain. Additionally, to examine the role of self-sovereign identity in the land registry system. To have the most comprehensive coverage of all published literature, our systematic review methods were carefully planned using the guidelines of Kitchenham and Charters [55] to identify the need for review and create a review plan. Our systematic review method includes the research questions, data sources used for retrieving papers, search strategy, inclusion and exclusion criteria, and screening and final selection description are summarized in Figure 2. 3.1. Research Questions. The first stage of the systematic literature review was to identify research questions (RQs) for a detailed review of available topics. The main research question addressed in this study is as follows. (RQ): how to select the most appropriate self-sovereign identity for the blockchain-based land registry? To answer the main research question of this study, we outlined three guiding questions. RQ1: how self-sovereign identity solves the issue of noncompliance with digital identity principles in the blockchain-based land registry? RQ2: which criteria can be used to compare the most appropriate blockchain-based self-sovereign identity solution? RQ3: what is the evaluation result of various blockchainbased self-sovereign identity solutions? To address the above guiding questions, we used the guidelines given by Kitchenham and Charters for a systematic review [55] and the standard procedure for selecting the literature for our research. 3.2. Data Sources. In this systematic research, material collection was performed through various scholarly databases such as Scopus, Web of Science, ACM Digital Library, and IEEE Xplore to collect more articles. These databases were chosen as they contain peer-reviewed papers and enable logical expressions (keywords, names, and/or abstracts) to be searched. Grey’s literature, such as reports on government projects, working papers, and documents on assessment, was also included. The blockchain-based self-sovereign identity implementation subject is a new study area, and the various blockchain-based firms are currently working on it. Including grey literature extends state-of-the-art research sources by using a broader research source. Each selected database was checked separately by the specified search words, and the results were combined after removing duplicates using Mendeley software. Table 1 shows the number of articles generated by search string in each database. Some found publications are available in more than one database. The total number of articles with duplication is 251. 3.3. Search Strategy. The search strategy is carried out between 2008 and 2021. This systematic review study took the starting point from 2008 when the first actual research in the blockchain was published. The grey literature includes magazines, company whitepapers, and books. To identify different blockchain-based self-sovereign identity solutions, and to be as generic as possible, the search string used to retrieve the articles from databases is (“self-sovereign identity” AND “Blockchain”) OR (“self-sovereign identity” AND “identity management “) OR (“self-sovereign identity” AND “Blockchain” AND “identity management”). Also, semantic search words were identified in the fields of digital identity, and self-sovereign identity and blockchain are also searched in the databases. Moreover, our search string is restricted only to the article’s title, abstract, and subject terms. It was done to exclude irrelevant articles referencing the search words only in the body’s text. The next step was to search for all related papers. A final search was carried out on 17 November 2020, covering years from 2008 to 2022. The search consists of conferences, journals, workshops, government project reports, working papers, review documents, and book sections. The searched terms are “blockchain”, “land registry”, “Identity model”, and “Law of identity” to check the title, keywords, and abstracts of academic papers. Some research papers use real estate in place of land registry, so we have modified the search strategy and used only the real estate & blockchain keywords. Additionally, some researchers use identity management in place of the identity model. As a result, we finally decided to discover all papers based on strings (“land registry” AND “Blockchain” or “real estate” AND “Blockchain” or “Identity model” AND “Law of identity”, “identity principle” or “Identity management” AND “Law of identity”, “identity principle”). Table 2 displays the search string and the results from scholarly databases. 3.4. Inclusion/Exclusion Criteria. Not all of the articles found were important to the subject, and thus, the next step was to identify the article that satisfies the scope of our study. We have done this by specifying criteria for inclusion and exclusion, as seen in Table 3. These criteria are applied to all titles, abstracts, and keywords of the identified article to classify them according to the scope of our study. Titles and abstracts in some cases have not been all appropriate; therefore, the whole paper has been examined to ensure the compactness of criteria for inclusion and exclusion. 3.5. Screening and Final Selection. The initial screening process was carried out on collected papers to verify compliance with our scope of the study. In this Systematic Literature Review, 251 articles were collected mainly from the scholarly databases (grey literature has been omitted from the ----- 6 Mobile Information Systems Database searching Grey literature Record excluded based on criteria (n = 51) |Col1|Database searching|Col3| |---|---|---| ||Web of IEEE Xplore ACM DL Scopus (n = 69) (n = 41) science (n = 69) (n = 44) Total (n = 251)|| |||| ||Records imported into citation manager (n = 251)|| |||| |Records for screening (n = 65)|Col2| |---|---| ||| |Title/abstract screening (n = 214)|Col2| |---|---| ||| ||| |Full text article assessed for eligibility (n = 64)|Col2|Col3| |---|---|---| |||| |Number of articles selected (n = 31)||| |||| |Number of report selected (n = 12)|Col2| |---|---| ||| |Database searching|Grey lite| |---|---| |Web of IEEE Xplore ACM DL Scopus (n = 69) (n = 41) science (n = 69) (n = 44) Total (n = 251) Records imported into Duplication removed citation manager (n = 251) (n = 37) Articles excluded based on title (n = 88) Title/abstract screening (n = 214) Articles excluded based on Abstract (n = 62) Full text article Full text articles excluded assessed for eligibility (n = 33) (n = 64) Number of articles selected (n = 31)|Records retrieved from other sources (n = 65) Records for screening (n = 65) Number of report selected (n = 12)| |Total number of articles selected (n = 43):31 article and 12 report|| Figure 2: Procedural steps for the selection process. Table 1: Search string and results for scholarly databases. Database Scopus Web of Science ACM Digital Library IEEE Xplore (“self-sovereign identity” AND “Blockchain”) 48 19 19 25 (“self-sovereign identity” AND “identity management”) 27 14 11 18 (“self-sovereign identity” AND “Blockchain” AND “identity management”) 22 11 11 26 Total with duplicates 97 44 41 69 Table 2: Search terms and results from different scholarly databases. Search terms IEEE Xplore Scopus ACM Science direct Web of Science “Land registry” AND “Block chain” 7 28 19 36 14 “Real estate” AND “Blockchain” 20 77 67 77 33 “Identity model”, “Identity” AND “Law of identity”, “identity principle” 7 9 5 8 2 “Identity management” AND “Law of identity”, “identity principle” 6 21 8 22 11 Total with duplicates 40 135 99 143 60 Table 3: Inclusion/exclusion criteria. Inclusion criteria Exclusion criteria (i) Publication between 2008 and 2022 (ii) papers with research scope of blockchain technology and subscope—the application of that technology for the domain related to the self-sovereign identity, identity management (iii) original research paper instead of review/survey paper (i) Duplicate (ii) not English language paper (iii) papers that had some other meaning other than one relevant to the blockchain-based self-sovereign identity (iv) articles addressing technical aspects of blockchain technology ----- Mobile Information Systems 7 descriptive analyzes for conformity). The number of articles chosen as primary studies has been reduced to 214 after eliminating (37) duplicate papers, resulting in 214 articles. Subsequently, we read each publication’s titles, abstracts, and keywords to keep them relevant to the next stage of screening. We also carefully reviewed whether they are inside or outside the scope through the inclusion and exclusion criteria by reading the abstract, conclusion, and discussion sections. Eightyeight articles are excluded based on the title, and 62 articles are excluded based on the abstract. A limited number of publications passed the primary screening stage for many factors. Finally, the first screening of the article ended with 64 articles. In the final screening, the remaining 64 articles were read in detail, thereby removing the publications that have little significance to the scope of our study. Finally, 31 papers and 12 reports have been selected for our study. ## 4. Research Questions and Analysis This section is further divided into three subsections (A, B, and C). Section A presents the issues of noncompliance with identity principles in the blockchain-based land registry system and how SSI solves this issue. Section B describes the criteria for evaluating the blockchain-based SSI solutions. Section C shows evaluation results of various blockchainbased SSI solutions based on the defined criteria. 4.1. RQ1: How Self-Sovereign Identity Solves the Issue of Noncompliance with Digital Identity Principles in the Blockchain-Based Land Registry? Through an increasingly digital world, robust, useful, and flexible digital identity management systems are critical to electronically identifying and authenticating ourselves and to know who we communicate. As per McKinsey1, “Good Digital ID contains a high level of digital channel protection” verification and authenticated identity, specially created with the user consent [56]. It helps us to decide with whom and for what reasons we choose to exchange data to ensure user’s privacy and control of personal data. “This would” unlock value by encouraging inclusion, formalization, and digitalization. For instance: (i) 45% of females aged around 15+ in low-income countries lack ID, and only 30% of males do (ii) Digital ID could increase 3-13 percent of GDP in 2030 In 2005, Cameron wrote The Law of Identity as an Identity and Access Architect at Microsoft Corporation [2]. A basic definition of identity requires concepts that can be focused on the design of additional services by involved parties. The principles can also be used as a goal to build trust and interoperability between services in the environment. This law consists of 7 principles that translate several guidelines on managing and disclosing a user’s identity and identifying various entities with different types of identification. These principles describe digital identity systems’ success and failure. These are briefly explained below. (i) Law 1: User Control and Consent. “Identity systems only disclose user identification with user consent” (ii) Law 2: Minimum Disclosure. The most successful long-term solution is one that discloses the lowest quantity of information and limits its use (iii) Law 3: Justifiable Parties. Digital identity systems should be established to limit information disclosure to parties with the necessary, justifiable position in a particular identity relationship (iv) Law 4: Directed Identification. The universal identity scheme must recognize omnidirectional identifiers for public entities and unidirectional identifications for private entities, simplifying discovery and preventing unnecessary correlation disclosures (v) Law 5: Pluralism of Operators and Technology. The identity system should manage multiple identity technologies run by different providers and allow them to communicate (vi) Law 6: Human Integration. The human user must be represented as part of the distributed system that can be integrated into communication mechanisms between people and machines to safeguard from identity attacks (vii) Law 7: Consistent Experience across Contexts. A unifying identity metasystem must ensure that its users have a clear and consistent experience, enabling operators and technologies to differentiate between different contexts The explanation principles of digital identity are extensive. Some of these principles may be more specific. For example, the first concept can be divided into user control and consent. Some identity solutions may satisfy one but not the other. Given that there was no self-sovereign identification at the time of writing these principles. It was all the more remarkable to have the majority of principles adopted from “The Evolution of Digital Identity Concepts guiding principles” by Allen [34]. In a well-known post, “The Path to Self-Sovereign Identity,” Allen outlined SSI principles, including specific guidelines from other sources such as Cameron and the W3C Verifiable Statements Task Force [57]. These ten principles are taken from Allen’s paper [34] and serve as guidelines for SSI adaptation. (1) Control: Users Must Control Their Identities. The user is the ultimate authority of his identity, subject to well-understood and safe algorithms that ensure that the identity and its arguments remain valid. He should be able to identify, update, or even hide it. The user is free to pick actors or privacy as he wishes. The user does not regulate all identity claims: other users can make claims about a user, but they should not be central to its identity ----- 8 Mobile Information Systems (2) Access: Users Must Have Access to Their Own Data. A user must always be able to easily access and recover all the claims and other identification details. There must be no hidden data and no gatekeepers (3) Transparency: Transparent Systems and Algorithms. The systems for managing and running an identity network must be transparent in terms of their functioning, management, and updating. The algorithms should be open source, well-documented, and autonomous from any particular architecture (4) Persistence: Identities Must Be Long-Lived. The user can only remove identities. Claims can be updated and removed, but the identity that belongs to these claims should be long-lived. Identities can ideally remain permanently or probably as long as the consumer wants. Although private keys could have to be rotated and data need to be changed, the identity remains. In the rapidly evolving world of the Internet, this goal may not be entirely feasible, but identities at least remain until new identity systems outdate them (5) Portability: Identity Information and Services Must Be Transportable. A trusted third-party entity should not hold the identity. It should be transportable, although a trusted entity behaves in the customer’s best interests. Transportable identities ensure that the individual stays in charge of their identity, which can increase identity persistence over time (6) Interoperability: Identities Should Be Used as Widely as Possible. Identity is of little benefit if used only in small niches. A modern-day digital identity system aims to access identity information widely and across international borders to create global identities without relinquishing user control (7) Consent: Users Must Agree to the Use of Their Identity. Any identity system is designed to share identity and claims, and an interoperable system improves the number of shares occurring. However, data sharing must only occur with user consent. While other users such as an employer, credit office, or spouse can make claims, the user must also confirm consent (8) Existence: Users Must Have an Independent Existence. An SSI fundamentally depends on the ineffable “I” at the core of identity. It will never fully exist in digital form. It needs to be the self-supporting kernel to support this (9) Minimalization: Disclosure of Claims Must Be Minimized. It should include the least amount of data required to perform the task when sharing data. It is supported by selective disclosure and zeroknowledge proof. However, noncorruptibility is a difficult task. The best possible way to solve this is to use minimization to promote privacy (10) Protection: The Rights of Users Must Be Protected. If the identity network priorities vary from those of the rights of individuals, the network should commit to protecting the rights and freedom of users over the network SSI is considered as a long-lasting identity possessed and controlled by the individual without any external authority sans the possibility of identity removal. It requires user consent for interoperability of user identity across several locations and ownership over the identity to provide user autonomy. SSI may prove to be the new normal in the era of digital identity. The self-sovereign identity is a potential solution since it provides people, organizations, and companies sovereignty over their identifiers and full control on how and to whom information is shared or utilized. Only the necessary information will be revealed to third parties in what is known as selective disclosure [12, 16]. Issuing identity credential built on the trusted network among two parties is the main objective of self-sovereign identity. Through the use of an easy, automated process and standard format, SSI can create a convenient communication method. 4.2. RQ2: What Are the Criteria for Evaluating BlockchainBased Self-Sovereign Identity Solutions? 4.2.1. Related Work. The various evaluation criteria taken by multiple researchers to evaluate self-sovereign identity and comparative studies of blockchain-based self-sovereign solutions are discussed below. Cameron (2005) explained the seven laws discussed in the earlier section, where he outlined the strengths and weaknesses of digital identity concepts [2]. These laws are vital to prevent any repercussions where the laws of identity and the requirement of self-sovereign identity are described in detail. Certain blockchain-based solutions may not satisfy certain properties of self-sovereign identity. Based on these seven laws, Christopher (2016) outlined ten principles to consider when implementing SSI solutions [34]. In the self-sovereign identity solution, these principles are aimed at user control besides providing the differences between the seven laws. Stokkink and Pouwelse (2018) used these ten self-sovereign identity principles to test blockchainbased SSI solutions: Sovrin and uPort. They included an additional property in the evaluation list that involves claims to be provable [58]. The problem with the current identity solution is identified as the individual is not the real owner of their identity. Besides, this problem can be overcome with the growth of the SSI solution. A DNS-Idm blockchain-based identity management system is developed using a smart contract to improve protection and privacy features [59]. In [43], the author compares various blockchain identity management systems and identifies challenges like trust, security, and privacy issues. Also, he discussed various trust, security, and privacy-based schemes that can be utilized to improve the blockchain-based identity nmangennet system. Shuaib et al. ----- Mobile Information Systems 9 (2022) compare the identity model, namely, centralized, federated, user-centric, and SSI based on laws of digital identity and suggested the SSI solution to be used for blockchain-based land registry system [60, 61]. Finally, a comparison of the available SSI solution, i.e., uPort, Sovrin, and Shocard, is made with the developed DNS-Idm using security and privacy criteria like ownership, user control and consent, human integration, privacy-friendly, and directed precise identity. Dunphy and Petitcolas (2018) made a comparison between blockchain solutions that used SSI based on the seven laws of identity [62], where they used trusted, decentralized identity where identity proofing relies on trustable existing credentials. They concluded that the usability (human integration) feature needs further improvement [63]. Similarly, Panait et al. (2020) evaluated ten current blockchain identity management solutions using SSI focusing on the implementation of the platform and long-term validity [64]. He emphasizes the need to improve the cryptography and usability aspect of the SSI’s current identity management solution. On the other hand, Van Bokkem et al.(2019) evaluated the seven blockchain-based self-sovereign identity solutions based on the eleven identity principles outlined by [34] alongside the provable property notion [65]. In [66], a comparative study of the popular identity management system is done using SSI like ShoCard, UPort, and Sovrin based on the seven laws of identity by [62]. Liu et al. (2020) compared blockchain identity systems that use SSI, namely, uPort, Sovrin, and Shocard, based on aspects like control, security, and privacy. Liu et al. compared the existing blockchain-based self-sovereign identity system such as UPort, ShoCard, and Sovrin [43] using the principles of self-sovereign identity given by [34]. The three self-sovereign identity solutions, namely, Everest, Evernym, and uPort, are analyzed using the SSI principle using the desk research and interview with company blockchain experts [19]. As the “consent” principle in developing countries is difficult to adopt, so it has been removed with the “Inclusion” principle. 4.2.2. Our Evaluation Criteria. The SSI requires the basic principle of identity given by [2]. The principle of SSI in an article by Allen is examined that provides an additional view on the digital identity liked to the seven identity principles given by Cameron. The ten essential principles for SSI are portability, access, transparency, persistence, control, transparency, existence, interoperability, protection, and minimization [34]. A similar classification of principle for SSI is given in (Ferdous et al., 2019), containing three properties: acceptance, zero cost, and controllability. Further, these principles were classified in [29], where the SSI principle is divided into three main groups: controllability, security, and portability. Additionally, the seven principles, namely, availability, approval, tenacity, approval, authority, autonomy, and confidentiality, were used to compare SSI solutions [11]. None of these classifications is complete since several properties are still missing. However, it appears that some principles under one group can be irrelevant, as described in [29], where they highlighted that the principle of persistence and existence in the context of controllability are mis matched. This study introduces the principle of “Inclusion” and the elimination of “Existence,” which is essential for implementation in developing countries. The “usability” principle was also incorporated in the assessment model, as customer service’s role is crucial in creating a better digital identity system. Therefore, a new taxonomy is categorized based on the classifications given by [11, 30] to compare the SSI solutions in SSI compliance principles. Figure 3 gives a mapping of principles of identity with the SSI principles. Based on all these classifications, new criteria for evaluating the SSI solutions have been proposed. The proposed principles to compare the SSI based solution in our study are described as follows: (1) Inclusion. Everyone possesses an individual identity and should have an identity from birth to death (2) User Control and Consent. Users must have ownership over their identity and can refer, update, trace, and access their personal data. Online data sharing of personal data should only be accomplished with user consent (3) Privacy and Protection. The user’s “right to privacy” should be secured on the protocol level (4) Portability. The identities should be available as long as the identity owner desires. The identity information will be portable, allowing users to access and control their identity, increasing identity persistence over time (5) Persistence. The identity system will be long-lasting, where identity owners can recover private keys and passwords if their primary device is damaged or stolen (6) Transparency. The system used to manage the identity network must be transparent in its processes, management, and updates (7) Interoperability. User identities are universally acceptable across various international boundaries and systems (8) Human Integration. The system interface meets the user’s needs where identity owners will add user experience in upcoming technology and services 4.3. RQ3: What Is the Evaluation Result of Various Blockchain-Based Self-Sovereign Identity Solutions? Secure user authentication and authorization are significant challenges for a reliable identity solution that needs to be addressed. SSI is a possible solution for resolving current identity models’ issues and providing permanent identity while providing full user control. Blockchain is an innovative technology to implement SSI solutions. The use of blockchain technology in the identity management system presents a possible solution for storing data on the blockchain. The stored data is secured using cryptographic tools and makes them immutable. The blockchain-based SSI solution foster trust among participants within the network without disclosing ----- 10 Mobile Information Systems Laws of identity Principle of self-sovereign identity (Camerons, 2005) (Christopher Allen, 2016) 1. User control & 1. Existence consent 2. Control 2. Minimal disclosure 3. Access 3. Justifiable 4. Transparency parties 5. Persistence 4. Directed entity 6. Portability 5. Pluralism of operators 7. Interoperability 6. Human 8. Consent integration 9. Minimalization 7. Consistent experience 10. Protection Figure 3: Mapping of principles of identity with the SSI principles. the actual data. Various blockchain-based SSI solutions will be discussed and compared in this section based on the criteria defined in Section 4.2.2. The comparative analysis of the existing SSI solutions has been given in Table 4. (1) Selfkey [26]. It has been created based on an SSI network where users can store data on personal devices [67]. Selfkey is a digital identity self-sovereign network [68] where user information is stored on a user-operated device, providing user ownership. If a third party needs to access identity data, the user will present information stored in the blockchain. SelfKey ensures that zero-knowledge proof gathers only a minimal amount of data, meeting acceptance and minimization requirements. It uses censorshipresistant and force-resistant algorithms to verify the identity where individuals can ascertain the identity claims of a customer. The portability in Selfkey is achieved using UPort. A significant weakness of Selfkey is a third-party dependency where no specific information about a trusted third party is available. Other inadequacies of Selfkey include lack of human integration and persistent identifier attributes that can last for a particular time [69] (2) Shocard [27]. The ShoCard offers a digital identity authentication platform designed based on the public blockchain. Identity owner authentication is achieved using a centralized database containing cryptographic hashes of digital identity users. The individual is responsible for initiating interaction with third parties to check identity. Data is ----- Mobile Information Systems 11 ----- 12 Mobile Information Systems ultimately stored in a protected data envelope that receivers can only decrypt. ShoCard was founded in 2015 and can include five million records within 30 minutes in a verified public blockchain [27, 63] It enables users and organizations to create identities secure and verified where end-users control personal information access and 3rd party sharing. The third party or ShoCard will not access the data without first sharing useful information. The blockchain network is used to store the identities, but it does not hold the user’s identity data. Additionally, ShoCard does not have decentralized login data storage, a target of hacking where the central servers are intermediaries among users and trusted third parties [63]. The partially centralized status of ShoCard creates instability in the existence of ShoCard ID. If ShoCard servers stop running, identity holders will not be able to use their own digital IDs and credentials [70]. Additionally, the cryptographic key management does not support users since ShoCard stores the identities on the public blockchain, which provides open access and transparency. Users will secure the private key in their personal device, where the service provider uses a public key to verify the ShoCard ID. Organizations may use a software development kit to integrate ShoCard technology with its current application or website. ShoCard supports multiple authentication and verification, such as KYC, encryption, traceable authorizing, and credential certification, besides offering an authentication mechanism using a phone app. The method of authentication involves downloading the application to establish its ShoCard ID. It requires a user to take a snapshot of a legitimate government-issued identity through which ShoCard gathers personal information. The user can then validate the details, create a password, or ask for a biometric scan. (3) Civic [24]. Civic is based on a blockchain-based identity authentication ecosystem where a third party wallet creates key pairs, storing identification information in the user’s computer. Civic and blockchain only accept data hashes stored on the Ethereum network as ERC20 Tokens. Civics support three independent groups in the network: consumers, validators, and service providers, based on the Ethereum blockchain and uses smart contracts to track the proof of attestation The Civic identity utilizes the validated identity for websites and mobile development without requiring the username and passwords for multiparty authentication. Users monitor their protected data and must only share information in which they are willing. The Civic app is used to store identity information on a mobile device in an encrypted form. The hash value of attached identity information is stored in the Merkel tree and collected in the blockchain. The Merkle tree sections can be exposed selectively, increasing user control by enabling identity owners to disclose personal details selectively. The Civic allows trustworthy identity authentication providers known as validators to par ticipate and sign transactions in public blockchain nodes. It reconfigures the centralization function and provides an interactive open system for the validator, but it is not entirely decentralized. Nevertheless, it has the same consensus mechanism as the Sovrin. The authenticator can revoke identity records. For instance, when a user changes their last name, the authenticating agency cancels the blockchain’s previous/invalid last name. Therefore, Civic users depend on authentication authorities to establish a protected digital identity, resulting in a lack of portability [71]. Civic is a transparent system that utilizes a permission-less blockchain and does not have software or infrastructure for its network [72]. The benefits of the Civic ecosystem include a strong relationship among financial institutions, public agencies, and utilities as it intends to build a market among banks, utility organizations, local, state or federal governments, etc., verifying individual or business identity attributes in a blockchain. The validators can price identity authentication and sell the identity to stakeholders using smart contracts. The Civic system remains effective, as it plays a vital role in its ecosystem and uses validators to verify identity data accessible through mobile apps. Civic also plans to launch the Civic wallet. By integrating identification with other applications, users can interact more securely and efficiently using standard cryptocurrency applications compared to other wallets. However, the development of this project is at an early stage. (4) Sovrin [27]. The Sovrin foundation started using blockchain to store distributed identities to formalize and create an SSI network. Theoretically, anyone can verify or issue the identity. The Sovrin is used to build identities, using centralized CA to create a trust model network, and using permissioned blockchain and Stewards nodes to achieve consensus. The Sovrin Foundation is a nonprofit organization with a board of twelve trustees, including the governance council Sovrin allows a user to have complete control over digital identities where the user can choose which information to be shared and with whom. This selective disclosure uses a unique technique, ZKPs. Additionally, Sovrin provides pair-named DIDs [73] and public keys to protect user privacy without compromising functionality. Since the Sovrin network only has central authority, users rely on agencies and stewards where trust and accountability are managed through stewards’ confidence, integrity, and noncollusion system. User data is stored in the user’s personal computer and cannot be stored in the network service provider’s database. Sovrin aims to establish a market for customers to incorporate data portability and restore private key loss using cryptographic accumulators. Semantic graphs, like JSON-LD, are often used to provide portability among providers. Sovrin protocol uses open-source software licenses built based on the Hyperledger Indy [74]. The Sovrin trust system will regulate the Sovrin network of digital identity, security, policies, and stewards [75]. Sovrin network contains stewards worldwide, including various financial institutions, start-ups, ----- Mobile Information Systems 13 charities, and authority for personal information. Sovrin foundation requires systems to comply with other digital identity systems where the user’s interaction is not clearly defined. Since Sovrin is in the early stage of development, developers and providers entering the identity ecosystem need to extensively discuss user experience [74]. (5) uPort [23]. The uPort uses an open ID system that enables customers to enrol their identities securely, sign transactions, send and request identity keys, as well as, accessing keys and data [76]. The identity owner appoints trustees to produce a public key through a controller for key recovery purposes. The controller consensus is achieved by replacing the missing public key when executing a proxy with a newly created key. Built using the Ethereum blockchain, the UPort connects attributes and stores them as a basic JSON structure [77]. The identity owner will obtain the ecosystem’s credentials without performing identity proofing when using the uPort framework. Users control UPortID and share personal information with third parties where users’ personal data is always available and stored on-chain or off-chain using IPFS. In uPort, the user has more responsibilities and authority over uPort IDs UPort identifiers can be created without disclosing personal information since the missing inherent connection between the UPort identities contributes to system robustness. The registry user’s JSON information is publicly available, which may violate user privacy. Users can claim ownership over uPort IDs without depending on a centralized entity. UPort also contains several centralized components such as transfer messaging service, push notification centre, and program manager attributes which are means for machine control or compliance. UPort allows users to store identity data, credentials, and keys in the self-sovereign wallet while the personal user key is stored on user devices. The key recovery protocol allows users a persistent digital identity in case of mobile loss or theft. The software also supports faster authentication singular-sign-on support for Dapps and other apps besides establishing a Decentralized Identity Foundation for a uniform user experience. Furthermore, The QR codescanning functionality allows communication with the other party [78]. Nevertheless, users consider UPort’s key protocol to recover and preserve personal data complex and lack comprehensibility [77]. (6) Blockstack [25]. Blockstack is a decentralized network of computers that handles identity and perhaps even users’ data. Blockstack ID is a decentralized user ID that connects decentralized applications (DApps). Blockstack public benefit corporation (PBC) is an open-source organization interested in developing core Blockstack protocols and applications [79] The application developed on Blockstack provides users control over their own identities and eliminates failure refer ence points. The user’s used data credentials cannot be stored at a centralized server where content sharing is carried out using encryption. However, the collection of profiles can be seen and tracked globally through a blockchain which may leak information and endangers users’ privacy. Blockstack business logic and data processing works on a computer rather than on centralized servers hosted by service providers. The decentralized storage current scheme, Gaia [80], ensures that users own and operate private data lockers. Cloud users may use these lockers as additional data storage platforms. In Blockstack, the key recovery protocol is unavailable; thus, users cannot reset their keys in the event of failure or stolen ID, thereby noncompliance with the persistence principle. Conceptually, Blockstack operates on the top of the Bitcoin network and is an open-source repository offering programming libraries on a variety of platforms. The portable nature of Blockstack allows developers to adapt and integrate other technologies. Blockstack involves a full-stack approach that provides all layers required to build decentralized applications besides allowing customers with a single username to operate across all applications without passwords. Nevertheless, the Blockstack environment is in its initial development stage and only offers desktop versions of the Blockstack browser [81]. (7) LifeID [82]. LifeID is an open digital identity platform that allows users to create a personal online identity. Users verify every online real-world transaction where authentication is required without third-party companies or government organizations. LifeID is often used combined with a biometric smartphone and app [83]. Only the user accepts the information request from third parties that need user consent. LifeID uses zero-knowledge proof where data is recorded on the user’s computer, and the necessary information is released whenever identity verification is needed. The LifeID Identity is backed up and recovered using three different methods: cold storage backup, trusted relatives or associates, and a reputable organization, combating theft by momentarily disabling or restoring identities (8) Evernym [84]. Evernym was established in 2013 by Jason Law, and Timothy Ruff is a well-known player and aims to facilitate SSI introduction within various industries [84]. Sovrin was explicitly designed for identity, and the company describes itself as the world’s first professionally authenticated and verifiable public service provider. The mobile application, Connect Me wallet, enables users to create private, peer-to-peer communication with other people. It also allows users to control digital keys and verifiable credentials of their digital identity Evernym achieves universal accessibility by using Sovrin to claim that SSI is a global public utility to meet everyone’s ----- 14 Mobile Information Systems Table 5: Comparative study of the Blockchain-based self-sovereign identity solutions. Blockchain-based self-sovereign identity solutions SSI principles (evaluation criteria) Sovrin ShoCard Selfkey uPort Civic Blockstack LifeID Evernym EverID Inclusion ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ User control and consent ✓ ✓ x ✓ ✓ ✓ ✓ x ✓ Privacy and protection ✓ x ✓ x ✓ x x ✓ x Portability ✓ x ✓ ✓ x ✓ ✓ ✓ ✓ Persistence ✓ x x ✓ x x ✓ ✓ x Transparency ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ x Interoperability ✓ ✓ ✓ ✓ ✓ ✓ ✓ x ✓ Human integration x ✓ x ✓ ✓ x x x ✓ identity needs. The firm will handle an identity on behalf of a vulnerable person or anyone else incapable of managing their digital wallet. Evernym can store all personal information on the customer’s smartphone while control in an Evernym solution is enabled by biometry, using the default biometrics on a particular device. The Evernym solution provides an easy way to import/import a private key and handle an SSI. An individual may usually import a private key into a digital wallet through a text file or QR code scanning. Using Sovrin, Evernym will have a concept of “guardian,” a trusted third party to protect an exposed individual’s identity. Evernym uses a hybrid open-source framework that provides access to a permission ledger where guardian organizations must behave according to the criteria set out in the Sovrin Trust Framework. The INF or Sovrin Foundation management and the secure implementation of blockchain may reduce the abuse of digital identity and personal identity information. Evernym observed that the Sovrin network architecture, management, and operation could provide members with the possible portability of their public and private data in compliance with other principles. Evernym connections within the Sovrin network will be connected by comparing a “fairly-pseudonymous identification,” or a single DID in each relation. The Evernym system is unable to provide flexibility which results in a lack of interoperability. Also, a small amount of information is available for the user control of the issuer’s credential [72]. In Evernym’s Connect. Me DApp, user biometrics is necessary to access a given identity and the related details in all situations. Individuals may also be expected to provide biometric information to establish peer-to-peer contact networks with other individuals and organizations in accepting credentials from an issuer to exchange credentials. (9) EverID [85]. EverID is a user-centric-based SSI and transitional solution built on blockchain [85]. The decentralized framework of EverID includes data, documents, and biometrics to store and validate user identities. EverID provides multiple third-party user verification and enables the secure transfer of value between network members [86]. The decentralized architecture provider ownership of personal data, which can be accessed only by the user. The individual’s personal details are stored so that the individual controls how with whom and for how long these details are shared (persistence). The EverID system is operated on a number of network supernodes. Such supernodes are the blockchain host Additionally, it hosts the bridge service to allow data transfer to an API server where SDK-enabled devices perform these transactions, making it portable. EverID differs from other approaches, as the user does not need a device because the digital computer identity (a combination of biometrics, government identification and third-party confirmations) is being saved on the cloud. However, EverID noncompliance with the minimization property as the data is required for a claim to be checked where the user must fully reveal it. For example, if the user is over 18, the user can choose to show his complete birthday or not. EverID is also not open-source; thus, the statements in the whitepapers cannot be provable. Its implementation details are also not available in the public domain, raising concerns about compliance with transparency [65]. Additionally, it hosts the bridge service to allow data transfer to an API server where SDK-enabled devices perform these transactions, making it portable. EverID differs from other approaches, as the user does not need a device because the digital computer identity (a combination of biometrics, government identification and third-party confirmations) is being saved on the cloud. However, EverID noncompliance with the minimization property as the data is required for a claim to be checked where the user must fully reveal it. For example, if the user is over 18, the user can choose to show his complete birthday or not. EverID is also not open-source; thus, the statements in the whitepapers cannot be provable. Its implementation details are also not available in the public domain, raising concerns about compliance with transparency [65]. ## 5. Discussion Based on the detailed analysis of available SSI solutions that can be used in the land registry environment, Table 5 provides a review of these selected SSI solutions. It shows that ----- Mobile Information Systems 15 ShoCard is not complying with the principle of privacy, portability, and persistence due to its partial dependence on a centralized server for attribute validation. Selfkey lacks user control and consent, which is a significant weakness, persistence, and human integration. Civic does not comply with portability and persistence due to its reliance on a third party. Evernym does not comply with the principles of user control and interoperability. EverID does not comply with the principles of privacy, persistence, and transparency. LifeID has a significant issue of privacy and security. Blockstack does not comply with privacy, persistence, and human integration principles. Among these available SSI solutions, Sovrin and UPort are the SSI models that comply with maximum SSI principles but noncompliance with human integration and privacy principles, respectively. The above assessment shows that none of the available SSI solutions fully comply with SSI principles. ## 6. Conclusion This paper highlights the limitations of existing identity solutions, advantages of SSI, and its application in the blockchain-based land registry system. This paper uses a systemic literature review (SLR) based on three defined research questions highlighting the role of SSI in solving the issue of noncompliance with identity principles, evaluation criteria for evaluating existing SSI solutions, and suggesting the best possible SSI solution in the case of Blockchain-based Land registry system. This SLR has selected 251 papers based on criteria and 65 articles from grey literature and finally used a total of 43 articles for review. A detailed study of SSI principles and evaluation criteria for existing SSI solutions have been defined. Based on the defined evaluation criteria, an extensive review of the existing SSI solutions has been done. This study highlights the strengths, limitations, and functioning of each SSI solution, and it concludes that none of the existing SSI solutions complies with all the SSI principles. Based on the defined evaluation mechanism, Sovrin is the best possible solution among the existing SSI solutions. It complies with most of the SSI principles but lacks the scale of human integration. It is the best possible SSI solution that can be applied in the case of a Blockchain-based land registry system. As the Sovrin lacks a human integration factor that is essential for ease of use and high adaptability, it provides a scope for further improvement and future research. ## Data Availability Data is available on reasonable request. ## Conflicts of Interest The authors declare that they have no conflicts of interest. ## References [1] J. Dempsey and M. Graglia, Case study: property rights and stability in Afghanistan, New America, 2017. [[2] K. Cameron, “The laws of identity,” August 2005, https://www](https://www.identityblog.com/stories/2005/05/13/TheLawsOfIdentity.pdf) [.identityblog.com/stories/2005/05/13/TheLawsOfIdentity.pdf.](https://www.identityblog.com/stories/2005/05/13/TheLawsOfIdentity.pdf) [3] K. Mintah, K. T. Baako, G. Kavaarpuo, and G. K. Otchere, “Skin lands in Ghana and application of blockchain technology for acquisition and title registration,” Journal of Property, Planning and Environmental Law, vol. 12, no. 2, pp. 147– 169, 2020. [4] B. Bundesverband, Blockchain Opportunities and Challenges of a New Digital Infrastructure for Germany, Blockchain Bundes[verband, Berlin,Germany, 2017, https://jolocom.io/wp-](https://jolocom.io/wp-content/uploads/2018/07/Blockchain-Opportunities-and-challenges-of-a-new-digital-infrastructure-for-Germany-_-Blockchain-Bundesverband-2018.pdf) [content/uploads/2018/07/Blockchain-Opportunities-and-](https://jolocom.io/wp-content/uploads/2018/07/Blockchain-Opportunities-and-challenges-of-a-new-digital-infrastructure-for-Germany-_-Blockchain-Bundesverband-2018.pdf) [challenges-of-a-new-digital-infrastructure-for-Germany-_-](https://jolocom.io/wp-content/uploads/2018/07/Blockchain-Opportunities-and-challenges-of-a-new-digital-infrastructure-for-Germany-_-Blockchain-Bundesverband-2018.pdf) [Blockchain-Bundesverband-2018.pdf.](https://jolocom.io/wp-content/uploads/2018/07/Blockchain-Opportunities-and-challenges-of-a-new-digital-infrastructure-for-Germany-_-Blockchain-Bundesverband-2018.pdf) [5] M. Kaczorowska, “Blockchain-based land registration: possibilities and challenges,” Masaryk University Journal of Law and Technology, vol. 13, no. 2, pp. 339–360, 2019. [6] N. Mehdi, “Blockchain: an emerging opportunity for sur[veyors?,” 2020, https://www.rics.org/globalassets/blockchain_](https://www.rics.org/globalassets/blockchain_insight-paper.pdf) [insight-paper.pdf.](https://www.rics.org/globalassets/blockchain_insight-paper.pdf) [7] G. Sylvester, E-Agriculture in action: blockchain for agriculture opportunities and challanges, Food and agriculture organization of the united nations and the international telecommunication union, Bangkok, 2018th edition, 2019. [8] J. Torres, M. Nogueira, and G. Pujolle, “A survey on identity management for the future network,” IEEE Communication Surveys and Tutorials, vol. 15, no. 2, pp. 787–802, 2013. [9] H. Ning, X. Liu, X. Ye, J. He, W. Zhang, and M. Daneshmand, “Edge computing-based ID and nID combined identification and resolution scheme in IoT,” IEEE Internet of Things Journal, vol. 6, no. 4, pp. 6811–6821, 2019. [10] M. Laurent, J. Denouël, C. Levallois-Barth, and P. Waelbroeck, “Digital identity,” Digital Identity Management, pp. 1–45, 2015. [11] M. A. Bouras, Q. Lu, F. Zhang, Y. Wan, T. Zhang, and H. Ning, “Distributed ledger technology for eHealth identity privacy: state of the art and future perspective,” Sensors, vol. 20, no. 2, p. 483, 2020. [12] M. Schaffner, Analysis and Evaluation of Blockchain-Based Self-Sovereign Identity Systems, Technical University of Munich, Munich, Germany, 2020. [13] I. T. Javed, F. Alharbi, B. Bellaj, T. Margaria, N. Crespi, and K. N. Qureshi, “Health-ID: a blockchain-based decentralized identity management for remote healthcare,” Health, vol. 9, no. 6, p. 712, 2021. [14] A. S. Podda and L. Pompianu, “An overview of blockchainbased systems and smart contracts for digital coupons,” in ICSEW'20: Proceedings of the IEEE/ACM 42nd International Conference on Software Engineering Workshops, pp. 770–778, Jun 2020. [15] D. Pavithran, J. N. Al-Karaki, and K. Shaalan, “Edge-based blockchain architecture for event-driven IoT using hierarchical identity based encryption,” Information Processing and Management, vol. 58, no. 3, p. 102528, 2021. [16] Y. Panfil and C. Mellon, The credential highway: how selfsovereign identity unlocks property rights for the bottom billion, New America Weekly, 2019. [17] K. Panetta, 5 trends drive the gartner hype cycle for emerging [technologies, 2020, Gartner, 2020, https://www.gartner.com/](https://www.gartner.com/smarterwithgartner/5-trends-drive-the-gartner-hype-cycle-for-emerging-technologies-2020/) [smarterwithgartner/5-trends-drive-the-gartner-hype-cycle-](https://www.gartner.com/smarterwithgartner/5-trends-drive-the-gartner-hype-cycle-for-emerging-technologies-2020/) [for-emerging-technologies-2020/.](https://www.gartner.com/smarterwithgartner/5-trends-drive-the-gartner-hype-cycle-for-emerging-technologies-2020/) [18] M. Van Wingerde, Blockchain-Enabled Self-Sovereign Identity, Tilburg University, Tilburg, Netherland, 2017. ----- 16 Mobile Information Systems [19] M. Graglia, C. Mellon, and T. Robustelli, “The nail finds a hammer self-sovereign identity, design principles, and property rights in the developing world,” New America Weekly, 2018. [20] serkan senturk, Future of property rights: self-sovereign identity and property rights, New America Weekly,, America, 2019. [21] Q. Shang and A. Price, “A blockchain-based land titling project for the Republic of Georgia,” Innovations, vol. 12, no. 3/ 4, pp. 72–78, 2019. [22] Sovrin Foundation, Sovrin™: A Protocol and Token for SelfSovereign Identity and Decentralized Trust, Sovrin Foundation, 2018. [23] C. Lundkvist, R. Heck, J. Torstensson, and Z. Mitton, Uport: A Platform for Self-Sovereign Identity, ConsenSys, 2016. [[24] Civic Technologies Inc, “Civic whitepaper,” 2017, https://](https://tokensale.civic.com/CivicTokenSaleWhitePaper.pdf) [tokensale.civic.com/CivicTokenSaleWhitePaper.pdf..](https://tokensale.civic.com/CivicTokenSaleWhitePaper.pdf) [25] M. Ali, R. Shea, J. Nelson, and M. J. Freedman, “Blockstack: a [new internet for decentralized applications,” 2017, http://](http://blockstack.org) [blockstack.org.](http://blockstack.org) [26] Self Key Foundation, “Self-sovereign identity for more free[dom and privacy-self key,” 2017, https://selfkey.org/.](https://selfkey.org/) [[27] S. Card, “Sho Card,” 2020, https://shocard.com/wp-content/](https://shocard.com/wp-content/uploads/2019/02/ShoCard-Whitepaper-2019.pdf) [uploads/2019/02/ShoCard-Whitepaper-2019.pdf.](https://shocard.com/wp-content/uploads/2019/02/ShoCard-Whitepaper-2019.pdf) [28] A. Nagy, K. A. Nyante, A. Peter, and Z. Hattyasy, Eds., Secure Identity Management on the Blockchain, University of Twente, 2018. [29] A. Tobin and D. Reed, The inevitable rise of self-sovereign identity A white paper from the Sovrin Foundation, Sovrin [Foundation, 2017, https://sovrin.org/wp-content/uploads/](https://sovrin.org/wp-content/uploads/2017/06/The-Inevitable-Rise-of-Self-Sovereign-Identity.pdf) [2017/06/The-Inevitable-Rise-of-Self-Sovereign-Identity.pdf.](https://sovrin.org/wp-content/uploads/2017/06/The-Inevitable-Rise-of-Self-Sovereign-Identity.pdf) [30] M. M. S. M. M. S. M. Ferdous, F. Chowdhury, and M. O. Alassafi, “In search of self-sovereign identity leveraging blockchain technology,” IEEE Access, vol. 7, pp. 103059– 103079, 2019. [31] P. de Marneffe, “Vice laws and self-sovereignty,” Criminal Law and Philosophy, vol. 7, no. 1, pp. 29–41, 2013. [32] M. H. Weik, Computer Science and Communications Dictionary, Springer US, Boston, MA, 2001. [33] C. S. Wright, “Bitcoin: a peer-to-peer electronic cash system,” SSRN Electronic Journal, 2008. [34] C. Allen, The path to self-sovereign identity, Coin Desk, 2016, [http://www.lifewithalacrity.com/2016/04/the-path-to-self-](http://www.lifewithalacrity.com/2016/04/the-path-to-self-soverereign-identity.html) [soverereign-identity.html.](http://www.lifewithalacrity.com/2016/04/the-path-to-self-soverereign-identity.html) [35] M. Marlinspike, What Is ‘Sovereign Source Authority’?, Moxy [Tougue, 2019, https://www.moxytongue.com/2012/02/what-](https://www.moxytongue.com/2012/02/what-is-sovereign-source-authority.html) [is-sovereign-source-authority.html.](https://www.moxytongue.com/2012/02/what-is-sovereign-source-authority.html) [36] A. Abraham, Whitepaper: self-sovereign identity, Graz, Austria, [2017http://www.egiz.gv.at.](http://www.egiz.gv.at) [37] Q. Stokkink, D. Epema, and J. Pouwelse, “A truly self-sovereign [identity system,” 2020, http://arxiv.org/abs/2007.00415.](http://arxiv.org/abs/2007.00415) [38] F. P. Hjalmarsson, G. K. Hreioarsson, M. Hamdaqa, and G. Hjalmtysson, “Blockchain-based E-voting system,” in 2018 IEEE 11th International Conference on Cloud Computing (CLOUD), pp. 983–986, San Francisco, CA, USA. [39] M. Shuaib, S. M. Daud, S. Alam, and W. Z. Khan, “Blockchainbased framework for secure and reliable land registry system,” TELKOMNIKA Telecommunication, Computing, Electronics and Control, vol. 18, no. 5, p. 2560, 2020. [40] M. Shuaib, S. Alam, and S. M. Daud, Improving the Authenticity of Real Estate Land Transaction Data Using BlockchainBased Security Scheme, Springer, Singapore, 2021. [41] R. Joosten, Self-sovereign identity framework and block[chain, ERCIM NEWS, 2017, https://ercim-news.ercim.eu/](https://ercim-news.ercim.eu/en110/special/self-sovereign-identity-framework-and-blockchain) [en110/special/self-sovereign-identity-framework-and-](https://ercim-news.ercim.eu/en110/special/self-sovereign-identity-framework-and-blockchain) [blockchain.](https://ercim-news.ercim.eu/en110/special/self-sovereign-identity-framework-and-blockchain) [42] Medici, “22 companies leveraging blockchain for identity management and authentication,” MEDICI, vol. 13, no. 21, pp. 2–5, 2017. [43] Y. Liu, D. He, M. S. Obaidat, N. Kumar, M. K. Khan, and K.K. Raymond Choo, “Blockchain-based identity management systems: a review,” Journal of Network and Computer Applications, vol. 166, p. 102731, 2020. [44] G. Kondova and J. Erbguth, “Self-sovereign identity on public blockchains and the GDPR,” in SAC '20: Proceedings of the 35th Annual ACM Symposium on Applied Computing, pp. 342–345, Brno, Czech Republic, Mar 2020. [45] R. Saia, S. Carta, D. Recupero, and G. Fenu, “Internet of entities (IoE): a Blockchain-based distributed paradigm for data exchange between wireless-based devices,” Proceedings of the 8th International Conference on Sensor Networks, pp., 201977–84, 2019. [46] S. T. Siddiqui, R. Ahmad, M. Shuaib, and S. Alam, “Blockchain security threats, attacks and countermeasures,” Adv. Intell. Syst. Comput., vol. 1097, pp. 51–62, 2020. [47] R. Jesse McWaters, “A blueprint for digital identity. The role of [financial institutions in building digital identity,” 2016, http://](http://www3.weforum.org/docs/WEF_A_Blueprint_for_Digital_Identity.pdf) [www3.weforum.org/docs/WEF_A_Blueprint_for_Digital_](http://www3.weforum.org/docs/WEF_A_Blueprint_for_Digital_Identity.pdf) [Identity.pdf.](http://www3.weforum.org/docs/WEF_A_Blueprint_for_Digital_Identity.pdf) [48] D. Sikeridis, A. Bidram, M. Devetsikiotis, and M. J. Reno, “A blockchain-based mechanism for secure data exchange in smart grid protection systems,” in 2020 IEEE 17th Annual Consumer Communications & Networking Conference (CCNC), pp. 1–6, Las Vegas, NV, USA, Jan 2020. [49] T. Mitani and A. Otsuka, “Traceability in permissioned blockchain,” IEEE Access, vol. 8, pp. 21573–21588, 2020. [50] M. Shuaib, S. Alam, M. Shahnawaz Nasir, and M. Shabbir Alam, “Immunity credentials using self-sovereign identity for combating COVID-19 pandemic,” Materials Today : Proceedings, 2021. [51] A. Piore, “Can blockchain finally give us the digital privacy we [deserve?,” 2019. https://www.newsweek.com/2019/03/08/can-](https://www.newsweek.com/2019/03/08/can-blockchain-finally-give-us-digital-privacy-we-deserve-1340689.html) [blockchain-finally-give-us-digital-privacy-we-deserve-](https://www.newsweek.com/2019/03/08/can-blockchain-finally-give-us-digital-privacy-we-deserve-1340689.html) [1340689.html.](https://www.newsweek.com/2019/03/08/can-blockchain-finally-give-us-digital-privacy-we-deserve-1340689.html) [52] J. A. Ketterer and G. Andrade, “Blockchain asset registries: [approaching enlightenment?- Coin Desk,” Dec 2017, https://](https://www.coindesk.com/blockchain-asset-registries-entering-slope-enlightenment) [www.coindesk.com/blockchain-asset-registries-entering-](https://www.coindesk.com/blockchain-asset-registries-entering-slope-enlightenment) [slope-enlightenment.](https://www.coindesk.com/blockchain-asset-registries-entering-slope-enlightenment) [53] International Finance Corporation, Secured Transactions Systems and Collateral Registries, World Bank, 2010. [54] M. Hendow, “Bridging refugee protection and development,” [2019, https://www.researchgate.net/publication/331530630_](https://www.researchgate.net/publication/331530630_Bridging_refugee_protection_and_development_Policy_Recommendations_for_Applying_a_Development-Displacement_Nexus_Approach) [Bridging_refugee_protection_and_development_Policy_](https://www.researchgate.net/publication/331530630_Bridging_refugee_protection_and_development_Policy_Recommendations_for_Applying_a_Development-Displacement_Nexus_Approach) [Recommendations_for_Applying_a_Development-](https://www.researchgate.net/publication/331530630_Bridging_refugee_protection_and_development_Policy_Recommendations_for_Applying_a_Development-Displacement_Nexus_Approach) [Displacement_Nexus_Approach.](https://www.researchgate.net/publication/331530630_Bridging_refugee_protection_and_development_Policy_Recommendations_for_Applying_a_Development-Displacement_Nexus_Approach) [55] B. Kitchenham and S. Charters, Guidelines for performing systematic literature reviews in software engineering, EBSE, Dur[ham, UK, 2007, https://www.elsevier.com/__data/promis_](https://www.elsevier.com/__data/promis_misc/525444systematicreviewsguide.pdf) [misc/525444systematicreviewsguide.pdf.](https://www.elsevier.com/__data/promis_misc/525444systematicreviewsguide.pdf) [56] McKinsey, Infographic: what is good digital ID?, McKinsey & [Company, 2019, https://www.mckinsey.com/business-](https://www.mckinsey.com/business-functions/mckinsey-digital/our-insights/infographic-what-is-good-digital-id) [functions/mckinsey-digital/our-insights/infographic-what-is-](https://www.mckinsey.com/business-functions/mckinsey-digital/our-insights/infographic-what-is-good-digital-id) [good-digital-id.](https://www.mckinsey.com/business-functions/mckinsey-digital/our-insights/infographic-what-is-good-digital-id) ----- Mobile Information Systems 17 [57] W3C Credentials Community Group, “Verifiable claims task [force,” May 2017, https://w3c.github.io/vctf/.](https://w3c.github.io/vctf/) [58] Q. Stokkink and J. Pouwelse, “Deployment of a blockchainbased self-sovereign identity,” in 2018 IEEE international conference on internet of things (iThings) and IEEE green computing and communications (green com) and IEEE cyber, physical and social computing (CPSCom) and IEEE smart data (smart data),, pp. 1336–1342, Halifax, NS, Canada, 2018. [59] J. Alsayed Kassem, S. Sayeed, H. Marco-Gisbert, Z. Pervez, and K. Dahal, “DNS-IdM: a blockchain identity management system to secure personal data sharing in a network,” Applied Sciences, vol. 9, no. 15, p. 2953, 2019. [60] M. Shuaib, N. Hafizah Hassan, S. Usman et al., “Identity model for blockchain-based land registry system: a comparison,” Wireless Communications and Mobile Computing, vol. 2022, Article ID 5670714, 17 pages, 2022. [61] M. Shuaib, S. Alam, M. Shabbir Alam, and M. Shahnawaz Nasir, “Self-sovereign identity for healthcare using blockchain,” Materials Today: Proceedings, 2021. [62] K. Cameron, The Laws of Identity, 2005. [63] P. Dunphy and F. A. P. Petitcolas, “A first look at identity management schemes on the blockchain,” IEEE Security and Privacy, vol. 16, no. 4, pp. 20–29, 2018. [64] A. E. Panait, R. F. Olimid, and A. Stefanescu, “Identity management on blockchain – privacy and security aspects,” [(2020), https://arxiv.org/abs/2004.13107.](https://arxiv.org/abs/2004.13107) [65] D. van Bokkem, R. Hageman, G. Koning, L. Nguyen, and N. Zarin, “Self-sovereign identity solutions: the necessity of [blockchain technology,” (2019), https://arxiv.org/abs/1904](https://arxiv.org/abs/1904.12816) [.12816.](https://arxiv.org/abs/1904.12816) [66] S. El Haddouti and M. D. Ech-Cherif El Kettani, “Analysis of identity management systems using blockchain technology,” in 2019 International Conference on Advanced Communication Technologies and Networking (Comm Net), pp. 1–7, Rabat, Morocco, Apr 2019. [[67] The Self Key Foundation, Self Key, Self key, 2017, https://](https://selfkey.org/) [selfkey.org/.](https://selfkey.org/) [68] S Foundation and Self Key Foundation, “Self key-the self key [foundation,” (2017), https://selfkey.org/wp-content/uploads/](https://selfkey.org/wp-content/uploads/2019/03/selfkey-whitepaper-en.pdf) [2019/03/selfkey-whitepaper-en.pdf.](https://selfkey.org/wp-content/uploads/2019/03/selfkey-whitepaper-en.pdf) [69] T. Koens and S. Meijer, “Matching identity management solutions to self-sovereign identity principles.pdf,” vol. 2, p. 32, 2018. [70] A.-E. Panait, R. F. Olimid, and A. Stefanescu, “Identity management on the blockchain,” in Proceedings of the Romanian Academy Series A-Mathematics Physics Technical Sciences Information Science, pp. 45–52, 2020. [71] M. Kuperberg, “Blockchain-based identity management: a survey from the enterprise and ecosystem perspective,” IEEE Transactions on Engineering Management, vol. 67, no. 4, pp. 1008–1027, 2020. [72] A. G. Nabi, Comparative Study on Identity Management Methods Using Blockchain, University of Zurich, 2017. [73] D. Reed, M. Sporny, D. Longley, C. Allen, R. Grant, and M. Sabadello, “Decentralized identifiers (DIDs): data model [and syntaxes for decentralized identifiers,” (2019), https://](https://w3c-ccg.github.io/did-spec/) [w3c-ccg.github.io/did-spec/.](https://w3c-ccg.github.io/did-spec/) [74] Hyperledger Contributors, Hyperledger Indy documentation, [Hyperledger, 2018, https://www.hyperledger.org/use/](https://www.hyperledger.org/use/hyperledger-indy) [hyperledger-indy.](https://www.hyperledger.org/use/hyperledger-indy) [75] M. Ali, R. Shea, and M. J. Freedman, Blockstack: a new decen[tralized internet, Whitepaper, 2017, http://blockstack.org.](http://blockstack.org) [76] uport, uPort, 2019. [77] C. Lundkvist, R. Heck, J. Torstensson, Z. Mitton, and M. Sena, Uport: A Platform for Self-Sovereign Identity, Sovrin Foundation, 2016. [78] N. Nizamuddin, H. R. Hasan, and K. Salah, “IPFS-blockchainbased authenticity of online publications,” in Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 10974 LNCS, pp. 199–212, 2018. [79] O. Labazova, T. Dehling, and A. Sunyaev, “From hype to reality: a taxonomy of blockchain applications,” Proceedings of the 52nd Hawaii International Conference on System Sciences, 2019, pp. 4555–4564, Grand Wailea, Maui, 2019. [80] vercel, A Decentralized Storage Architecture, Stacks docs, 2018, [https://docs.blockstack.org/storage/overview.html.](https://docs.blockstack.org/storage/overview.html) [81] M. Ali, J. Nelson, R. Shea, B. Labs, and M. J. Freedman, “Blockstack: a global naming and storage system secured by blockchains,” in usenix Annual technical conferencepp. 181–194, Denver, Colorado, 2016. [[82] I. D. Life, “Digital identity simple & secure,” (2018), https://](https://lifeid.io/) [lifeid.io/.](https://lifeid.io/) [83] L Foundation, “Life ID-an open-source, blockchain-based platform for self-sovereign identity,” p. 34, 2019. [[84] M. Nijjar, “Evernym,” (2019), https://www.evernym.com/.](https://www.evernym.com/) [85] B. Reid, B. Witteman, and W. Brad, “Ever ID whitepaper,” [May 2018, https://neironix.io/documents/whitepaper/6176/](https://neironix.io/documents/whitepaper/6176/EverID_Whitepaper_v1.0.2_July2018.pdf) [EverID_Whitepaper_v1.0.2_July2018.pdf.](https://neironix.io/documents/whitepaper/6176/EverID_Whitepaper_v1.0.2_July2018.pdf) [86] A. Aloraini and M. Hammoudeh, “A survey on data confidentiality and privacy in cloud computing,” in Proceedings of the International Conference on Future Networks and Distributed Systems, pp. 1–7, Cambridge United Kingdom, Jul. 17. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1155/2022/8930472?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1155/2022/8930472, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://downloads.hindawi.com/journals/misy/2022/8930472.pdf" }
2,022
[ "Review" ]
true
2022-04-04T00:00:00
[ { "paperId": "ac6f4643e97a50fea7520ee5ec5b396bdf8ba46b", "title": "The credential highway: how self-sovereign identity unlocks property rights for the bottom billion" }, { "paperId": "e63a9d5738e4c3f9d6ea71f290b8d72099bf8be7", "title": "Health-ID: A Blockchain-Based Decentralized Identity Management for Remote Healthcare" }, { "paperId": "9e5a1a024db905c47d89b60c526c8b43cd06aa20", "title": "Edge-Based Blockchain Architecture for Event-Driven IoT using Hierarchical Identity Based Encryption" }, { "paperId": "5a54034b2509ec4e32c822fc5df1f532f66e5e26", "title": "Immunity credentials using self-sovereign identity for combating COVID-19 pandemic" }, { "paperId": "4fe8862cbde84d4a91e968a7b75b5cea3f805bda", "title": "Self-sovereign identity for healthcare using blockchain" }, { "paperId": "17e4f79fa03cc0d9bfd5928a98c38f7f00f41d46", "title": "Blockchain-Based Identity Management: A Survey From the Enterprise and Ecosystem Perspective" }, { "paperId": "368f1d90ae7a5d92dfca814b84d54de8068b492a", "title": "Blockchain-based framework for secure and reliable land registry system" }, { "paperId": "10a65ae7d8ea1ffa96148e1d36c9a69aa7285b26", "title": "A Truly Self-Sovereign Identity System" }, { "paperId": "6105c1c53f866864fabfd27c98f03192da24dbee", "title": "An overview of blockchain-based systems and smart contracts for digital coupons" }, { "paperId": "1c5cc9f976fb58be1de112b95cffd81a743a1453", "title": "Skin lands in Ghana and application of blockchain technology for acquisition and title registration" }, { "paperId": "496a450d5921c13a369fb885c8c19400d5d27c00", "title": "Identity Management on Blockchain - Privacy and Security Aspects" }, { "paperId": "83906464c306290dfd399e744f0ff1888e7f245d", "title": "Self-sovereign identity on public blockchains and the GDPR" }, { "paperId": "d324c7d451061edb5f2d8e04afb051058bd5b4da", "title": "Distributed Ledger Technology for eHealth Identity Privacy: State of The Art and Future Perspective" }, { "paperId": "cfbffdbd298a57cd99c52fc8c308e3e6b2f5fce9", "title": "A blockchain-based mechanism for secure data exchange in smart grid protection systems" }, { "paperId": "084813a5b1faa4d66f9cb3818994b9e680be3a2c", "title": "Blockchain-based Land Registration: Possibilities and Challenges" }, { "paperId": "f74c0d01b0481ed21ec5e9dfc8f7a76d8e988a5d", "title": "In Search of Self-Sovereign Identity Leveraging Blockchain Technology" }, { "paperId": "e9747eace33602d3f4b5f5c0fd35e33c31b022cc", "title": "DNS-IdM: A Blockchain Identity Management System to Secure Personal Data Sharing in a Network" }, { "paperId": "9a94f839b960382f3c4dc7da4c50d17a5cb2d0fa", "title": "Traceability in Permissioned Blockchain" }, { "paperId": "77460667e38f21092b42addefcd4049a965dd559", "title": "Self-Sovereign Identity Solutions: The Necessity of Blockchain Technology" }, { "paperId": "632b76ec1d55f7b13dafe8b4ce7b34eaab96afb4", "title": "Edge Computing-Based ID and nID Combined Identification and Resolution Scheme in IoT" }, { "paperId": "621975a39f59d555fbe7f517cf1486479435edc9", "title": "Analysis of Identity Management Systems Using Blockchain Technology" }, { "paperId": "4964c80f15fc7347aa82481fd29e1b33983b6e9f", "title": "Internet of Entities (IoE): A Blockchain-based Distributed Paradigm for Data Exchange between Wireless-based Devices" }, { "paperId": "3c22ae83d1f6f029835f7a529a79ec3016592d4e", "title": "From Hype to Reality: A Taxonomy of Blockchain Applications" }, { "paperId": "80a80be07321fb6354b7cc64b4765d56a338e8c5", "title": "A Blockchain-Based Land Titling Project in the Republic of Georgia: Rebuilding Public Trust and Lessons for Future Pilot Projects" }, { "paperId": "54d50269928dafc6a0744e46044c17d973fdb01c", "title": "Blockchain-Based E-Voting System" }, { "paperId": "866306df18c325ccea7d9cfae027e407489f5e3b", "title": "IPFS-Blockchain-Based Authenticity of Online Publications" }, { "paperId": "9de7f1057a235570be559205ce5203f09e09af81", "title": "Deployment of a Blockchain-Based Self-Sovereign Identity" }, { "paperId": "bb46aeb975545b581c7301c786de3075bcceb471", "title": "Identity Management with Blockchain" }, { "paperId": "d4cf27fc7484eac5069d647a734d26df67b05c41", "title": "A First Look at Identity Management Schemes on the Blockchain" }, { "paperId": "2158248201cbd43a23fb5134208bcec4150e0c9a", "title": "A Survey on Data Confidentiality and Privacy in Cloud Computing" }, { "paperId": "1f42fdecd70a7d72f0f108e80511320f7204316c", "title": "Blockstack: A Global Naming and Storage System Secured by Blockchains" }, { "paperId": "042e3227bf4bb39d15aadcbce00d051663dea30e", "title": "A Survey on Identity Management for the Future Network" }, { "paperId": "55bdaa9d27ed595e2ccf34b3a7847020cc9c946c", "title": "Performing systematic literature reviews in software engineering" }, { "paperId": "c7916abb8b13c54ba532edd3edcde745f2d722f3", "title": "Digital identity" }, { "paperId": "ddde481484b674f451b3c5bdbda407e16a02503a", "title": "Open Source Biotechnology" }, { "paperId": "1bfd09796a2cc89b09053e79510c6de6b995dac1", "title": "Consent" }, { "paperId": "eddd1a2ff14e4a90ee138533e323d12f24e23284", "title": "International Finance Corporation" }, { "paperId": null, "title": "Identity model for blockchain-based land registry system: a comparison" }, { "paperId": "0e438c9f0686d9d56e2cb822cf77be0e0e7871da", "title": "Improving the Authenticity of Real Estate Land Transaction Data Using Blockchain-Based Security Scheme" }, { "paperId": "877db6de2b97d491482f118374eb9f8be8010697", "title": "Blockchain Security Threats, Attacks and Countermeasures" }, { "paperId": null, "title": "5 trends drive the gartner hype cycle for emerging technologies, 2020" }, { "paperId": null, "title": "“ Blockchain: an emerging opportunity for sur-veyors?, ”" }, { "paperId": null, "title": "uport" }, { "paperId": null, "title": "“ Decentralized identi fi ers (DIDs): data model and syntaxes for decentralized identi fi ers, ”" }, { "paperId": null, "title": "Infographic: what is good digital ID?" }, { "paperId": null, "title": "Future of property rights: self-sovereign identity and property rights" }, { "paperId": null, "title": "E-Agriculture in action: blockchain for agriculture opportunities and challanges, Food and agriculture organization of the united nations and the international telecommunication union" }, { "paperId": "f5a77c5a4dd43e76ea11dd7f92793a98aff24153", "title": "Secure Identity Management on the Blockchain" }, { "paperId": null, "title": "2018 IEEE international conference on internet of things (iThings) and IEEE green computing and communications (green com) and IEEE cyber, physical and social computing (CPSCom) and IEEE smart data" }, { "paperId": null, "title": "Hyperledger Contributors" }, { "paperId": null, "title": "“ The nail fi nds a hammer self-sovereign identity, design principles, and property rights in the developing world, ”" }, { "paperId": null, "title": "A Decentralized Storage Architecture" }, { "paperId": null, "title": "Matching identity management solutions to self-sovereign identity principles.pdf" }, { "paperId": null, "title": "Sovrin™: A Protocol and Token for Self-Sovereign Identity and Decentralized Trust" }, { "paperId": "03a396becd6c730c6142204e9429ce4503649bf7", "title": "The Path to Self-Sovereign Identity" }, { "paperId": "8dc1c83c80e628a6ba6d1dfd3f22972c82ede0c5", "title": "Blockstack : A New Internet for Decentralized Applications" }, { "paperId": "5817acc7ab5181c78893283ce1dc48b2294781b1", "title": "Self-Sovereign Identity Framework and Blockchain" }, { "paperId": "606b2c57cfed7328dedf88556ac657e9e1608311", "title": "Blockstack : A New Decentralized Internet" }, { "paperId": null, "title": "Civic Technologies Inc" }, { "paperId": null, "title": "The inevitable rise of self-sovereign identity A white paper from the Sovrin Foundation, Sovrin Foundation" }, { "paperId": null, "title": "22 companies leveraging blockchain for identity management and authentication" }, { "paperId": null, "title": "Blockchain Opportunities and Challenges of a New Digital Infrastructure for Germany" }, { "paperId": null, "title": "The Self Key Foundation" }, { "paperId": null, "title": "Wingerde, Blockchain-Enabled Self-Sovereign Identity, Tilburg University, Tilburg, Netherland, 2017" }, { "paperId": null, "title": "Graglia, Case study: property rights and stability in Afghanistan" }, { "paperId": "2a4c9e9e396f2be969b3af6bf8fa23afc5d017ae", "title": "Vice Laws and Self-Sovereignty" }, { "paperId": null, "title": "What Is ‘Sovereign Source Authority’" }, { "paperId": "213475e88c019a2bb8afd673f84d0a62285efbea", "title": "Secured transactions systems and collateral registries" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": "3b46aef1526699c2f79ab5fb13a0a2e4afde5c48", "title": "The Laws of Identity" }, { "paperId": "9f54677fc97e2d51a7e4f6c9c783dc53055811c8", "title": "Privacy" }, { "paperId": "f0f05e01c828536afa2c3f2fd0a7bb3b3df8c00a", "title": "Computer Science and Communications Dictionary" }, { "paperId": "3240ee72999cb51b82b5820c36cae18cbc171d9a", "title": "Mobile information systems" }, { "paperId": "667cc1ac858c058b6f55797d427bb5174f712dba", "title": "Journal of Network and Computer Applications" }, { "paperId": null, "title": "Sho Card" }, { "paperId": null, "title": "Facilitate Access to Finance . Self-sovereign identity-based land registers can also provide more detailed and trusted information about potential borrowers in developing countries" }, { "paperId": null, "title": "Existence: Users Must Have an Independent Exis-tence . An SSI fundamentally depends on the ine ff able “ I ” at the core of identity" }, { "paperId": null, "title": "Transparency . The system used to manage the identity network must be transparent in its processes, management, and updates" }, { "paperId": null, "title": "v) Natural Disaster Resilience . Land ownership is important for preparing for disasters and can improve the restoration process. New programs for disaster preparedness" }, { "paperId": null, "title": "Blockchain asset registries: approaching enlightenment?- Coin Desk" }, { "paperId": null, "title": "improve protection and privacy features [ author compares various blockchain identity systems and identi fi es challenges like trust, privacy issues" }, { "paperId": null, "title": "Bridging refugee protection and development" }, { "paperId": null, "title": "Persistence . The identity system will be long-lasting, where identity owners can recover private keys and passwords if their primary device is damaged or stolen" }, { "paperId": null, "title": "Portability . The identities should be available as long as the identity owner desires" }, { "paperId": null, "title": "Persistence: Identities Must Be Long-Lived" }, { "paperId": null, "title": "Minimalization" }, { "paperId": null, "title": "Interoperabilit y" }, { "paperId": null, "title": "Evernym" }, { "paperId": null, "title": "Ever ID whitepaper" }, { "paperId": null, "title": "Can blockchain finally give us the digital privacy we deserve?" }, { "paperId": null, "title": "Verifiable claims task force" } ]
22,070
en
[ { "category": "Business", "source": "external" }, { "category": "Medicine", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/012dbb6ffa748446949f36b1e1551bc593a440bb
[ "Business" ]
0.847333
Critical Dimensions of Blockchain Technology Implementation in the Healthcare Industry: An Integrated Systems Management Approach
012dbb6ffa748446949f36b1e1551bc593a440bb
Sustainability
[ { "authorId": "30797251", "name": "S. Aich" }, { "authorId": "13379591", "name": "S. Tripathy" }, { "authorId": "66938976", "name": "Moon-Il Joo" }, { "authorId": "40478718", "name": "Hee-Cheol Kim" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "http://mdpi.com/journal/sustainability", "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-172127" ], "id": "8775599f-4f9a-45f0-900e-7f4de68e6843", "issn": "2071-1050", "name": "Sustainability", "type": "journal", "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-172127" }
In the digital era, almost every system is connected to a digital platform to enhance efficiency. Although life is thus improved, security issues remain important, especially in the healthcare sector. The privacy and security of healthcare records is paramount; data leakage is socially unacceptable. Therefore, technology that protects data but does not compromise efficiency is essential. Blockchain technology has gained increasing attention as it ensures transparency, trust, privacy, and security. However, the critical factors affecting efficiency require further study. Here, we define the critical factors that affect blockchain implementation in the healthcare industry. We extracted such factors from the literature and from experts, then used interpretive structural modeling to define the interrelationships among these factors and classify them according to driving and dependence forces. This identified key drivers of the desired objectives. Regulatory clarity and governance (F2), immature technology (F3), high investment cost (F6), blockchain developers (F9), and trust among stakeholders (F12) are key factors to consider when seeking to implement blockchain technology in healthcare. Our analysis will allow managers to understand the requirements for successful implementation.
## sustainability _Article_ # Critical Dimensions of Blockchain Technology Implementation in the Healthcare Industry: An Integrated Systems Management Approach **Satyabrata Aich** **[1]** **, Sushanta Tripathy** **[2], Moon-Il Joo** **[1]** **and Hee-Cheol Kim** **[3,]*** 1 Institute of Digital Anti-Aging Healthcare, Inje University, Gimhae 50834, Korea; satyabrataaich@gmail.com (S.A.); joomi@inje.ac.kr (M.-I.J.) 2 School of Mechanical Engineering, KIIT Deemed to be University, Bhubaneswar 751024, Odisha, India; sushant.tripathy@gmail.com 3 College of AI Convergence/Institute of Digital Anti-aging Healthcare/u-AHRC, Inje University, Gimhae 50834, Korea ***** Correspondence: heeki@inje.ac.kr; Tel.: +82-55-320-3720 [����������](https://www.mdpi.com/article/10.3390/su13095269?type=check_update&version=1) **�������** **Citation: Aich, S.; Tripathy, S.; Joo,** M.-I.; Kim, H.-C. Critical Dimensions of Blockchain Technology Implementation in the Healthcare Industry: An Integrated Systems Management Approach. Sustainability **[2021, 13, 5269. https://doi.org/](https://doi.org/10.3390/su13095269)** [10.3390/su13095269](https://doi.org/10.3390/su13095269) Academic Editor: Nicu Bizon Received: 28 February 2021 Accepted: 15 April 2021 Published: 8 May 2021 **Publisher’s Note: MDPI stays neutral** with regard to jurisdictional claims in published maps and institutional affil iations. **Copyright: © 2021 by the authors.** Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **Abstract: In the digital era, almost every system is connected to a digital platform to enhance** efficiency. Although life is thus improved, security issues remain important, especially in the healthcare sector. The privacy and security of healthcare records is paramount; data leakage is socially unacceptable. Therefore, technology that protects data but does not compromise efficiency is essential. Blockchain technology has gained increasing attention as it ensures transparency, trust, privacy, and security. However, the critical factors affecting efficiency require further study. Here, we define the critical factors that affect blockchain implementation in the healthcare industry. We extracted such factors from the literature and from experts, then used interpretive structural modeling to define the interrelationships among these factors and classify them according to driving and dependence forces. This identified key drivers of the desired objectives. Regulatory clarity and governance (F2), immature technology (F3), high investment cost (F6), blockchain developers (F9), and trust among stakeholders (F12) are key factors to consider when seeking to implement blockchain technology in healthcare. Our analysis will allow managers to understand the requirements for successful implementation. **Keywords: blockchain; healthcare; critical factors; digital healthcare; interpretive structural modeling** **1. Introduction** Recently, blockchain (BC) technology has attracted increasing attention from industry and academia. BC technology allows users to preserve, certify, and synchronize the contents of a transaction ledger, which are available to multiple users. Transactions are decentralized; the data are not controlled by a third party. Within the system, transactions are timestamped in a ledger; data modifications/alterations are generally impossible without changing the ledger. Figure 1 shows the key components of a BC. BC technology ensures that trust and security are maintained during any transaction [1,2]. The healthcare, financial, and educational industries perceive the advantages afforded. Figure 2 describes the working principles of a BC. As BC technology reduces fraudulent activity and protects privacy, healthcare providers would like to implement it [3]. Breaches of healthcare data are increasing rapidly; in 2017, the number of people affected exceeded 300 records; from 2010 to 2017, this number rose to 37 million records [4,5]. There are growing concerns regarding healthcare data sharing, secure data storage, and data ownership, as digitization becomes the norm [3]. A BC ensures transparency, security, and speed during data storage and distribution; it also solves the security, privacy, and integrity issues that arise in the field of healthcare technology [6–9]. A BC is decentralized, thus eliminating the accuracy and security concerns associated ----- _Sustainability 2021, 13, 5269_ 2 of 17 with dependence on a central authority. BC technology is inter-operator-based, ensuring _Sustainability Sustainability 20212021,, 1313, x FOR PEER REVIEW, x FOR PEER REVIEW_ 2 of 17 2 of 17 a high standard of data exchange among healthcare associates. This boosts innovation, coordination among associates, market competition, and care quality [10–13]. **Figure 1.Figure 1. Figure 1. Components of a BC.Components of a BC. Components of a BC.** **Figure 2.Figure 2. Figure 2. Principles of a BC: a step-by-step flowchart.Principles of a BC: a step-by-step flowchart. Principles of a BC: a step-by-step flowchart.** In the past (when no BC was available), healthcare data interoperability among As BC technology reduces fraudulent activity and protects privacy, healthcare pro-As BC technology reduces fraudulent activity and protects privacy, healthcare pro different institutions was categorized as push, pull, and view. In the push model, data viders would like to implement it [3]. Breaches of healthcare data are increasing rapidly; viders would like to implement it [3]. Breaches of healthcare data are increasing rapidly; transfer is possible between two providers; no third provider has access. For example, data in 2017, the number of people affected exceeded 300 records; from 2010 to 2017, this num-in 2017, the number of people affected exceeded 300 records; from 2010 to 2017, this num transfer is possible between departments within the same hospital; however, data cannot ber rose to 37 million records [4,5]. There are growing concerns regarding healthcare data ber rose to 37 million records [4,5]. There are growing concerns regarding healthcare data be accessed by a different hospital, regardless of patient transfer to the other hospital. In sharing, secure data storage, and data ownership, as digitization becomes the norm [3]. A sharing, secure data storage, and data ownership, as digitization becomes the norm [3]. A the push model, it is very difficult to ensure data integrity during transfer. During a pull, one provider informally seeks data from another provider; there is no standardized auditBC ensures transparency, security, and speed during data storage and distribution; it also BC ensures transparency, security, and speed during data storage and distribution; it also trail. For example, an orthopedic surgeon can informally ask a cardiologist for information.solves the security, privacy, and integrity issues that arise in the field of healthcare tech-solves the security, privacy, and integrity issues that arise in the field of healthcare techDuring a view, one provider sees the record of another provider. For example, a surgeonnology [6nology [6––9]. A BC is decentralized, thus eliminating the accuracy and security concerns 9]. A BC is decentralized, thus eliminating the accuracy and security concerns can access an X-ray taken in the emergency department. The security approaches areassociated with dependence on a central authority. BC technology is inter-operator-based, associated with dependence on a central authority. BC technology is inter-operator-based, not based on the relationship that exists between a patient and a provider; thus, they areensuring a high standard of data exchange among healthcare associates. This boosts inno-ensuring a high standard of data exchange among healthcare associates. This boosts inno vation, coordination among associates, market competition, and care quality [10vation coordination among associates market competition and care quality [10 13]–13] ----- _Sustainability 2021, 13, 5269_ 3 of 17 largely ad hoc. The relevant policies are also subject to the laws of the local and federal governments. A BC-based model for a healthcare market creates a new dimension by considering the safety of data integrity and the use of standardized formal contracts for data accession. When an electronic health record (EHR) (which stores data from multiple workers) is accessed, it is difficult to determine the identity of the person who performed a task and when the work was performed. BC timestamps all work and identify the worker; the data are also distributed to all participating nodes. If a modification or update appears in any node, this is distributed to all nodes and is thus visible systemwide. Data integrity is maintained without the need for human intervention [14]. Although BC affords many benefits, it has never been implemented in real-time healthcare. Adoption is inevitable. Our literature review revealed only limited empirical evidence for BC use, despite its many possible benefits [15]. Very few studies have investigated the benefits, deficits, and functionalities of BC technology [16–19]. Most studies have sought to explain how BC works and to determine its current real-world implementation status [20]. However, critical factors affecting BC implementation in healthcare have not been addressed; knowledge of these factors is essential. Thus, we sought to identify these factors. Our findings can remove the confusion associated with real-time BC implementation. We offer a better understanding of the challenges imposed by implementation of BC technology in healthcare and the factors affecting such implementation. Our objectives are: (1) To identify factors that critically impact the implementation of BC technology in the healthcare industry; (2) To build a structured framework that depicts the interrelationships among such factors; (3) To define the motivation and reliance powers of such factors. Based on past works and the opinions of experts in BC technology, we define 13 factors that greatly affect the implementation of such technology in healthcare. We used interpretive structural modeling (ISM) to explore the relationships among such factors. We performed Matrice d’Impact Croise’s Multiplication Appliquée a UN Classement (MICMAC) analysis to define the motivation and reliance powers of the factors. We sought to encourage industries that wish to implement BC technology. The remainder of this paper is organized as follows. The literature regarding applications of BC technology in healthcare is reviewed in Section 2. Section 3 describes the methods used to achieve our research objectives. The research approach is discussed in Section 4. Managerial implications are discussed in Section 5. Practical implications are discussed in Section 6. The outcomes are summarized and conclusions are drawn in Section 7. **2. Related Works** This section is divided into two subsections. Past works regarding applications of BC in healthcare are covered in the first subsection. The second subsection discusses critical factors influencing BC implementation in healthcare. _2.1. Past Studies Regarding Applications of BC in Healthcare_ BC use in healthcare scenarios has focused on smart healthcare management, useroriented medical research, and prevention of drug counterfeiting. In terms of healthcare management, health networks allow medical experts to obtain detailed information regarding current patient status (described in the reports of physicians or healthcare centers, as well as various studies). Analysis of medical records creates an ecosystem that transparently reduces the merit costs of patient records. Moreover, medical experts can monitor the treatment activities of stakeholders, such as physicians and healthcare centers. These systems facilitate insurance claim settling if insurance companies are permitted (by patients) to access data. As healthcare records are increasingly stored digitally, security elements must be incorporated in such digital systems. Any digital platform must be scalable and ----- _Sustainability 2021, 13, 5269_ 4 of 17 adaptable, thus capable of handling large numbers of records and adaptable to many types of changes [21]. One practical solution is Live Interactive FramE (LIFE), which ensures that all media streaming in a healthcare domain are appropriately secured with minimal video quality loss during immersive applications [22]. In the context of user-oriented medical research, several authors have focused on the workings and structures of health banks; these studies have led to major breakthroughs in medical research. A company may accumulate data from wearable devices and may provide a user platform for data storage and management. The stored medical data facilitate high-quality research by trusted organizations. Patients are not financially compensated for their data. In terms of preventing drug counterfeiting, various authors have shown that the typical counterfeiting level today ranges from approximately 10% to 30% in developed countries. Counterfeiting is not confined to lifestyle medicines or drugs; it can include drugs that treat major medical issues, such as cardiovascular diseases. Recently, the Hyperledger research network described how drug counterfeiting could be reduced by timestamping. This problem could be addressed using a BC: all data have an address based on the stored information, thereby preventing drug counterfeiting [23]. Other authors [24] identified a simple but very robust BC system for patient data storage by the healthcare sector. The system encourages the storage of all data (beginning at birth), including lipid profile data yielded by wearable devices and magnetic resonance imaging information. All information is stored in a data lake, which facilitates simple querying, advanced analytics, and machine learning. This is a simple form of data warehousing; the stored materials include documents, images, .pdf files, and key values. The BC of each user serves as an index catalog containing a unique identification number, an encrypted link to the health record stored in the lake, and a timestamp that shows all data modifications. The user enjoys robust access control and can allow or restrict access using an audit log that shows every visit to the data repository. Such systems will greatly aid medical professionals (ranging from students to doctors) and governmental agencies. The data shared from the lake are tagged with the unique patient identification numbers, are up-to-date and accurate, and can be used for longitudinal healthcare research. Electronic records contain data from wearable devices, medical records, and/or medical imaging; all are private. Cloud storage in a central regulatory database may be associated with data leakage; unauthorized data transfer can be avoided by enabling a BC architecture wherein a user or a patient has full control over data access by third parties (expert consultants or medical researchers). BC use in healthcare will aid in the development of a consensus system that verifies the appropriateness of prescribed medication [25]. As healthcare data are very sensitive, the type of BC to be used is critical. One report [26] discussed the various types of BC available, their bases, and their uses in terms of maintenance, validation, and storage of medical data. Out-of-the-box consortium BCs were considered optimal; both the node owner and miners control access. A consensus can be determined regarding the optimal number of validations imparted by a healthcare provider, a clinic, or an insurance company; the stored data are thus very accurate. Although BC implementation in the healthcare sector is vital, this has not yet happened. It is essential to define factors that would aid implementation. This is the topic of the next subsection. _2.2. Crucial Factors Affecting Implementation of BC in Healthcare_ Here, we review works regarding crucial factors that affect BC implementation in healthcare. Ekblaw et al. [14] regarded data security and privacy assurance as a major concern. Although BC data are safely stored, Shi et al. [27] presumed that healthcare data in a public database would be vulnerable to statistical attack. Thus, frequent encryption key changes would be required, increasing key storage costs. To address this problem, Zhao et al. [28] developed a smart key recovery scheme based on a body sensor network. ----- _Sustainability 2021, 13, 5269_ 5 of 17 In this system, there is no need for a stored encryption key. If a key must be recovered, it is retrieved by the body sensor network. Khan et al. [29] stated that medical data interoperability was another key barrier; medical data must be communicated, shared, and used by employing various information communication technologies. Although creation of data interoperability has been discussed by several authors [11,30], its implementation remains a major concern. Agarwal and Prabakaran [31] presumed that data fragmentation could compromise EHRs. Thus, there is a need to meticulously maintain EHRs [32]. Errors compromise diagnosis and patient safety [33]. Data unavailability is unacceptable, thus posing a problem when implementing BC technology. The compatibility of existing information with BC technology is also critical; incompatibility reduces system performance. Incompatible information technology interfaces [23] greatly slow the system. The deployment and maintenance of dedicated technology are essential components [34]. Regulation plays a major role in healthcare; lives are at stake [1]. Schwerin [35] questioned the compatibility of BC with general data protection regulations. BC must protect consumer rights according to law [36]. Trust is a critical consideration when using BC; data are decentralized, transparent, and uploaded [37]. Illegal bitcoin activities have compromised trust in the technology [38]. However, in the healthcare sector, the technology can never be completely trusted due to the demand for healthcare data privacy [39–42]. Regulators have stated that a key issue preventing large-scale implementation of BC technology is the lack of standardization. Only a few well-known infrastructures are available [38]. Different countries seek to apply BC in various manners. For example, Estonia wishes to use BC to award e-residency to new citizens [43]. Other countries seek to use BC for transparent taxation and other social needs [44]. Standardization is urgently required [45]. The creativity and intricacy of BC increase adoption costs. Furthermore, there are few service providers, which leads to increasing costs [38]. As BC implementation reduces transaction costs [46], the initial costs may be worthwhile. However, the technology may not be suitable for small- and medium-size industries that lack workers with skills to manage the system [38]. There is a need for more skilled operators. BC became well-known due to bitcoin, but disagreements are increasing in the bitcoin community due to frequent code forking, supporting the notion that the technology is immature [47]. Many technical limitations remain [38]. Scalability is essential for robust system performance. In its present form, the network of a public BC is comparatively expensive and slow and thus difficult to adopt at a large scale. However, new scaling techniques such as plasma chains, lightning networks, zksnarks, and state channels enhance performance and scalability (both in and out of the healthcare context) [1]. Although BC is newer than the Internet of Things (IoT), cloud computing, and artificial intelligence (AI), the IoT exhibits many security issues. BC has greatly improved IoT applications, restoring trust [48]. The use of AI in healthcare affords excellent accuracy and precision, facilitating rapid decision-making. In the current coronavirus disease 2019 era, AI-based decisions categorize patients within a few minutes, greatly aiding clinicians. However, privacy remains important; patients fear that their data may be misplaced during AI analysis in the cloud. Prior to BC, cloud security was enhanced using specific applications [49–51]. After BC integration, data can be stored securely. As each transaction is recorded, patients can easily identify where the data are used. All involved groups experience benefits [52]. The integration of BC into the cloud resolves issues regarding location-based storage and analysis, while guaranteeing security [53]. BC must be integrated with more recent cutting-edge technologies. All industries accept that modern technology will render them smarter, but safety concerns remain. BC is already safe; the addition of other technologies can render it safer, faster, energy-saving, and cheaper [54–56]. ----- _Sustainability 2021, 13, 5269_ 6 of 17 **3. Solutions** We first identified critical factors affecting BC introduction; we reviewed past works and sought 15 expert opinions (inputs to structured self-interaction matrices (SSIMs)). These opinions were collected during a workshop concerning digital technology in the healthcare sector held at KIIT University, Bhubaneswar, India, in 2020. The 15 experts included nine senior medical practitioners with at least 10 years of experience in reputable hospitals with digital platforms hosting patient records and managing medicine supplies, as well as six academics with at least 10 years of research experience in BC (all academics were at or above professor/associate professor level in their medical colleges/universities). There was no limit on the number of experts that had completed S exploring remanufacturing and green campus operations (Singhal et al., 2020 [57], Gholami et al., 2020 [58]); 10 had completed SIMs concerning researcher selection (Nilashi et al., 2019 [59]). Our 15 experts were thus adequate. Next, ISM was used to develop a baseline model of associations among critical factors, and MICMAC analysis was performed to group the factors. ISM seeks to determine relationships among factors identified through literature review or expert opinion as an issue or a problem (Jharkharia and Shankar 2005 [60], Ravi and Shankar 2005 [61], Raj and Attri 2011 [62]). ISM techniques include brainstorming, nominal group techniques, and face-to-face interviews, yielding expert views regarding how to develop a contextual relationship among selected key factors (Ravi et al., 2005 [63], Barve et al., 2007 [64], Hasan et al., 2007 [65], Raj et al., 2007 [66]). Here, we addressed the complex barriers to BC implementation in healthcare. Factors determined through a literature review were reviewed by experts. No limit was imposed on the number of factors (Singhal et al., 2020 [57], Nayak et al., 2019 [67]). Table 1 lists the 13 factors identified and Table 2 lists the ISM steps. The flowchart of the solution (i.e., research framework and sequential steps) is shown in Figure 3. The critical dimensions of BC in healthcare commences with SIM completion and concludes with MICMAC policy recommendations. A strong correlation is evident between the ISM model and the critical factors identified. **Table 1. Numbers of factors evaluated in various reports.** **Source** **No. of Factors Used** **Research Objective** Singhal et al. 2020 [57] 15 Factors affecting electronic remanufacturing Nayak et al. 2019 [67] 14 Factors affecting rail safety performance Ahmed et al. 2019 [68] 15 Benchmarking of significant factors in seismic soil liquefaction Nayak et al., 2018 [69] 17 Factors affecting nontechnical human skills in engineering Aich and Tripathy, 2014 [70] 13 Factors affecting green supply chain management Tripathy et al., 2013 [71] 14 Factors affecting manufacturing R&D **Table 2. ISM steps.** **Steps** **Focus** Define pairwise relationships among identified critical 1: Establishment of a structural self-interaction matrix (SSIM) dimensions of healthcare BC technology 2: Create a reachability matrix Determine driving and dependent factors 3: Level partitioning Define structural levels (factor level partitioning) Develop an ISM model using a reachability matrix and 4: ISM modeling level partitioning Classify critical dimensions of healthcare BC technology into 5: MICMAC analysis four categories (drivers, dependents, autonomous factors, and linked factors) via MICMAC analysis ----- _Sustainability 2021, 13, 5269_ Classify critical dimensions of healthcare BC technology into four categories (drive7 of 17 #### 5: MICMAC analysis autonomous factors, and linked factors) via MICMAC analysis **Figure 3. Flowchart of solution methodology.** #### Figure 3. Flowchart of solution methodology. _3.1. Data Collection_ ### 3.1. Data Collection We reviewed all BC papers in Web of Science and Scopus in terms of critical factors influencing the adoption of BC in healthcare. With assistance from experts, we selected the 13 factors listed in Table 3. ----- _Sustainability 2021, 13, 5269_ 8 of 17 **Table 3. Factors affecting the implementation of BC in healthcare.** **Code** **Factor** **References** F1 Data unavailability (DU) [33,34,36] F2 Regulatory clarity and governance (RCG) [38–40] F3 Immature technology (IMT) [42,51] F4 Safer and smarter organization (SSO) [52–57] F5 Compatibility with other IT systems (CIT) [36,37] F6 High investment cost (HIC) [42,50] F7 Privacy and security of stored data (PSD) [27–29] F8 Scalability and accessibility (SA) [1,42] F9 Blockchain developers (BD) [42] F10 Interoperability of electronic health records (IEH) [30,32,34] F11 Data standardization (DS) [42,47–49] F12 Trust among stakeholders (TAS) [41–46] F13 Encouragement of integration (EI) [50,53–57] _3.2. ISM_ ISM is old and widely used by researchers in knowledge management, energy conservation, supplier selection, and green supply chain management; it is also used by strategic decisionmakers in various organizations [72–74]. ISM seeks to recognize/construct associations between factors affecting decision-making when a particular problem arises, then to solve the problem by considering the driving and dependency powers of each factor [75]. The framework features associations among factors, as identified by experts [76]. Fewer experts are required, compared with structural equation modeling or the Delphi method. ISM nonetheless builds models that solve decision-making problems [77,78]. Table 4 lists the various applications of ISM. Modeling proceeds as follows: (1) recognition of relevant factors based on past studies and expert opinion; (2) development of an SSIM and then a reachability matrix; (3) creation of a partition level table using a reachability matrix; (4) characterization of relationships among various factors; and (5) identification of uncertainties and consequent modifications. **Table 4. Applications of ISM.** **Techniques** **Application** ISM Adoption of IoT services [70] ISM Challenges posed by BC adoption within the Indian public sector [79] ISM BC as a disruptive technology in the construction industry [80] ISM and DEMATEL Modeling BC-enabled traceability in an agriculture supply chain [81] ISM Factors influencing lean implementation in healthcare organizations [82] 3.2.1. The SSIM The SSIMs completed by experts served as the ISM inputs. The contextual relationships among the 13 factors were determined by the majority opinions of the 15 experts expressed in a brainstorming session conducted during a 2020 workshop. The contextual relationships were finalized after considering the nature of each problem, the objective, and the majority opinion concerning the relationships between factors. The contextual association between two elements (i and j) is represented in one of four manners: (a) if i influences j, this is represented by “V”; (b) if j influences i, this is represented by “A”; (c) if i and j influence each other, this is represented by “X”; and (d) if i and j are independent, this is represented by “O”. For example, the interoperability of electronic health records F10 (IEH) influences the BC developers F9 (BD); the symbol used is V. Compatibility with other IT systems F5 (CIT) influences high investment cost F6 (HIC); the symbol used is A. The interoperability of electronic health records F10 (IEH) and privacy and security of storage data F7 (PSD) interact; the symbol used is X. Scalability and accessibility F8 (SA) has no ----- _Sustainability 2021, 13, 5269_ 9 of 17 relationship with data unavailability F1 (DU); the symbol used is O. The SSIM summary is presented in Table 5. The reachability matrix associated with the SSIMs is addressed below. **Table 5. SSIM summary.** **F1** **F2** **F3** **F4** **F5** **F6** **F7** **F8** **F9** **F10** **F11** **F12** **F13** F1 A A V O A V O A V O A V F2 V V V V V V V V V V V F3 V V A V V A V V A V F4 A A A A A A A A A F5 A V O A V O A V F6 V V O V V X V F7 A A X A A X F8 A V O A V F9 V V O V F10 A A X F11 A V F12 V 3.2.2. Reachability Matrix The four SSIM representations, V, A, X, and O, were replaced by 1 or 0 in a reachability matrix, as follows: (a) the symbol “V” in the (i, j) position of the SSIM matrix is substituted by 1 and 0 in the (i, j) and (j, i) positions of the reachability matrix; (b) the symbol “A” in the (i, j) position of the SSIM matrix is substituted by 0 and 1 in the (i, j) and (j, i) positions of the reachability matrix; (c) the symbol “X” in the (i, j) position of the SSIM matrix is substituted by 1 in both the (i, j) and (j, i) positions of the reachability matrix; and (d) the symbol “O” in the (i, j) position in the SSIM matrix is substituted by 0 in both the (i, j) and (j, i) positions of the reachability matrix. Next, the transitivity of the reachability matrix was checked. Transitivity means that if factor F1 influences F2 and F2 influences F3, then F1 impacts F3. If the position (i, j) of F1 impacts F3, the value becomes 1. The driving power (DVP) of a factor is calculated by adding all values in the accommodating row and the dependence power (DNP) is calculated by adding all values in the accommodating column. After considering transitivity, the final version of the reachability matrix is shown in Table 6. The subsequent step (i.e., partition of different levels) uses the reachability matrix. **Table 6. Reachability matrix.** **F1** **F2** **F3** **F4** **F5** **F6** **F7** **F8** **F9** **F10** **F11** **F12** **F13** **DVP** F1 1 0 0 1 1 0 1 0 0 1 0 0 1 5 F2 1 1 1 1 1 1 1 1 1 1 1 1 1 13 F3 1 0 1 1 1 0 1 1 0 1 1 0 1 9 F4 0 0 0 1 0 0 0 0 0 0 0 0 0 1 F5 0 0 0 1 1 0 1 0 0 1 0 0 1 5 F6 1 0 1 1 0 1 1 1 0 1 1 1 1 11 F7 0 0 0 1 1 0 1 0 0 1 0 0 1 4 F8 0 0 0 1 1 0 1 1 0 1 0 0 1 5 F9 1 0 1 1 1 0 1 1 1 1 1 0 1 10 F10 0 0 0 1 1 0 1 0 0 1 0 0 1 4 F11 0 0 0 1 0 0 1 0 0 1 1 0 1 5 F12 1 0 1 1 1 1 1 1 0 1 1 1 1 11 F13 0 0 0 1 0 0 1 0 0 1 0 0 1 4 DNP 6 1 5 13 6 3 12 6 2 12 6 3 12 1 3.2.3. Level Partition The antecedent and reachability sets for each element were developed based on the reachability matrix [83]. The reachability set contains the factors themselves and factors impacted by other factors, and the antecedent set consists of the factors themselves and ----- _Sustainability 2021, 13, 5269_ 10 of 17 factors impacting those factors. The intersection set is the group of elements common to the antecedent and reachability sets. The procedure was iterated; when the antecedent and reachability sets were equal, the top factor was identified. For example, level I is occupied by F13 due to the equality of the antecedent and reachability sets. Five iterations were performed when identifying the level of a factor. The level partition is shown in Table 7. All 13 factors are split into six levels. F2 occupies the sixth level and F13 occupies the first level; the other factors lie between these levels. **Table 7. Level partition.** **Factors** **Reachability Set** **Antecedent Set** **Intersection Set** **Level** F1 1,4,7,10,13 1,2,3,6,9,12 1 III F2 1,2,3,4,5,6,7,8,9,10,11,12,13 2 2 VI F3 1,3,4,5,7,8,10,11,13 2,3,6,9,12 3 IV F4 4 1,2,3,4,5,6,7,8,9,10,11,12,13 4 I F5 4,5,7,10,13 2,3,5,6,9,12 5 III F6 1,3,4,5,6,7,8,10,11,12,13 2,6,12 6,12 V F7 4,7,10,13 1,2,3,5,6,7,8,9,10,11,12,13 7,10,13 II F8 4,7,8,10,13 2,3,6,8,9,12 8 III F9 1,3,4,5,7,8,9,10,11,13 2,9 9 V F10 4,7,10,13 1,2,3,5,6,7,8,9,10,11,12,13 7,10,13 II F11 4,7,10,11,13 2,3,6,9,11,12 11 III F12 1,3,4,5,6,7,8,10,11,12,13 2,6,12 6,12 V F13 4,7,10,13 1,2,3,5,6,7,8,9,10,11,12,13 7,10,13 II 3.2.4. ISM The ISM of Figure 4 was developed based on the digraph and level partition table. A digraph exemplifies the interrelationships among elements at edges and nodes. Digraphs _3, x FOR PEER REVIEW_ 11 of 17 remove the transitive relationships between elements. The ISM is extracted from the combinative information of the digraph [84]. ##### Figure 4. The ISM. #### 3 3 MICMAC Analysis **Figure 4. The ISM.** ----- _Sustainability 2021, 13, 5269_ 11 of 17 **Figure 4. The** ISM. _3.3. MICMAC Analysis3.3. MICMAC Analysis_ |Figure 4.|The|ISM.| |---|---|---| 11 of 17 MICMAC requires factor dependence and driving powers as inputs [MICMAC requires factor dependence and driving powers as inputs [85] and then 85] and then categorizes the factors into four types (Figurecategorizes the factors into four types (Figure 5). Autonomous variables (factors with 5). Autonomous variables (factors with weak dependence and driving powers) are shown in the first quadrant. Dependent variablesweak dependence and driving powers) are shown in the first quadrant. Dependent vari(factors with strong dependence but weak driving powers) are shown in the secondables (factors with strong dependence but weak driving powers) are shown in the second quadrant. Linkage variables (factors with strong dependence and driving powers) arequadrant. Linkage variables (factors with strong dependence and driving powers) are shown in the third quadrant. Driving variables (factors with weak dependence powers but shown in the third quadrant. Driving variables (factors with weak dependence powers strong driving powers) appear in the fourth quadrant [66]. but strong driving powers) appear in the fourth quadrant [66]. **Figure 5.Figure 5. MICMAC analysis.MICMAC analysis.** **4. Results and Discussion4. Results and Discussion** We systematically analyzed and constructed the relationships among factors affecting We systematically analyzed and constructed the relationships among factors affect the adoption of BC in healthcare. We derived factor dependence and driving powers ing the adoption of BC in healthcare. We derived factor dependence and driving powers through MICMAC analysis. We first identified 13 critical factors affecting the adoption of through MICMAC analysis. We first identified 13 critical factors affecting the adoption of BC in healthcare (Table 1). Experts from academia and industry chose these factors. We BC in healthcare (Table 1). Experts from academia and industry chose these factors. We used ISM to construct the model. The ISM split all factors into six levels. Regulatory clarity and governance (F2) (at the bottom of the hierarchy) was the key driver of BC adoption in healthcare. Daluwathumullagamage and Sims found that BC would ensure better corporate governance if development was accompanied by changes in regulatory frameworks [86]. Healthcare industries must encourage governments to regulate appropriately; BC use is essential. Level IV included trust among stakeholders (F12), high investment cost (F6), and BC developers (F9); all were strong drivers of adoption. Senior healthcare managers must enthusiastically adopt BC and consumers must understand the great benefits afforded by BC use. Gomez-Trujillo et al. emphasized that BC guarantees trust and transparency; if all individuals, industries, and other stakeholders maintain confidence in BC, long-term success is ensured [87]. Koster and Borgman found that BC adoption required the support of senior authorities and trust among partners [88]. Level IV contained only immature technology (F3), which strongly influenced BC adoption. BC is new, not standardized, and has seldom been implemented in governmental agencies. Compatibility issues affecting performance may arise [48,89]. Level III included data standardization (F11), compatibility with other IT systems (F5), data unavailability (F1), and scalability and accessibility (SA); all strongly influenced BC adoption. The technology remains immature, and therefore the above factors must all be upgraded to enhance performance in the healthcare sector [90]. Level II comprises interoperability of electronic health records (F10), privacy and security of storage (F7), and encouragement of integration (F13); all dynamically influence adoption. Finally, the factor at level I is affected by all other factors ----- _Sustainability 2021, 13, 5269_ 12 of 17 and thus exhibits the highest dependence power. Encouragement of integration is the key driver; BC can be combined with many cutting-edge technologies that render organizations more efficient and smarter [51–53]. Secure data storage makes organizations safer; this is one of our long-term healthcare objectives. After MICMAC analysis, quadrant 4 hosted five factors with strong driving powers and quadrant 2 hosted four factors with strong dependence powers. Quadrant 3 was empty; there was no linkage variable. Quadrant 1 hosted four autonomous variables. Dependent variables comprised privacy and security of storage data (F7), interoperability of EHRs (F10), encouragement of integration (F13), and safer and smarter organizations (F4). MICMAC analysis revealed that regulatory clarity and governance (F2), immature technology (F3), high investment cost (F6), BC developers (F9), and trust among stakeholders (F12) exhibited strong driving powers and were thus the most important factors in terms of BC adoption in healthcare. Data unavailability (F1), compatibility with other IT systems (F5), scalability and accessibility (F8), and data standardization (F11) (autonomous variables) exhibited weaker dependences and driver powers, suggesting that they were less important than other factors. However, all identified factors affect the adoption of BC in healthcare. We shared our analysis with stakeholders in healthcare industries. Surprisingly, many managers were unaware of many factors. We hope that our analysis will help them to prepare for successful BC adoption. **5. Managerial Implications** Our work will allow regulators, policymakers, governments, healthcare industrialists, and consumers to recognize the critical factors that affect BC incorporation in healthcare. Managers and decisionmakers should focus on the inputs and outputs of the ISM model. The inputs were based on a literature review and expert opinions. The outputs identify the interdependencies and the short- and long-term importances of various factors. The model will be implemented and tested in a cross-sectional manner in multiple industries. Managers will be interested in the outcomes; they should prepare the resources for successful implementation. Managers must offer staff workshops and training regarding BC and its benefits. Existing educational institutes and special training schools may be involved. Managers must be careful when sharing information; a competitive advantage must not be lost. Petersson and Baur [91] emphasized that an organization is not required to reorganize its business model during BC integration. Furthermore, BC is possible in a traditional system; a new system is unnecessary. During organizational preparation, knowledge of the technical aspects will be helpful. All organizations must now adopt cutting-edge BC technology. Its basic features include smart contracts, privacy, and data security; it is easy to switch to new (improved) future platforms. Existing open-source platforms are expensive if they are expected to serve as proprietorial infrastructure [92]. Organizations should implement BC technology immediately; the “wait-and-see” period is over. Early acceptance of the technology will afford competitive advantages [93]. **6. Practical Implications** Healthcare decisionmakers must implement BC to protect the privacy of healthcare data. Such privacy supports the implementation of AI and federated learning, which enhance organizational efficiency. Kumar et al. [52] used BC technology for data authentication, allowing efficient use of AI and federated learning. In the era of coronavirus disease 2019, cutting-edge AI can rapidly identify an infection and is applicable worldwide; this could be combined with BC technology. With increasing data digitization, the need for privacy increases, along with the desire for societal betterment. BC technology can serve as the foundation of the required systems. ----- _Sustainability 2021, 13, 5269_ 13 of 17 **7. Conclusions** In our digital age, it is essential to protect healthcare data, but appropriate technology is lacking. BC technology can achieve the desired objectives. AI and federated learning enhance efficiency. BC systems would improve greatly if organizations were to successfully implement the technology. Large organizations (e.g., NVIDIA) have commenced research regarding AI and federated learning, motivated by societal betterment. Here, we recognized 13 factors that influence successful BC implementation in the healthcare industry. We used ISM to divide these 13 factors into six levels. An inappropriate regulatory environment greatly hinders BC adoption in the healthcare industry. Firms are reluctant to adopt this intricate and immature technology. Compatibility, investment cost, and security concerns are equally important. Our work has the following strengths. First, no similar formal study has appeared. Second, we have highlighted the key obstacles hindering the implementation of BC technology and have proposed methods to eliminate them. We offer useful tips for specialists in cutting-edge technology. BC will greatly advance organization in our digital era. Nonetheless, this study had the following limitations. First, we evaluated only a few critical factors emphasized in the literature. Second, as this technology is emerging, there are few skilled experts; we canvassed only 15. In the future, we plan to validate the results obtained after implementing BC technology and to combine the findings with AI and federated learning to create a useful, real-time generalized model. As previously suggested, and as reinforced by current demand, reliable security solutions must be integrated into all digital platforms and must be capable of adaptation to new environments [94,95]. We will seek BC technology that is secure across all applications. We will share the corresponding implementations in future articles. We will also perform cross-sectional studies to identify factors that can enhance the impact (i.e., strength) of BC implementation. Finally, we suggest that others could implement our approach in their diverse sectors by combining longitudinal and cross-sectional studies. We hope that our work may serve as a reference. It should be shared and may aid other industries. **Author Contributions: Conceptualization, S.A. and S.T.; methodology, S.A. and S.T.; validation,** S.A., M.-I.J. and H.-C.K.; formal analysis, S.A., M.-I.J. and H.-C.K.; data curation, S.A. and S.T.; writing—original draft preparation, S.A.; writing—review and editing, S.A. and S.T.; supervision, H.-C.K.; project administration, H.-C.K.; funding acquisition, H.-C.K. All authors have read and agreed to the published version of the manuscript. **Funding: This research was supported by Basic Science Research Program through the National** Research Foundation of Korea (NRF), supported by the Ministry of Science, ICT & Future Planning (NRF2017R1D1A3B04032905). **Informed Consent Statement: All the participants gave their consent to participate in this study.** **Conflicts of Interest: The authors declare no conflict of interest.** **References** 1. Katuwal, G.J.; Pandey, S.; Hennessey, M.; Lamichhane, B. Applications of blockchain in healthcare: Current landscape & challenges. arXiv 2018, arXiv:1812.02776. 2. Zhang, P.; Schmidt, D.C.; White, J.; Lenz, G. Blockchain technology use cases in healthcare. In Advances in Computers; Elsevier: Amsterdam, The Netherlands, 2018; Volume 111, pp. 1–41. 3. Meinert, E.; Alturkistani, A.; Foley, K.A.; Osama, T.; Car, J.; Majeed, A.; Van Velthoven, M.; Wells, G.; Brindley, D. Blockchain [implementation in health care: Protocol for a systematic review. JMIR Res. Protoc. 2019, 8, e10994. [CrossRef]](http://doi.org/10.2196/10994) 4. Talesh, S.A. Data breach, privacy, and cyber insurance: How insurance companies act as “compliance managers” for businesses. _[Law Soc. Inq. 2018, 43, 417–440. [CrossRef]](http://doi.org/10.1111/lsi.12303)_ 5. McCoy, T.H.; Perlis, R.H. Temporal trends and characteristics of reportable health data breaches, 2010–2017. JAMA 2018, 320, [1282–1284. [CrossRef] [PubMed]](http://doi.org/10.1001/jama.2018.9222) 6. Yaeger, K.; Martini, M.; Rasouli, J.; Costa, A. Emerging blockchain technology solutions for modern healthcare infrastructure. J. _[Sci. Innov. Med. 2019, 2. [CrossRef]](http://doi.org/10.29024/jsim.7)_ 7. Khezr, S.; Moniruzzaman, M.; Yassine, A.; Benlamri, R. Blockchain technology in healthcare: A comprehensive review and [directions for future research. Appl. Sci. 2019, 9, 1736. [CrossRef]](http://doi.org/10.3390/app9091736) ----- _Sustainability 2021, 13, 5269_ 14 of 17 8. McGhin, T.; Choo, K.K.R.; Liu, C.Z.; He, D. Blockchain in healthcare applications: Research challenges and opportunities. J. Netw. _[Comput. Appl. 2019, 135, 62–75. [CrossRef]](http://doi.org/10.1016/j.jnca.2019.02.027)_ 9. Casino, F.; Dasaklis, T.K.; Patsakis, C. A systematic literature review of blockchain-based applications: Current status, classification [and open issues. Telemat. Inform. 2019, 36, 55–81. [CrossRef]](http://doi.org/10.1016/j.tele.2018.11.006) 10. Park, J.H.; Park, J.H. Blockchain security in cloud computing: Use cases, challenges, and solutions. Symmetry 2017, 9, 164. [[CrossRef]](http://doi.org/10.3390/sym9080164) 11. Zhang, P.; Walker, M.A.; White, J.; Schmidt, D.C.; Lenz, G. Metrics for assessing blockchain-based healthcare decentralized apps. In Proceedings of the 2017 IEEE 19th International Conference on e-Health Networking, Applications and Services (Healthcom), Dalian, China, 12–15 October 2017; pp. 1–4. 12. Ahram, T.; Sargolzaei, A.; Sargolzaei, S.; Daniels, J.; Amaba, B. Blockchain technology innovations. In Proceedings of the 2017 IEEE Technology & Engineering Management Conference (TEMSCON), Santa Clara, CA, USA, 8–10 June 2017; pp. 137–141. 13. Kuo, T.T.; Zavaleta Rojas, H.; Ohno-Machado, L. Comparison of blockchain platforms: A systematic review and healthcare [examples. J. Am. Med. Inform. Assoc. 2019, 26, 462–478. [CrossRef]](http://doi.org/10.1093/jamia/ocy185) 14. Ekblaw, A.; Azaria, A.; Halamka, J.D.; Lippman, A. A Case Study for Blockchain in Healthcare: “MedRec” prototype for electronic health records and medical research data. In Proceedings of the IEEE Open & Big Data Conference, Vienna, Austria, 22–24 August 2016; Volume 13, p. 13. 15. Batubara, F.R.; Ubacht, J.; Janssen, M. Challenges of blockchain technology adoption for e-government: A systematic literature review. In Proceedings of the 19th Annual International Conference on Digital Government Research: Governance in the Data Age, Delft, The Netherlands, 30 May–1 June 2018; pp. 1–9. 16. Ali, O.; Ally, M.; Dwivedi, Y. The state of play of blockchain technology in the financial services sector: A systematic literature [review. Int. J. Inf. Manag. 2020, 54, 102199. [CrossRef]](http://doi.org/10.1016/j.ijinfomgt.2020.102199) 17. Alketbi, A.; Nasir, Q.; Talib, M.A. Blockchain for government services—Use cases, security benefits and challenges. In Proceedings of the 2018 15th Learning and Technology Conference (L&T), Jeddah, Saudi Arabia, 25–26 February 2018; pp. 112–119. 18. Lindman, J.; Rossi, M.; Tuunainen, V.K. Opportunities and risks of blockchain technologies in payments—A research agenda. In Proceedings of the 50th Hawaii International Conference on System Sciences, Hilton Waikoloa Village, HI, USA, 4–7 January 2017; pp. 1533–1542. 19. Swan, M. Blockchain: Blueprint for a New Economy; O’Reilly Media, Inc.: Sebastopol, CA, USA, 2015. 20. Tama, B.A.; Kweka, B.J.; Park, Y.; Rhee, K.H. A critical review of blockchain and its current applications. In Proceedings of the 2017 International Conference on Electrical Engineering and Computer Science (ICECOS), Palembang, Indonesia, 22–23 August 2017; pp. 109–113. 21. Antón, P.; Munoz, A.; Mana, A.; Koshutanski, H. Security-enhanced ambient assisted living supporting school activities during [hospitalisation. J. Ambient Intell. Humaniz. Comput. 2012, 3, 177–192. [CrossRef]](http://doi.org/10.1007/s12652-010-0039-6) 22. Antón, P.; Maña, A.; Muñoz, A.; Koshutanski, H. An immersive view approach by secure interactive multimedia proof-of-concept [implementation. Multimed. Tools Appl. 2015, 74, 8401–8420. [CrossRef]](http://doi.org/10.1007/s11042-013-1682-7) 23. Mettler, M. Blockchain technology in healthcare: The revolution starts here. In Proceedings of the 2016 IEEE 18th International Conference on E-Health Networking, Applications and Services (Healthcom), Munich, Germany, 14–16 September 2016; pp. 1–3. 24. Linn, L.A.; Koo, M.B. Blockchain for health data and its potential use in health it and health care related research. In ONC/NIST _Use of Blockchain for Healthcare and Research Workshop. Gaithersburg, Maryland, United States: ONC/NIST; NIST: Gaithersburg, MD,_ USA, 2016; pp. 1–10. 25. Joshi, A.P.; Han, M.; Wang, Y. A survey on security and privacy issues of blockchain technology. Math. Found. Comput. 2018, 1, [121. [CrossRef]](http://doi.org/10.3934/mfc.2018007) 26. Alhadhrami, Z.; Alghfeli, S.; Alghfeli, M.; Abedlla, J.A.; Shuaib, K. Introducing blockchains for healthcare. In Proceedings of the 2017 International Conference on Electrical and Computing Technologies and Applications (ICECTA), Ras Al Khaimah, United Arab Emirates, 21–23 November 2017; pp. 1–4. 27. Shi, S.; He, D.; Li, L.; Kumar, N.; Khan, M.K.; Choo, K.K.R. Applications of blockchain in ensuring the security and privacy of [electronic health record systems: A survey. Comput. Secur. 2020, 97, 101966. [CrossRef] [PubMed]](http://doi.org/10.1016/j.cose.2020.101966) 28. Zhao, H.; Zhang, Y.; Peng, Y.; Xu, R. Lightweight backup and efficient recovery scheme for health blockchain keys. In Proceedings of the 2017 IEEE 13th International Symposium on Autonomous Decentralized System (ISADS), Bangkok, Thailand, 22–24 March 2017; pp. 229–234. 29. Khan, F.A.; Asif, M.; Ahmad, A.; Alharbi, M.; Aljuaid, H. Blockchain technology, improvement suggestions, security challenges [on smart grid and its application in healthcare for sustainable development. Sustain. Cities Soc. 2020, 55, 102018. [CrossRef]](http://doi.org/10.1016/j.scs.2020.102018) 30. Zhang, P.; White, J.; Schmidt, D.C.; Lenz, G. Applying software patterns to address interoperability in blockchain-based healthcare apps. arXiv 2017, arXiv:1706.03700. 31. Agrawal, R.; Prabakaran, S. Big data in digital healthcare: Lessons learnt and recommendations for general practice. Heredity **[2020, 124, 525–534. [CrossRef]](http://doi.org/10.1038/s41437-020-0303-2)** 32. Kruse, C.S.; Kothman, K.; Anerobi, K.; Abanaka, L. Adoption factors of the electronic health record: A systematic review. JMIR _[Med. Inform. 2016, 4, e19. [CrossRef]](http://doi.org/10.2196/medinform.5525)_ 33. Tanner, C.; Gans, D.; White, J.; Nath, R.; Pohl, J. Electronic health records and patient safety: Co-occurrence of early EHR implementation with patient safety practices in primary care settings. Appl. Clin. Inform. 2015, 6, 136. ----- _Sustainability 2021, 13, 5269_ 15 of 17 34. Randall, D.; Goel, P.; Abujamra, R. Blockchain applications and use cases in health information technology. J. Health Med. Inform. **[2017, 8, 8–11. [CrossRef]](http://doi.org/10.4172/2157-7420.1000276)** 35. Schwerin, S. Blockchain and privacy protection in the case of the european general data protection regulation (GDPR): A delphi [study. J. Br. Blockchain Assoc. 2018, 1, 3554. [CrossRef]](http://doi.org/10.31585/jbba-1-1-(4)2018) 36. Abramova, S.; Böhme, R. Perceived Benefit and Risk as Multidimensional Determinants of Bitcoin Use: A Quantitative Exploratory Study; ICIS: Dublin, Ireland, 2016. 37. Sas, C.; Khairuddin, I.E. Design for trust: An exploration of the challenges and opportunities of bitcoin users. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, Denver, CO, USA, 6–11 May 2017; pp. 6499–6510. 38. [Sadhya, V.; Sadhya, H. Barriers to Adoption of Blockchain Technology. Available online: https://aisel.aisnet.org/amcis2018/](https://aisel.aisnet.org/amcis2018/AdoptionDiff/Presentations/20/) [AdoptionDiff/Presentations/20/ (accessed on 27 February 2021).](https://aisel.aisnet.org/amcis2018/AdoptionDiff/Presentations/20/) 39. Tanwar, S.; Parekh, K.; Evans, R. Blockchain-based electronic healthcare record system for healthcare 4.0 applications. J. Inf. Secur. _[Appl. 2020, 50, 102407. [CrossRef]](http://doi.org/10.1016/j.jisa.2019.102407)_ 40. Yaqoob, S.; Khan, M.M.; Talib, R.; Butt, A.D.; Saleem, S.; Arif, F.; Nadeem, A. Use of blockchain in healthcare: A systematic [literature review. Int. J. Adv. Comput. Sci. Appl. 2019, 10, 644–653. [CrossRef]](http://doi.org/10.14569/IJACSA.2019.0100581) 41. Esmaeilzadeh, P.; Mirzaei, T. The potential of blockchain technology for health information exchange: Experimental study from [patients’ perspectives. J. Med. Internet Res. 2019, 21, e14184. [CrossRef]](http://doi.org/10.2196/14184) 42. Pandey, P.; Litoriya, R. Implementing healthcare services on a large scale: Challenges and remedies based on blockchain [technology. Health Policy Technol. 2020, 9, 69–78. [CrossRef]](http://doi.org/10.1016/j.hlpt.2020.01.004) 43. [Sullivan, C.; Burger, E. E-residency and blockchain. Comput. Law Secur. Rev. 2017, 33, 470–481. [CrossRef]](http://doi.org/10.1016/j.clsr.2017.03.016) 44. Pokrovskaia, N.N. Tax, financial and social regulatory mechanisms within the knowledge-driven economy. Blockchain algorithms and fog computing for the efficient regulation. In Proceedings of the 2017 XX IEEE International Conference on Soft Computing and Measurements (SCM), IEEE, St. Petersburg, Russia, 24–26 May 2017; pp. 709–712. 45. Ølnes, S.; Ubacht, J.; Janssen, M. Blockchain in government: Benefits and implications of distributed ledger technology for [information sharing. Gov. Inf. Q. 2017, 34, 355–364. [CrossRef]](http://doi.org/10.1016/j.giq.2017.09.007) 46. Thakre, A.; Thabtah, F.; Shahamiri, S.R.; Hammoud, S. A novel block chain technology publication model proposal. Appl. Comput. _[Inform. 2019. [CrossRef]](http://doi.org/10.1016/j.aci.2019.10.003)_ 47. Andersen, J.V.; Bogusz, C.I. Patterns of Self-Organising in the Bitcoin Online Community: Code Forking as Organising in Digital Infrastructure. In Proceedings of the International Conference on Information Systems (ICIS 2017), Seoul, Korea, 10–13 December 2017. 48. Reyna, A.; Martín, C.; Chen, J.; Soler, E.; Díaz, M. On blockchain and its integration with IoT. Challenges and opportunities. _[Future Gener. Comput. Syst. 2018, 88, 173–190. [CrossRef]](http://doi.org/10.1016/j.future.2018.05.046)_ 49. Waller, A.; Sandy, I.; Power, E.; Aivaloglou, E.; Skianis, C.; Muñoz, A.; Maña, A. Policy based management for security in cloud computing. In FTRA International Conference on Secure and Trust Computing, Data Management, and Application; Springer: Berlin/Heidelberg, Germany, 2011; pp. 130–137. 50. Muñoz, A.; Maña, A.; González, J. Dynamic Security Properties Monitoring Architecture for Cloud Computing. In Security _Engineering for Cloud Computing: Approaches and Tools; IGI Global: Hershey, PA, USA, 2013; pp. 1–18._ 51. Muñoz, A.; Gonzalez, J.; Maña, A. A performance-oriented monitoring system for security properties in cloud computing [applications. Comput. J. 2012, 55, 979–994. [CrossRef]](http://doi.org/10.1093/comjnl/bxs042) 52. Kumar, R.; Khan, A.A.; Zhang, S.; Wang, W.; Abuidris, Y.; Amin, W.; Kumar, J. Blockchain-federated-learning and deep learning models for covid-19 detection using ct imaging. arXiv 2020, arXiv:2007.06537. 53. Nguyen, D.C.; Pathirana, P.N.; Ding, M.; Seneviratne, A. Integration of blockchain and cloud of things: Architecture, applications [and challenges. IEEE Commun. Surv. Tutor. 2020, 22, 2521–2549. [CrossRef]](http://doi.org/10.1109/COMST.2020.3020092) 54. Kshetri, N. Blockchain’s roles in strengthening cybersecurity and protecting privacy. Telecommun. Policy 2017, 41, 1027–1038. [[CrossRef]](http://doi.org/10.1016/j.telpol.2017.09.003) 55. Sun, M.; Zhang, J. Research on the application of block chain big data platform in the construction of new smart city for low [carbon emission and green environment. Comput. Commun. 2020, 149, 332–342. [CrossRef]](http://doi.org/10.1016/j.comcom.2019.10.031) 56. [Shen, C.; Pena-Mora, F. Blockchain for cities—a systematic literature review. IEEE Access 2018, 6, 76787–76819. [CrossRef]](http://doi.org/10.1109/ACCESS.2018.2880744) 57. Singhal, D.; Tripathy, S.; Jena, S.K. Remanufacturing for the circular economy: Study and evaluation of critical factors. Resour. _[Conserv. Recycl. 2020, 156, 104681. [CrossRef]](http://doi.org/10.1016/j.resconrec.2020.104681)_ 58. Gholami, H.; Bachok, M.F.; Saman, M.Z.M.; Streimikiene, D.; Sharif, S.; Zakuan, N. An ISM Approach for the Barrier Analysis in [Implementing Green Campus Operations: Towards Higher Education Sustainability. Sustainability 2020, 12, 363. [CrossRef]](http://doi.org/10.3390/su12010363) 59. Nilashi, M.; Dalvi, M.; Ibrahim, O.; Zamani, M.; Ramayah, T. An interpretive structural modelling of the features influencing [researchers’ selection of reference management software. J. Librariansh. Inf. Sci. 2019, 51, 34–46. [CrossRef]](http://doi.org/10.1177/0961000616668961) 60. Jharkharia, S.; Shankar, R. IT-enablement of supply chains: Understanding the barriers. J. Enterp. Inf. Manag. 2005, 18, 11–27. [[CrossRef]](http://doi.org/10.1108/17410390510571466) 61. Ravi, V.; Shankar, R. Analysis of interactions among the barriers of reverse logistics. Technol. Forecast. Soc. Chang. 2005, 72, [1011–1029. [CrossRef]](http://doi.org/10.1016/j.techfore.2004.07.002) 62. Raj, T.; Attri, R. Identification and modelling of barriers in the implementation of TQM. Int. J. Product. Qual. Manag. 2011, 8, [153–179. [CrossRef]](http://doi.org/10.1504/IJPQM.2011.041844) ----- _Sustainability 2021, 13, 5269_ 16 of 17 63. Ravi, V.; Shankar, R.; Tiwari, M.K. Productivity improvement of a computer hardware supply chain. Int. J. Product. Perform. _[Manag. 2005, 54, 239–255. [CrossRef]](http://doi.org/10.1108/17410400510593802)_ 64. Barve, A.; Kanda, A.; Shankar, R. Analysis of interaction among the barriers of third party logistics. Int. J. Agil. Syst. Manag. 2007, _[2, 109–129. [CrossRef]](http://doi.org/10.1504/IJASM.2007.015684)_ 65. [Hasan, M.A.; Shankar, R.; Sarkis, J. A study of barriers to agile manufacturing. Int. J. Agil. Syst. Manag. 2007, 2, 1–22. [CrossRef]](http://doi.org/10.1504/IJASM.2007.015679) 66. Raj, T.; Shankar, R.; Suhaib, M. An ISM approach for modelling the enablers of flexible manufacturing system: The case for India. _[Int. J. Prod. Res. 2008, 46, 6883–6912. [CrossRef]](http://doi.org/10.1080/00207540701429926)_ 67. Nayak, S.; Tripathy, S.; Dash, A. Non-technical skill development strategy to enhance safety performance of railway system: An [interpretive structural modelling approach. Int. J. Bus. Excell. 2019, 19, 168–188. [CrossRef]](http://doi.org/10.1504/IJBEX.2019.102233) 68. Ahmad, M.; Tang, X.W.; Qiu, J.N.; Ahmad, F. Interpretive structural modeling and MICMAC analysis for identifying and [benchmarking significant factors of seismic soil liquefaction. Appl. Sci. 2019, 9, 233. [CrossRef]](http://doi.org/10.3390/app9020233) 69. Nayak, S.; Tripathy, S.; Dash, A. Role of non technical skill in human factor engineering: A crucial safety issue in Indian Railway. _[Int. J. Syst. Assur. Eng. Manag. 2018, 9, 1120–1136. [CrossRef]](http://doi.org/10.1007/s13198-018-0715-z)_ 70. Aich, S.; Tripathy, S. An interpretive structural model of green supply chain management in Indian computer and its peripheral [industries. Int. J. Procure. Manag. 2014, 7, 239–256. [CrossRef]](http://doi.org/10.1504/IJPM.2014.060774) 71. Tripathy, S.; Sahu, S.; Ray, P.K. Interpretive structural modelling for critical success factors of R&D performance in Indian manufacturing firms. J. Model. Manag. 2013, 8, 212–240. 72. Lim, M.K.; Tseng, M.L.; Tan, K.H.; Bui, T.D. Knowledge management in sustainable supply chain management: Improving [performance through an interpretive structural modelling approach. J. Clean. Prod. 2017, 162, 806–816. [CrossRef]](http://doi.org/10.1016/j.jclepro.2017.06.056) 73. Saxena, J.P.; Vrat, P. Scenario building: A critical study of energy conservation in the Indian cement industry. Technol. Forecast. Soc. _[Chang. 1992, 41, 121–146. [CrossRef]](http://doi.org/10.1016/0040-1625(92)90059-3)_ 74. Girubha, J.; Vinodh, S.; Vimal, K.E.K. Application of interpretative structural modelling integrated multi criteria decision making [methods for sustainable supplier selection. J. Model. Manag. 2016, 11, 358–388. [CrossRef]](http://doi.org/10.1108/JM2-02-2014-0012) 75. Diabat, A.; Govindan, K. An analysis of the drivers affecting the implementation of green supply chain management. Resour. _[Conserv. Recycl. 2011, 55, 659–667. [CrossRef]](http://doi.org/10.1016/j.resconrec.2010.12.002)_ 76. Mangla, S.K.; Luthra, S.; Mishra, N.; Singh, A.; Rana, N.P.; Dora, M.; Dwivedi, Y. Barriers to effective circular supply chain [management in a developing country context. Prod. Plan. Control 2018, 29, 551–569. [CrossRef]](http://doi.org/10.1080/09537287.2018.1449265) 77. Chakraborty, K.; Mondal, S.; Mukherjee, K. Critical analysis of enablers and barriers in extension of useful life of automotive products through remanufacturing. J. Clean. Prod. 2019, 227, 1117–1135. 78. Kim, Y.; Park, Y.; Song, G. Interpretive Structural Modeling in the Adoption of IoT Services. KSII Trans. Internet Inf. Syst. 2019, 13. [[CrossRef]](http://doi.org/10.3837/tiis.2019.03.004) 79. Rana, N.P.; Dwivedi, Y.K.; Hughes, D.L. Analysis of Challenges for Blockchain Adoption within the Indian Public Sector: An [Interpretive Structural Modelling Approach. Inf. Technol. People 2021. [CrossRef]](http://doi.org/10.1108/ITP-07-2020-0460) 80. Sharma, M.G.; Kumar, S. The Implication of Blockchain as a Disruptive Technology for Construction Industry. IIM Kozhikode Soc. _[Manag. Rev. 2020, 9, 177–188. [CrossRef]](http://doi.org/10.1177/2277975220932343)_ 81. Kamble, S.S.; Gunasekaran, A.; Sharma, R. Modeling the blockchain enabled traceability in agriculture supply chain. Int. J. Inf. _[Manag. 2020, 52, 101967. [CrossRef]](http://doi.org/10.1016/j.ijinfomgt.2019.05.023)_ 82. Patri, R.; Suresh, M. Factors influencing lean implementation in healthcare organizations: An ISM approach. Int. J. Healthc. Manag. **[2018, 11, 25–37. [CrossRef]](http://doi.org/10.1080/20479700.2017.1300380)** 83. Bhosale, V.A.; Kant, R. An integrated ISM fuzzy MICMAC approach for modelling the supply chain knowledge flow enablers. _[Int. J. Prod. Res. 2016, 54, 7374–7399. [CrossRef]](http://doi.org/10.1080/00207543.2016.1189102)_ 84. [Sushil, S. Interpreting the interpretive structural model. Glob. J. Flex. Syst. Manag. 2012, 13, 87–106. [CrossRef]](http://doi.org/10.1007/s40171-012-0008-3) 85. Sivaprakasam, R.; Selladurai, V.; Sasikumar, P. Implementation of interpretive structural modelling methodology as a strategic [decision making tool in a Green Supply Chain Context. Ann. Oper. Res. 2015, 233, 423–448. [CrossRef]](http://doi.org/10.1007/s10479-013-1516-z) 86. Jayasuriya Daluwathumullagamage, D.; Sims, A. Blockchain-Enabled Corporate Governance and Regulation. Int. J. Financ. Stud. **[2020, 8, 36. [CrossRef]](http://doi.org/10.3390/ijfs8020036)** 87. Gomez-Trujillo, A.M.; Velez-Ocampo, J.; Gonzalez-Perez, M.A. Trust, Transparency, and Technology: Blockchain and Its Relevance in the Context of the 2030 Agenda. In The Palgrave Handbook of Corporate Sustainability in the Digital Era; Palgrave Macmillan: Cham, Switzerland, 2021; pp. 561–580. 88. Koster, F.; Borgman, H. New Kid on The Block! Understanding Blockchain Adoption in the Public Sector. In Proceedings of the 53rd Hawaii International Conference on System Sciences, Maui, HI, USA, 7–10 January 2020. 89. Monrat, A.A.; Schelén, O.; Andersson, K. A survey of blockchain from the perspectives of applications, challenges, and [opportunities. IEEE Access 2019, 7, 117134–117151. [CrossRef]](http://doi.org/10.1109/ACCESS.2019.2936094) 90. Hasselgren, A.; Kralevska, K.; Gligoroski, D.; Pedersen, S.A.; Faxvaag, A. Blockchain in healthcare and health sciences—A scoping [review. Int. J. Med. Inform. 2020, 134, 104040. [CrossRef] [PubMed]](http://doi.org/10.1016/j.ijmedinf.2019.104040) 91. Petersson, E.; Baur, K. Impacts of Blockchain Technology on Supply Chain Collaboration: A Study on the Use of Blockchain [Technology in Supply Chains and How It Influences Supply Chain Collaboration. Available online: https://www.diva-portal.](https://www.diva-portal.org/smash/get/diva2:1215210/FULLTEXT01.pdf) [org/smash/get/diva2:1215210/FULLTEXT01.pdf (accessed on 19 March 2019).](https://www.diva-portal.org/smash/get/diva2:1215210/FULLTEXT01.pdf) ----- _Sustainability 2021, 13, 5269_ 17 of 17 92. Satyavolu, P.; Herridge, M. Blockchain in Manufacturing: Enhancing Trust, Cutting . . . (n.d.). [Available online: https:](https://www.cognizant.com/whitepapers/blockchain-in-manufacturing-enhancing-trust-cuttingcosts-and-lubricating-processes-across-the-value-chain-codex3239.pdf) [//www.cognizant.com/whitepapers/blockchain-in-manufacturing-enhancing-trust-cuttingcosts-and-lubricating-processes-](https://www.cognizant.com/whitepapers/blockchain-in-manufacturing-enhancing-trust-cuttingcosts-and-lubricating-processes-across-the-value-chain-codex3239.pdf) [across-the-value-chain-codex3239.pdf (accessed on 18 January 2021).](https://www.cognizant.com/whitepapers/blockchain-in-manufacturing-enhancing-trust-cuttingcosts-and-lubricating-processes-across-the-value-chain-codex3239.pdf) 93. Choi, D.; Chung, C.Y.; Seyha, T.; Young, J. Factors Affecting Organizations’ Resistance to the Adoption of Blockchain Technology [in Supply Networks. Sustainability 2020, 12, 8882. [CrossRef]](http://doi.org/10.3390/su12218882) 94. Sánchez-Cid, F.; Mana, A.; Spanoudakis, G.; Kloukinas, C.; Serrano, D.; Munoz, A. Representation of security and dependability solutions. In Security and Dependability for Ambient Intelligence; Springer: Boston, MA, USA, 2009; pp. 69–95. 95. Serrano, D.; Ruíz, J.F.; Muñoz, A.; Maña, A.; Armenteros, A.; Crespo, B.G.N. Development of applications based on security patterns. In Proceedings of the 2009 Second International Conference on Dependability, IEEE, Athens, Greece, 18–23 June 2009; pp. 111–116. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/SU13095269?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/SU13095269, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/2071-1050/13/9/5269/pdf" }
2,021
[]
true
2021-05-08T00:00:00
[ { "paperId": "46ed63d40964e6aa648b6e4be838938f1ef33711", "title": "Analysis of challenges for blockchain adoption within the Indian public sector: an interpretive structural modelling approach" }, { "paperId": "7a645e1ccabc0e83cb3cc12486e6e3d0f7fd5991", "title": "Factors Affecting Organizations’ Resistance to the Adoption of Blockchain Technology in Supply Networks" }, { "paperId": "5b3711cf8dd639f02f275f51f35c45adc988051e", "title": "Improving Security on Blockchain and Its Integration with IoT" }, { "paperId": "27701fa918fb540565b1f841a8174c54f1dc67f4", "title": "The state of play of blockchain technology in the financial services sector: A systematic literature review" }, { "paperId": "88b299a252821ba8997ba2cd6390d8396fba1511", "title": "A novel block chain technology publication model proposal" }, { "paperId": "76b35c04677dd488e1bed4dca50720c21c63e992", "title": "Applications of blockchain in ensuring the security and privacy of electronic health record systems: A survey" }, { "paperId": "238878e4b4d7ed4a830be1598e610b2a9bec563a", "title": "Blockchain-Federated-Learning and Deep Learning Models for COVID-19 Detection Using CT Imaging" }, { "paperId": "fc178e2b4259eafa5c6a9dac3bd7218a2cfd772c", "title": "The Implication of Blockchain as a Disruptive Technology for Construction Industry" }, { "paperId": "e25a1cb79cd19497aabe9d9b51a17f13f32e7830", "title": "Blockchain-Enabled Corporate Governance and Regulation" }, { "paperId": "21b2c213b10b88bdd9976c64bf44bafa49b1c849", "title": "Modeling the blockchain enabled traceability in agriculture supply chain" }, { "paperId": "463266ec018df5acc164b3f7832779a327f3a7c3", "title": "Remanufacturing for the circular economy: Study and evaluation of critical factors" }, { "paperId": "bc1cb28eb152806acb587a1fc3dfe5fc7df4a1e3", "title": "Blockchain technology, improvement suggestions, security challenges on smart grid and its application in healthcare for sustainable development" }, { "paperId": "fb30e18bbfc8de7bf7df55af7d40c0d757d1942e", "title": "Big data in digital healthcare: lessons learnt and recommendations for general practice" }, { "paperId": "d448f6ad3717e0a4f8a3c6d1e43dcd57462277b9", "title": "Implementing healthcare services on a large scale: Challenges and remedies based on blockchain technology" }, { "paperId": "3665a3eaebed837f09b21a0141594bd654b18170", "title": "Blockchain-based electronic healthcare record system for healthcare 4.0 applications" }, { "paperId": "16f5ed6a95099276f0d7770040f6b3012ec533dd", "title": "New Kid On The Block! Understanding Blockchain Adoption in the Public Sector" }, { "paperId": "3e92f40d412c9f27ee07353a2793ffe1e6db457b", "title": "An ISM Approach for the Barrier Analysis in Implementing Green Campus Operations: Towards Higher Education Sustainability" }, { "paperId": "31812ed442a283a46e6e4e6a5895df2a1025a5a1", "title": "Research on the application of block chain big data platform in the construction of new smart city for low carbon emission and green environment" }, { "paperId": "e21700b05fc480d4f2c18462dbeaaf90dadb378d", "title": "Blockchain in healthcare and health sciences - A scoping review" }, { "paperId": "5f37190f9df8825d09f335110617e71a209d6a9d", "title": "Non-technical skill development strategy to enhance safety performance of railway system: an interpretive structural modelling approach" }, { "paperId": "b815f782662787e121611e5a74b59047feca7efd", "title": "Integration of Blockchain and Cloud of Things: Architecture, Applications and Challenges" }, { "paperId": "627341e1a5872676798c3ab56355157c1fe78bcc", "title": "A Survey of Blockchain From the Perspectives of Applications, Challenges, and Opportunities" }, { "paperId": "835505396982a1ae0b28a656194bd7634ed6cee6", "title": "Critical analysis of enablers and barriers in extension of useful life of automotive products through remanufacturing" }, { "paperId": "c1e98fa629cd65080afd919eecb8628829fe66ce", "title": "Blockchain in healthcare applications: Research challenges and opportunities" }, { "paperId": "04b805999f4055e9087f7d927cd0af44866a1022", "title": "Blockchain Technology in Healthcare: A Comprehensive Review and Directions for Future Research" }, { "paperId": "c098d87e026092fe0c0e941505eebb79122a6815", "title": "Interpretive Structural Modeling in the Adoption of IoT Services" }, { "paperId": "6c6b4d43df1529b969910f7e112e831360f12132", "title": "The Potential of Blockchain Technology for Health Information Exchange: Experimental Study From Patients’ Perspectives" }, { "paperId": "6934fddef7a37cec01e595394541abd5f2bced0a", "title": "Comparison of blockchain platforms: a systematic review and healthcare examples" }, { "paperId": "e4eac826a4afce44ea1de42f78d9dde755ccfef7", "title": "An interpretive structural modelling of the features influencing researchers’ selection of reference management software" }, { "paperId": "4c0945cb52d0734b25ecea49e3ae1c1b243fca66", "title": "A systematic literature review of blockchain-based applications: Current status, classification and open issues" }, { "paperId": "8bb40227ce2981bb9f462f70f81815021c3f801b", "title": "Blockchain Implementation in Health Care: Protocol for a Systematic Review" }, { "paperId": "be05eb406a5040efe19d244170a5e7c6f2d86529", "title": "Emerging Blockchain Technology Solutions for Modern Healthcare Infrastructure" }, { "paperId": "d50a82091e30c81f3d80c5d3d1db36b496535aa4", "title": "Interpretive Structural Modeling and MICMAC Analysis for Identifying and Benchmarking Significant Factors of Seismic Soil Liquefaction" }, { "paperId": "62f1da816679c1a0500906b97b8aeaafc5fa50b5", "title": "Applications of Blockchain in Healthcare: Current Landscape & Challenges" }, { "paperId": "e036da854e1fa5fbe9de6a8b25406e6af3900bae", "title": "Blockchain for Cities—A Systematic Literature Review" }, { "paperId": "f5ecc63b6d74a58413ecf8bc98c835d5d1488ea1", "title": "Temporal Trends and Characteristics of Reportable Health Data Breaches, 2010-2017" }, { "paperId": "d3fd9e741f161696363171ef683123dd1823251e", "title": "Challenges of blockchain technology adoption for e-government: a systematic literature review" }, { "paperId": "e68053f7e09e4d0a665fd03729f4a71c80d42538", "title": "A survey on security and privacy issues of blockchain technology" }, { "paperId": "107d0efeb55738282180ae3c7408e9be6b2b89a4", "title": "Barriers to effective circular supply chain management in a developing country context" }, { "paperId": "89debf2d924199c8d8147f6f365447bd4b1f81eb", "title": "Blockchain and Privacy Protection in the Case of the European General Data Protection Regulation (GDPR): A Delphi Study" }, { "paperId": "d8bfcdc081d5c95cfe128c8dfab0117ede9aac11", "title": "Role of non technical skill in human factor engineering: a crucial safety issue in Indian Railway" }, { "paperId": "2d09a943e9ea24803dcfea50c7c003a57a964d38", "title": "Data Breach, Privacy, and Cyber Insurance: How Insurance Companies Act as “Compliance Managers” for Businesses" }, { "paperId": "4196cdac0e779ea8a79e1192e32f7c8115123517", "title": "Blockchain for government services — Use cases, security benefits and challenges" }, { "paperId": "9fc4af5757abd1465578855417c2a7760212c263", "title": "Factors influencing lean implementation in healthcare organizations: An ISM approach" }, { "paperId": "a45b58c66c90ec4d3435e358fd46ee509dfea05b", "title": "Introducing blockchains for healthcare" }, { "paperId": "7f1c3c97a639a93796e935add3665c3ef329c0c8", "title": "Blockchain's roles in strengthening cybersecurity and protecting privacy" }, { "paperId": "f99878f6f3df724ffaf648f7515be053347d19de", "title": "Metrics for assessing blockchain-based healthcare decentralized apps" }, { "paperId": "1967549a611e7dd34a875aa220e534a0075521c1", "title": "Knowledge management in sustainable supply chain management: Improving performance through an interpretive structural modelling approach" }, { "paperId": "488ebe4db7190efe445c225aa67a10f70bc46d8d", "title": "Blockchain in government: Benefits and implications of distributed ledger technology for information sharing" }, { "paperId": "5d7bf180157709f2515ea7b596bb0bf231e83559", "title": "Blockchain Security in Cloud Computing: Use Cases, Challenges, and Solutions" }, { "paperId": "c4f78541eff05e539927d17ece67f239603b18a1", "title": "A critical review of blockchain and its current applications" }, { "paperId": "a89287356a6d16997157ec9f91a35fb59360d903", "title": "E-residency and blockchain" }, { "paperId": "54c6cb0a6ebdbc349e08b9c2e7c6bda3b72e1607", "title": "Blockchain Applications and Use Cases in Health Information Technology" }, { "paperId": "9092a7802f6e56dd5b6d1be30c8b5588a22e53fe", "title": "Blockchain technology innovations" }, { "paperId": "d73d60fb82318e91a8a66a1f8f2a5a1578d31f02", "title": "Applying Software Patterns to Address Interoperability in Blockchain-based Healthcare Apps" }, { "paperId": "0e7be5a5e9814db82ade63b9aac80de5d7da8476", "title": "Tax, financial and social regulatory mechanisms within the knowledge-driven economy. Blockchain algorithms and fog computing for the efficient regulation" }, { "paperId": "20bf5c9d575f19eaf465c43555b72ffd092d68fe", "title": "Design for Trust: An Exploration of the Challenges and Opportunities of Bitcoin Users" }, { "paperId": "10268e4e102ce9c3635708b9242079c26414ba90", "title": "Lightweight Backup and Efficient Recovery Scheme for Health Blockchain Keys" }, { "paperId": "310e677ce23004fdf0a549c2cfda2ef15420d6ec", "title": "Blockchain technology in healthcare: The revolution starts here" }, { "paperId": "5a76824d7ddc85e2f49b86b6575cde8d3806247e", "title": "Application of interpretative structural modelling integrated multi criteria decision making methods for sustainable supplier selection" }, { "paperId": "de8f0cec6e709ae7c0dc499343c6f0f815517460", "title": "Adoption Factors of the Electronic Health Record: A Systematic Review" }, { "paperId": "cd17afa7165fc0a39e8d8f5b46c9b6c897f94f52", "title": "An integrated ISM fuzzy MICMAC approach for modelling the supply chain knowledge flow enablers" }, { "paperId": "1b5d1095dcb43c8d6b3c1457a1e8b9fbaa62eed7", "title": "Implementation of interpretive structural modelling methodology as a strategic decision making tool in a Green Supply Chain Context" }, { "paperId": "a381a43e9345028b52211f65f003dfebbf25d43d", "title": "Electronic Health Records and Patient Safety" }, { "paperId": "97fddbbfd681bce9eeb8e0a013353b4d5b2ba0db", "title": "Blockchain: Blueprint for a New Economy" }, { "paperId": "c0dd8f735f44ed53c9c6fc28b30ae784181a3dc6", "title": "An interpretive structural model of green supply chain management in Indian computer and its peripheral industries" }, { "paperId": "ede1d4debfb33f626af2b611973fc1f74184449f", "title": "Interpretive structural modelling for critical success factors of R&D performance in Indian manufacturing firms" }, { "paperId": "2b0f9aa8f60148e9f8786f3f1dfe10fd8eab8a0d", "title": "An immersive view approach by secure interactive multimedia proof-of-concept implementation" }, { "paperId": "f49b834d24a521314c6a9f2f302cbe0dc70820b0", "title": "Security Engineering for Cloud Computing: Approaches and Tools" }, { "paperId": "f84c059fb5660d6ea1e67f6f40327d2d2b4c355f", "title": "Security-enhanced ambient assisted living supporting school activities during hospitalisation" }, { "paperId": "4985bfae6a1b69b8ef40bcb14bb9967ebe597a5f", "title": "A Performance-Oriented Monitoring System for Security Properties in Cloud Computing Applications" }, { "paperId": "dfe48a870bc2238fc47916599c2aa30fd9c4c999", "title": "Interpreting the Interpretive Structural Model" }, { "paperId": "4e72b82ab96b24defe9e21913657579a3518669a", "title": "Identification and modelling of barriers in the implementation of TQM" }, { "paperId": "b9fd67f0a90ab547e3a946cf95d77fe6562ec5a4", "title": "Policy Based Management for Security in Cloud Computing" }, { "paperId": "1d40975a335f302acbe015d41e183501dc655717", "title": "An analysis of the drivers affecting the implementation of green supply chain management" }, { "paperId": "0ba0c4a9e2eb1329dd8412323c9f0efe7eebbf6b", "title": "Development of Applications Based on Security Patterns" }, { "paperId": "88fd825a1b2f732835d55d95da3b9ffedf25f17b", "title": "An ISM approach for modelling the enablers of flexible manufacturing system: the case for India" }, { "paperId": "132e301b035d7bd046a6a2a56e2be66546ea0231", "title": "Analysis of interaction among the barriers of Third Party Logistics" }, { "paperId": "9abf8076e1639f12c304ffda8562bdaeb7f447f0", "title": "A study of barriers to agile manufacturing" }, { "paperId": "787fc0b665ef8b696286202e34a5ae4de067a3db", "title": "ANALYSIS OF INTERACTIONS AMONG THE BARRIERS OF REVERSE LOGISTICS" }, { "paperId": "53842e647f5e65df8edc3c8d71509ac07653a01f", "title": "Productivity improvement of a computer hardware supply chain" }, { "paperId": "dc63bb8d856edd9ce0a575664348a30ae047aa7a", "title": "IT-enablement of supply chains: understanding the barriers" }, { "paperId": "05445c3defa10aa44e2ecd49a91008e9241cb892", "title": "Scenario building: A critical study of energy conservation in the Indian cement industry" }, { "paperId": null, "title": "Blockchain in Manufacturing: Enhancing Trust, Cutting . ." }, { "paperId": "2b0befc4714c376c1d7e1d2d21d3d123c9fb59ec", "title": "Trust, Transparency, and Technology: Blockchain and Its Relevance in the Context of the 2030 Agenda" }, { "paperId": "79a33b6b7d2ae17d55659faa084d37d6e7f930f4", "title": "Use of Blockchain in Healthcare: A Systematic Literature Review" }, { "paperId": "b3bca8039c72847c2fe5f7f848c0955c30662273", "title": "Chapter One - Blockchain Technology Use Cases in Healthcare" }, { "paperId": "3006545e79f50126dc421a5dfe1f118c05dd82e9", "title": "Impacts of Blockchain technology on Supply Chain Collaboration : A study on the use of blockchain technology in supply chains and how it influences supply chain collaboration" }, { "paperId": "b43ecf4409b56ab5786ed12fac711ae6249dd39d", "title": "Barriers to Adoption of Blockchain Technology" }, { "paperId": "508b1c18ac074428c33b0e023d3e3110e718991c", "title": "Patterns of Self-Organising in the Bitcoin Online Community: Code Forking as Organising in Digital Infrastructure" }, { "paperId": "3ed0db58a7aec7bafc2aa14ca550031b9f7021d5", "title": "A Case Study for Blockchain in Healthcare : “ MedRec ” prototype for electronic health records and medical research data" }, { "paperId": "f15ddad79b91c084b64dc920344e23fdbbe9e0ce", "title": "Perceived Benefit and Risk as Multidimensional Determinants of Bitcoin Use: A Quantitative Exploratory Study" }, { "paperId": "cab41a7bec3b928cc13ca920d893f449cabb1599", "title": "Opportunities and risks of Blockchain Technologies in payments – a research agenda" }, { "paperId": null, "title": "Blockchain for health data and its potential use in health it and health care related research" }, { "paperId": "b88ca6bf59f16c8e13e21ba2f37012b027f7253b", "title": "Dynamic Security Properties Monitoring Architecture for Cloud Computing" }, { "paperId": "94afa897bd6daad207942fff481d922c1d9cbf12", "title": "Representation of Security and Dependability Solutions" }, { "paperId": "ac5a34da896e64c113369ea49d3e6691523e06a0", "title": "Future Generation Computer Systems" } ]
18,225
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/012e174c148901bedf28ae161518f3fa5c2eee4c
[ "Computer Science" ]
0.87742
An overview of the OMNeT++ simulation environment
012e174c148901bedf28ae161518f3fa5c2eee4c
International ICST Conference on Simulation Tools and Techniques
[ { "authorId": "145381357", "name": "A. Varga" }, { "authorId": "102064564", "name": "R. Hornig" } ]
{ "alternate_issns": null, "alternate_names": [ "Simul Tool Tech Commun Netw Syst", "Simulation Tools and Techniques for Communications, Networks and System", "SIMUTools", "SimuTools", "Int ICST Conf Simul Tool Tech" ], "alternate_urls": [ "https://eudl.eu/proceedings" ], "id": "dbfba120-41ed-4e2b-966a-442cb212437c", "issn": "2310-9831", "name": "International ICST Conference on Simulation Tools and Techniques", "type": "conference", "url": "http://www.icst.org/" }
null
# AN OVERVIEW OF THE OMNeT++ SIMULATION ENVIRONMENT ## András Varga ### OpenSim Ltd. Sz l köz 11, 1032 ő ő Budapest, Hungary ## andras.varga@omnest.com ABSTRACT The OMNeT++ discrete event simulation environment has been publicly available since 1997. It has been created with the simulation of communication networks, multiprocessors and other distributed systems in mind as application area, but instead of building a specialized simulator, OMNeT++ was designed to be as general as possible. Since then, the idea has proven to work, and OMNeT++ has been used in numerous domains from queuing network simulations to wireless and ad-hoc network simulations, from business process simulation to peer-to-peer network, optical switch and storage area network simulations. This paper presents an overview of the OMNeT++ framework, recent challenges brought about by the growing amount and complexity of third party simulation models, and the solutions we introduce in the next major revision of the simulation framework.[1] ## KEYWORDS discrete simulation, network simulation, simulation tools, performance analysis, computer systems, telecommunications, hierarchical, integrated development environment ## 1. INTRODUCTION OMNeT++[1][2] is a C++-based discrete event simulator for modeling communication networks, multiprocessors and other distributed or parallel systems. OMNeT++ is public-source, and can be used under the Academic Public License that makes the software free for non-profit use. The motivation of developing OMNeT++ was to produce a powerful open-source discrete event simulation tool that can be used by academic, educational and research-oriented commercial institutions for the simulation of computer networks and distributed or parallel systems. OMNeT++ attempts to fill the gap between open-source, research-oriented simulation software such as NS-2 [11] and expensive commercial alternatives like OPNET [16]. A later section of this paper presents a comparison with other simulation packages. OMNeT++ Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are Permission to make digital or hard copies of all or part of this work for not made or distributed for profit or commercial advantage and that personal or classroom use is granted without fee provided that copies copies bear this notice and the full citation on the first page. Copyrights are not made or distributed for profit or commercial advantage and that for components of this work owned by others than ICST must be copies bear this notice and the full citation on the first page. To copy honored. Abstracting with credit is permitted. To copy otherwise, to otherwise, to republish, to post on servers or to redistribute to lists, republish, to post on servers or to redistribute to lists, requires prior requires prior specific permission and/or a fee. specific permission and/or a fee. SIMUTOOLS 2008, March 03-07, Marseille, France Copyright © 2008 ICST 978-963-9799-20-2 SIMUTools, March 03 – 07, 2008, Marseille, France. DOI 10.4108/ICST.SIMUTOOLS2008.3027ISBN 978-963-9799-20-2 1 The 4.0 release is scheduled to appear in Q1 2008. ## Rudolf Hornig ### OpenSim Ltd. Sz l köz 11, 1032 ő ő Budapest, Hungary ## rudolf.hornig@omnest.com is available on all common platforms including Linux, Mac OS/X and Windows, using the GCC tool chain or the Microsoft Visual C++ compiler. OMNeT++ represents a framework approach. Instead of directly providing simulation components for computer networks, queuing networks or other domains, it provides the basic machinery and tools to write such simulations. Specific application areas are supported by various simulation models and frameworks such as the Mobility Framework or the INET Framework. These models are developed completely independently of OMNeT++, and follow their own release cycles. Since its first release, simulation models have been developed by various individuals and research groups for several areas including: wireless and ad-hoc networks, sensor networks, IP and IPv6 networks, MPLS, wireless channels, peer-to-peer networks, storage area networks (SANs), optical networks, queuing networks, file systems, high-speed interconnections (InfiniBand), and others. Some of the simulation models are ports of real-life protocol implementations like the Quagga Linux routing daemon or the BSD TCP/IP stack, others have been written directly for OMNeT++. A later section of this paper will discuss these projects in more detail. In addition to university research groups and non-profit research institutions, companies like IBM, Intel, Cisco, Thales and Broadcom are also using OMNeT++ successfully in commercial projects or for in-house research. ## 2. THE DESIGN OF OMNeT++ OMNeT++ was designed from the beginning to support network simulation on a large scale. This objective lead to the following main design requirements: - To enable large-scale simulation, simulation models need to be hierarchical, and built from reusable components as much as possible. - The simulation software should facilitate visualizing and debugging of simulation models in order to reduce debugging time, which traditionally takes up a large percentage of simulation projects. (The same feature set is also useful for educational use of the software.) - The simulation software itself should be modular, customizable and should allow embedding simulations into larger applications such as network planning software. (Embedding brings additional requirements about the memory management, restartability, etc. of the simulation). ----- - Data interfaces should be open: it should be possible to generate and process input and output files with commonly available software tools. - Should provide an Integrated Development Environment that largely facilitates model development and analyzing results. The following sections go through the most important aspects of OMNeT++, highlighting the design decisions that helped achieve the above main goals. ## 2.1 Model Structure An OMNeT++ model consists of modules that communicate with message passing. The active modules are termed simple modules; they are written in C++, using the simulation class library. Simple modules can be grouped into compound modules and so forth; the number of hierarchy levels is not limited. Messages can be sent either via connections that span between modules or directly to their destination modules. The concept of simple and compound modules is similar to DEVS [46][47] atomic and coupled models. Both simple and compound modules are instances of module _types. While describing the model, the user defines module types;_ instances of these module types serve as components for more complex module types. Finally, the user creates the system module as a network module which is a special compound module type without gates to the external world. When a module type is used as a building block, there is no distinction whether it is a simple or a compound module. This allows the user to transparently split a module into several simple modules within a compound module, or do the opposite, re-implement the functionality of a compound module in one simple module, without affecting existing users of the module type. The feasibility of model reuse is proven by the model frameworks like INET Framework [1] and Mobility Framework [17][18], and their extensions. ### network simple modules compound module Figure 1. Model Structure in OMNeT++. Boxes represent simple modules (thick border), and compound modules (thin border). Arrows connecting small boxes represent connections and gates. Modules communicate with messages which – in addition to usual attributes such as timestamp – may contain arbitrary data. Simple modules typically send messages via gates, but it is also possible to send them directly to their destination modules. Gates are the input and output interfaces of modules: messages are sent out through output gates and arrive through input gates. An input and an output gate can be linked with a connection. Connections are created within a single level of module hierarchy: within a compound module, corresponding gates of two submodules, or a gate of one submodule and a gate of the compound module can be connected. Connections spanning across hierarchy levels are not permitted, as it would hinder model reuse. Due to the hierarchical structure of the model, messages typically travel through a chain of connections, to start and arrive in simple modules. Compound modules act as 'cardboard boxes' in the model, transparently relaying messages between their inside and the outside world. Properties such as propagation delay, data rate and bit error rate, can be assigned to connections. One can also define connection types with specific properties (termed channels) and reuse them in several places. Modules can have parameters. Parameters are mainly used to pass configuration data to simple modules, and to help define model topology. Parameters may take string, numeric or boolean values. Because parameters are represented as objects in the program, parameters – in addition to holding constants – may transparently act as sources of random numbers with the actual distributions provided with the model configuration, they may interactively prompt the user for the value, and they might also hold expressions referencing other parameters. Compound modules may pass parameters or expressions of parameters to their submodules. ## 2.2 The Design of the NED Language The user defines the structure of the model (the modules and their interconnection) in OMNeT++'s topology description language, _NED. Typical ingredients of a NED description are simple module_ declarations, compound module definitions and network definitions. Simple module declarations describe the interface of the module: gates and parameters. Compound module definitions consist of the declaration of the module's external interface (gates and parameters), and the definition of submodules and their interconnection. Network definitions are compound modules that qualify as self-contained simulation models. The NED language has been designed to scale well, however, recent growth in the amount and complexity of OMNeT++-based simulation models and model frameworks made it necessary to improve the NED language as well. In addition to a number of smaller improvements, the following major features have been introduced: **Inheritance. Modules and channels can now be subclassed.** Derived modules and channels may add new parameters, gates, and (in the case of compound modules) new submodules and connections. They may set existing parameters to a specific value, and also set the gate size of a gate vector. This makes it possible, for example, to take a GenericTCPClientApp module and derive an FTPApp from it by setting certain parameters to a fixed value; or derive a WebClientHost compound module from a BaseHost compound module by adding a WebClientApp submodule and connecting it to the inherited TCP submodule. **Interfaces. Module and channel interfaces can be used as a** placeholder where normally a module or channel type would be used, and the concrete module or channel type is determined at network setup time by a parameter. Concrete module types have to “implement” the interface they can substitute. For example, the module types `ConstSpeedMobility` and ``` RandomWayPointMobility need to implement IMobility ``` to be able to be plugged into a MobileHost that contains an ``` IMobility submodule. ``` **Packages. To address name clashes between different models and** to simplify specifying which NED files are needed by a specific ----- simulation model, a Java-like package structure was introduced into the NED language. **Inner types. Channel types and module types used locally by a** compound module can now be defined within the compound module, in order to reduce namespace pollution. **Metadata annotations. It is possible to annotate module or** channel types, parameters, gates and submodules by adding _properties. Metadata are not used by the simulation kernel_ directly, but they can carry extra information for various tools, the runtime environment, or even for other modules in the model. For example, a module's graphical representation (icon, etc) or the prompt string and unit (milliwatt, etc) of a parameter are specified using properties. The NED language has an equivalent XML representation, that is, NED files can be converted to XML and back without loss of data, including comments. This lowers the barrier for programmatic manipulation of NED files, for example extracting information, refactoring and transforming NED, generating NED from information stored in other system like SQL databases, and so on. ## 2.3 Graphical Editor The OMNeT++ package includes an Integrated Development Environment which contains a graphical editor using NED as its native file format; moreover, the editor can work with arbitrary, even hand-written NED code. The editor is a fully two-way tool, i.e. the user can edit the network topology either graphically or in NED source view, and switch between the two views at any time. This is made possible by design decisions about the NED language itself. First, NED is a declarative language, and as such, it does not use an imperative programming language for defining the internal structure of a compound module. Allowing arbitrary programming constructs would make it practically infeasible to write two-way graphical editors which could work directly with both generated and hand-made NED files. (Generally, the editor would need AI capability to understand the code.) Most graphical editors only allow the creation of fixed topologies. However, NED contains declarative constructs (resembling loops and conditionals in imperative languages), which enable parametric topologies: it is possible to create common regular topologies such as ring, grid, star, tree, hypercube, or random interconnection whose parameters (size, etc.) are passed in numeric-valued parameters. The potential of parametric topologies and associated design patterns have been investigated in [7][9]. With parametric topologies, NED holds an advantage in many simulation scenarios both over OPNET where only fixed model topologies can be designed, and over NS-2 where building model topology is programmed in Tcl and often intermixed with simulation logic, so it is generally impossible to write graphical editors which could work with existing, hand-written code. ## 2.4 Separation of Model and Experiments It is always a good practice to try to separate the different aspects of a simulation as much as possible. Model behavior is captured _in C++ files as code, while model topology (and of course the_ parameters defining this topology) is defined by the NED files. This approach allows the user to keep the different aspects of the model in different places which in turn allows having a cleaner model and better tooling support. In a generic simulation scenario, one usually wants to know how the simulation behaves with different inputs. These variables neither belong to the behavior (code) nor the topology (NED files) as they can change from run to run. INI files are used to store these values. INI files provide a great way to specify how these parameters change and enable us to run our simulation for each parameter combination we are interested in. The generated simulation results can be easily harvested and processed by the built in analysis tool. We will explore later, in the Result Analysis paragraph, how the INI files are organized and how they can make experimenting with our model a lot easier. ## 2.5 Simple Module Programming Model _Simple modules are the active elements in a model. They are_ atomic elements in the module hierarchy: they cannot be divided any further. Simple modules are programmed in C++, using the OMNeT++ simulation class library. OMNeT++ provides an Integrated C++ Development Environment so it is possible to write, run and debug the code without leaving the OMNeT++ IDE. The simulation kernel does not distinguish between messages and events – events are also represented as messages. Simple modules are programmed using the process-interaction method. The user implements the functionality of a simple module by subclassing the cSimpleModule class. Functionality is added via one of two alternative programming models: (1) _coroutine-based, and (2)_ _event-processing function. When using_ _coroutine-based programming, the module code runs in its own_ (non-preemptively scheduled) thread, which receives control from the simulation kernel each time the module receives an event (=message). The function containing the coroutine code will typically never return: usually it contains an infinite loop with _send and receive calls._ When using event-processing function, the simulation kernel simply calls the given function of the module object with the message as argument – the function has to return immediately after processing the message. An important difference between the _coroutine-based and event-processing function programming_ models is that with the former, every simple module needs an own CPU stack, which means larger memory requirements for the simulation program. This is of interest when the model contains a large number of modules (over a few ten thousands). It is possible to write code which executes on module _initialization and finalization: the latter takes place on successful_ simulation termination, and finalization code is mostly used to save scalar results into a file. OMNeT++ also supports multi_stage initialization: situations where model initialization needs to_ be done in several "waves". Multi-stage initialization support is missing from most simulation packages, and it is usually emulated with broadcast events scheduled at zero simulation time, which is a less clean solution. Message sending and receiving are the most frequent tasks in simple modules. Messages can be sent either via output gates, or directly to another module. Modules receive messages either via one of the several variations of the receive call (when using coroutine-based programming), or messages are delivered to the module in an invocation from the simulation kernel (when using the event-processing function). Messages can be defined by specifying their content in an MSG file. OMNeT++ takes care of ----- creating the necessary C++ classes. MSG files allow the OMNeT++ kernel to generate reflection code which enables us to peek into messages and explore their content at runtime. It is possible to modify the topology of the network dynamically: one can create and delete modules and rearrange connections while the simulation is executing. Even compound modules with parametric internal topology can be created on the fly. ## 2.6 Design of the Simulation Library The OMNeT++ provides a rich object library for simple module implementers. There are several distinguishing factors between this library and other general-purpose or simulation libraries. The OMNeT++ class library provides reflection functionality which makes it possible to implement high-level debugging and tracing capability, as well as automatic animation on top of it (as exemplified by the Tkenv user interface, see later). Memory leaks, pointer aliasing and other memory allocation problems are common in C++ programs not written by specialists; OMNeT++ alleviates this problem by tracking object ownership and detecting bugs caused by aliased pointers and misuse of shared objects. The requirements for ease of use, modularity, open data interfaces and support of embedding also heavily influenced the design of the class library. The consistent use of object-oriented techniques makes the simulation kernel compact and slim. This makes it relatively easy to understand its internals, which is a useful property for both debugging and educational use. Recently it has become more common to do large scale network simulations with OMNeT++, with several ten thousand or more network nodes. To address this requirement, aggressive memory optimization has been implemented in the simulation kernel, based on shared objects and copy-on-write semantics. Until recently, simulation time has been represented as with C's ``` double type (IEEE double precision). Well-known precision ``` problems with floating point calculations however, have caused problems in simulations from time to time. To address this issue, simulation time has been recently changed to 64-bit integer-based fixed-point representation. One of the major problems that had to be solved here was how to detect numeric overflows, as the C and C++ languages, despite their explicit goals of being “close to the hardware”, lack any support to detect integer overflows. ## 2.7 Contents of the Simulation Library This section provides a very brief catalog of the classes in the OMNeT++ simulation class library. The classes were designed to cover most of the common simulation tasks. OMNeT++ has the ability to generate random numbers from several independent streams. The common distributions are supported, and it is possible to add new distributions programmed by the user. It is also possible to load user distributions defined by histograms. The class library offers queues and various other container _classes. Queues can also operate as priority queues._ _Messages are objects which may hold arbitrary data structures and_ other objects (through aggregation or inheritance), and can also embed other messages. OMNeT++ supports routing traffic in the network. This feature provides the ability to explore actual network topology, extract it into a graph data structure, then navigate the graph or apply algorithms such as Dijkstra to find shortest paths. There are several statistical classes, from simple ones which collect the mean and the standard deviation of the samples to a number of distribution estimation classes. The latter include three highly configurable histogram classes and the implementations of the P[2] [10] and the k-split [8] algorithms. It is also supported to write time series result data into an output file during simulation execution, and there are tools for post-processing the results. ## 2.8 Parallel Simulation Support OMNeT++ also has support for parallel simulation execution. Very large simulations may benefit from the parallel distributed simulation (PDES) feature, either by getting speedup, or by distributing memory requirements. If the simulation requires several Gigabytes of memory, distributing it over a cluster may be the only way to run it. For getting speedup (and not actually slowdown, which is also easily possible), the hardware or cluster should have low latency and the model should have inherent parallelism. Partitioning and other configuration can be configured in the INI file, the simulation model itself doesn't need to be changed (unless, of course, it contains global variables that prevents distributed execution in the first place.) The communication layer is MPI, but it's actually configurable, so if the user does not have MPI it is still possible to run some basic tests over named pipes. The figure below explains the logical architecture of the parallel simulation kernel: Simulation Model Simulation Kernel Parallel simulation subsystem Synchronization Event scheduling, sending, receiving Partition (LP) Communication communications library (MPI, sockets, etc.) Figure 2. Logical Architecture of the OMNeT++ Parallel Simulation kernel ## 2.9 Internal Architecture |architecture of the parallel simulation kernel:|Col2| |---|---| |Simulation Model|| ||| |Col1|Col2|Col3| |---|---|---| |Partition (||LP)| |||| |Simulation Kernel Parallel simulation subsystem Synchronization Event scheduling, sending, receiving Partition (LP) Communication|Col2|Col3| |---|---|---| ||Parallel simulation subsystem Synchronization Partition (LP) Communication|| |||| **OMNeT++ executable** Simulation Model |Col1|SIM (simulation kernel)|Col3| |---|---|---| |||| |||| |||| Figure 3. Logical Architecture of an OMNeT++ Simulation Program ----- OMNeT++ simulation programs possess a modular structure. The logical architecture is shown on Figure 3. The Model Component Library consists of the code of compiled simple and compound modules. Modules are instantiated and the concrete simulation model is built by the simulation kernel and class library (Sim) at the beginning of the simulation execution. The simulation executes in an environment provided by the user interface libraries (Envir, Cmdenv and Tkenv) – this environment defines where input data come from, where simulation results go to, what happens to debugging output arriving from the simulation model, controls the simulation execution, determines how the simulation model is visualized and (possibly) animated, etc. **Embedding Application** OMNeT++ subsystem Simulation Model other parts of the embedding application |Col1|SIM (sim. kernel)|Col3|Col4| |---|---|---|---| ||||| ||||| ||||| Figure 4. Embedding OMNeT++ By replacing the user interface libraries, one can customize the full environment in which the simulation runs, and even embed an OMNeT++ simulation into a larger application (Figure 4). This is made possible by the existence of a generic interface between Sim and the user interface libraries, as well as the fact that all Sim, _Envir, Cmdenv and Tkenv are physically separate libraries. It is_ also possible for the embedding application to assemble models from the available module types on the fly – in such cases, model topology will often come from a database. ## 2.10 Real-Time Simulation, Network Emulation Network emulation, together with real-time simulation and hardware-in-the-loop like functionality, is available because the event scheduler in the simulation kernel is pluggable too. The OMNeT++ distribution contains a demo of real-time simulation and a simplistic example of network emulation. Interfacing OMNeT++ with other simulators (hybrid operation) or HLA is also largely a matter of implementing one's own scheduler class. ## 2.11 Animation and Tracing Facility An important requirement for OMNeT++ was easy debugging and traceability of simulation models. Associated features are implemented in Tkenv, the GUI user interface of OMNeT++. _Tkenv uses three methods: automatic animation, module output_ windows and object inspectors. Automatic animation (i.e. animation without any programming) in OMNeT++ is capable of animating the flow of messages on network charts and reflecting state changes of the nodes in the display. Automatic animation perfectly fits the application area, as network simulation applications rarely need fully customizable, programmable animation capabilities. Figure 5. Screenshot of the Tkenv User Interface of OMNeT++ Simple modules may write textual debugging or tracing information to a special output stream. Such debug output appears in module output windows. It is possible to open separate windows for the output of individual modules or module groups, so compared to the traditional printf()-style debugging, module output windows make it easier to follow the execution of the simulation program. Further introspection into the simulation model is provided by _object inspectors. An object inspector is a GUI window_ associated with a simulation object. Object inspectors can be used to display the state or contents of an object in the most appropriate way (i.e. a histogram object is displayed graphically, with a histogram chart), as well as to manually modify the object. In OMNeT++, it is automatically possible to inspect every simulation object; there is no need to write additional code in the simple modules to make use of inspectors. It is also possible to turn off the graphical user interface altogether, and run the simulation as a pure command-line program. This feature is useful for batched simulation runs. ## 2.12 Visualizing Dynamic Behavior The behavior of large and complex models is usually hard to understand because of the complex interaction between different modules. OMNeT++ helps to reduce complexity by mandating the communication between modules using predefined connections. The graphical runtime environment allows the user to follow module interactions to a certain extent: one can animate, slow down or single-step the simulation, but sometimes it is still hard to see the exact sequence of the events, or to grasp the timing relationships (as, for practical reasons, simulation time is not proportional to real time; also, when single-stepping through events, events with the same timestamp get animated sequentially). OMNeT++ helps the user to visualize the interaction by logging interactions between modules to a file. This log file can be processed after (or even during) the simulation run and can be used to draw interaction diagrams. The OMNeT++ IDE has a sequence chart diagramming tool which provides a sophisticated view of how the events follow each other. One can focus on all, or ----- just selected modules, and display the interaction between them. The tool can analyze and display the causes or consequences of an event, and display all of them (using a non-linear time axis) on a single screen even if time intervals between events are of different magnitudes. One can go back and forth in time and filter for modules and events. Figure 6. Screenshot of a Sequence Chart from the OMNeT++ IDE ## 2.13 Organizing and Performing Experiments The ultimate goal of running a simulation is to obtain results and to get some insight into the system by analyzing the results. Thorough simulation studies very often produce large amounts of data, which are nontrivial to organize in a meaningful way. OMNeT++ organizes simulation runs (and the results they generate) around the following concepts: **_model – the executable (C++ model files, external libraries, etc.)_** and NED files. (INI files are considered to be part of the study and _experiment rather than the model.) Model files are considered to_ be invariant for the purposes of experimentation, meaning that if a C++ source or NED file gets modified, then it will count as a different model. **_study – a series of experiments to study some phenomenon on one_** or more models; e.g. “handover optimization for mobile IPv6”. For a study one usually performs a number of experiments from which conclusions can be drawn. One study may contain _experiments on different models, but one experiment is always_ performed on one specific model. **_experiment – exploration of a parameter space on a model, e.g._** “the `adhocNetwork` _model’s_ _behavior_ _with_ _numhosts=5,10,20,50,100 and load=2..5 step 0.1 (Cartesian_ _product)”; consists of several measurements._ **_measurement – a set of simulation runs on the same model with_** the same parameters (e.g. “numhosts=10, load=3.8”), but potentially different seeds. May consist of several replications of whose results get averaged to supply one data point for the _experiment. A measurement can be characterized with the_ parameter settings and simulation kernel settings in the INI file, minus the seeds. **_replication – one repetition of a measurement. Very often, one_** would perform several replications, all with different seeds. A replication can be characterized by the seed values it uses. **_run – or actual run: one instance of running the simulation; that_** is, a run can be characterized with an exact time/date and the computer (e.g. the host name). OMNeT++ supports the execution of whole (or partial) experiments as a single batch. After specifying the model (executable file + NED files) and the experiment parameters (in the INI file) one can further refine which measurements one is interested in. The simulation batch can be executed and its progress monitored from the IDE. Multiple CPUs or CPU cores can be exploited by letting the launcher run more than one simulation at a time. The significance of running multiple independent simulations concurrently is often overlooked, but it is not only a significantly easier way of reducing overall execution time of an experiment than distributed parallel simulation (PDES) but also more efficient (as it guarantees linear speedup which is not possible with PDES). ## 2.14 Result Analysis Analyzing the simulation result is a lengthy and time consuming process. In most cases the user wants to see the same type of data for each run of the simulation or display the same graphs for different modules in the model, so automation is very important. (The user does not want to repeat the steps of re-creating charts every time simulations have to be re-run for some reason.) The lack of automation support drives many users away from existing GUI analysis tools, and forces them to write scripts. OMNeT++ solves this by making result analysis rule-based. Simulations and series of simulations produce various result files. The user selects the input of the analysis by specifying file names or file name patterns (e.g. "adhoc-*.vec"). Data of interest can be selected into datasets by further pattern rules. The user completes datasets by adding various processing, filtering and charting steps, all using the GUI (Figure 7). Whenever the underlying files or their contents change, dataset contents and charts are recalculated. The editor only saves the "recipe" and not the actual numbers, so when simulations are re-run and so result files get replaced, charts are automatically up-to-date. Data in result files are tagged with meta information: experiment, measurement and replication labels are added to the result files to make the filtering process easy. It is possible to create very sophisticated filtering rules, for example, “all 802.11 retry counts of host[5..10] in experiment X, averaged over replications”. In addition datasets can use other datasets as their input so datasets can build on each other. Figure 7. Rule based processing ----- OMNeT++ supports several fully customizable chart and graph types which are rendered directly from datasets (Figure 8). The visual properties of the charts are also stored in the “recipe”. Figure 8. Charts in the OMNeT++ IDE ## 3. CONTRIBUTIONS TO OMNeT++ Currently there are two major network simulation model frameworks for OMNeT++: the Mobility Framework [17][18] and the INET Framework [1]. The Mobility Framework was designed at TU Berlin to provide solid foundations for creating wireless and mobile networks within OMNeT++. It provides a detailed radio model, several mobility models, MAC models including IEEE 802.11b, and several other components. Other model frameworks for mobile, ad-hoc and sensor simulations [26][33][13] have also been published (LSU SenSim [25][26] and Castalia [19][20], for example), but they have so far failed to make significant impact. Further related simulation models are NesCT for TinyOS [21] simulations, MACSimulator and Positif [13] which are continued in the MiXiM [5] project, EWsnSim, SolarLEACH, ChSim [27], AdHocSim, AntNet, etc. The INET Framework has evolved from the IPSuite originally developed at the University of Karlsruhe. It provides detailed protocol models for TCP, IPv4, IPv6, Ethernet, Ieee802.11b/g, MPLS, OSPFv4, and several other protocols. INET also includes the Quagga routing daemon directly ported from Linux code base. Several authors have developed various extensions for the INET Framework. OverSim [22][23][24] is used to model P2P protocols on top of the INET Framework. AODV-UU, DSR is also available as an add-on for the INET Framework. IPv6Suite [45] (discontinued by 2007) supported MIPv6 and HMIPv6 simulations over wired and wireless networks. The OppBSD [44] model allows using the FreeBSD kernel TCP/IP protocol stack directly inside an OMNeT++ simulation. Other published simulation models include Infiniband [28], FieldBus [14] and SimSANs [43]. A very interesting application area of OMNeT++ is the modeling of dynamic behavior of software systems based on the UML standard, by translating annotated UML diagrams into OMNeT++ models. A representative of this idea is the SYNTONY project [30][31][32]; similar approach have been reported in [35] where the authors used UML-RT, and in [34] where performance characteristics of web applications running on the JBoss Application Server were studied. The Simulation Library API can be mapped to programming languages other than C++. There is already 3[rd] party support for Java and C# which makes it possible to write simple module behavior in these languages. ## 4. COMPARISON WITH OTHER SIMULATION TOOLS The network simulation scene has changed a lot in the past ten years, simulation tools coming and going. This section presents an overview of various commercial and noncommercial network simulation tools in wide use today, and compares them to OMNeT++. Specialized network simulators (like TOSSIM, for TinyOS simulations), and simulation packages not or rarely used for network simulations (such as Ptolemy or Ptolemy II) are not considered. Also, the discussion only covers the features and services of the simulation environments themselves, but not the availability or characteristics of specific simulation models like IPv6 or QoS (the reason being that they do not form part of the OMNeT++ simulation package.) ## 4.1 NS NS-2 [11] is currently the most widely used network simulator in academic and research circles. NS-2 does not follow the same clear separation of simulation kernel and models as OMNeT++: the NS-2 distribution contains the models together with their supporting infrastructure, as one inseparable unit. This is a key difference: the NS-2 project goal is to build a network simulator, while OMNeT++ intends to provide a simulation _platform, on_ which various research groups can build their own simulation frameworks. The latter approach is what called the abundance of OMNeT++-based simulation models and model frameworks into existence, and turned OMNeT++ into a kind of an “ecosystem”. NS-2 lacks many tools and infrastructure components that OMNeT++ provides: support for hierarchical models, a graphical editor, GUI-based execution environment (except for nam), separation of models from experiments, graphical analysis tools, simulation library features such as multiple RNG streams with arbitrary mapping and result collection, seamlessly integrated parallel simulation support, etc. This is because the NS-2 project concentrates on developing the simulation models, and much less on simulation infrastructure. NS-2 is a dual-language simulator: simulation models are Tcl scripts[2], while the simulation kernel and various components (protocols, channels, agents, etc) are implemented in C++ and are made accessible from the Tcl language. Network topology is expressed as part of the Tcl script, which usually deals with several other things as well, from setting parameters to adding application behavior and recording statistics. This architecture makes it practically impossible to create graphical editors for NS-2 models[3]. NS-3 is an ongoing effort to consolidate all patches and recently developed models into a new version of NS. Although work includes refactoring of the simulation core as well, the concepts 2 In fact, OTcl, which is an object-oriented extension to Tcl. 3 Generating a Tcl script from a graphical representation is of course possible, but not the other way round: no graphical editor will ever be able to understand an arbitrary NS-2 script, and let the user edit it graphically. ----- are essentially unchanged. The NS-3 project goals [36] include some features (e.g. parallel simulation, use of real-life protocol implementations as simulation models) that have already proven to be useful with OMNeT++. ## 4.2 J-Sim J-Sim [37][38] (formerly known as JavaSim) is a componentbased, compositional simulation environment, implemented in Java. J-Sim is similar to OMNeT++ in that simulation models are hierarchical and built from self-contained components, but the approach of assembling components into models is more like NS-2: J-Sim is also a dual-language simulation environment, in which classes are written in Java, and glued together using Tcl (or Java). The use of Tcl in J-Sim has the same drawback as with NS-2: it makes implementing graphical editors impossible. In fact, J-Sim does provide a graphical editor (gEditor), but its native format is XML. Although gEditor can export Tcl scripts, developers recommend that XML files are directly loaded into the simulator, bypassing Tcl. This way, XML becomes the equivalent of OMNeT++ NED. However, the problem with XML as native file format is that it is hard to read and write by humans. Simulation models are provided in the Inet package, which contains IPv4, TCP, MPLS and other protocol models. The fact that J-Sim is Java-based has some implications. On one hand, model development and debugging can be significantly faster than C++, due to existence of excellent Java development tools. However, simulation performance is significantly weaker than with C++, and it is also not possible to reuse existing real-life protocol implementations written in C as simulation models. (The feasibility and usefulness of the latter has been demonstrated with OMNeT++, where simulation models include port of the Quagga Linux routing daemon, the TCP stack from the FreeBSD kernel, the port of the UU-AODV routing package, etc. The NS-3 team has similar plans as well.) Development of the J-Sim core and simulation models seem to have stalled after 2004 when version 1.3 was published; later entries on the web site are patches and contributed documents only. There are no independent (3rd party) simulation models for J-Sim. ## 4.3 SSFNet SSFNet [39] (Scalable Simulation Framework) is defined as a “public-domain standard for discrete-event simulation of large, complex systems in Java and C++.” The SSFNet standard defines a minimalist API (which, however, was designed with parallel simulation in mind). The topology and configuration of SSFNet simulations are given in DML files. DML is a text-based format comparable to XML, but has its own syntax. DML can be considered the SSFNet equivalent of NED, however it lacks expressing power and features to scale up to support large model frameworks built from reusable components. SSFNet also lacks OMNeT++'s INI files, all parameters need to be given in the DML. SSFNet has four implementations: DaSSF and CSSF in C++, and two Java implementations (Renesys Raceway and JSSF). There were significantly more simulation models developed for the Java versions than for DaSSF. Advantages and disadvantages of using Java in SSFNet are the same as discussed with J-Sim. As with J-Sim, development of the SSFNet simulation framework and models seem to have stalled after 2004 (date of the SSFNet for Java 2.20 release), and little activity can be detected outside the main web site as well. ## 4.4 JiST and SWANS JiST [42][6] represents a very interesting approach to building a high performance Java based simulation environment. It modifies the Java Virtual Machine to run the programs in simulation time instead of real time. JiST is basically just a simulation kernel, and as such, it lacks most of the features present in the OMNeT++ package. SWANS is a scalable wireless network simulator built atop the JiST platform as a proof of concept model, to prove the efficiency of the virtual machine based approach. It appears that no further simulation models have been created by the JiST team or independent groups. Development of JiST/SWANS seems to be halted after 2005. ## 4.5 OPNET Modeler OPNET Modeler is the flagship product of OPNET Technologies Inc. [16]. OPNET Modeler is a commercial product which is freely available worldwide to qualifying universities. OPNET has probably the largest selection of ready-made protocol models (including IPv6, MIPv6, WiMAX, QoS, Ethernet, MPLS, OSPFv3 and many others). OPNET and OMNeT++ provide rich simulation libraries of roughly comparable functionalities. The OPNET simulation library is based on C, while the one in OMNeT++ is a C++ class library. OPNET's architecture is similar to OMNeT++ as it allows hierarchical models with arbitrarily deep nesting (like OMNeT++), but with some restrictions (namely, the "node" level cannot be hierarchical). A significant difference from OMNeT++ is that OPNET models are always of fixed topology, while OMNeT++'s NED and its graphical editor allow parametric topologies. In OPNET, the preferred way of defining network topology is by using the graphical editor. The editor stores models in a proprietary binary file format, which means in practice that OPNET models are usually difficult to generate by program (it requires writing a C program that uses an OPNET API, while OMNeT++ models are simple text files which can be generated e.g. with Perl). Both OPNET and OMNeT++ provide a graphical debugger and some form of automatic animation which is essential for easy model development. OPNET does not provide source code to the simulation kernel (although it ships with the sources of the protocol models). OMNeT++ – like NS-2 and most other non-commercial tools – is fully public-source allowing much easier source level debugging. OPNET's main advantage over OMNeT++ is definitely its large protocol model library, while its closed nature (proprietary binary file formats and the lack of source code) makes development and problem solving harder. ## 4.6 Qualnet Qualnet [41] is a commercial simulation environment mainly for wireless networks, which has a significant client base in the ----- military. Qualnet has evolved from the Parsec parallel simulation “language”[4] [12] developed at the UCLA Parallel Computing Laboratory (PCL), and the GloMoSim (Global Mobile system Simulation) model written on top of Parsec. The Parsec language divides the simulation model into entities, and provides a minimalistic simulation API (timers, etc) for them. Entities are implemented with coroutines. Because coroutine CPU stacks require relatively large amounts of memory (the manual recommends reserving 200KByte each), it is rarely feasible to map the natural units of the simulation (say, hosts and routers, or protocols) one-to-one onto entities. What GloMoSim and Qualnet models do is implement the equivalent of the OMNeT++ model structure in model space, above the Parsec runtime. The Parsec kernel is only used to provide event scheduling and parallel simulation services. Parsec provides a very efficient parallel simulation infrastructure, and models (GloMoSim and Qualnet simulation models) have been written with parallel execution in mind[5], resulting in an excellent parallel performance for wireless network simulations. ## 4.7 Summary In this section we have examined the simulation packages most relevant for analysis of telecommunication networks, and compared them to OMNeT++. NS-2 is still the most widely used network simulator in the Academia, but it lacks much of the infrastructure provided by OMNeT++. The other three open-source network simulation packages examined (J-Sim, SSFNet and JiST/SWANS), have failed to gain significant acceptance, and their project web pages indicate near inactivity since 2004. We have examined two commercial products as well. Qualnet emphasizes wireless simulations. OPNET has similar foundations as OMNeT++, but ships with an extensive model library and provides several additional programs and GUI tools. ## 5. CONCLUSIONS In this paper we presented an overview of the OMNeT++ discrete event simulation platform, designed to support the simulation of telecommunication networks and other parallel and distributed systems. The OMNeT++ approach significantly differs from that of NS-2, the most widely used network simulator in academic and research circles: while the NS-2 (and NS-3) project goal is to build a network simulator, OMNeT++ aims at providing a rich simulation platform, and leaves creating simulation models to independent research groups. The last ten years have shown that the OMNeT++ approach is viable, and several OMNeT++-based open-source simulation models and model frameworks have been published by various research groups and individuals. ## 6. REFERENCES [1] OMNeT++ Home Page. http://www.omnetpp.org [accessed on September, 2007] [2] Varga, A. 2001. The OMNeT++ Discrete Event Simulation System. In the Proceedings of the European Simulation Multiconference (ESM2001. June 6-9, 2001. Prague, Czech Republic). 4 It extends the C language with some constructs, and Parsec programs are translated into C before compilation. 5 Lookahead annotations, avoiding central components, etc. [3] Kaage, U., V. Kahmann, F. Jondral. 2001. An OMNeT++ TCP Model. To appear in Proceedings of the European _Simulation Multiconference (ESM 2001), June 7-9, Prague._ [4] Wehrle, K, J. Reber, V. Kahmann. 2001. “A Simulation Suite for Internet Nodes with the Ability to Integrate Arbitrary Quality of Service Behavior”. In Proceedings of _the Communication Networks and Distributed Systems_ _Modeling and Simulation Conference 2001, Phoenix (AZ),_ USA, January 7-11. [5] [MiXiM home page. http://sourceforge.net/projects/mixim/](http://sourceforge.net/projects/mixim/) [accessed on September, 2007] [6] [JiST home page. http://jist.ece.cornell.edu [accessed on](http://jist.ece.cornell.edu/) September, 2007] [7] Varga, A. and Gy. Pongor. 1997. Flexible Topology Description Language for Simulation Programs. In _Proceedings of the 9th European Simulation Symposium_ _(ESS'97), pp.225-229, Passau, Germany, October 19-22._ [8] Varga, A and B. Fakhamzadeh. 1997. The K-Split Algorithm for the PDF Approximation of MultiDimensional Empirical Distributions without Storing Observations. In Proc. of the 9th European Simulation _Symposium (ESS'97), pp.94-98. October 19-22, Passau,_ Germany. [9] Varga, A. 1998. Parameterized Topologies for Simulation Programs. In Proceedings of the Western Multiconference _on Simulation (WMC'98), Communication Networks and_ _Distributed Systems (CNDS'98). San Diego, CA, January_ 11-14. [10] Jain, R, and I. Chlamtac. 1985. The P[2] Algorithm for Dynamic Calculation of Quantiles and Histograms Without Storing Observations. Communications of the ACM, 28, no. 10 (Oct.): 1076-1085. [11] Bajaj, S., L. Breslau, D. Estrin, K. Fall, S. Floyd, P. Haldar, M. Handley, A. Helmy, J. Heidemann, P. Huang, S. Kumar, S. McCanne, R. Rejaie, P. Sharma, K. Varadhan, Y. Xu, H. Yu and D. Zappala. 2000. Improving simulation for network research. _IEEE Computer. (to appear, a preliminary draft is_ currently available as USC technical report 99-702) [12] Bagrodia, R, R. Meyer, M. Takai, Y. Chen, X. Zeng, J. Martin, B. Park, H. Song. 1998. Parsec: A Parallel Simulation Environment for Complex Systems. Computer, Vol. 31(10), October, pp. 77-85. [13] Consensus home page. [http://www.consensus.tudelft.nl/software.html [accessed on](http://www.consensus.tudelft.nl/software.html) September, 2007] [14] FieldBus home page. [http://developer.berlios.de/projects/fieldbus [accessed on](http://developer.berlios.de/projects/fieldbus) September, 2007] [15] Davis, J, M. Goel, C. Hylands, B. Kienhuis, E.A. Lee, J. Liu, X. Liu, L. Muliadi, S. Neuendorffer, J. Reekie, N. Smyth, J. Tsay and Y. Xiong. 1999. Overview of the Ptolemy Project. ERL Technical Report UCB/ERL No. M99/37, Dept. EECS, University of California, Berkeley, CA 94720, July. [16] OPNET Technologies, Inc. OPNET Modeler. [http://www.opnet.com [accessed on September, 2007]](http://www.opnet.com/) [[17] Mobility Framework. http://mobility-fw.sourceforge.net](http://mobility-fw.sourceforge.net/) [accessed on September, 2007] [18] W. Drytkiewicz, S. Sroka, V. Handziski, A. Koepke, and H. Karl, A Mobility Framework for OMNeT++. 2003. 3rd International OMNeT++ Workshop (Budapest University of Technology and Economics, Department of ----- Telecommunications Budapest, Hungary, January 2003). [http://www.tkn.tu-berlin.de/~koepke/](http://www.tkn.tu-berlin.de/~koepke/) [19] D. Pediaditakis, S. H. Mohajerani, and A. Boulis. 2007. Poster Abstract: Castalia: the Difference of Accurate Simulation in WSN. 4th European Conference on Wireless Sensor Networks, (Delft, The Netherlands, 29-31 January 2007). [20] Castalia: A Simulator for WSN. [http://castalia.npc.nicta.com.au. [accessed on September,](http://castalia.npc.nicta.com.au/) 2007] [[21] NesCT: A language translator. http://nesct.sourceforge.net](http://nesct.sourceforge.net/) [accessed on September, 2007] [22] OverSim:The Overlay Simulation Framework [http://www.oversim.org [accessed on September, 2007]](http://www.oversim.org/) [23] Ingmar Baumgart and Bernhard Heep and Stephan Krause. 2007. OverSim: A Flexible Overlay Network Simulation Framework. Proceedings of 10th IEEE Global Internet Symposium (May, 2007). p.79-84. [24] Ingmar Baumgart and Bernhard Heep and Stephan Krause. 2007. A P2PSIP Demonstrator Powered by OverSim. Proceedings of 7th IEEE International Conference on Peerto-Peer Computing (P2P2007, Galway, Ireland. Sep, 2007). pp. 243-244, [25] C. Mallanda, A. Suri, V. Kunchakarra, S.S. Iyengar, R. Kannan, A. Durresi, and S. Sastry. 2005. Simulating Wireless Sensor Networks with OMNeT++, submitted to IEEE Computer, 2005 [http://csc.lsu.edu/sensor_web/publications.html](http://csc.lsu.edu/sensor_web/publications.html) [[26] Sensor Simulator. http://csc.lsu.edu/sensor_web [accessed](http://csc.lsu.edu/sensor_web) on September, 2007] [27] S. Valentin. 2006. ChSim - A wireless channel simulator for OMNeT++, (TKN TU Berlin Simulation workshop, Sep. [2006) http://www.cs.uni-paderborn.de/en/research-](http://www.cs.uni-paderborn.de/en/research-group/research-group-computer-networks/projects/chsim.html) [group/research-group-computer-](http://www.cs.uni-paderborn.de/en/research-group/research-group-computer-networks/projects/chsim.html) [networks/projects/chsim.html](http://www.cs.uni-paderborn.de/en/research-group/research-group-computer-networks/projects/chsim.html) [28] Mellanox Technologies: InfiniBand model: [http://www.omnetpp.org/filemgmt/singlefile.php?lid=133](http://www.omnetpp.org/filemgmt/singlefile.php?lid=133) [29] I. Dietrich, C. Sommer, F. Dressler. Simulating DYMO in OMNeT++. Erlangen-Nürnberg : Friedrich-AlexanderUniversität. 2007 Internal report. [30] Isabel Dietrich, Volker Schmitt, Falko Dressler and Reinhard German, 2007. "SYNTONY: Network Protocol Simulation based on Standard-conform UML 2 Models," Proceedings of 1st ACM International Workshop on Network Simulation Tools (NSTools 2007), Nantes, France, October 2007. [31] I. Dietrich, C. Sommer, F. Dressler, and R. German. 2007. Automated Simulation of Communication Protocols Modeled in UML 2 with Syntony. Proceedings of GI/ITG Workshop Leistungs-, Zuverlässigkeits- und Verlässlichkeitsbewertung von Kommunikationsnetzen und verteilten Systemen (MMBnet 2007), Hamburg, Germany, September 2007. pp. 104-115. [[32] Syntony home page. http://www7.informatik.uni-](http://www7.informatik.uni-erlangen.de/syntony) [erlangen.de/syntony](http://www7.informatik.uni-erlangen.de/syntony) [accessed on September, 2007] [33] Feng Chen, Nan Wang, Reinhard German and Falko Dressler, 2008. "Performance Evaluation of IEEE 802.15.4 LR-WPAN for Industrial Applications," Proceedings of 5th IEEE/IFIP Conference on Wireless On demand Network Systems and Services (IEEE/IFIP WONS 2008), GarmischPartenkirchen, Germany, January 2008. [34] A. Hennig, D. Revill and M. Pönitsch. 2003. From UML to Performance Measures - Simulative Performance Predictions of IT-Systems using the JBoss Application Server with OMNET++. Proceedings of ESS2003 conference. Siemens AG, Corporate Technology, CT SE 1. [35] Michael, J. B., Shing, M., Miklaski, M. H., and Babbitt, J. D. 2004. Modeling and Simulation of System-of-Systems Timing Constraints with UML-RT and OMNeT++. In Proceedings of the 15th IEEE international Workshop on Rapid System Prototyping (Rsp'04) - Volume 00 (June 28 30, 2004). RSP. IEEE Computer Society, Washington, DC, 202-209. DOI= http://dx.doi.org/10.1109/RSP.2004.30. [36] T. R. Henderson, S. Roy, S. Floyd, G. F. Riley. ns3 Project Goals. WNS2 ns-2: The IP Network Simulator, Pisa, Italy Oct. 10, 2006. [http://www.nsnam.org/docs/meetings/wns2/wns2-ns3.pdf](http://www.nsnam.org/docs/meetings/wns2/wns2-ns3.pdf) [37] Ahmed Sobeih, Wei-Peng Chen, Jennifer C. Hou, Lu-Chuan Kung, Ning Li, Hyuk Lim, Hung-Ying Tyan, and Honghai Zhang,.J-Sim: a simulation and emulation environment for wireless sensor networks. IEEE Wireless Communications Magazine, Vol. 13, No. 4, pp. 104--119, August 2006. [[38] J-SIM home page: http://www.j-sim.org [accessed on](http://www.j-sim.org/) [September, 2007]](http://www.j-sim.org/) [39] Cowie, J. H., Nicol, D. M., and Ogielski, A. T. 1999. Modeling the Global Internet. Computing in Science and Engg. 1, 1 (Jan. 1999), 42-50. [DOI=http://dx.doi.org/10.1109/5992.743621](http://dx.doi.org/10.1109/5992.743621) [40] X. Zeng, R. Bagrodia, M. Gerla.GloMoSim: a Library for Parallel Simulation of Large-scale Wireless Networks. PADS '98, May 26-29, 1998 in Banff, Alberta, Canada. [[41] Qualnet home page: http://www.qualnet.com [accessed on](http://www.qualnet.com/) September, 2007] [42] R. Barr, Z. J. Haas, R. van Renesse. 2004. JiST: Embedding Simulation Time into a Virtual Machine. Proceedings of EuroSim Congress on Modelling and Simulation, September 2004. Computer Science and Electrical Engineering, Cornell University, Ithaca NY 14853. [[43] SimSAN home page. http://simsan.storwav.com/ [accessed](http://simsan.storwav.com/) on September, 2007] [44] OppBSD home page. [https://projekte.tm.uka.de/trac/OppBSD [accessed on](https://projekte.tm.uka.de/trac/OppBSD) September, 2007] [45] E. Wu, S. Woon, J. Lai and Y. A. Sekercioglu, 2005. "IPv6Suite: A Simulation Tool for Modeling Protocols of the Next Generation Internet", In Proceedings of the Third International Conference on Information Technology: Research and Education (ITRE 2005), June 2005, Taiwan. [46] Zeigler, B. 1990. Object-oriented Simulation with Hierarchical, Modular Models. Academic Press, 1990. [47] Chow, A and Zeigler, B. 1994. Revised DEVS: A Parallel, Hierarchical, Modular Modeling Formalism. In Proceedings _of the Winter Simulation conference 1994._ -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1145/1416222.1416290?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1145/1416222.1416290, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "http://eudl.eu/pdf/10.4108/ICST.SIMUTOOLS2008.3027" }
2,008
[ "JournalArticle", "Review" ]
true
2008-03-03T00:00:00
[]
14,675
en
[ { "category": "Business", "source": "s2-fos-model" }, { "category": "Economics", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/012e3df4305cbe5ca918f6f7f87aa80a93593662
[]
0.880203
The importance of financial management in small and medium-sized enterprises (SMEs): an analysis of challenges and best practices
012e3df4305cbe5ca918f6f7f87aa80a93593662
Technology audit and production reserves
[ { "authorId": "2260075770", "name": "Eugine Nkwinika" }, { "authorId": "151429448", "name": "Segun Akinola" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
The object of research is the importance of monetary management in Small and Medium-sized Enterprises (SMEs), specializing in challenges, best practices, and future trends. Financial management in SMEs is an important aspect that influences their growth, sustainability, and competitiveness. The paper begins by defining SMEs and highlighting the significance of financial management for their success. It emphasizes the need for SME owners to understand financial concepts, make informed decisions, and prioritize financial planning to ensure sound business operations. Insights from real-world case studies showcase successful financial management practices adopted by SMEs. Government policies and support for SME financial management are also explored, with a focus on initiatives, tax incentives, and access to financial advisory services. These government interventions play a crucial position in empowering SMEs with the necessary sources and steerage for powerful financial management. Moreover, the evaluation delves into destiny developments, such as rising technology (AI, blockchain, IoT) and regulatory adjustments, and their capacity impact on economic management for SMEs. It discusses the challenges and possibilities in monetary forecasting, highlighting using information analytics and predictive modeling for improved accuracy. In conclusion, this review assessment underscores the significance of financial control for SMEs, emphasizing the want for monetary literacy, era adoption, and compliance with regulatory adjustments. By embracing first-class practices and authorities’ help, SMEs can reap long-term financial balance and thrive in dynamic commercial enterprise environments. As SMEs preserve to evolve within digital technology, powerful economic control remains vital for his or her sustainable increase and achievement.
#### ECONOMICS OF ENTERPRISES: ECONOMICS AND MANAGEMENT OF ENTERPRISE ISSN 2664-9969 ### UDC 336.02:336.64 JEL Classification: G32 DOI: 10.15587/2706-5448.2023.285749 ## Eugine Nkwinika, Segun Akinola # THE IMPORTANCE OF FINANCIAL MANAGEMENT IN SMALL AND MEDIUM-SIZED ENTERPRISES (SMEs): AN ANALYSIS OF CHALLENGES AND BEST PRACTICES #### The object of research is the importance of monetary management in Small and Medium-sized Enterprises (SMEs), specializing in challenges, best practices, and future trends. Financial management in SMEs is an important aspect that influences their growth, sustainability, and competitiveness. The paper begins by defining SMEs and highlighting the significance of financial management for their success. It emphasizes the need for SME owners to understand financial concepts, make informed decisions, and prioritize financial planning to ensure sound business opera­ tions. Insights from real-world case studies showcase successful financial management practices adopted by SMEs. Government policies and support for SME financial management are also explored, with a focus on initiatives, tax incentives, and access to financial advisory services. These government interventions play a crucial position in empowering SMEs with the necessary sources and steerage for powerful financial management. Moreover, the evaluation delves into destiny developments, such as rising technology (AI, blockchain, IoT) and regulatory adjustments, and their capacity impact on economic management for SMEs. It discusses the challenges and possibilities in monetary forecasting, highlighting using information analytics and predictive modeling for improved accuracy. In conclusion, this review assessment underscores the significance of financial control for SMEs, emphasizing the want for monetary literacy, era adoption, and compliance with regulatory adjustments. By embracing first- class practices and authorities’ help, SMEs can reap long-term financial balance and thrive in dynamic commercial enterprise environments. As SMEs preserve to evolve within digital technology, powerful economic control remains vital for his or her sustainable increase and achievement. Keywords: financial management, financial literacy, cashflow management, financial risk management, finan­ cial technology, financial resources. _Received date: 08.08.2023_ _Accepted date: 22.09.2023_ _Published date: 28.09.2023_ **_How to cite_** _© The Author(s) 2023_ _This is an open access article_ _under the Creative Commons CC BY license_ _Nkwinika, E., Akinola, S. (2023). The importance of financial management in small and medium-sized enterprises (SMEs): an analysis of challenges and_ _best practices. Technology Audit and Production Reserves, 5 (4 (73)), 12–20. doi: https://doi.org/10.15587/2706-5448.2023.285749_ ### 1. Introduction Small and Medium-sized Enterprises (SMEs) play a cru­ cial position within the worldwide economic system, riding innovation, employment, and economic growth [1]. The definition of SMEs varies across international locations, however, in general, they’re characterized via their es­ pecially small scale and constrained assets compared to larger companies. SMEs frequently face unique challenges in managing their economic affairs, making monetary control a vital aspect of their sustainability and success [2]. The importance of financial management in SMEs cannot be overstated. Efficient financial control is crucial for opti­ mizing useful resource allocation, making sure of liquidity, and improving the general economic performance of those firms [3]. Effective monetary management practices enable SMEs to make informed selections, control risks, and seize boom opportunities [4]. It additionally contributes to build­ ing investor confidence, attracting external funding, and preserving a competitive area inside the market [5]. The aim _of this study is to explore the demanding situations and best_ practices associated with financial management in SMEs. By identifying and knowledge of those key elements, re­ searchers, policymakers, and commercial enterprise prac­ titioners can gain treasured insights into the dynamics of economic management in SMEs and formulate strategies to enhance their economic fitness and sustainability. ### 2. Materials and Methods This overview paper focuses on the economic manage­ ment practices of SMEs throughout various industries and **12** TECHNOLOGY AUDIT AND PRODUCTION RESERVES — №5/4(73) 2023 ----- ISSN 2664-9969 geographical areas. Its goals are to analyze the demand­ ing situations that SMEs stumble upon in dealing with their price range and to provide an in-depth examination of high-quality practices followed with the aid of successful SMEs. The review will embody a complete evaluation of applicable academic literature, empirical research, and reviews from legitimate assets. The examination will encompass subjects related to monetary planning, budgeting, coins go with the flow control, financ­ ing options, danger control, and financial reporting, among others. Furthermore, the review will assess the effect of macroeconomic factors and government policies on the economic control of SMEs. Understanding the outside influences that have an effect on SMEs’ economic decision-making will offer a broader con­ text for the challenges and opportunities they face. The insights derived from this evaluation can function as a foundation for destiny research on the subject of SME financial management. Moreover, policymakers and practitioners can benefit from the proof-primarily based recommendations to devise powerful support mechanisms and rules that foster the financial growth and stability of SMEs. In the end, this overview paper aims to shed mild on the importance of financial management in SMEs and offer a complete assessment of the challenges and nice practices in this area. By knowledge of the intricacies of financial control in SMEs, stakeholders can make a contribution to enhancing their monetary resilience, thereby promoting financial development and prosperity. ### 3. Results and Discussion 3.1. The role of financial management in SMEs Financial control is a fundamental factor of SMEs’ operations, encompassing the tactics of planning, orga­ nizing, controlling, and directing the economic assets to attain the company’s objectives [6]. It includes studying financial information, making informed choices, and impos­ ing strategies to ensure the efficient utilization of funds. SMEs face precise challenges because of their limited as­ sets and exposure to marketplace uncertainties, making financial control all the extra important for sustenance and boom [7]. 3.1.1. Financial management defined. One of the few primary functional areas of management, financial manage­ ment is often regarded as the foundation for the growth and success of any enterprise. Activities aimed at managing a business’s finances in order to meet its financial goals are included in financial management [8]. The definition of financial management is based on how funding sources are mobilized and used. Financial management is important for obtaining the cash needed to finance an enterprise’s assets and commercial operations, distributing funds among competing uses, and ensuring effective and efficient use of funds in order to accomplish the organization’s main aim and goal. Although one of several functional areas of management, financial management is crucial to the success of SMEs. The primary function and location of financial management in relation to other niche areas of business management are the emphasis of this definition [9]. #### ECONOMICS OF ENTERPRISES: ECONOMICS AND MANAGEMENT OF ENTERPRISE Fig. 1 illustrates the primary function and position of financial management in relation to particular company management domains. Fig. 1. The central position and role of financial management [10] 3.1.2. Significance of financial management for SMEs. Financial management has a favorable impact on com­ petitiveness, the sufficiency of company records, and the survival of SMEs. Business records are crucial to SMEs’ existence and their ability to acquire funding from investors and/or financial institutions. Poor cash management has an impact on the financial position and liquidity of businesses, and financial management helps to eliminate it. In order for SMEs to maintain liquidity and continue operating, financial management is crucial. Financial management gives SME owners the information and foresight necessary to anticipate future cash flow issues, which are essential for the sustainability of the company [11]. Allocating financial resources with the use of financial management increases the likelihood that a business will survive. 3.2. Challenges in financial management for SMEs 3.2.1. Limited financial resources and capital. Resource limitations make it difficult for SMEs to innovate because they can’t afford to experiment, which is essential for the creation of new items. The shortage of resources faced by SMEs also prevents the creation of routines and organiza­ tional structures that would be helpful in attracting and developing human potential as well as improving corpo­ rate operations. Compared to individuals working for large companies, SMEs are often regarded to be more financially limited and have less access to formal financing (long-term loans) [12]. SMEs frequently view access to finance as one of the major obstacles keeping them from operating effectively in both developed and developing nations. A key factor in the growth and development of African businesses is hav­ ing enough access to financing. The lack of financial assets and capital is one of the biggest problems SMEs encounter while managing their finances. Unlike huge businesses, SMEs frequently operate on a smaller scale with limited access to resources. This obstacle makes it difficult for them to expand their operations, invest in research and development, and explore new opportunities. As a result, SMEs must carefully prioritize their financial obligations and use price-effective strategies to make the most of their limited resources [13]. 3.2.2. Access to financing and credit. Giving SMEs credit encourages and boosts economic growth; more credit TECHNOLOGY AUDIT AND PRODUCTION RESERVES — №5/4(73) 2023 **13** ----- #### ECONOMICS OF ENTERPRISES: ECONOMICS AND MANAGEMENT OF ENTERPRISE encour­ages entrepreneurship in the form of increased firm formation and expansion [14]. Comparing Middle Eastern and Central Asian countries at comparable economic de­ velopment levels to other nations, the biggest barrier to SME access to funding is in these regions. The failure of SMEs to recognize the value of having appropriate fi­ nance and inadequate control of net working capital are examples of related concepts [15]. Despite the SMEs’ cru­ cial contribution to the socioeconomic development of the nation, it has become difficult for them to secure sources of short- to long-term, flexible finance. There are a number of reasons why SMEs do not have access to credit and finance, including high-risk lending to SMEs, information asymmetry caused by SMEs lending, high administrative transaction costs associated with SMEs financing, and weak institutional and legal structures [16]. 3.2.3. Cash flow management. Most SMEs struggle with inadequate cash management since there are no cash bud­ gets in place to anticipate future cash flow issues and other financial concerns [17]. The majority of SMEs are unable to distinguish between cash flow and profit. SMEs prioritize profit over cash flow, which is essential for the long-term viability of the company. When a company’s cash flow balance is declared to be positive, it immediately makes substantial purchases without thinking about the post-dated cheques that were issued or the payments that would need to be made later. Businesses only realize they lack the money afterward to meet their responsibilities. SMEs find it difficult to track inflows and outflows of cash because of inadequate cash budgeting and a lack of a business bank account [18]. SMEs concentrate on in­ creasing sales and lowering inventories without taking into account the fact that rising sales are also accompanied by rising debtor levels. If payments are not received as specified in the invoice, the increase in debtors may result in liquidity restrictions. SMEs face a number of difficul­ ties, including bad debts and difficulty paying creditors, which have an impact on the cash flows, profit margins, and liquidity of the company [19]. Operating costs will increase as a result of the difficulties in collecting money from debtors, and the increase in operating costs will af­ fect cash flow. The difficulty of collecting payment from debtors causes an increase in write-off revenue, which places financial limits on the business because of variable expenses and unrecoverable inventories [20]. 3.2.4. Financial risk management. A key component of financial management for SMEs is managing financial risk. Among these dangers might be market risk, which is in­ fluenced by a number of factors with an emphasis on the level of market competition as a whole. Market risk is the strategic risk that SMEs face when it comes to the longterm retention of current consumers, the acquisition and retention of new clients, the creation of novel products, or the provision of novel services [21]. The only thing that enables SMEs to execute a suitable sales volume that allows them to sustain their market position is a sufficient number of clients. The two primary components of the competitive environment-customers and competitors – have a significant impact on how competitive a company is. The ability to manage finite resources that are challenging to replace gives organizations a competitive advantage, which SMEs must develop if they are to thrive. Among ISSN 2664-9969 them is human capital [22]. To succeed and develop, as well as meet customer demand, businesses must be able to innovate new items. Any sort of company’s principal objective is to maximize business performance, and man­ agers place high importance on this goal. Because risk is viewed as a crucial component of a company’s financial management, the amount of financial risk must be evalua­ ted in terms of how well a firm manages risk in order to make successful financial risk management decisions. Financial risk is one of the main threats facing SMEs [23]. The main indicators of SME financial risk are challenges in business financing and a lack of funds because the ma­ jority of a company’s operations are funded by the capital of the owners or managers themselves. This could result in a rise in operational costs and corporate debt due to worries about debt repayment and the ensuing high finan­ cial risk. Access to capital is anticipated to raise the bar for a business environment by motivating organizations to pursue more fruitful business prospects. Competitive advantage and internal SME capabilities have a positive and significant correlation [24]. For SMEs that have been in operation for under five years, this rela­ tionship is weaker than it is for more seasoned companies. Innovation has a significant and lasting effect on the com­ pany’s competitiveness via increasing productivity. As levels of customer happiness and loyalty climb, so does support for the purchasing procedures. One of the key signs that could affect the tighter business environment is support from suppliers and customers in the commercial sector. In a more constricted business climate, the help of corporate clients has a greater impact. Improved risk management is not always a result of long-term relationships between SMEs and their suppliers. Some of the main reasons for business failure include a lack of management planning activities, a lack of working capital, offering customers too much credit, failure to implement rapid outsourcing, market competition, and insufficient monitoring of corporate finances [25]. Other factors may also contribute to an organization’s failure. The risk of failure decreases as a manager age and managerial ownership becomes clear. However, if larger management boards and managers are present in other organizations, the likelihood of failure will increase. 3.3. Best practices in financial management for SMEs Financial management provides SMEs with essential financial competencies, such as knowledge, attitude, and awareness [26]. The ability to make financial decisions grows as a result of the understanding of the financial markets. The knowledge competency helps the owners of SMEs to successfully balance their assets and liabilities, which is a requirement for business liquidity, and for an appropriate financial history, which is essential for the inability to ob­ tain external finance. Owners of SMEs must consider their attitude when deciding whether to take risks. Owners of small and medium-sized businesses can efficiently allocate financial resources to initiatives with higher risks and have expertise in how to use those resources. Financial manage­ ment is crucial for risk management risk diversification, aiding insufficient financial mixing, and financial mana­ gement. Analyzing firm financial circumstances and managing financial resources are made easier with awareness. Financial management gives SME owners the knowledge and foresight they need to anticipate future cash flow issues, which is essential to their ability to stay in business [27]. Poor cash **14** TECHNOLOGY AUDIT AND PRODUCTION RESERVES — №5/4(73) 2023 ----- ISSN 2664-9969 management has an impact on the financial position and liquidity of the organization, and financial management helps to eliminate it. Financial management serves as an analytical tool for future sales estimates, the assets needed to meet demand in the future, and operational costs. Fi­ nancial management gives SME owners in-depth knowledge of the relationship between the supplier chain, production process, and operational costs. The owners of SMEs who are trained in financial management are able to create a connection between costs and appropriate activity as well as effective cash flow management. Monitoring cash inflows and outflows is crucial to the management of SMEs with constrained financial resources [28]. The tracking of cash inflows and outflows provides SMEs with the neces­ sary data to assess their company’s competitiveness and provides a means of ensuring their survival during the first year of operation. 3.4. Technology and financial management for SMEs The application of technology assists SMEs to remain competitive and plays a vital role in financial management and sustainability. The financial management of SMEs is related to many new developments in different ways [10]. An association between SMEs’ financial management and innovation has been found in earlier studies. The impact of innovation on an SME’s financial management can be demonstrated using both financial and non-financial metrics. Some advantages of innovation include the capacity for competition, financial accessibility, connectivity, communica­ tion, marketing, and export success. However, other com­ mentators hold a different point of view. It has also been suggested that ignoring innovation’s potential downsides can ultimately have a negative effect on the environment and lead to uncontrollable firm expansion. Despite wor­ ries about possible negative effects, there is a wealth of research showing that innovation has a positive impact on SMEs’ financial management [29]. 3.4.1. Role of financial software and tools. Financial software and equipment are essential for optimizing financial management procedures for SMEs [30]. These technologies encompass a wide range of applications, such as account­ ing software, budgeting tools, financial analytics platforms, and cash flow control systems. Financial software saves time and effort by automating repetitive operations, en­ abling SMEs to concentrate on making strategic decisions. Real-time access to financial records made possible by financial software offers better financial reporting and analysis [31]. Giving SMEs the knowledge they need to make wise business decisions, it provides insights into important financial indicators including revenue, costs, and profitability. Additionally, these systems usually come with capabilities that can be altered to accommodate particular demands of SMEs, offering flexibility and scalability as organizations grow. 3.4.2. Fintech solutions for SMEs. FinTech makes it simple for SMEs to get finance that will enable them to expand their businesses [32]. FinTech firms are essen­ tial for providing SMEs with financial support. FinTech today provides more than simply capital finance; it also includes a wide range of other services including digital payments and financial mechanisms. FinTech is essential in enhancing SMEs’ success since it increases operational #### ECONOMICS OF ENTERPRISES: ECONOMICS AND MANAGEMENT OF ENTERPRISE efficiency. By providing services like non-cash transactions utilizing applications, fintech lowers operating expenses by relieving firms of bank administrative fees. Furthermore, non-collateral loans will give business owners easier access to finance. Financial technology considerably and favo­ rably affects the asset value and capital growth of SMEs. FinTech, however, has no appreciable effects on financial inclusion and stability [33]. 3.4.3. Benefits and challenges of adopting financial tech­ nology. For SMEs, using financial technology has a number of advantages. First off, by automating manual financial operations, it improves efficiency and production. By re­ ducing processes, errors are less likely to occur and SMEs are given more time to concentrate on their main business operations. The accuracy of financial data and financial transparency are also improved by financial technology. It is simpler to spot potential financial hazards and op­ portunities when SMEs have access to real-time data that keeps them informed about their financial performance [34]. Thirdly, cost savings are offered by fintech solutions to SMEs. Paperless operations and the usage of digital plat­ forms lower administrative expenses and improve overall cost-effectiveness. Adopting financial technology, though, also poses difficulties for SMEs. The initial cost associated with implementing and integrating these technologies is one of the main obstacles. SMEs may find it challenging to invest in new systems and receive the requisite training to operate them efficiently [35]. The digitization of financial data and transactions also raises cybersecurity issues. To safeguard confidential financial information from potential online dangers, SMEs must give data security first priority. And last, some SMEs can experience opposition to change from staff members accustomed to conventional finance procedures. Successful technology adoption depends on appropriate training and change management techniques. Finally, as technology and financial management merge more and more, SMEs have access to a wide range of tools and solutions to improve their financial management procedures. With the automation, accessibility, and costeffectiveness that financial software and fintech solutions offer, SMEs may make better decisions, increase transpa­ rency, and enhance their financial outcomes [5]. Although using new technology might be difficult, the advantages outweigh the disadvantages, setting up SMEs for greater financial success and growth in the digital age. 3.5. Financial literacy for SME owners Financial literacy is regarded on a global scale as the basic determinant in business performance, growth, and financial efficiency for SMEs. Financial strategies are goals, patterns, or other approaches designed to improve and op­ timize financial management in order to achieve corporate objectives [36]. Financial management is made up of these strategies. SME owners find it challenging to differentiate between profit and money in the bank without financial knowledge. SMEs find it challenging to build their busi­ nesses because of a lack of financial literacy and inadequate financial planning, which makes it difficult for them to manage their cash inflow and outflow. Owners of SMEs in general lack a thorough understanding of financial ac­ counts and the amount of money required to raise financing. In order to run and grow their operations, SMEs lack the managerial skills necessary [37]. It has a high failure TECHNOLOGY AUDIT AND PRODUCTION RESERVES — №5/4(73) 2023 **15** ----- #### ECONOMICS OF ENTERPRISES: ECONOMICS AND MANAGEMENT OF ENTERPRISE rate for new business owners because there aren’t many organizations that train and assist SME owners in manag­ ing and growing their businesses. It is difficult for SMEs to acquire these skills because there are few institutions and little available space. In the context of the current corporate environment, financial literacy is the capacity to effectively manage financial resources across their life cycles and interact with financial products and services. By learning more about financial products and how to evaluate risks and opportunities, investors and SMEs can both gain from improved financial literacy [38]. SMEs who are financially literate and have sufficient resources typically gain access to the loan markets. Because they are financially literate, SMEs can increase their market share, increase profits and sales, and retain more employees. It could be possible to get beyond the financial barriers that prevent SMEs from succeeding through financial literacy. The existence of SMEs depends on being able to provide owners with the knowledge needed to perform financial forecasting and utilize resources efficiently [39]. When one has the correct financial mix and understands how to lower risks, such as through asset diversification and gearing, financial challenges are simpler to solve. Financial literacy aids a company’s liquidity by maintaining the right ratio of assets to liabilities. Finan­ cial literacy’s primary objectives are to increase an SME’s assets, decrease their obligations, and increase their net profit. SMEs that are financially literate are better equipped to comprehend how finances and operational success are related. These skills help SMEs better manage their debt, make timely payments to creditors, and maintain correct financial records [40]. SMEs who are financially literate are better equipped to manage their debt, improve their credit status with current and/or potential creditors, and benefit from paying off their loans early. Business owners who are financially literate may be better able to maintain adequate financial accounts and correct accounting records, which is advantageous when trying to access the credit markets. The level of literacy of SMEs in terms of their understanding of all of their financial possibilities has some bearing on their aspirations. Financial literacy is a crucial component in boosting a SMEs performance. The ability of SMEs to create appropriate goals and plans is what defines their performance [41]. The relationship between financial literacy and company resources has a direct im­ pact on how well SMEs perform. Internal assets known as strategic resources are employed to take advantage of opportunities that develop outside of the organization and give it a competitive edge. These resources, which come in both tangible and intangible forms, can be bought with the help of financial resources. As opposed to business ex­ pertise, which is regarded an intangible resource, financial and physical resources are considered tangible resources. The performance of SMEs is greatly impacted by financial resources [14]. SMEs lack the financial resources necessary to compete in the market for the newest technologies. Busi­ ness performance is also impacted by the management and allocation of financial resources. The growth and performance of SMEs depend on financial resource availability, which has a direct impact on how well the business employs its strategic resources [42]. For a company to succeed and thrive, competitive human resources are crucial. Human capital comes in two forms: knowledge and experience. SMEs perform well when their human capital includes financial knowledge. SMEs with a grasp of finance have the skill ISSN 2664-9969 set necessary to communicate with a variety of investors, including angel investors, capital investors, and financial institutions. Human resources play an important role in the growth and success of SMEs. Strategic, financial, and human resource management that is efficient is essential for SMEs to be innovative and productive [43]. 3.6. Government policies and support for SME A business cannot survive successfully in its early stages for a variety of reasons, including a lack of experience, be­ ing new to the market, and having a small customer base. The main determinant in the growth of newly founded businesses is government support. It is hardly unexpect­ ed that governments all across the world have expressed a keen interest in funding projects. Additionally, SMEs may rely on government assistance at different points in their business cycles, including for starting, ongoing operational activities, and even process innovation [44]. The effective­ ness of SMEs in terms of innovation and new technology is impacted by government incentives, both directly and indirectly. As a result, it is frequently advised that SMEs strengthen their networking by forming positive relation­ ships with political and governmental institutions in order to gain access to valuable resources or to avoid the negative effects of government policies that can negatively affect SMEs [45]. According to social network theory, a company with close relationships with suppliers, political organiza­ tions, and customers may be able to access rare resources more affordably, improving its performance. Despite the fact that historically government support for industrial growth was disregarded, recently the government has shown a keen interest in industrial development by investing in R&D and technology. Recent research has demonstrated the impor­ tance of government support from this perspective in SMEs’ performance. For instance, it has been thoroughly explored that SMEs operating in developing markets like Pakistan are encouraged to retain their financial performance and competitive position. Entrepreneurial characteristics do not directly influence how well a firm performs; government support does. Although financial assistance from the Korean government helped Korean SMEs survive over the long term, it wasn’t always beneficial for increased productivity and profitability [46]. Government assistance also enables businesses to take advantage of entrepreneurial prospects on a local and global scale, greatly improving business success. In transition economies, government support has a strong beneficial impact on SME performance and a significant negative impact on new venture performance. The effective­ ness of a corporation is directly correlated with govern­ ment backing, and this link has increased the advantages of various market entry techniques. In addition, a number of other studies have addressed how crucial government assistance is to the development and profitability of newly founded businesses [47]. 3.7. Case studies: successful financial management in SMEs 3.7.1. Real-world examples of SMEs with effective finan­ cial management practices 3.7.1.1. Case study 1: Fanella (a software development startup). Fanella is a promising software development startup that exemplifies effective financial management practices. The employer, founded by way of three young marketers, com­ menced as a small-scale operation with restrained financial sources. However, via prudent economic making plans and **16** TECHNOLOGY AUDIT AND PRODUCTION RESERVES — №5/4(73) 2023 ----- ISSN 2664-9969 strategic selection-making, Fanella effectively navigated the aggressive tech industry and completed a widespread increase. One key aspect of their financial management success was their emphasis on cash flow management. They implemented rigorous invoicing and payment tracking systems to en­ sure timely collections and payments. This allowed them to maintain a healthy cash flow, enabling them to fund their operations and investments without relying heavily on external financing. Additionally, Fanella adopted a conservative debt management approach. Rather than burdening themselves with high-interest loans, they utilized bootstrapping and reinvested profits into their business expansion. By main­ taining a low debt-to-equity ratio, they safeguarded their financial stability and minimized financial risks [48]. 3.7.1.2. Case study 2: Wandegeya business centre, Kam­ pala (a family-owned SME). Wandegeya business centre, Kampala, a family-owned SME specializing in custom metal fabrication, illustrates the significance of financial planning and budgeting in achieving sustainable growth. Despite facing market fluctuations and economic downturns, Wandegeya business centre, Kampala maintained steady financial performance over the years. The company’s suc­ cess can be attributed to its rigorous financial planning process. They formulated detailed budgets, setting revenue and expense targets for each quarter and year. Regular monitoring and analysis of financial performance against budgeted figures allowed them to identify cost overruns and revenue shortfalls promptly. This enabled Wandegeya business centre, Kampala to implement corrective measures, such as cost-cutting initiatives or diversifying revenue streams, to ensure financial stability [49]. 3.7.2. Lessons learned from these case studies 1. _Emphasize Cash Flow Management: The case studies_ demonstrate the critical importance of effective cash flow management. SMEs have to prioritize timely invoicing, green price collections, and prudent cash go with the flow forecasting to ensure enough liquidity for daily operations and funding opportunities. 2. _Conservative Debt Management: SMEs need to be cau­_ tious approximately taking on immoderate debt, particularly inside the early degrees of their operations. Maintaining a healthy debt-to-fairness ratio and exploring alternative investment sources can mitigate monetary dangers and en­ hance lengthy-term monetary balance. 3. _Rigorous Financial Planning and Budgeting: Case re­_ search emphasize the importance of thorough monetary mak­ ing plans and budgeting. SMEs should set clear financial goals, create detailed budgets, and regularly monitor their financial performance to make data-driven decisions and adapt to market dynamics. 4. _Proactive Decision-Making: Successful SMEs take a pro­_ active approach to financial management. They identify poten­ tial risks, seize growth opportunities, and implement strategic measures to optimize their financial resources effectively. 5. _Long-Term Focus: Both case studies exemplify the_ value of long-term financial focus. SMEs with a vision for sustainable growth and financial stability are more likely to endure challenges and achieve success over time. In conclusion, the case studies of Fanella and Wandegeya business centre, Kampala provide valuable insights into effective financial management practices for SMEs. The training discovered underscore the significance of coins flow #### ECONOMICS OF ENTERPRISES: ECONOMICS AND MANAGEMENT OF ENTERPRISE management, conservative debt control, rigorous financial making plans, proactive selection-making, and a protractedterm monetary recognition. By adopting these practices, SMEs can improve their monetary overall performance, reap sustainable growth, and navigate the complexities of the commercial enterprise landscape with confidence. 3.8. Future trends in financial management for SMEs 3.8.1. Emerging technologies and their impact. The future of financial management for SMEs will be significantly influenced by emerging technologies that are transforming the financial landscape. One such technology is Artificial Intelligence (AI), which is revolutionizing various financial processes. AI-powered financial management software can automate bookkeeping, financial analysis, and budgeting tasks, enabling SMEs to streamline their financial opera­ tions and improve accuracy. Another crucial technology is blockchain, offering secure and transparent transactional systems. Blockchain can en­ hance financial data integrity, facilitate faster cross-border transactions, and reduce the need for intermediaries. This technology holds potential for transforming payment sys­ tems and improving supply chain finance, benefiting SMEs with improved efficiency and reduced costs. Moreover, the Internet of Things (IoT) can impact finan­ cial control by offering actual-time statistics from connected gadgets, including stock control systems or manufacturing equipment. These records can enhance stock forecasting, maintenance planning, and operational performance, main to better economic making plans and useful resource allo­ cation for SMEs. 3.8.2. Regulatory changes affecting SMEs. The future of financial management for SMEs will also be shaped by evolving regulatory environments. Governments worldwide are focusing on strengthening financial regulations and pro­ moting transparency, particularly after the global financial crisis. SMEs need to stay abreast of changing compliance requirements, tax regulations, and reporting standards to ensure full compliance and avoid potential penalties. Furthermore, environmental, social, and governance (ESG) reporting is gaining prominence. Regulators and investors are increasingly emphasizing sustainability and responsible business practices. SMEs that incorporate ESG principles into their financial management and reporting will not only enhance their reputation but also attract socially respon­ sible investors. 3.8.3. Forecasting challenges and opportunities. Future trends in financial management for SMEs will present both challenges and opportunities in forecasting. As markets become more complex and volatile, forecasting becomes challenging due to the uncertainty of economic conditions and consumer behavior. However, advanced data analytics and predictive modeling tools offer SMEs opportunities to improve their forecasting accuracy. Big data analytics can help SMEs identify patterns and trends, making better-informed predictions about mar­ ket demand, customer preferences, and revenue projections. Moreover, the integration of AI and machine learning algorithms into financial forecasting can provide SMEs with more sophisticated insights. These technologies can analyze historical financial data, market trends, and external fac­ tors to generate accurate forecasts and scenario planning. TECHNOLOGY AUDIT AND PRODUCTION RESERVES — №5/4(73) 2023 **17** ----- #### ECONOMICS OF ENTERPRISES: ECONOMICS AND MANAGEMENT OF ENTERPRISE Additionally, the rise of alternative data sources, such as social media data or satellite imagery, provides new avenues for gathering insights and improving forecasting models. SMEs that embrace these technologies and data sources can gain a competitive advantage in their financial planning and decision-making. In conclusion, future trends in financial management for SMEs will be shaped by emerging technologies, regulatory changes, and advancements in forecasting techniques. Embrac­ ing AI, blockchain, IoT, and other transformative technolo­ gies can enhance financial efficiency and decision-making for SMEs. Staying updated with regulatory changes and adopting ESG principles will ensure compliance and repu­ tation enhancement. Despite forecasting challenges, SMEs can leverage data analytics and predictive modeling tools to make better-informed financial projections. By embrac­ ing these future trends, SMEs can position themselves for success in an ever-evolving financial landscape. 3.9. Discussion 3.9.1. Recap of the importance of financial management in SMEs. Financial control plays a pivotal function within the success and sustainability of Small and Medium-sized Enterprises (SMEs). Effective monetary control lets in SMEs to allocate their constrained assets efficaciously, make in­ formed enterprise choices, and navigate economic demanding situations with self-assurance. It enables SME owners to assess their financial health, set clear financial goals, and monitor progress toward achieving them. By maintaining sound financial practices, SMEs can enhance their credi­ bility among stakeholders, attract external financing, and seize growth opportunities. Ultimately, financial management empowers SMEs to achieve long-term financial stability and thrive in a competitive business environment. 3.9.2. Key challenges and best practices highlighted. Throughout this review paper, several key challenges and best practices in financial management for SMEs were highlighted. _Challenges:_ 1. _Limited financial resources and capital: SMEs face_ constraints in accessing sufficient funds for growth and development. 2. _Access to financing and credit: SMEs encounter dif­_ ficulties in securing loans and financing from traditional institutions. 3. _Cash flow management: Maintaining a steady cash_ flow is challenging for SMEs, affecting day-to-day operations. 4. _Debt management: Managing debt responsibly while_ balancing growth ambitions is a delicate challenge for SMEs. 5. _Financial risk management: SMEs must identify and_ mitigate various financial risks to safeguard their financial stability. _Best Practices:_ 1. _Budgeting and forecasting: Creating comprehensive_ budgets and forecasts aids SMEs in resource allocation and strategic planning. 2. _Financial reporting and analysis: Regular financial_ reporting and analysis enable data-driven decision-making. 3. _Working capital management: Efficiently managing_ working capital ensures smooth business operations. 4. _Investment appraisal and decision-making: Rigorous_ evaluation of investment opportunities helps prioritize growth initiatives. ISSN 2664-9969 5. _Financial planning and strategy: Aligning financial_ planning with business strategy enhances financial per­ formance and competitiveness. 3.9.3. Recommendations for improving financial manage­ ment in SMEs. To further improve financial management in SMEs, several recommendations can be implemented: 1. _Enhance Financial Literacy: Governments and industry_ associations should invest in financial education programs for SME owners to improve their financial literacy and decision-making abilities. 2. _Leverage Technology: SMEs should embrace financial_ software, fintech solutions, and data analytics to stream­ line financial processes, access real-time data, and enhance forecasting accuracy. 3. _Seek Financial Advisory Services: SMEs should seek_ professional financial advisory services to gain expert in­ sights and guidance on financial planning, risk manage­ ment, and growth strategies. 4. _Foster Collaboration: Governments, financial insti­_ tutions, and industry associations should collaborate to establish mentorship programs and financing initiatives that benefit SMEs. 5. _Monitor Regulatory Changes: SMEs should remain_ informed about regulatory changes and adapt their finan­ cial practices to ensure compliance and take advantage of incentives. 6. _Strengthen Financial Controls: Implementing sturdy_ economic controls, such as normal audits and segregation of duties, ensures transparency and reduces the risk of economic mismanagement. 7. _Prioritize Long-Term Planning: SMEs must attention_ on long-term economic planning and sustainable increase strategies, avoiding short-term decision-making which could compromise their financial stability. 3.9.4. Limitation of this study 1. Limited Generalizability: The paper may focus pri­ marily on a specific region or industry, which can limit the generalizability of its findings. SMEs in different geo­ graphical locations or industries may face unique challenges and opportunities not addressed in the paper. 2. Data Sources: The paper relies on case studies and insights from real-world examples. These sources may not provide a comprehensive and representative view of all SMEs, as successful practices and government support can vary widely across different contexts. 3. _Future Trends Speculation: While the paper discusses_ future trends, such as emerging technologies and regulatory changes, it does not provide concrete evidence or data on how these trends currently impact financial management in SMEs. It may benefit from more empirical research in this regard. 4. Lack of Quantitative Analysis: The paper primarily focuses on qualitative aspects of financial management in SMEs and lacks quantitative analysis or statistical data. Quantitative data could provide a more robust foundation for the paper’s conclusions and recommendations. ### 4. Conclusions So, it is possible to conclude that monetary management is a critical thing of SMEs’ achievement. By addressing challenges and imposing first-rate practices, SMEs can **18** TECHNOLOGY AUDIT AND PRODUCTION RESERVES — №5/4(73) 2023 ----- ISSN 2664-9969 optimize their monetary performance, make informed deci­ sions, and function themselves for lengthy-term increase and prosperity. Embracing technological advancements, seeking financial advice, and staying updated on regulatory changes will further enhance SMEs’ financial management practices. As governments and industry stakeholders con­ tinue to support SMEs with resources and programs, the future looks promising for SME financial management, driving economic growth and innovation. ### Conflict of interest The authors declare that they have no conflict of inte­ rest in relation to this study, including financial, personal, authorship, or any other, that could affect the study and its results presented in this article. ### Financing The research was performed without financial support. ### Data availability The manuscript has no associated data. References **1.** Olowofela, O., Kuforiji, O., Odekeye, O., Olaiya, K. I. (2022). Financial Inclusion and Growth of Small and Medium Sized Enterprises: Evidence from Nigeria. _Izvestiya Journal of the_ _University of Economics – Varna, 66 (3-4), 198–212. doi: https://_ doi.org/10.56065/ijuev2022.66.3-4.198 **2.** Dewi, G. C., Yulianah, Y., Alimbudiono, R. S., Kurniawan, D. (2023). Application of Business Strategy to Create Competitive Advantage in Indonesian Micro, Small and Medium Enterprises. _Jurnal Manajemen Bisnis, 10 (1), 77–83._ **3.** Adjei-Boateng, E. S. (2023). A Literature Review on Management Practices among Small and Medium-Sized Enterprises. _Journal_ _of Engineering Applied Science and Humanities, 8 (1),_ 1–23. **4.** Mbaye, M. H. (2023). _Effective working Capital Management_ _Practice and SMEs’ financial performance: The case of SMEs_ _operating in the service and construction sectors in Senegal. Uni­_ versity of Wales Trinity Saint David. **5.** Hendayani, N., Muzakir, M., Yuliana, Y., Asir, M., Wahab, A. (2022). Best Practice of Financial Management in SMEs Ope­ ration in Digital times. _Budapest International Research and_ _Critics Institute-Journal (BIRCI-Journal), 5 (1), 3350–3361._ **6.** Nunden, N., Abbana, S., Marimuthu, F., Sentoo, N. (2022). An assessment of management skills on capital budgeting planning and practices: evidence from the small and medium enterprise sector. Cogent Business &amp; Management, 9 (1). doi: https:// doi.org/10.1080/23311975.2022.2136481 **7.** Casagranda, F. (2020). _The Chinese online market: opportunities_ _and challenges for Italian SMEs. Universit Ca’ Foscari Venezia._ **8.** Helmold, M., Samara, W. (2019). _Progress in performance_ _management._ Springer International Publishing. doi: https:// doi.org/10.1007/978-3-030-20534-8 **9.** Lu, J., Shon, J., Zhang, P. (2019). Understanding the Disso­ lution of Nonprofit Organizations: A Financial Management Perspective. _Nonprofit and Voluntary Sector Quarterly, 49 (1),_ 29–52. doi: https://doi.org/10.1177/0899764019872006 **10.** Zada, M., Yukun, C., Zada, S. (2019). Effect of financial mana­ gement practices on the development of small-to-medium size forest enterprises: insight from Pakistan. _GeoJournal, 86 (3),_ 1073–1088. doi: https://doi.org/10.1007/s10708-019-10111-4 **11.** Halim, H. A., Zainal, S. R. M., Ahmad, N. H. (2022). Strategic Foresight and Agility: Upholding Sustainable Competitive­ ness Among SMEs During COVID-19 Pandemic. International _Journal of Economics & Management, 16, 81–97. doi: https://_ doi.org/10.47836/ijeamsi.16.1.006 #### ECONOMICS OF ENTERPRISES: ECONOMICS AND MANAGEMENT OF ENTERPRISE **12.** Teka, B. M. (2022). Determinants of the sustainability and growth of micro and small enterprises (MSEs) in Ethiopia: literature review. _Journal of Innovation and Entrepreneurship,_ _11 (1)._ doi: https://doi.org/10.1186/s13731-022-00261-0 **13.** Haji Karimian, S. (2023). _Productivity in road pavement main­_ _tenance & rehabilitation projects: perspectives of New Zealand_ _roading contractors on the constraints and improvement measures._ Massey University. **14.** Zarrouk, H., Sherif, M., Galloway, L., Ghak, T. E. (2020). En­ trepreneurial Orientation, Access to Financial Resources and SMEs’ Business Performance: The Case of the United Arab Emirates. _The Journal of Asian Finance, Economics and Busi­_ _ness, 7 (12), 465–474. doi: https://doi.org/10.13106/jafeb.2020._ vol7.no12.465 **15.** Simon-Oke, O. O. (2020). Working Capital Management – Performance Relationship: A Study of Small and Medium En­ terprises in Akure, Nigeria. International Journal of Small Busi­ _ness and Entrepreneurship Research, 8 (2), 32–42. doi: https://_ doi.org/10.37745/ijsber.vol8.no.2p32-42.2020 **16.** Baloyi, F., Khanyile, M. B. (2022). Innovative mechanisms to improve access to funding for the black-owned small and me­ dium enterprises in South Africa. The Southern African Journal _of Entrepreneurship and Small Business Management, 14 (1)._ doi: https://doi.org/10.4102/sajesbm.v14i1.578 **17.** Ramli, A., Yekini, L. S. (2022). Cash Flow Management among Micro-Traders: Responses to the COVID-19 Pandemic. Sustain­ _ability, 14 (17), 10931. doi: https://doi.org/10.3390/su141710931_ **18.** Wadesango, N., Tinarwo, N., Sitcha, L., Machingambi, S. (2019). The impact of cash flow management on the profitability and sustainability of small to medium sized enterprises. International _Journal of Entrepreneurship, 23 (3), 1–19._ **19.** Nkwinika, S. E. R., Mashau, P. (2020). Evaluating the financial challenges affecting the competitiveness of small businesses in South Africa. _Gender and Behaviour, 18 (1),_ 15151–15162. **20.** Fred, M. U. G. A. R. U. R. A. (2021). _Effects of accounts re­_ _ceivable management on the financial performance of construction_ _companies in Rwanda: A case of NPD Ltd. University of Rwanda._ **21.** Yeon, G., Hong, P. C., Elangovan, N., Divakar, G. M. (2022). Implementing strategic responses in the COVID-19 market crisis: a study of small and medium enterprises (SMEs) in India. _Journal of Indian Business Research, 14 (3), 319–338._ doi: https://doi.org/10.1108/jibr-04-2021-0137 **22.** Dwikat, S. Y., Arshad, D., Mohd Shariff, M. N. (2023). Ef­ fect of Competent Human Capital, Strategic Flexibility and Turbulent Environment on Sustainable Performance of SMEs in Manufacturing Industries in Palestine. Sustainability, 15 (6), 4781. doi: https://doi.org/10.3390/su15064781 **23.** Grondys, K., lusarczyk, O., Hussain, H. I., Androniceanu, A. (2021). Risk Assessment of the SME Sector Operations during the COVID-19 Pandemic. International Journal of Environmental _Research and Public Health, 18 (8), 4183. doi: https://doi.org/_ 10.3390/ijerph18084183 **24.** Farida, I., Setiawan, D. (2022). Business Strategies and Com­ petitive Advantage: The Role of Performance and Innovation. _Journal of Open Innovation: Technology, Market, and Complexity,_ _8 (3), 163. doi: https://doi.org/10.3390/joitmc8030163_ **25.** Cant, A., Agui aga, E., Scheel, C. (2021). Learning from Failure and Success: The Challenges for Circular Economy Implementa­ tion in SMEs in an Emerging Economy. _Sustainability, 13 (3),_ 1529. doi: https://doi.org/10.3390/su13031529 **26.** Rosyadah, K. (2020). The influence of financial knowledge, financial attitudes and personality to financial management behavior for micro, small and medium enterprises typical food of coto makassar. JHSS (Journal of Humanities and Social Studies), _4 (2),_ 152–156. doi: https://doi.org/10.33751/jhss.v4i2.2468 **27.** Morales, R. P. (2023). Financial Literacy of Micro Entrepre­ neurs in Daet, Camarines Norte. Iconic research and engineering _journals, 6 (12),_ 114–141. **28.** Islam, A., Mansoor, A., Rahman, M., Abd Wahab, S. (2020). Аdjust­ ing a strategic cash-flow model for bangladeshi small and medium enterprises: the art of surviving COVID-19 emergency. Business _Excellence and Management, S.I. (1), 194–213. doi: https://_ doi.org/10.24818/beman/2020.s.i.1-16 TECHNOLOGY AUDIT AND PRODUCTION RESERVES — №5/4(73) 2023 **19** ----- #### ECONOMICS OF ENTERPRISES: ECONOMICS AND MANAGEMENT OF ENTERPRISE **29.** Du, L., Razzaq, A., Waqas, M. (2022). The impact of COVID-19 on small- and medium-sized enterprises (SMEs): empirical evidence for green economic implications. _Environmental Sci­_ _ence and Pollution Research, 30 (1), 1540–1561. doi: https://_ doi.org/10.1007/s11356-022-22221-7 **30.** Gao, J. (2022). Research on Financial Management Informa­ tization Mode of SME under Cloud Computing. International _Journal of Science and Research (IJSR), 11 (7),_ 793–796. doi: https://doi.org/10.21275/sr22712093816 **31.** Kotios, D., Makridis, G., Walser, S., Kyriazis, D., Monferrino, V. (2022). Personalized finance management for smes. _Big Data_ _and Artificial Intelligence in Digital Finance. Springer, 215–232._ doi: https://doi.org/10.1007/978-3-030-94590-9_12 **32.** Utami, N., Sitanggang, M. L. (2021). The Effect of Fintech Implementation on The Performance of SMEs. Journal of Inter­ _national Conference Proceedings, 4 (3), 407–417. doi: https://_ doi.org/10.32535/jicp.v4i3.1342 **33.** Muthuswamy, V. V., Sharma, A. (2023). Role of Emerging Financial Technology on Environmental and Social Governance of Textile Companies in Saudi Arabia. _Cuadernos de Econom a,_ _46 (130), 64–72._ **34.** Alkhawaldeh, B. Y., Alhawamdeh, H., Al-Afeef, M. A. M., Al-Smadi, A. W., Almarshad, M., Fraihat, B. A. M., Sou­ madi, M. M., Nawasra, M., Alaa, A. A. (2023). The effect of financial technology on financial performance in Jordanian SMEs: The role of financial satisfaction. Uncertain Supply Chain _Management, 11 (3), 1019–1030. doi: https://doi.org/10.5267/_ j.uscm.2023.4.020 **35.** Masood, T., Sonntag, P. (2020). Industry 4.0: Adoption chal­ lenges and benefits for SMEs. _Computers in Industry, 121,_ 103261. doi: https://doi.org/10.1016/j.compind.2020.103261 **36.** Tuffour, J. K., Amoako, A. A., Amartey, E. O. (2020). Assessing the Effect of Financial Literacy Among Managers on the Perfor­ mance of Small-Scale Enterprises. Global Business Review, 23 (5), 1200–1217. doi: https://doi.org/10.1177/0972150919899753 **37.** Utami, E. S., Aprilia, M. R., Putra, I. C. A. (2021). Financial literacy of micro, small, and medium enterprises of consumption sector in Probolinggo city. _Jurnal Manajemen Dan Kewirausa­_ _haan, 23 (1), 10–17. doi: https://doi.org/10.9744/jmk.23.1.10-17_ **38.** Oppong, C., Salifu Atchulo, A., Akwaa-Sekyi, E. K., Grant, D. D., Kpegba, S. A. (2023). Financial literacy, investment and per­ sonal financial management nexus: Empirical evidence on pri­ vate sector employees. _Cogent Business & Management, 10 (2)._ doi: https://doi.org/10.1080/23311975.2023.2229106 **39.** Balen, J., Nojeem, L., Bitala, W., Junta, U., Browndi, I. (2023). Essential Determinants for Assessing the Strategic Agility Frame­ work in Small and Medium-sized Enterprises (SMEs). European _Journal of Scientific and Applied Sciences, 10 (2023), 2124–2129._ **40.** Mutamimah, M., Tholib, M., Robiyanto, R. (2021). Corporate governance, credit risk, and financial literacy for small medium enterprise in Indonesia. Business: Theory and Practice, 22 (2), 406–413. doi: https://doi.org/10.3846/btp.2021.13063 ISSN 2664-9969 **41.** Kitsios, F., Kamariotou, M. (2019). Strategizing information systems: An empirical analysis of IT alignment and success in SMEs. Computers, 8 (4), 74. doi: https://doi.org/10.3390/ computers8040074 **42.** Memon, A., Yong An, Z., Memon, M. Q. (2019). Does financial availability sustain financial, innovative, and environmental performance? Relation via opportunity recognition. _Corporate_ _Social Responsibility and Environmental Management, 27 (2),_ 562–575. Portico. doi: https://doi.org/10.1002/csr.1820 **43.** Rodrigues, M., Franco, M., Silva, R., Oliveira, C. (2021). Suc­ cess Factors of SMEs: Empirical Study Guided by Dynamic Capabilities and Resources-Based View. Sustainability, 13 (21), 12301. doi: https://doi.org/10.3390/su132112301 **44.** Bakhtiari, S., Breunig, R., Magnani, L., Zhang, J. (2020). Fi­ nancial Constraints and Small and Medium Enterprises: A Re­ view. Economic Record, 96 (315), 506–523. doi: https://doi.org/ 10.1111/1475-4932.12560 **45.** Garc a-P rez-de-Lema, D., Madrid-Guijarro, A., Dur ndez, A. (2022). Operating, financial and investment impacts of Covid-19 in SMEs: Public policy demands to sustainable recovery con­ sidering the economic sector moderating effect. _International_ _Journal of Disaster Risk Reduction, 75, 102951. doi: https://_ doi.org/10.1016/j.ijdrr.2022.102951 **46.** Agwaniru, A. (2023). _ICT as a Strategy for Sustainable Small_ _and Medium Enterprises in Nigeria. California Baptist University._ **47.** Balaji, M., Dinesh, S. N., Raja, S., Subbiah, R., Manoj Ku­ mar, P. (2022). Lead time reduction and process enhancement for a low volume product. _Materials Today: Proceedings, 62,_ 1722–1728. doi: https://doi.org/10.1016/j.matpr.2021.12.240 **48.** Aruho, A. (2021). _Impact of financial management practices_ _on the performance of small and medium enterprises (SMEs)_ _in Uganda: case study of Wandegeya business centre, Kampala._ Makerere University. **49.** Nurmadewi, D., Mahendrawathi, E. R. (2019). Analyzing Link­ age Between Business Process Management (BPM) Capability and Information Technology: A Case Study in Garment SMEs. _Procedia Computer Science, 161, 935–942. doi: https://doi.org/_ 10.1016/j.procs.2019.11.202 *Eugine Nkwinika, _Doctor of Business Administration, Johan­_ _nesburg Business School, University of Johannesburg, Johannesburg,_ _South Africa, e-mail: sthembison@uj.ac.za, ORCID: https://orcid.org/_ _0000-0001-7626-4051_ **_Segun Akinola,_** _PhD in Electrical/Electronic Engineering, Johan­_ _nesburg Business School, University of Johannesburg, Johannesburg,_ _South Africa, ORCID: https://orcid.org/0000-0003-1565-7825_ *Corresponding author **20** TECHNOLOGY AUDIT AND PRODUCTION RESERVES — №5/4(73) 2023 -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.15587/2706-5448.2023.285749?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.15587/2706-5448.2023.285749, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://journals.uran.ua/tarp/article/download/285749/282012" }
2,023
[ "JournalArticle", "Review" ]
true
2023-09-28T00:00:00
[ { "paperId": "38eaf7a0a7a9333527511ac2e3c08bcdfaf899cc", "title": "Financial literacy, investment and personal financial management nexus: Empirical evidence on private sector employees" }, { "paperId": "c3eef88ac3fee5d00d5233d43ea820c2bd421bf4", "title": "Application of Business Strategy to Create Competitive Advantage in Indonesian Micro, Small and Medium Enterprises" }, { "paperId": "8699dcce8a4d996c73fabf9c0723f9cb47987f79", "title": "Effect of Competent Human Capital, Strategic Flexibility and Turbulent Environment on Sustainable Performance of SMEs in Manufacturing Industries in Palestine" }, { "paperId": "8d3bd064fd7cc513d6be74e96836beed306996b7", "title": "Financial Inclusion and Growth of Small and Medium Sized Enterprises: Evidence from Nigeria" }, { "paperId": "d7ac66fad800b71f42af33d998724fb967df6730", "title": "Determinants of the sustainability and growth of micro and small enterprises (MSEs) in Ethiopia: literature review" }, { "paperId": "8f04be6eb4e410ef04d7c1901509e8799db77f9f", "title": "Strategic Foresight and Agility: Upholding Sustainable Competitiveness Among SMEs During COVID-19 Pandemic" }, { "paperId": "0fabb5684d5c1de42f0dc50434b0edaa98c3385c", "title": "Innovative mechanisms to improve access to funding for the black-owned small and medium enterprises in South Africa" }, { "paperId": "c0ad4d585649898c249386b322396550a742d217", "title": "An assessment of management skills on capital budgeting planning and practices: evidence from the small and medium enterprise sector" }, { "paperId": "f7282b2d7647150b7f8f1a7494cda2aee43e0e90", "title": "Business Strategies and Competitive Advantage: The Role of Performance and Innovation" }, { "paperId": "06c347e709ad9612565e78079d6494147b35895c", "title": "Cash Flow Management among Micro-Traders: Responses to the COVID-19 Pandemic" }, { "paperId": "525ff01657557ce76ed63e194fdff665876e275f", "title": "The impact of COVID-19 on small- and medium-sized enterprises (SMEs): empirical evidence for green economic implications" }, { "paperId": "09a9e760d4d9fef188d737800c0024546c54626a", "title": "Research on Financial Management Informatization Mode of SME under Cloud Computing" }, { "paperId": "39c407cb34fa9bd7599f7c6021ec13c6888457e3", "title": "Operating, financial and investment impacts of Covid-19 in SMEs: Public policy demands to sustainable recovery considering the economic sector moderating effect" }, { "paperId": "ad61b081dbfb99d299fc9099c074b7757ad52dbe", "title": "Implementing strategic responses in the COVID-19 market crisis: a study of small and medium enterprises (SMEs) in India" }, { "paperId": "0f4e9f4071b0087c5e015e20bc884ba7dddbfef5", "title": "The Effect of Fintech Implementation on The Performance of SMEs" }, { "paperId": "67af2390a482aea6d81ee5f442fa85021730cb03", "title": "Lead time reduction and process enhancement for a low volume product" }, { "paperId": "901b364ad8ac58d3e11028f92941cac9ae614c9a", "title": "Success Factors of SMEs: Empirical Study Guided by Dynamic Capabilities and Resources-Based View" }, { "paperId": "4efe10a28a5ad68d326ec11cca6946cdb3fd8260", "title": "CORPORATE GOVERNANCE, CREDIT RISK, AND FINANCIAL LITERACY FOR SMALL MEDIUM ENTERPRISE IN INDONESIA" }, { "paperId": "9c9e99cbfd0e1b6ed8777e39dc7988e57934b805", "title": "Risk Assessment of the SME Sector Operations during the COVID-19 Pandemic" }, { "paperId": "4afc6b0404d8baf24370390bf79789c9dfae66d3", "title": "FINANCIAL LITERACY OF MICRO, SMALL, AND MEDIUM ENTERPRISES OF CONSUMPTION SECTOR IN PROBOLINGGO CITY" }, { "paperId": "293f6f63dc7405b0ea1540d872344e034afa5879", "title": "Learning from Failure and Success: The Challenges for Circular Economy Implementation in SMEs in an Emerging Economy" }, { "paperId": "c011a7029032bb935793c60dbcbf3371c5b95bdb", "title": "ADJUSTING A STRATEGIC CASH-FLOW MODEL FOR BANGLADESHI SMALL AND MEDIUM ENTERPRISES: THE ART OF SURVIVING COVID-19 EMERGENCY" }, { "paperId": "d5c37a19ad8b21099e2c4549ceb39420107d04fb", "title": "Industry 4.0: Adoption challenges and benefits for SMEs" }, { "paperId": "5f8d95a628f3812b4030ced35549a5bf0683efe8", "title": "THE INFLUENCE OF FINANCIAL KNOWLEDGE, FINANCIAL ATTITUDES AND PERSONALITY TO FINANCIAL MANAGEMENT BEHAVIOR FOR MICRO, SMALL AND MEDIUM ENTERPRISES TYPICAL FOOD OF COTO MAKASSAR" }, { "paperId": "0df3ea830575baa7de553f12330f8fed976162ec", "title": "Financial Constraints and Small and Medium Enterprises: A Review" }, { "paperId": "6a9bbf497eff7aca040562b1f3f0dfa415a41f2b", "title": "Working Capital Management –Performance Relationship: A Study of Small and Medium Enterprises in Akure, Nigeria" }, { "paperId": "73a820ea21bd5a1aa13362f593aac5876aab5aa8", "title": "Does financial availability sustain financial, innovative, and environmental performance? Relation via opportunity recognition" }, { "paperId": "1b9b30b77c38cbb64c141ca1eaaf8454986b1ac5", "title": "Assessing the Effect of Financial Literacy Among Managers on the Performance of Small-Scale Enterprises" }, { "paperId": "291e1aea5106955e5c467a85ab4aeea2af71dafa", "title": "Understanding the Dissolution of Nonprofit Organizations: A Financial Management Perspective" }, { "paperId": "99e8623cbb4843e4ca22bea514c99cd91c84691e", "title": "Effect of financial management practices on the development of small-to-medium size forest enterprises: insight from Pakistan" }, { "paperId": "4b66ac98df1b6f1ca03b00fc769f9209f6d63fb2", "title": "Strategizing Information Systems: An Empirical Analysis of IT Alignment and Success in SMEs" }, { "paperId": "f2e36efee4f8f5641b8424ded1ec7038eec47bf9", "title": "The effect of financial technology on financial performance in Jordanian SMEs: The role of financial satisfaction" }, { "paperId": null, "title": "Productivity in road pavement maintenance & rehabilitation projects: perspectives of New Zealand roading contractors on the constraints and improvement measures" }, { "paperId": null, "title": "Essential Determinants for Assessing the Strategic Agility Framework in Small and Medium-sized Enterprises (SMEs)" }, { "paperId": null, "title": "Role of Emerging Financial Technology on Environmental and Social Governance of Textile Companies in Saudi Arabia" }, { "paperId": null, "title": "Effective working Capital Management Practice and SMEs’ financial performance: The case of SMEs operating in the service and construction sectors in Senegal" }, { "paperId": null, "title": "A Literature Review on Management Practices among Small and Medium-Sized Enterprises" }, { "paperId": null, "title": "Personalized finance management for smes" }, { "paperId": null, "title": "Best Practice of Financial Management in SMEs Operation in Digital times" }, { "paperId": null, "title": "TECHNOLOGY AUDIT AND PRODUCTION RESERVES — № 5/4(73), 11" }, { "paperId": null, "title": "Effects of accounts re-ceivable management on the financial performance of construction companies in Rwanda: A case of NPD" }, { "paperId": null, "title": "The Chinese online market: opportunities" }, { "paperId": null, "title": "Evaluating the financial challenges affecting the competitiveness of small businesses in South Africa" }, { "paperId": null, "title": "En-trepreneurial Orientation, Access to Financial Resources and SMEs’ Business Performance: The Case of the United Arab Emirates" }, { "paperId": "8b196510ed0750af56aa3dd80279b1a91284dd15", "title": "Analyzing Linkage Between Business Process Management (BPM) Capability and Information Technology: A Case Study in Garment SMEs" }, { "paperId": "646ac926411ce211dbf6ce36b18eab6eae576171", "title": "The Impact of Cash Flow Management on the Profitability and Sustainability of Small to Medium Sized Enterprises" }, { "paperId": null, "title": "and challenges for Italian SMEs" }, { "paperId": "385104d888b02b57c6fcb9177ea9bb0b91c7757b", "title": "Financial Literacy of Micro Entrepreneurs in Daet, Camarines Norte" }, { "paperId": null, "title": "working capital ensures smooth business operations. 4. Investment appraisal and decision-making: Rigorous evaluation of investment opportunities helps prioritize growth initiatives" } ]
14,821
en
[ { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0130f7d0c517b61a63e1018d5a1ba30c13ab16ab
[]
0.821035
Optimal Locating and Sizing of BESSs in Distribution Network Based on Multi-Objective Memetic Salp Swarm Algorithm
0130f7d0c517b61a63e1018d5a1ba30c13ab16ab
Frontiers in Energy Research
[ { "authorId": "2072713343", "name": "Sui Peng" }, { "authorId": "121139968", "name": "Xianfu Gong" }, { "authorId": "2110656546", "name": "Xinmiao Liu" }, { "authorId": "2116676328", "name": "Xun Lu" }, { "authorId": "2849453", "name": "X. Ai" } ]
{ "alternate_issns": null, "alternate_names": [ "Front Energy Res" ], "alternate_urls": [ "https://www.frontiersin.org/journals/energy-research", "http://www.frontiersin.org/Energy_Research/archive" ], "id": "757ec547-4fbf-4010-b6ae-c6b833ccd3a4", "issn": "2296-598X", "name": "Frontiers in Energy Research", "type": "journal", "url": "http://www.frontiersin.org/Energy_Research" }
Battery energy storage systems (BESSs) are a key technology to accommodate the uncertainties of RESs and load demand. However, BESSs at an improper location and size may result in no-reasonable investment costs and even unsafe system operation. To realize the economic and reliable operation of BESSs in the distribution network (DN), this paper establishes a multi-objective optimization model for the optimal locating and sizing of BESSs, which aims at minimizing the total investment cost of BESSs, the power loss cost of DN and the power fluctuation of the grid connection point. Firstly, a multi-objective memetic salp swarm algorithm (MMSSA) was designed to derive a set of uniformly distributed non-dominated Pareto solutions of the BESSs allocation scheme, and accumulate them in a retention called a repository. Next, the best compromised Pareto solution was objectively selected from the repository via the ideal-point decision method (IPDM), where the best trade-off among different objectives was achieved. Finally, the effectiveness of the proposed algorithm was verified based on the extended IEEE 33-bus test system. Simulation results demonstrate that the proposed method not only effectively improves the economy of BESSs investment but also significantly reduces power loss and power fluctuation.
Edited by: Bo Yang, Kunming University of Science and Technology, China Reviewed by: Yixuan Chen, The University of Hong Kong, Hong Kong, SAR China Yang Li, Northeast Electric Power University, China *Correspondence: Xinmiao Liu [lxm2021@foxmail.com](mailto:lxm2021@foxmail.com) Specialty section: This article was submitted to Smart Grids, a section of the journal Frontiers in Energy Research Received: 10 May 2021 Accepted: 07 June 2021 Published: 27 July 2021 Citation: Peng S, Gong X, Liu X, Lu X and Ai X (2021) Optimal Locating and Sizing of BESSs in Distribution Network Based on Multi-Objective Memetic Salp Swarm Algorithm. Front. Energy Res. 9:707718. [doi: 10.3389/fenrg.2021.707718](https://doi.org/10.3389/fenrg.2021.707718) p y [doi: 10.3389/fenrg.2021.707718](https://doi.org/10.3389/fenrg.2021.707718) # Optimal Locating and Sizing of BESSs in Distribution Network Based on Multi-Objective Memetic Salp Swarm Algorithm Sui Peng [1], Xianfu Gong [1], Xinmiao Liu [2]*, Xun Lu [2] and Xiaomeng Ai [3] 1Grid Planning and Research Center, Guangdong Power Grid Corporation, China Southern Power Grid Company Limited, Guangzhou, China, [2]Guangdong Power Grid Corporation, China Southern Power Grid Company Limited, Guangzhou, China, 3State Key Laboratory of Advanced Electromagnetic Engineering and Technology, School of Electrical and Electronic Engineering, Huazhong University of Science and Technology, Wuhan, China ### Battery energy storage systems (BESSs) are a key technology to accommodate the uncertainties of RESs and load demand. However, BESSs at an improper location and size may result in no-reasonable investment costs and even unsafe system operation. To realize the economic and reliable operation of BESSs in the distribution network (DN), this paper establishes a multi-objective optimization model for the optimal locating and sizing of BESSs, which aims at minimizing the total investment cost of BESSs, the power loss cost of DN and the power fluctuation of the grid connection point. Firstly, a multi-objective memetic salp swarm algorithm (MMSSA) was designed to derive a set of uniformly distributed non-dominated Pareto solutions of the BESSs allocation scheme, and accumulate them in a retention called a repository. Next, the best compromised Pareto solution was objectively selected from the repository via the ideal-point decision method (IPDM), where the best trade-off among different objectives was achieved. Finally, the effectiveness of the proposed algorithm was verified based on the extended IEEE 33- bus test system. Simulation results demonstrate that the proposed method not only effectively improves the economy of BESSs investment but also significantly reduces power loss and power fluctuation. Keywords: distribution networks, battery energy storage systems, optimal locating and sizing, multi-objective memetic salp swarm algorithm, ideal-point decision method ## INTRODUCTION In recent years, distributed generators (DGs) and controllable load in the distribution network (DN) have continued to increase, meaning that the traditional DN faces many challenges (Sepulveda Rangel et al., 2018; Liu et al., 2020; Peng et al., 2020). At present, one obvious tendency is that the rapid-developed photovoltaic (PV) and wind turbine (WT) power generation technologies make the permeability of distributed PV and WT in the DN higher. A series of problems ensue, such as voltage quality declination and power supply reliability reduction, etc (Wang et al., 2014; Yu et al., 2016; Sun et al., 2020). The active power through the line increases at the peak of power load, the loss increases, and a large voltage offset appears at the end of the line (Kerdphol et al., 2016a; Zhou et al., 2021). Battery energy storage systems (BESSs) have the characteristics of flexibility and fast response and are an effective way to solve the above problems. The application of BESSs can greatly improve the ----- connection of renewable energy sources (RESs) (Kerdphol et al., 2016b; Gan et al., 2019; Hlal et al., 2019). BESSs can effectively solve the problems of enlarging the load peak and off-peak difference, delay in the power grid upgrading, alleviate the power supply capacity shortage in the transition phase of the power grid, improve the reliability and stability of the power grid, and optimize the power flow of the grid, as well as improving the economic benefits of system operation (Chong et al., 2016; Chong et al., 2018; Murty and Kumar., 2020). BESSs could provide a new direction for large-scale RESs integration, which is one of the most effective ways to solve renewable energy grid access (Trovão and Antunes, 2015; Liu et al., 2018; Wu et al., 2019). However, prudent BESSs allocation and sizing in DN determine the satisfactory performance of BESSs applications. The optimal allocation and sizing of BESSs are crucial for the power quality improvement of DN and transmission system protection settings. Once BESSs are connected to the DN, the dispatching system of DN sends dispatching instructions to the BESSs according to the real-time running state of the system load, and then BESSs absorbs or sends power to the parallel network through its two-way energy flow (He et al., 2017; Jia et al., 2017; He et al., 2020). This two-way power regulation can save investment and improve the reliability and economy of BESSs. If the location and sizing of BESSs are not set reasonably, or the operation strategy adopted fails to efficiently play the role of BESSs, the voltage quality may deteriorate, and further increase investment and operation costs (Li et al., 2020). To enable us to take full advantage of distributed BESSs and make their access to the DN have a positive impact, it is important to select the appropriate location and sizing of BESSs based on the appropriate operation strategy (Li et al., 2018). Recently, a large number of scholars have performed studies in this field (Yang et al., 2020). The literature (Oudalov et al., 2007) tends to optimize the location and power capacity of BESSs by calculating the sensitivity of network loss, and then reduce the power loss of DN. In one study (Pang et al., 2019), a semi-definite relaxation method was proposed to solve the optimal BESSs allocation problem. Another study (Wong et al., 2019) introduces a whale optimization algorithm for the optimal location and sizing of BESSs, while the optimization results do not achieve a significant breakthrough. This paper devises a multi-objective optimization model considering total investment cost, power loss cost, and power fluctuation for optimal BESSs locating and sizing. For the sake of solving this model, a multi-objective memetic salp swarm algorithm (MMSSA) is proposed to search the non-dominated solutions of BESSs allocation strategy, which reach significant improvement and better balance on the global exploration and local exploitation abilities compared with the salp swarm algorithm (SSA). Furthermore, the ideal-point decision method (IPDM) is adapted to objectively determine the optimal weight coefficients of each objective function and then select the best compromised solution. To verify the effectiveness, the proposed model and algorithm are implemented in the extended IEEE-33 bus test system. The rest of this paper is organized as follows: Problem Formulation develops the multi-objective optimization model. TABLE 1 | The economic parameters of BESSs. Parameters Values Installation cost 1470000 ($/per BESS) Equipment cost 175,000 ($/MW) 225,000 ($/MW h) O&M cost 4,000 ($/MW year) 2000 ($/(MW h) year) Lifetime 20 (year) In Multi-Objective Memetic Salp Swarm Algorithm Based on Pareto, MMSSA based on IPDM is introduced. Case studies are undertaken in Case Studies. Finally, Conclusion summarizes the main contributions of this study. ## PROBLEM FORMULATION Objective Functions The optimal allocation of BESSs is a multi-objective optimization problem with multiple variables and constraints. To realize the economic and reliable operation of BESSs in the DN, a multiobjective optimization model is established based on the Pareto principle, where minimizing the total investment cost of BESSs, power loss cost, and power fluctuation are the main objectives. ### Total Investment Cost This paper focuses on the DN that has been built and operated, so the investment and construction costs of DN other than BESSs are not included in the cost model. The economic parameters of BESSs are provided in Table 1, extracted from a previous study (Behnam and Sanna, 2015). The total investment cost is considered as the annual costs of BESSs, which can be mathematically formulated as follows Min F1 � Cins + Cequ + Com (1) where F1 is the annual total investment cost of BESSs; Cins, Cequ, and Com represent the annual installation cost, equipment cost, and operation and maintenance (O&M) cost, respectively. The annual installation cost of BESSs is expressed as Cins � Ccap · NBESS · μCRF (2) where Ccap means the cost of per BESS for installation; NBESS is the number of BESSs deployed in DN; μCRF denotes the capital recovery factor (CRF) that is the knowing present worth. The CRF translates the costs throughout the useful life of BESSs to the initial moment of the investment, which is obtained by μCRF � ([r]1[ · (] +[1] r[ +])[y][ r]−[)][y]1 (3) where y is the economic life cycle of BESSs; r means the discount rate, which is calculated by the weighted average cost of capital as follows (Harvey, 2020) r � fd · id + �1 − fd� - ie (4) ----- where fd and ie represent the debt ratio and the return on equity, is respectively, 80 and 50%; id denotes the interest rate of 4.165%. The annual equipment cost of BESSs is calculated by NBESS Cequ � � �α · PBESS,i + β · EBESS,i� - μCRF (5) i�1 where α and β mean the costs per unit power and per unit capacity, respectively; PBESS,i and EBESS,i are the power capacity and energy capacity of the ith BESS. The annual O&M cost of BESSs is expressed as NBESS COM � � �λ · PBESS,i + c · EBESS,i� - μCRF (6) i�1 where λ and c are respectively the O&M cost of per unit power and per unit energy of BESSs. Note that the O&M costs of rectifier, inverter, and charge regulator are neglected. ### Power Loss Cost BESSs grid-connected will change the power flow of DN (Injeti and Thunuguntla, 2020). Furthermore, the different locations and sizes of BESSs will have different influences on power losses. For the sake of minimizing the total active power losses, the power losses index is established in the optimization model, as follows T Min F2 � ��ρloss(t) · Ploss(t)� (7) t�1 L Ploss(t) � ��RjIj[2][(][t][)][�] (8) j�1 where F2 is the daily cost of power losses; ρloss(t) and Ploss(t) represent the time of use (TOU) electricity prices and power losses at time t; L is the total number of lines in the DN; Rj means the resistance on the jth line; Ij(t) denotes the current on the jth line at time t. The lower F2 is that the greater positive effect of BESSs deployment in reducing power loss. ### Power Fluctuation Owing to the intermittent nature of RESs, the integration of them into power grids poses significant power fluctuation in the grid connection point. However, BESSs can provide an effective supplement for RESs in smoothing power fluctuation to improve power quality. The power quality index can be expressed as ## Constraints ### Power Balance ⎪⎧⎪⎪⎪⎪⎨ ⎪⎪⎪⎪⎪⎩ N Pi(t) � Vi(t) � Vj(t)�Gij cos θij(t) + Bij sin θij(t)� j�N1 (10) Qi(t) � Vi(t) � Vj(t)�Gij sin θij(t) − Bij cos θij(t)� j�1 where Pi(t) and Qi(t) represent the injected active power and reactive power at ith node in the DN at time t, respectively; Vi(t) is the voltage of the ith node at time t; Gij and Bij represent the admittance and susceptance between the ith node and the jth node; θij(t) is the power angle between the ith node and the jth node at time t. ### Range of Node Voltages Vi[min] < Vi < Vi[max] (11) where Vi[min] and Vi[max] represent the upper and lower limits of the voltages of the ith node. ### Charging and Discharging Power Limits of BESSs (12) � −[0]P[ ≤]BESS[P][cha],i ·[,][i] η[(][t]dis[)][ ≤] ≤[P]P[BESS]dis,i[,]([i][ ·]t[ η]) ≤[cha]0 where Pcha,i(t) andPdis,i(t) represent the charging and discharging power of BESSs at time t, respectively; ηcha and ηdis are respectively the charging and discharging efficiency of BESSs. ### State of Charge Limits SOC[min] < SOC(t) < SOC[max] (13) where SOC[min] and SOC[max], respectively, mean the upper and lower limits of SOC, is that 20 and 90%. ## Multi-Objective Optimization Model ### Establishment of the Optimization Model In terms of multi-objective optimization problems such as BESSs allocation, all objectives generally conflict with each other, and optimizing one of the objectives leads to the deterioration of other objectives in most cases. It is difficult to objectively evaluate the superiority-inferiority of all solutions because there is no absolute optimal solution for the overall objective (Huang et al., 2020). Nevertheless, there exists an optimal solution set, elements of which are named Pareto optimal solutions, realizing the optimum matching among objectives (Fonseca and Fleming, 1993). In this paper, the multi-objective optimization model of BESSs locating and sizing is designed to simultaneously meet investment economy and operation reliability requirements, as follows Min F3 � ����������������� � T 2 � �Pgrid(t) − Pgrid� t�1 (9) where F3 is the daily total power fluctuation of the grid connection point; Pgrid(t) represents the power fluctuation at time t; Pgrid means the mean power fluctuation over a day. ----- exploration and local exploitation abilities. Therefore, there are two important search mechanisms in MSSA, namely the local search in a single chain and the global coordination in the whole population. In MSSA, multiple salp chains are arranged in parallel, where each salp chain is regarded as a swarm of salps that independently perform local searches similar to SSA. Meanwhile, all salp chains are regrouped by information communication among all salps for the improvement of convergence stability. The optimization framework of MSSA is illustrated in Figure 1. ### Mathematical Model In the single chain, the salps can be divided into two roles, including the leaders and the followers. As illustrated in Figure 1, the leader is regarded as the salp at the front of each salp chain, while the rest of the salps are followers. In each iteration, the leading salp seeks the food source, while the follower salps follow each other in a row. Note that the best salp with the best fitness is considered to be the food source, and will be chased by the whole salp chain. The position of the leading salp and follower salps can be updated as follows (Mirjalili et al., 2017) xm[j] 1 [�] [�] [F]F[ j]mm[j] [+][−][ c][c][1][1][�][�][c][c][2][2][�][�][ub][ub][ j][ j][ −][ −] [lb][lb][ j][ j][�][�] [+][+][ lb][ lb][ j][ j][�][�][,][,][ if c][ if c][3][3][ ≥][ <][ 0][0] (16) xmi[j] [�] [1]2 [�][x]mi[ j] [+][ x][ j]m,i−1[�][,] i � 2, 3, . . ., n; m � 1, 2, . . ., M (17) where the j means the jth dimension of searching space; xm[j] 1 [and] xmi[j] [respectively denote the positions of the leading salp and the] ith follower salp in the mth salp chain; Fm[j] [is the position of a food] source; ub [j] and lb [j] are respectively the upper and lower limits of the jth dimension variables; n and M represent the population size of a single salp chain and the number of salp chains, respectively; c2, and c3 are both the uniform random numbers from 0 to 1; c1 is a random number that is related to the iteration number, as follows (Mirjalili et al., 2017) ⎪⎧⎨ ⎪⎩ min F(x) �[F1(x), F2(x), F3(x)][T] s.t.E(x) � 0 (14) I(x) ≤ 0 where F(x) represents the target space consists of all objective functions; x denotes the decision space that is constituted by all optimization variables; E(x) and I(x) are respectively, equality and inequality constraints that need to be satisfied in the multiobjective optimization model. ### Design of Optimization Variables Optimization variables include the installation locations, power, and energy capacities of two BESSs, all of which need to be constructed in a reasonable range, otherwise, some negative effects on the power flow, relay protection, voltage, and waveform of the original power grid raise. In this paper, nodes in the range of (Mirjalili et al., 2017; Liu et al., 2020) were selected as the installation locations, in which environmental and geographical factors need to be considered in engineering practice. The limits of power and energy capacities are determined to consider the topology of DN, the power limit of the interconnection point, especially the total load power. Therefore, the power capacity allowed to access the power grid of a BESS is determined as 90% of the total active power load of the power grid, and the numerical value of energy capacity limit is equal to power capacity limit, as follows BESS (15) � E[P][BESS]BESS[,],[i]i[ ≤] ≤ [P]EBESS[max][max] where PBESS,i and EBESS,i are the power capacity and energy capacity of the ith BESS; PBESS[max] [and][ E]BESS[max] [denote the upper] limits of the energy capacity and power capacity of BESSs, are respectively, 3375 and 3375 kWh. Note that the power and energy capacities of two BESSs are continuous, while installation locations are discrete. In this paper, continuous variables can converge to the optimal value in the iteration process while the optimal value of discrete variables needs to be rounded in continuous space (Zhang et al., 2017). ## MULTI-OBJECTIVE MEMETIC SALP SWARM ALGORITHM BASED ON PARETO Memetic Salp Swarm Algorithm ### Optimization Framework SSA is inspired by the swarming motility and foraging behavior of salps, which successfully solves varieties of optimization problems since it has a simple search mechanism and high optimization efficiency (Mirjalili et al., 2017). In recent years, the memetic algorithm has developed into a broad class of algorithms and can properly combine global search and local search mechanisms (Moscato, 1989; Neri and Cotta, 2012). In this paper, the memetic computing framework first proposed by Moscato (Moscato, 1989) is adopted in the memetic salp swarm algorithm (MSSA) to improve the searching ability of SSA. Then, multiple slap chains were employed to better balance global c1 � 2e−�kmax[4][k] [�] 2 (18) where k and kmax are the current iteration number and maximum iteration number, respectively. In the salp population, each salp is taken as an individual of the virtual salp population. At each iteration, the population can be regrouped into multiple new salp chains based on the descending order of all salps’ fitness values. In the regroup operation, the global coordination among different salp swarms is achieved, as shown in Figure 2. It can be seen that the best solution is assigned to salp chain #1, and then the second-best solution is assigned to salp chain #2, and so on. Therefore, the mth salp chain can be updated by (Eusuff and Lansey, 2015) Y [m] � �xmi, fmi|xmi � X(m + M(i − 1)), fmi � F(m + M(i − 1)), i � 1, 2, /, n�, m � 1, 2, /, M (19) where xmi and fmi are the position vector and fitness value of the ith salp in the mth chain, respectively; X and F denote a position vector set and a fitness value set of all the salps, respectively. ----- food sources with a repository to restore the non-dominated solutions obtained by MSSA so far (Coello et al., 2004). In the optimization process, every new non-dominated solution needs to be compared against all residents in the repository using the Pareto dominance operators, as follows (Faramarzi et al., 2020). ### • If a new solution dominates a set of solutions in the repository, they have to be swapped; ### • If at least one of the solutions in the repository dominates the new solution, this new solution should be discarded straight away; ### • If a new solution is non-dominated in comparison with all repository residents, this new solution will be added to the repository. The repository can just store limited solutions. Therefore, a wise method adopted to remove the similar nondominated solutions in the repository, is that the one in the populated region is identified as the best candidate to be removed from the repository to improve the distribution diversity of the Pareto optimal solution set. The solutions that are removed from the repository need to satisfy the following equation |Fh(xm) − Fh(xn) < Dh|, h � 1, 2, 3 h − Fh[min] Dh � [F][max]Nr (20) ## Multi-Objective Memetic Salp Swarm Algorithm As discussed in Problem Formulation, the solutions for a multi-objective problem should be a set of Pareto optimal solutions. MSSA can drive salps towards the food source with the best solution for the optimization problem and update it at each iteration. The design of MMSSA is first to equip the ⎪⎧⎪⎨ ⎪⎪⎩ where Fh(xm) and Fh(xn) denote the hth fitness value of the mth salp and the nth salp, respectively; Dh is the distance threshold of the Pareto solution set; Fh[max] and Fh[min], respectively, represent the maximum and minimum of the hth objective function obtained as far; Nr is the maximum size of the repository to store the nondominated solutions. ----- problem. Crucially, the squared Euclidean distance between different solutions and the ideal point is taken as an important basis for ranking all non-dominated solutions and then decide the best compromised solution from them. The squared Euclidean distance can be calculated by 3 EUi � � � fh(xm) − 0�2 · ω2h (22) h�1 ωh means the weights of the hth objective function, as follows 1 ωh � 2 3 �m[N][r]�1 [�] [f][h][(][x][m][) −] [0][�] - �h�1 (23) 1 Nr 2 �m�1 [[][f][h][(][x][m][)−][0][]] Owing to the weights of each objective function obtained by IPDM, it does not rely on the evaluation and preference of experts so that the decision is credible. In the end, the best compromised solution is expressed as 3 xbest � arg minm�1,2,...,Nr h[�]�1 � fh(xm) − 0�2 · ω2h (24) To sum up, the flowchart of MMSSA to solve the optimal locating and sizing of BESSs is shown in Figure 3. ## CASE STUDIES Test System In this section, the optimal locating and sizing of BESSs based on MMSSA is implemented in the extended IEEE-33 bus system for verifying the effectiveness of the proposed algorithm. The topology structure of the test system with a total load of 3,715 + j2300 kVA is depicted in Figure 4. It is assumed that the resource units include one PV and three WT, where the maximum generation limits of PV and WT both are 0.2 MW. The typical daily curves of load, wind and PV power are demonstrated in Figure 5. In addition, multi-objective particle swarm optimization (MOPSO) (Hlal et al., 2019) is used for In this paper, the IPDM was adopted to filter out the best compromised solution of the Pareto non-dominated solution set, which is often used in multiple attribute decision making. Firstly, the objective functions of all Pareto non-dominated solutions obtained by MPOA are normalized as follows fh(xm) � [y][h][(][x][m][) −] [y]h[min] (21) yh[max] − yh[min] where yh(xm) is the hth objective function value of the nondominated solution xm; fh(xm) represents the normalized value of the hth objective function; yh[min] and yh[max] mean the maximum and minimum of the hth objective function. Thus, an ideal point can be selected in the target decisionmaking region formed by all Pareto non-dominated solutions. It is worth mentioning that the objective functions of the ideal point can be normalized to be (0, 0, 0) in terms of the minimization ----- comparison. For the sake of a relatively fair comparison, the population size of MMSSA and other algorithms all are set to 100, and the maximum iterations are set to be 500. The size of the repository was chosen to equal 100 for multi-objective optimization. Some specific parameters of all comparison algorithms were set to the default values. If the parameters are not chosen properly, the convergence time will be too long or the local optimum will be trapped. It is worth mentioning that the key parameters in the MMSSA algorithm, such as c1, the most important parameter since they can directly influence the trade-off between exploration and exploitation. To achieve a proper balance, it was designed according to the iteration number. ## Simulation Results Figure 6 and Figure 7, respectively, exhibit the bi-objective Pareto front curves by two algorithms, including the total investment cost versus the power loss cost, the total investment cost versus the power fluctuation, as well as the power loss cost versus power fluctuation, which demonstrates these three bi-objective Pareto fronts obtained by MMSSA are ----- more uniform than MOPSO from the perspective of distribution. Figure 8 shows the three-objective Pareto front obtained by two algorithms. As can be seen from Figure 8, MMSSA can acquire the Pareto solution set with higher quality compared with MOPSO. Moreover, the schematic diagram of the IPDM based on MMSSA is illustrated in Figure 9. Figure 9 shows the normalized objective function curve based on MMSSA, as well as the decision-making schematic for the best compromise solution of BESSs allocation. IPDM based on MMSSA can obtain the objective weight coefficients and select the best compromise solution by means of the sum of squares of Euclidean distance. To better compare the convergence and diversity of the Pareto solution set obtained by two algorithms, the performance indexes are evaluated in Table 2, including coverage over the Pareto front (CPF) (Tian et al., 2019), spread (Wang et al., 2010), spacing (Schott, 1995), and execution time. It is worth mentioning that CPF defines the TABLE 2 | Comparison of performance metrics of two algorithms. Algorithm Performance metric CPF Spread Spacing Time (s) MOPSO 0.4996 0.4753 9,075.45 1.5428e+04 MMSSA 0.1636 0.4481 3.3858 1.4676e+04 diversity of Pareto solution set as its coverage over the Pareto front in an (M-1) dimensional hypercube (Wang et al., 2010), while spread and spacing respectively denote the diversity and the evenness of the Pareto solution set, which are all the negative indexes. In addition, Table 3 shows the best compromise decision scheme of BESSs allocation from two algorithms, along with the objective function values. It is evident that the MMSSA outperforms the MOPSO in the multi-objective optimization model for optimal locating and sizing of BESSs: ----- TABLE 3 | Optimization results of two algorithms. Algorithm The best compromise allocation scheme of BESSs Objective function values under the best compromise allocation scheme Bus location Power capacity Energy capacity Total investment Power loss Power fluctuation (MW) (MW·h) cost ($/year) cost ($/year) (MW/year) MOPSO Trovão and Antunes (2015); Mirjalili et al. (2017) [0.1972, 0.2786] [1.6203, 1.6204] 2.0873e+05 1.3698e+05 292.9054 MMSSA Pang et al. (2019); Injeti and Thunuguntla (2020) [0.0849, 0.0618] [0.6535, 0.3943] 1.7417e+05 1.3679e+05 32.1682 ### • It has the smallest CPF value, indicating that MMSSA owns better diversity; ### • It gains the smallest spread and spacing value, which indicates that the Pareto solutions obtained by MMSSA are evenly and widely distributed on the Pareto front; ### • It also has the smallest execution time, which means that MMSSA can converge to the Pareto front much faster than conventional MOPSO; ### • It has the least investment cost, meaning that MMSSA can improve the economy of BESSs investment; ### • It slightly reduces power loss cost, and ensures a higher operation economy of DN; ### • It significantly gains lower power fluctuation of the grid connection point, which means MMSSA can contribute to power supply reliability. ## CONCLUSION In this paper, a multi-objective optimization model based on the Pareto principle was established. This study proposes MMSSA as a method for solving the optimal location and size of BESSs in DN. The contributions of the proposed approach are as follows: ### • The multi-objective optimization model takes the economic criteria, incorporates time value into cost, and the technical criteria relate to system reliability and take it into consideration, which aims to make BESSs more costeffective and ensure the reliable operation of DN; ### • The proposed MMSSA has a strong global search ability and convergence ability under complex multi-objective functions, which can quickly search high-quality nondominated solutions, and then objectively select the best compromised solution with the help of IPDM; ### • The simulation results based on the extended IEEE-33 bus test system effectively verify that the best-compromised solution of BESSs allocation scheme obtained by MMSSA owns the lowest investment cost, power loss cost, and power fluctuation, which is beneficial for DN to increase economic efficiency and improve system reliability. ## REFERENCES Behnam, Z., and Sanna, S. (2015). Electrical Energy Storage Systems: A Comparative Life Cycle Cost Analysis. Renew. Sust. Energ. Rev. 42, 569–596. [doi:10.1016/j.rser.2014.10.011](https://doi.org/10.1016/j.rser.2014.10.011) However, there are several limitations to this work, including the inapplicability of the proposed MMSSA for the highdimension optimization problem, and the limited scenario design in terms of validating its effectiveness. Therefore, the MMSSA can be further enhanced to improve the accuracy for high-dimensional objective optimization. Meanwhile, a multiscenario design that combines different typical daily data in a year should be conducted to capture the time-variable nature and uncertainties related to RESs and load demand. ## DATA AVAILABILITY STATEMENT The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author. ## AUTHOR CONTRIBUTIONS SP and XG contributed to conception and design of the study. XiL and XuL performed the case analysis. SP wrote the first draft of the manuscript. XG, XiL, XuL, and XA wrote sections of the article. All authors contributed to article revision, read, and approved the submitted version. ## FUNDING The paper is funded by the project supported by the China Southern Power Grid Science and Technology Project under Project 037700KK52190011 (GDKJXM20198273). ## ACKNOWLEDGMENTS The authors gratefully acknowledge that this project was supported by the China Southern Power Grid Science and Technology Project under Project 037700KK52190011 (GDKJXM20198273). Chong, L. W., Wong, Y. W., Rajkumar, R. K., and Isa, D. (2018). An Adaptive Learning Control Strategy for Standalone PV System with BatterySupercapacitor Hybrid Energy Storage System. J. Power Sourc. 394, 35–49. [doi:10.1016/j.jpowsour.2018.05.041](https://doi.org/10.1016/j.jpowsour.2018.05.041) Chong, L. W., Wong, Y. W., Rajkumar, R. K., and Isa, D. (2016). An Optimal Control Strategy for Standalone PV System with Battery-Supercapacitor Hybrid ----- Energy Storage System. J. Power Sourc. 331, 553–565. [doi:10.1016/](https://doi.org/10.1016/j.jpowsour.2016.09.061) [j.jpowsour.2016.09.061](https://doi.org/10.1016/j.jpowsour.2016.09.061) Coello, C. A. C., Pulido, G. T., and Lechuga, M. S. (2004). Handling Multiple Objectives with Particle Swarm Optimization. Evol. Comput. IEEE Trans. 8, [256–279. doi:10.1109/tevc.2004.826067](https://doi.org/10.1109/tevc.2004.826067) Eusuff, M. M., and Lansey, K. E. (2015). Optimization of Water Distribution Network Design Using the Shuffled Frog Leaping Algorithm. J. Water Resour. [Plann. Manag. 129 (3), 210–225. doi:10.1061/40569(2001)412](https://doi.org/10.1061/40569(2001)412) Faramarzi, A., Heidarinejad, M., and Stephens, b. (2020). Equilibrium Optimizer: a Novel Optimization Algorithm. Knowledge Based Syst. 191, 105190. [doi:10.1016/j.knosys.2019.105190](https://doi.org/10.1016/j.knosys.2019.105190) Fonseca, C. M., and Fleming, P. J. (1993). Genetic Algorithms for Multiobjective Optimization: Formulation Discussion and Generalization. Icga 93, 416–423. Gan, W., Ai, X., Fang, J., Yan, M., Yao, W., Zuo, W., et al. (2019). Security Constrained Co-planning of Transmission Expansion and Energy Storage. [Appl. Energ. 239, 383–394. doi:10.1016/j.apenergy.2019.01.192](https://doi.org/10.1016/j.apenergy.2019.01.192) Harvey, H. L. D. (2020). Clarifications of and Improvements to the Equations Used to Calculate the Levelized Cost of Electricity (LCOE), and Comments on the [Weighted Average Cost of Capital (WACC). Energy 207, 118340. doi:10.1016/](https://doi.org/10.1016/j.energy.2020.118340) [j.energy.2020.118340](https://doi.org/10.1016/j.energy.2020.118340) He, X., Ai, Q., Qiu, R. C., Huang, W. T., Piao, L. J., and Liu, H. C. (2017). A Big Data Architecture Design for Smart Grids Based on Random Matrix Theory. IEEE Trans. Smart Grid. 8 (2), 674–686. He, X., Qiu, R. C., Ai, Q., and Zhu, T. Y. (2020). A Hybrid Framework for Topology Identification of Distribution Grid with Renewables Integration. IEEE Trans. [Power Syst. 36 (2), 1493–1503. doi:10.1109/tpwrs.2020.3024955](https://doi.org/10.1109/tpwrs.2020.3024955) Hlal, M. I., Ramachandaramurthya, V. K., Padmanaban, S., Kaboli, H. R., Pouryekta, A., and Tuan Abdullah, T. A. R. b. (2019). NSGA-II and MOPSO Based Optimization for Sizing of Hybrid PV/wind/battery Energy Storage System. Ijpeds 10 (1), 463–478. [doi:10.11591/](https://doi.org/10.11591/ijpeds.v10.i1.pp463-478) [ijpeds.v10.i1.pp463-478](https://doi.org/10.11591/ijpeds.v10.i1.pp463-478) Huang, Z., Fang, B. L., and Deng, J. (2020). Multi-objective Optimization Strategy for Distribution Network Considering V2G Enabled Electric Vehicles in Building Integrated Energy System. Prot. Control. Mod. Power Syst. 5 (1), [48–55. doi:10.1186/s41601-020-0154-0](https://doi.org/10.1186/s41601-020-0154-0) Injeti, S. K., and Thunuguntla, V. K. (2020). Optimal Integration of DGs into Radial Distribution Network in the Presence of Plug-In Electric Vehicles to Minimize Daily Active Power Losses and to Improve the Voltage Profile of the System Using Bioinspired Optimization Algorithms. Prot. Control. Mod. Power Syst. 5 [(1), 21–35. doi:10.1186/s41601-019-0149-x](https://doi.org/10.1186/s41601-019-0149-x) Jia, K., Chen, Y., Bi, T., Lin, Y., Thomas, D., and Sumner, M. (2017). Historicaldata-based Energy Management in a Microgrid with a Hybrid Energy [Storage System. IEEE Trans. Ind. Inf. 13 (5), 2597–2605. doi:10.1109/](https://doi.org/10.1109/tii.2017.2700463) [tii.2017.2700463](https://doi.org/10.1109/tii.2017.2700463) Kerdphol, T., Fuji, K., Mitani, Y., Watanabe, M., and Qudaih, Y. (2016). Optimization of a Battery Energy Storage System Using Particle Swarm Optimization for Stand-Alone Microgrids. Int. J. Electr. Power Energ. Syst. [81, 32–39. doi:10.1016/j.ijepes.2016.02.006](https://doi.org/10.1016/j.ijepes.2016.02.006) Kerdphol, T., Qudaih, Y., and Mitani, Y. (2016). Optimum Battery Energy Storage System Using PSO Considering Dynamic Demand Response for Microgrids. [Int. J. Electr. Power Energ. Syst. 83, 58–66. doi:10.1016/j.ijepes.2016.03.064](https://doi.org/10.1016/j.ijepes.2016.03.064) Li, R., Wang, W., Chen, Z., and Wu, X. (2018). Optimal Planning of Energy Storage System in Active Distribution System Based on Fuzzy Multi-Objective Bi-level [Optimization. J. Mod. Power Syst. Clean. Energ. 6 (2), 342–355. doi:10.1007/s40565-](https://doi.org/10.1007/s40565-017-0332-x) [017-0332-x](https://doi.org/10.1007/s40565-017-0332-x) Li, Y., Vilathgamuwa, M., Choi, S. S., Xiong, B., Tang, J., Su, Y., et al. (2020). Design of Minimum Cost Degradation-Conscious Lithium-Ion Battery Energy Storage System to Achieve Renewable Power Dispatchability. Appl. Energ. 260, 114282. [doi:10.1016/j.apenergy.2019.114282](https://doi.org/10.1016/j.apenergy.2019.114282) Liu, J., Yao, W., Wen, J., Fang, J., Jiang, L., He, H., et al. (2020). Impact of Power Grid Strength and PLL Parameters on Stability of Grid-Connected DFIG Wind Farm. IEEE Trans. Sustain. Energ. 11 (1), 545–557. [doi:10.1109/](https://doi.org/10.1109/tste.2019.2897596) [tste.2019.2897596](https://doi.org/10.1109/tste.2019.2897596) Liu, Z., Chen, Y., Zhuo, R., and Jia, H. (2018). Energy Storage Capacity Optimization for Autonomy Microgrid Considering CHP and EV [Scheduling. Appl. Energ. 210, 1113–1125. doi:10.1016/j.apenergy.2017.07.002](https://doi.org/10.1016/j.apenergy.2017.07.002) Mirjalili, S., Gandomi, A. H., Mirjalili, S. Z., Saremi, S., Faris, H., and Mirjalili, S. M. (2017). Salp Swarm Algorithm: A Bio-Inspired Optimizer for Engineering Design Problems. Adv. Eng. Softw. 114, 163–191. [doi:10.1016/](https://doi.org/10.1016/j.advengsoft.2017.07.002) [j.advengsoft.2017.07.002](https://doi.org/10.1016/j.advengsoft.2017.07.002) Moscato, P. (1989). On Evolution, Search, Optimization, Genetic Algorithms and Martial Arts: Towards Memetic Algorithms. Caltech Concurrent Computation Program, Technical Reports 826. Murty, V. V. S. N., and Kumar, A. (2020). Multi-objective Energy Management in Microgrids with Hybrid Energy Sources and Battery Energy Storage [Systems. Prot. Control. Mod. Power Syst. 5 (1), 1–20. doi:10.1186/s41601-](https://doi.org/10.1186/s41601-019-0147-z) [019-0147-z](https://doi.org/10.1186/s41601-019-0147-z) Neri, F., and Cotta, C. (2012). Memetic Algorithms and Memetic Computing [Optimization: A Literature Review. Swarm Evol. Comput. 2, 1–14. doi:10.1016/](https://doi.org/10.1016/j.swevo.2011.11.003) [j.swevo.2011.11.003](https://doi.org/10.1016/j.swevo.2011.11.003) Oudalov, A., Chartouni, D., and Ohler, C. (2007). Optimizing a Battery Energy Storage System for Primary Frequency Control. IEEE Trans. Power Syst. 22 (3), [1259–1266. doi:10.1109/tpwrs.2007.901459](https://doi.org/10.1109/tpwrs.2007.901459) Pang, M., Shi, Y., Wang, W., and Pang, S. (2019). Optimal Sizing and Control of Hybrid Energy Storage System for Wind Power Using Hybrid Parallel PSO-GA [Algorithm. Energy. explorat & exploitat. 37 (1), 558–578. doi:10.1177/](https://doi.org/10.1177/0144598718784036) [0144598718784036](https://doi.org/10.1177/0144598718784036) Peng, X., Yao, W., Yan, C., Wen, J., and Cheng, S. (2020). Two-stage Variable Proportion Coefficient Based Frequency Support of Grid-Connected [DFIG-WTs. IEEE Trans. Power Syst. 35 (2), 962–974. doi:10.1109/](https://doi.org/10.1109/tpwrs.2019.2943520) [tpwrs.2019.2943520](https://doi.org/10.1109/tpwrs.2019.2943520) Schott, J. R. (1995). Fault Tolerant Design Using Single and Multicriteria Genetic Algorithm Optimization. Cambridge, MA: Massachusetts Institute of Technology. Sepulveda Rangel, C. A., Canha, L., Sperandio, M., and Severiano, R. (2018). Methodology for ESS-type Selection and Optimal Energy Management in Distribution System with DG Considering Reverse Flow Limitations and Cost Penalties. IET Generation, Transm. Distribution. 12 (5), 1164–1170. [doi:10.1049/iet-gtd.2017.1027](https://doi.org/10.1049/iet-gtd.2017.1027) Sun, K., Yao, W., Fang, J., Ai, X., Wen, J., and Cheng, S. (2020). Impedance Modeling and Stability Analysis of Grid-Connected DFIG-Based Wind Farm with a VSC-HVDC. IEEE J. Emerg. Sel. Top. Power Electron. 8 (2), 1375–1390. [doi:10.1109/jestpe.2019.2901747](https://doi.org/10.1109/jestpe.2019.2901747) Tian, Y., Cheng, R., Zhang, X., Li, M., and Jin, Y. (2019). Diversity Assessment of Multi-Objective Evolutionary Algorithms: Performance Metric and Benchmark [Problems. IEEE Comput. Intelligence Mag. 14 (3), 61–74. doi:10.1109/](https://doi.org/10.1109/mci.2019.2919398) [mci.2019.2919398](https://doi.org/10.1109/mci.2019.2919398) Trovão, J. P., and Antunes, C. H. (2015). A Comparative Analysis of MetaHeuristic Methods for Power Management of a Dual Energy Storage System for Electric Vehicles. Energ. Convers. Manag. 95, 281–296. [doi:10.1016/](https://doi.org/10.1016/j.enconman.2015.02.030) [j.enconman.2015.02.030](https://doi.org/10.1016/j.enconman.2015.02.030) Wang, B., Yang, Z., Lin, F., and Zhao, W. (2014). An Improved Genetic Algorithm for Optimal Stationary Energy Storage System Locating and Sizing. Energies 7 [(10), 6434–6458. doi:10.3390/en7106434](https://doi.org/10.3390/en7106434) Wang, Y., Wu, L., and Yuan, X. (2010). Multi-objective Self-Adaptive Differential Evolution with Elitist Archive and Crowding Entropy-Based Diversity Measure. [Soft Comput. 14 (3), 193–209. doi:10.1007/s00500-008-0394-9](https://doi.org/10.1007/s00500-008-0394-9) Wong, L. A., Ramachandaramurthy, V. K., Walker, S. L., Taylor, P., and Sanjari, M. J. (2019). Optimal Placement and Sizing of Battery Energy Storage System for Losses Reduction Using Whale Optimization Algorithm. J. Energ. Storage. 26, [100892. doi:10.1016/j.est.2019.100892](https://doi.org/10.1016/j.est.2019.100892) Wu, T., Shi, X., Liao, L., Zhou, C., Zhou, H., and Su, Y. (2019). A Capacity Configuration Control Strategy to Alleviate Power Fluctuation of Hybrid Energy Storage System Based on Improved Particle Swarm Optimization. [Energies 12 (4), 642. doi:10.3390/en12040642](https://doi.org/10.3390/en12040642) Yang, B., Wang, J., Chen, Y., Li, D., Zeng, C., Guo, Z., et al. (2020). Optimal Sizing and Placement of Energy Storage System in Power Grids: a State-Of-The-Art ----- One-Stop Handbook. J. Energ. Storage. 32, 101814. [doi:10.1016/](https://doi.org/10.1016/j.est.2020.101814) [j.est.2020.101814](https://doi.org/10.1016/j.est.2020.101814) Yu, H., Tarsitano, D., Hu, X., and Cheli, F. (2016). Real Time Energy Management Strategy for a Fast Charging Electric Urban Bus Powered [by Hybrid Energy Storage System. Energy 112, 322–331. doi:10.1016/](https://doi.org/10.1016/j.energy.2016.06.084) [j.energy.2016.06.084](https://doi.org/10.1016/j.energy.2016.06.084) Zhang, X. S., Yu, T., Yang, B., and Cheng, L. F. (2017). Accelerating Bio-Inspired Optimizer with Transfer Reinforcement Learning for Reactive Power [Optimization. Knowledge-Based Syst. 116, 26–38. doi:10.1016/j.knosys.2016.10.024](https://doi.org/10.1016/j.knosys.2016.10.024) Zhou, B., Fang, J. K., Ai, X. M., Yang, C. X., Yao, W., and Wen, J. Y. (2021). Dynamic Var Reserve-Constrained Coordinated Scheduling of LCCHVDC Receiving-End System Considering Contingencies and Wind [Uncertainties. IEEE Trans. Sust. Energ. 12 (01), 469–481. doi:10.1109/](https://doi.org/10.1109/tste.2020.3006984) [tste.2020.3006984](https://doi.org/10.1109/tste.2020.3006984) Conflict of Interest: Authors SP, XG, XiL, and XuL were employed by the company China Southern Power Grid Company Limited. The remaining author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Publisher’s Note: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher. Copyright © 2021 Peng, Gong, Liu, Lu and Ai. This is an open-access article [distributed under the terms of the Creative Commons Attribution License (CC BY).](https://creativecommons.org/licenses/by/4.0/) The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. ----- ## GLOSSARY ### BESSs battery energy storage systems CRF capital recovery factor DN distribution network IPDM ideal-point decision method MMSSA multi-objective memetic salp swarm algorithm MOPSO multi-objective particle swarm optimization O&M operation and maintenance PV photovoltaic RESs renewable energy sources SOC state of charge SSA salp swarm algorithm TOU time of use WT wind turbines ### Variables PBESS,i power capacity of the ith BESSs. EBESS,i energy capacity of the ith BESSs. Pcha,i(t) charging power of the ith BESSs at time t Pdis,i(t) discharging power of the ith BESSs at time t ρloss TOU electricity prices Ploss(t) power loss at time t Pgrid(t) power fluctuation of the grid connection point at time t x[j]mi [positions of the][ i][th follower salp in the][ m][th salp chain] F[j]m [position of food source] ωh weights of the hth objective function n population size of single salp chain M the number of salp chains N r the maximum size of the repository -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3389/fenrg.2021.707718?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3389/fenrg.2021.707718, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.frontiersin.org/articles/10.3389/fenrg.2021.707718/pdf" }
2,021
[]
true
2021-07-27T00:00:00
[ { "paperId": "db5e3e4a61e799fe2b75f4d2aac434a64673c15c", "title": "Dynamic Var Reserve-Constrained Coordinated Scheduling of LCC-HVDC Receiving-End System Considering Contingencies and Wind Uncertainties" }, { "paperId": "cfb29ed8f75329bd3aae7da8ee9bf70a97f88870", "title": "Optimal sizing and placement of energy storage system in power grids: A state-of-the-art one-stop handbook" }, { "paperId": "4ffcd25e146e022c0e0c6694e30625d3069d1464", "title": "Clarifications of and improvements to the equations used to calculate the levelized cost of electricity (LCOE), and comments on the weighted average cost of capital (WACC)" }, { "paperId": "84f8ea26638bbbb95faaf13043745fbda78d99e2", "title": "A Hybrid Framework for Topology Identification of Distribution Grid With Renewables Integration" }, { "paperId": "0cd6ccae312339957f27b63eba409755e7243b7b", "title": "Impedance Modeling and Stability Analysis of Grid-Connected DFIG-Based Wind Farm With a VSC-HVDC" }, { "paperId": "76275f466a409a87d050a0ce9cdeb3827b3647f6", "title": "Two-Stage Variable Proportion Coefficient Based Frequency Support of Grid-Connected DFIG-WTs" }, { "paperId": "65b178fcf2fa537755a40db47bb2a5652b4d1e2f", "title": "Equilibrium optimizer: A novel optimization algorithm" }, { "paperId": "7658a0511537f14941f5f23b444b0e3896357b69", "title": "Design of minimum cost degradation-conscious lithium-ion battery energy storage system to achieve renewable power dispatchability" }, { "paperId": "bc65f033fd96777959e04f489c13057642e5f772", "title": "Multi-objective optimization strategy for distribution network considering V2G-enabled electric vehicles in building integrated energy system" }, { "paperId": "04a86f467bde4ff182d32be80431a72fa53a9f2a", "title": "Optimal integration of DGs into radial distribution network in the presence of plug-in electric vehicles to minimize daily active power losses and to improve the voltage profile of the system using bio-inspired optimization algorithms" }, { "paperId": "616998c81f1fe8fa26bc6273e7af7e56bcc64736", "title": "RETRACTED ARTICLE: Multi-objective energy management in microgrids with hybrid energy sources and battery energy storage systems" }, { "paperId": "b9ba90834b61e8e378bcbb36ea5abf0585da73a4", "title": "Impact of Power Grid Strength and PLL Parameters on Stability of Grid-Connected DFIG Wind Farm" }, { "paperId": "b12945cd80dbcfa593093ba292ebb1649dc6ced8", "title": "Optimal placement and sizing of battery energy storage system for losses reduction using whale optimization algorithm" }, { "paperId": "7cc01fd5ff07da3688e8cdd2236a0788e33b2f7f", "title": "Diversity Assessment of Multi-Objective Evolutionary Algorithms: Performance Metric and Benchmark Problems [Research Frontier]" }, { "paperId": "ef82a3ce236bb81aed19cdc2bcb74f2ce347e347", "title": "Security constrained co-planning of transmission expansion and energy storage" }, { "paperId": "7ef3ed9198de3ddd5f207485f63e39781ffbf670", "title": "NSGA-II and MOPSO based optimization for sizing of hybrid PV/wind/battery energy storage system" }, { "paperId": "3e781240b4ecaf8f1fdb45be7be9e6ffa1ec9837", "title": "A Capacity Configuration Control Strategy to Alleviate Power Fluctuation of Hybrid Energy Storage System Based on Improved Particle Swarm Optimization" }, { "paperId": "f8ae130acf9b858bf3101c0e56ac76c4be720974", "title": "An adaptive learning control strategy for standalone PV system with battery-supercapacitor hybrid energy storage system" }, { "paperId": "7d4d83894ca153fb070390839783feb9c36099ba", "title": "Optimal sizing and control of hybrid energy storage system for wind power using hybrid Parallel PSO-GA algorithm" }, { "paperId": "f4402fb740019d901dd81a8ac7e6ac840dafab5e", "title": "Optimal planning of energy storage system in active distribution system based on fuzzy multi-objective bi-level optimization" }, { "paperId": "21aef7c41cd90a84f8e2e47b83923a218cfe4a23", "title": "Energy storage capacity optimization for autonomy microgrid considering CHP and EV scheduling" }, { "paperId": "2096fc288ef32942b52602fc671a8fc1bca5c001", "title": "Salp Swarm Algorithm: A bio-inspired optimizer for engineering design problems" }, { "paperId": "9d6504270f62b47976b99ea8597a096a602957fc", "title": "Methodology for ESS-type selection and optimal energy management in distribution system with DG considering reverse flow limitations and cost penalties" }, { "paperId": "3397acfb90c9058ad41df3d2e065c00ee2f8dd1c", "title": "Historical-Data-Based Energy Management in a Microgrid With a Hybrid Energy Storage System" }, { "paperId": "45199bf25e7041cabbbcf7b6efe34f99e724c7ef", "title": "Accelerating bio-inspired optimizer with transfer reinforcement learning for reactive power optimization" }, { "paperId": "382bc5f2e4b9d934bd277edab3f24c41eef42088", "title": "Optimum battery energy storage system using PSO considering dynamic demand response for microgrids" }, { "paperId": "60e3cebdc25a0ae06e1e5680d23f3b2537818e03", "title": "An optimal control strategy for standalone PV system with Battery-Supercapacitor Hybrid Energy Storage System" }, { "paperId": "f95e4e3458530eea5d1e48b6add13346e96f7908", "title": "Optimization of a battery energy storage system using particle swarm optimization for stand-alone microgrids" }, { "paperId": "71db2286f08c02c645b124f63b78599ec61fa7ef", "title": "Real time energy management strategy for a fast charging electric urban bus powered by hybrid energy storage system" }, { "paperId": "cf4fbd38465721cfa3d4c43a3291620ab87d6423", "title": "A comparative analysis of meta-heuristic methods for power management of a dual energy storage system for electric vehicles" }, { "paperId": "e3315a2156b2f27484f8df4d722d638328d84847", "title": "Electrical energy storage systems: A comparative life cycle cost analysis" }, { "paperId": "aec372fd2af3de28e896165e04ec86a26d0ae61d", "title": "A Big Data Architecture Design for Smart Grids Based on Random Matrix Theory" }, { "paperId": "075ef9a314d99c57d2c7fa89cd6237da432f150d", "title": "An Improved Genetic Algorithm for Optimal Stationary Energy Storage System Locating and Sizing" }, { "paperId": "df37beb5bbda42f039a059e571de253b043a221c", "title": "Memetic algorithms and memetic computing optimization: A literature review" }, { "paperId": "9618c8c9debef6f3c8e613641df49b68f47b542d", "title": "Multi-objective self-adaptive differential evolution with elitist archive and crowding entropy-based diversity measure" }, { "paperId": "16494c8936b6d8c2916aeaa7e0f84ccdecfeff4f", "title": "Optimizing a Battery Energy Storage System for Primary Frequency Control" }, { "paperId": "73ee298302fab0d92ec445266b46ad3b57dde858", "title": "Handling multiple objectives with particle swarm optimization" }, { "paperId": "461f133dce9728aa66029efb0b233ab31ba7edb8", "title": "Water distribution network design using the shuffled frog leaping algorithm" }, { "paperId": "339a43d4eee1452921b51ffe6ecf9b8224cd2a82", "title": "Fault Tolerant Design Using Single and Multicriteria Genetic Algorithm Optimization." }, { "paperId": "7cc2bc3b1c6a673d3751e26c3c6b0e3feebc1cdb", "title": "Genetic Algorithms for Multiobjective Optimization: FormulationDiscussion and Generalization" }, { "paperId": "8b9a748ae77f9235396e04301b82143feb1167fe", "title": "On Evolution, Search, Optimization, Genetic Algorithms and Martial Arts : Towards Memetic Algorithms" } ]
13,344
en
[ { "category": "Medicine", "source": "external" }, { "category": "Medicine", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01324b0a808e3440c60626f6cacec48ecc261d44
[ "Medicine" ]
0.915568
A pay for performance scheme in primary care: Meta-synthesis of qualitative studies on the provider experiences of the quality and outcomes framework in the UK
01324b0a808e3440c60626f6cacec48ecc261d44
BMC Family Practice
[ { "authorId": "7268962", "name": "N. Khan" }, { "authorId": "2272463225", "name": "David Rudoler" }, { "authorId": "40581660", "name": "Mary McDiarmid" }, { "authorId": "2249555", "name": "S. Peckham" } ]
{ "alternate_issns": null, "alternate_names": [ "BMC Fam Pract" ], "alternate_urls": [ "https://link.springer.com/journal/12875", "http://www.pubmedcentral.nih.gov/tocrender.fcgi?journal=29", "https://bmcfampract.biomedcentral.com/" ], "id": "7faf8088-9241-4d06-8503-9d72da732aab", "issn": "1471-2296", "name": "BMC Family Practice", "type": "journal", "url": "http://www.biomedcentral.com/bmcfampract/" }
Background The Quality and Outcomes Framework (QOF) is an incentive scheme for general practice, which was introduced across the UK in 2004. The Quality and Outcomes Framework is one of the biggest pay for performance (P4P) scheme in the world, worth £691 million in 2016/17. We now know that P4P is good at driving some kinds of improvement but not others. In some areas, it also generated moral controversy, which in turn created conflicts of interest for providers. We aimed to undertake a meta-synthesis of 18 qualitative studies of the QOF to identify themes on the impact of the QOF on individual practitioners and other staff. Methods We searched 5 electronic databases, Medline, Embase, Healthstar, CINAHL and Web of Science, for qualitative studies of the QOF from the providers’ perspective in primary care, published in UK between 2004 and 2018. Data was analysed using the Schwartz Value Theory as a theoretical framework to analyse the published papers through the conceptual lens of Professionalism. A line of argument synthesis was undertaken to express the synthesis. Results We included 18 qualitative studies that where on the providers’ perspective. Four themes were identified; 1) Loss of autonomy, control and ownership; 2) Incentivised conformity; 3) Continuity of care, holism and the caring role of practitioners’ in primary care; and 4) Structural and organisational changes . Our synthesis found, the Values that were enhanced by the QOF were power, achievement, conformity, security, and tradition. The findings indicated that P4P schemes should aim to support Values such as benevolence, self-direction, stimulation, hedonism and universalism, which professionals ranked highly and have shown to have positive implications for Professionalism and efficiency of health systems. Conclusions Understanding how practitioners experience the complexities of P4P is crucial to designing and delivering schemes to enhance and not compromise the values of professionals. Future P4P schemes should aim to permit professionals with competing high priority values to be part of P4P or other quality improvement initiatives and for them to take on an ‘influencer role’ rather than being ‘responsive agents’. Through understanding the underlying Values and not just explicit concerns of professionals, may ensure higher levels of acceptance and enduring success for P4P schemes.
p g ## RESEARCH ARTICLE Open Access # A pay for performance scheme in primary care: Meta-synthesis of qualitative studies on the provider experiences of the quality and outcomes framework in the UK ### Nagina Khan[1*], David Rudoler[2], Mary McDiarmid[3] and Stephen Peckham[4] Abstract Background: The Quality and Outcomes Framework (QOF) is an incentive scheme for general practice, which was introduced across the UK in 2004. The Quality and Outcomes Framework is one of the biggest pay for performance (P4P) scheme in the world, worth £691 million in 2016/17. We now know that P4P is good at driving some kinds of improvement but not others. In some areas, it also generated moral controversy, which in turn created conflicts of interest for providers. We aimed to undertake a meta-synthesis of 18 qualitative studies of the QOF to identify themes on the impact of the QOF on individual practitioners and other staff. Methods: We searched 5 electronic databases, Medline, Embase, Healthstar, CINAHL and Web of Science, for qualitative studies of the QOF from the providers’ perspective in primary care, published in UK between 2004 and 2018. Data was analysed using the Schwartz Value Theory as a theoretical framework to analyse the published papers through the conceptual lens of Professionalism. A line of argument synthesis was undertaken to express the synthesis. Results: We included 18 qualitative studies that where on the providers’ perspective. Four themes were identified; 1) Loss of autonomy, control and ownership; 2) Incentivised conformity; 3) Continuity of care, holism and the caring role of practitioners’ in primary care; and 4) Structural and organisational changes. Our synthesis found, the Values that were enhanced by the QOF were power, achievement, conformity, security, and tradition. The findings indicated that P4P schemes should aim to support Values such as benevolence, self-direction, stimulation, hedonism and universalism, which professionals ranked highly and have shown to have positive implications for Professionalism and efficiency of health systems. (Continued on next page) [* Correspondence: nkhan786can@gmail.com](mailto:nkhan786can@gmail.com) 1Independent Researcher, Ontario, Canada Full list of author information is available at the end of the article © The Author(s). 2020 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain [permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.](http://creativecommons.org/licenses/by/4.0/) [The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the](http://creativecommons.org/publicdomain/zero/1.0/) data made available in this article, unless otherwise stated in a credit line to the data. ----- Background Internationally, there has been substantial interest in the use of pay for performance (P4P) schemes for primary care in high, medium and low-income countries. The longest standing and most comprehensive scheme, is the Quality and Outcomes Framework (QOF) for United Kingdom (UK) general practice. However, in the UK there have been increasing calls for the QOF to be abolished and in 2016 Scotland ended the scheme. The QOF now continues only in England, Wales and Northern Ireland [1, 2]. In early 2017, the British Medical Association (BMA) called for the QOF to be suspended to reduce bureaucratic pressures and free up clinical time [3]. In April 2016, National Health Service (NHS) England commenced a review of the QOF, acknowledging that it may have ‘served its purpose’ and may be ‘a barrier to holistic management’ [4, 5]. Published in July 2018 the Review of the Quality and Outcomes Framework in England [6], concluded that the scheme should be revised with a greater emphasis on an approach that would “… increase the likelihood of improved patient outcomes, decrease the likelihood of harm from over-treatment and improve personalisation of care” (p11). Among the recommended changes, the report outlined an approach that included supporting practices to undertake quality improvement activities set out in the GP contract for 2019/20 [6]. It also supported the development of pooled incentive schemes or shared savings programs for networks of practices [6]. In England the proposal for shared savings and financial incentive schemes signals a shift from the focus on individual practices with new incentive schemes seeking to influence primary care professional behaviour through more collective and quality improvement approaches to “… facilitate achievement of system efficiencies and increase income for reinvestment to primary care networks” [6]. While QOF has predominantly had a clinical practice focus (some process and organisational criteria were dropped after just a few years), it has always had a practice wide impact and studies suggest it has had a significant influence on the functioning and organisation of practices [6, 7]. Though the QOF has had an impact on clinical practice, it has also had some unintended consequences. Understanding the importance and impact of these consequences is useful for decision-makers designing P4P schemes [7]. To date, most studies of the QOF have used quantitative methods to evaluate the impact of QOF on clinical performance [8–10] and the universally high QOF achievement means that practices have little motivation to improve achievement further. However, ‘high performance’ does not necessarily mean ‘high quality’ [6]. Motivation to deliver high quality care among health professionals is complex, but it is likely that other motivational factors other than financial rewards may be effective [6]. Therefore, it is important to consider other ways of motivating health professionals to deliver high quality care [6]. MacDonald and others have argued that it is possible to avoid unintended consequences of P4P systems if they are designed with the involvement of clinicians and aligned with their underlying values [11, 12]. As governments are developing schemes for quality improvement, they need relevant and context-sensitive evidence to support policy interventions, which means that there is significant ambiguity over the optimal design of such schemes to maximise efficiency and tolerability. Decisionmakers are increasingly using qualitative evidence to understand various socioeconomic contexts, health systems and communities [13]. Furthermore, this type of evidence is useful to assess the needs, values, perceptions and experiences of stakeholders, including policymakers, providers, communities and patients, and is thus crucial for complex health decision-making [7, 13]. For this paper, we conducted a meta-synthesis of the available qualitative research on QOF to identify lessons that will be useful for decision-makers in designing and implementing new incentive schemes. Drawing on evidence from the UK provides the widest range of studies on one scheme from which to develop clear lessons for those factors that might support or hinder particular behaviours and outcomes within P4P schemes. Method For this review, we sought to understand impacts of QOF on the individual clinicians and other groups of ----- professionals in primary care, using a Lines-of-argument (LOA) synthesis. The LOA synthesis involves building up a picture of the whole from the studies of its parts [14] and assists knowledge synthesis through a process of re-conceptualisation of themes across several published qualitative papers [14, 15] and is a interpretative approach. We then applied the Schwartz Value Theory as a theoretical framework to our synthesis. Schwartz proposes that there are ten broad Value Domains that are universal and fairly comprehensive [16]. The theory defines these ten broad Values according to the motivation that underlies each of them (described in Table 1) [17]. Although the theory discriminates ten Values, it postulates that, Values form a continuum of related motivations (the circular structure in Fig. 1 portrays the total pattern of relations of conflict and congruity among Values, the closer the Values are on the circular structure then that indicates that they are more congruent and the further away they are, indicates that they are more conflicting [20]. The theory explains that among some Values there is conflict with one another (e.g., benevolence and power) whereas others are congruent (e.g., conformity and security) [18]. One basis of the Value structure is the fact that actions in pursuit of any Value have consequences that conflict with some Values but are congruent with others. Also actions in pursuit of Values have practical, psychological, and social consequences for professionals [17] and their profession [21]. Professionalism is fundamental to good medical practice and so Professor Dame Judy Dacre also states Medical Professionalism has changed and must keep up to date with the demands of modern day clinical practice Table 1 Schwartz Value Theory: The Ten Basic Values Openness to change Self-Direction: Independent thought and action—choosing, creating, exploring. Stimulation: Excitement, novelty and challenge in life. Hedonism: Pleasure or sensuous gratification for oneself. Self-enhancement Achievement: Personal success through demonstrating competence according to social standards. Power: Social status and prestige, control or dominance over people and resources. Conservation Security: Safety, harmony, and stability of society, of relationships, and of self. Conformity: Restraint of actions, inclinations, and impulses likely to upset or harm others and violate social expectations or norms. Tradition: Respect, commitment, and acceptance of the customs and ideas that one’s culture or religion provides. Self-transcendence Benevolence: Preserving and enhancing the welfare of those with whom one is in frequent personal contact (the ‘in-group’). Universalism: Understanding, appreciation, tolerance, and protection for the welfare of all people and for nature. [22]. It has been postulated that the professional organisation of medical work no longer reflects the changing health needs caused by the growing number of complex and chronically ill patients [21]. The Royal College of Physicians (RCP) redefined Professionalism in 2018, advising its benefits for patients, that it increases the job satisfaction of doctors, makes for superior organisations, and improves the productivity of health systems. The RCP defined Professionalism as ‘a set of values, behaviours and relationships that underpin the trust the public has in doctors’ [22]. They described seven professional roles; doctor as healer, patient partner, team worker, manager and leader, advocate, learner and teacher and as an innovator (Table. 3). The importance of Medical Professionalism has been well documented in the literature [31], together with its effects on the doctors’ relationships with their patients, quality of care, and ultimately health and illness outcomes [32]. For that reason, we further include Professionalism as a conceptual lens to contextualise our analysis in this review [33]. Search strategy and data extraction To identify relevant studies, we searched for peerreviewed empirical research on QOF using the electronic database searching, hand-searching and web-based searching. The following databases were initially searched: Medline, Embase, Healthstar, CINAHL, and Web of Science. We also searched the reference lists of obtained papers. The details of our electronic search are included in the Additional file 1. We included studies that reported primary qualitative research (in-depth interviews, focus groups, ethnography, observation, reflective diaries, case-studies and reviews containing qualitative analysis) of the QOF published in English between 2004 (when QOF was introduced) and 2018. We excluded studies that did not specifically focus on the QOF, UK and did not involve primary qualitative research methods. The search of electronic databases identified 33 relevant papers (see Fig. 2, PRISMA flowchart, including reasons for exclusion). We excluded 15 papers and the 18 papers included were independently reviewed by two researchers (N.K and D.R) and any disagreements discussed. We erred on the side of caution and endeavoured to keep all the 18 studies in until the researcher (N.K.) had independently extracted data from these papers and applied the exclusion criteria. The researcher (N.K.) extracted data and assessed the eligibility criteria for all retrieved papers, which were then appraised by a second researcher (D.R.). Disagreements between researchers’ were resolved through discussion with S.P. Differences between researchers tended to arise because of different understandings of some of the study questions and because of different ----- interpretations of what authors of the papers had written and generally related to the qualitative research methods used. The qualitative papers were initially quality assessed by N.K. using the British Sociological Association for the evaluation of qualitative research papers [34] and if any discrepancies arose then they were discussed with S.P. The scale comprises 20 questions about the relevance of the study question, appropriateness of qualitative method, transparency of procedures, and ethics. In order to make judgements about the quality of papers, we dichotomised each question to yes or no, in a separate table. All the qualitative papers included in this synthesis were published in peer reviewed journals and adhered to transparency of high quality work. Following the systematic steps of the metaethnography approach, we included 18 qualitative research studies for the final qualitative synthesis. Data analysis and interpretation Meta-ethnography is a systematic but interpretative approach to analysis that begins with noting verbatim and coded text in terms of first-order and second-order constructs. Then translation of these constructs were synthesised across papers to form third-order constructs, and finally constructing the synthesis using either reciprocal, refutational, or line of argument approaches [15, 35]. Our data analysis was undertaken using a ‘line of argument’ synthesis which serves to reveal what is hidden in individual studies and to discover a ‘whole’ among a set of parts [15]. This method has previously been adapted for utility in the syntheses of qualitative data in healthcare research [35, 36]. We placed the 18 papers identified in a table that included relevant details of the study setting and research design (see Additional file 2 Table. 4). Our first-order constructs represented the primary data reported in each paper (see Additional File 3 [A3], Table. 5). The emergent themes from the papers represented our second-order constructs (A3, Table. 5). They were extracted utilising a more fine-grained approach, in which the researcher (N.K.) went through each paper in a detailed and line-by-line manner and the papers were reviewed for common and recurring concepts. As a way of remaining faithful to the meanings and concepts of each study; we preserved the terminology used in the original papers in the grids. We then combined and synthesised these themes (taken from the published papers) to create our third-order constructs (see Additional file 4, Table. 6). Each cell of the table was considered in turn, from this, we identified our key concepts and consequent themes and once these where identified (see in an Additional file 5, Table. 7), we simultaneously mapped the concepts against the ten Values using the ten Values as our theoretical framework (Fig. 3). These were then compared with Professionalism as defined by the seven professional roles in Table. 3. ----- Results Main findings from the synthesis The 18 papers were published between 2008 and 2018 in the UK. The 18 papers included where of the providers perspective; general practitioners (GP) including GP leads, principals, partners and salaried [23–29, 37– 42], nurses [25, 26, 37, 38, 43, 44] (practice and condition specialist) [28–30, 37, 40–42, 44, 45], healthcare assistants [25, 37, 45] and administrative staff [25, 26, 30, 37, 38, 41, 45] (practice managers, IT) on their views and experiences of the QOF. The majority of papers utilised one to one retrospective semi-structured interviews [23, 24, 26, 27, 30, 40–44, 46, 47], focus groups [6, 28, 37, 45, 48], observations [25, 39], using thematic analysis [27, 37, 38, 41], framework approach [26, 44], constant comparison [29, 43] (additionally, see supplementary material 2 for the summary of the sample size, research questions and individual participant characteristics, including the findings from the studies). The synthesis identified four themes (Table. 2): 1) Loss of autonomy, control and ownership; 2) Incentivised conformity; 3) Continuity of care, holism and the caring role of practitioners in primary care; and 4) Structural and organisational changes. In the next section we present the thematic analysis (summarised in Table. 3) which includes the application of the synthesis to the ten Values with implications for aspects of Professionalism. Loss of autonomy, control and ownership We found that this theme identified from the published papers [6, 25–28, 30, 42] included professionals’ submission to the QOF targets despite their applied concerns. Such as the ethical distress caused by a reductionist approach to managing markers of chronic disease and its ----- being incompatible with the humanitarian values of general practice [49]. For instance most health professionals believed that they needed to place biomedical care in the context of their patients’ concerns and life experience [50]. We also found that professionals wanted to retain control and clinical autonomy; however on closer examination and within the context of the QOF this took the form of modifying the way structured tools of the QOF were utilised by the professionals. “The more templates that get introduced, it takes away the clinicians freedom and that sort of rapport that you can build with a patient is much more difficult when you have to go through set (depression score) questions.” (p. 413 ) [42] “…but I don’t particularly like them... because I tend to write my notes and then do everything on the computer when the patients gone.” (p. 57) [45] Both, professionals and patients were aware of the QOF targets acting as an independent mechanism of control, which essentially changed the nature of the discussion between patient and professionals [26, 29, 44, 45]. “Some patients will come to you and they’ll plead with you, ‘Please don’t give me any tablets, I’ll bring my blood pressure down, I’ll do anything. I’ll bring it down’, and again they’re not horrendously high, they’re like say 140/90 or whatever … but we’re saying to them ‘well, look we’ve checked it three times now and it remains raised, you’re clinically classed as hypertensive, we follow these guidelines and this is what we should be doing with you.” (p. 143) [25] Nearly all the published papers showed that the main motivation for practice staff to follow the QOF targets was the link with income loss [23, 24, 26, 29, 40–45]. “So if you deviate from that [QOF] because of the individual need. You have complete autonomy to, but there are financial implications to you because of that…So you still have autonomy, but you lose income.” (p. 57) [24] This created a conflict for practice staff and suggested a decreasing sense of clinical autonomy. Especially in areas that were clinical and easy to measure and were bound by templates or driven through the use of IT tools [24, 42]. Respondents in one of the ----- Table 2 The Impact of the QOF Mapped to the Ten Motivational Values QOF modifications Synthesis of the main findings Influence on ten basic values [18] Congruent Power Conformity Security Achievement, Conflict Self-direction Stimulation Benevolence, Universalism Hedonism, Tradition Congruent Achievement Conformity Security Power Tradition Conflict Self-direction Stimulation Benevolence Hedonism Universalism Congruent Conformity Power Security Achievement power Conflict Benevolence Universalism Self-direction Stimulation Tradition Congruent Power Conformity Achievement Security Stimulation Self-direction Universalism Conflict Tradition Benevolence Hedonism Templates Guidelines Indicators Governmental goals Raised standards in basic care Drove provider care Systemized and standardised care Neglected areas of care targeted Focus on chronic disease management Certain aspects of professionalism threatened Indicators conflict -patient advocate Information technology (IT) Practice managers Increased skill mix Monitoring systems Recording performance Surveillance (a) Loss of autonomy, control and ownership Most papers described a sense of decreased clinical autonomy and loss of professionalism [39]. They also described a sense of micromanagement from above [28] and frequently cited the late communication about changes to the wider QOF and year-on-year variability in the occurrence and timing of changes to indicators as politically motivated [28, 39]. (b) Incentivised conformity In the papers reviewed professionals recognized that QOF had led to considerable extra income at the practice level [29]. As the owners of their organizations, economic factors were more salient and apparent in principals’ accounts. Subsequently the finance and achieving maximum income became an increasingly key issue in participants’ beliefs about QOF and their adherence to QOF work [28]. (c) Continuity of care, holism and the caring role of clinicians in primary care Although participants in the papers reviewed emphasised the importance of traditional general practice values, such as holism and continuity, the majority felt that the 2004 changes had negatively impacted on these values. Participants related that patients now experienced less continuity with their GPs [41]. (d) Structural & organisational changes All the practices that were studied in the papers included in the review had changed their modes of operation in response to the QOF [27, 29, 43, 45]. Role of monitoring compliance with the coding regime which feeds into the contract monitoring system and of highlighting deficient coding and recording performance amongst staff, contributed to on-one-hand to increased surveillance and on the other to the doctors sense of self-worth [45]. papers suggested that most of the internationally agreed attributes of medical professionalism were not perceived or described as being threatened by the introduction of the QOF [42]. Although, on further analysis we found that acquiring a say in the development of indicators was important to GPs and was linked to freedom to practise in the patient’s best interest, indicating that aspects of Professionalism were being affected [22]. Incentivised conformity The papers indicated that extensive improvement in QOF scores was perceived as a result of consistency and recording of incentivised activities, the outcomes and new protocols being introduced within practices and that these were now connected to the wider governmental objectives through the mechanism of the QOF [23, 24, 26, 29, 40–45]. “…There are lots of systems in operation here that other people are operating.” (p. 53) [45] “It’s raised standards, narrowed health inequalities, and introduced evidence-based medicine and err the rest of the world look up on err us and our implementation of QOF with a degree of envy. Its ----- Table 3 Application of the Synthesis to the Values and Implication for Professionalism Main Themes Application of the findings to the Values Implications for aspects of professionalism [22] (a) Loss of autonomy Activated values When values are stimulated, they become infused with feeling. Therefore, GPs for whom independence is an important value may experience provocation if their independence (self-direction) seemed to be threatened, discouraged when they are helpless to keep their professional autonomy (power), and would be happy when they can enjoy their freedom as self-regulated practitioners (security). Control and ownership Professionals appeared preoccupied by their lack of control in achieving indicator targets (achievement), especially if dependent upon patient cooperation, quality of care (security), and implementation of outsider perceived changes (power) [23, 24]. (b) Incentivised conformity Motivating actions Those GPs for whom social order, justice, and medical superiority (power, achievement, and security) are important values are motivated to pursue these incentivised goals (self-satisfaction) in the context of pay for performance schemes. GPs’ values form an ordered system of priorities that characterise them as individuals and general practitioners (professionalism) with specialist set of values, behaviours and relationships that underpin the trust the public has in doctors [22] (tradition, benevolence, universalism, tradition). GPs that hold expert positions as generalist medical practitioners are seen as first point of contact for patients in healthcare services (power, security). They offer a doctor patient relationship with mutual understanding of problems that are brought into the practice (tradition, benevolence, universalism). (c) Continuity of care, holism and the caring role of clinicians in primary care Consequences of cherished values Holism and continuity of care (benevolence) for example are relevant in the workplace for GPs (universalism). There was a tension between the standardised QOF driven care, being ‘patient-centred’ with clinicians reporting that “it’s not always easy to deal with disregarding, or setting aside a patient’s’ perceived need or to move onto a more pressing practice target (conformity) during personal discussions” [23, 25–30]. The trade-off between relevant, competing values guides attitudes and behaviours. When values are shown to be in conflict, not corresponding to the cherished value, then do practitioners attribute more importance to their achievement (completing QOF targets, case finding etc.) or justice (work in best interest of others, benevolence, universalism), and to novelty or tradition (medical model). Any attitude or behaviour typically has implications for more than one value. For example, A ‘tick box’ approach to medicine encouraged by pay for performance indicators might express and promote EBM and conformity values at the expense of hedonism and stimulation values for GPs. Values influence action when they are relevant in the context (specific) – such as in pay for performance (hence likely to be activated) and important to the GPs (Status, professional progression, and EBM – achievement, power, security) and bureaucrats (focus on GPs performance to the QOF targets-conformity). Doctor as manager and leader Loss of autonomy impacts clinical engagement and leadership which is pivotal to the success of health systems. Doctors make decisions that determine where resources flow. Yet there is a conflict experienced between doctors as employees of huge complex systems and the autonomy of individual doctors. Autonomy is crucial for the delivery of care, but modern autonomy is more complex and nuanced and needs greater judgement [22]. Doctor as team worker Relinquishing control is important to allow an important component of teamwork as professional satisfaction, engagement, and effective teamwork improves patient outcomes and satisfaction, as well as organisational performance and productivity. Teamwork has become more important because of the growing complexity of patients’ problems and health systems, and the increasing range of possible interventions [22]. Doctor as advocate Professionalism requires that doctors’ advocate on behalf of their patients, all patients and future patients, yet incentivised conformity and indicators conflicted with this aspect. However, this was one concern that should be given the highest priority to advocate on patient safety. Raising concerns about poor care, or the potential for poor care, is a professional duty for all doctors but is not easy; such advocacy needs training, practice, and mentorship [22]. Doctor as patient partner The patient–doctor relationship is at the core of the doctor’s work. The traditional relationship of patient deference to doctors has been replaced by an equal partnership. Values, including integrity, respect, and compassion must underpin the partnership with patients. Integrity involves staying up to date, but also being willing to admit one’s limitations. Doctors can show respect for patients by listening to them actively, involving them in decisions, and respecting their choices (patient centred). Compassion means not just recognising the suffering of the patient, acting to reduce the suffering [22]. ----- Table 3 Application of the Synthesis to the Values and Implication for Professionalism (Continued) Main Themes Application of the findings to the Values Implications for aspects of professionalism [22] (d) Structural & Multiple values organisational changes Values guide the selection or evaluation of actions, policies, people, and events in practice organisations. Hence, GPs in self-regulated disciplines (self-direction) decide what is good or bad, justified or illegitimate, worth doing or avoiding, based on possible consequences for their cherished values. But the impact of values in everyday decisions is rarely conscious and activates a multiple set of values. The results show GP values entered awareness when the QOF actions or judgments GPs were considering had antagonistic or conflicting implications for multiple values they also cherished. Such as undertaking templates use (IT) during consultations. GPs are guided by professional practice which is regulated by the guidelines agreed by GPs. They work to a degree, autonomously although subject to audit and some monitoring. QOF impinges by directing activity in a standardised way (conformity, power). Doctor as innovator The challenge for doctors is how to innovate amid the innovation happening all around them. The use of machine (in this context -template) learning was feared could lead to the diluted face-to-face patient doctor consultations with a collaboration in which the machine (template) becomes effectively an independent actor. It is doctors, rather than machines, who can provide solidarity, understanding, and compassion to patients [22]. evidence-based medicine, standardised care.” (p. 412) [42] Respondents in papers reviewed, also stated that the incentive payments attached to QOF did drive provider behaviour and that it encouraged them to work towards performance targets [23–26, 29, 40–45]. “They’re trying to control our income and we’re trying to get as much money out of them as we can.” (p. 412) [42] Financial rewards in return for extra work was felt to have increased morale for some within the profession [23, 25, 30, 40]. “We’re so hard up at the moment, so desperate for income wherever we can get it, you can’t afford to pass up a chance of income, so that’s probably as much a driver . . . even if we didn’t necessarily buy into the clinical benefit, it was worth doing to try and earn the money because we needed to.” (p. 7) [26] Practices had experienced rising practice income and our synthesis findings indicated that certain Values were enhanced by this, particularly power, achievement, conformity, security, and tradition values (Fig. 4). Future P4P schemes should aim to support Values such as benevolence, self-direction, stimulation, hedonism and universalism, which professionals ranked highly and have shown to have positive implications for Professionalism and the efficiency of health systems (Fig. 5). Correspondingly, lower job satisfaction was associated with intention to leave general practice [51]. The papers in the synthesis suggest that the rising income was also linked to the practices adherence to the QOF as a factor that led to the gradual routinisation of the scheme into everyday practice increasing systematised and standardised care [25, 26, 29, 30]. It was also acknowledged that some aspects of neglected clinical activity were appropriately targeted by QOF. “Patient care has definitely improved because we’ve been doing that, and so I think some people believe we’re number crunching but I don’t think we are in this practice, I think we are actually meeting targets the patients’ care is benefiting.” (p. 52) [45] Therefore, any changes to QOF are and will be controversial mainly because they represent a substantial proportion of general practitioners’ incomes [52]. Setting the political machinations to one side (the Department of Health has been clawing back from the original settlement since 2004); Gillam and Steel believe that the incentive payments in the QOF also comprise too large a proportion of general practice income. They suggest that money should be taken out of the QOF and redirected to supporting general practice in other ways [52]. However, there is no link between the size of the financial incentive and likely health gain from the activity incentivised [53]. There was also a greater acceptance of standardised approaches [23, 25, 29, 30, 40] which may have restricted personalised care for the individual patient. Even complicated the management of multiple conditions over time [52] and narrowed the focus of the consultation, reducing the time to deal with the wider context of the illness [37]. Further confounded by very limited access to specialist input for patients with more complex treatment resistant or recurrent mental health problems [37]. “We developed this zero tolerance to blood pressure a while ago, no one is allowed to say it’s a little bit up leave it, it’s not acceptable so it has to be if it’s ----- up do something about it, if you’re not doing something about it because if we go and find they’re not on target and you look and they’ve seen somebody and they’ve not acted on it yeh, I’ll have a little word.” (p. 55) [45] “…the interesting thing for me is that since the introduction of PHQ-9 I find in terms of material I’m treating the score, not the patient. Because, you know, it’s such a short barrier in the consultation.” (p. 282) [37] Yet, this does not diminish the ethical imperative to practise in the light of best evidence and the challenge is to deliver good quality technical care for medical conditions while simultaneously considering what is in the best interests of the whole person [52]. Some of the QOF’s design flaws are inherent to all pay for performance schemes [54]. As such, areas of high performance will continue to elicit negative feelings, arising from scepticism about achieving maximum points [23, 25, 29, 30, 40]. “I think it’s anyone who gets maximum points is probably bent, I think it’s almost impossible to get maximum points without some kind of fudge. That maybe unkind but we haven’t got maximum points. . . I think its easy just to tick the boxes when you haven’t done it.” (p. 137) [44] Time pressures were reported to be the motivating factor for prioritising areas of care that were financially incentivised [30]. “I think because there is limited time and if you have to focus on something in order to get the money, obviously if you don’t have the time, then it’s going to be ignored automatically.” (p. 1059) [30] Continuity of care, holism and the caring role of practitioners in primary care Continuity of care, was a central feature of both doctor and practice nurse roles. Organisational and structural changes were attributed to the loss of continuity of care; consequently, accessing the same GP was difficult for patients. “Increased staff numbers and changed working patterns had contributed to a loss in continuity of care and choice of who to see. The appointment targets paradoxically seemed to have made access worse in many practices, due to requirements to book on the same day . . . We’ve had to have increased staff and then you very quickly lose continuity if you’ve got a lot of people waiting.” (p. 136) [44] ----- For most nurses, interpersonal continuity was described as a relatively new feature of their role as they assumed further responsibility for patients with chronic conditions [40]. “…with asthma, the patients are beginning to see the same nurse, you know, rather than a different GP… I will see the diabetics and they know that I’ve been trying to say to them, ‘Can you come, you know you can always come back,’ and I always try and make it so that there is open access for them if they have got a problem.” (p. 230) [40] Holistic care and the caring role of GP practitioners was not recognised in the QOF despite this being seen as a core component of clinical professional roles [22]. Patient-centredness was deemed to be of pronounced significance in the papers reviewed [28, 29, 39, 43]. However, there was a tension between the standardised QOF driven care and being ‘patient-centred’ with practitioners reporting that ‘it’s not always easy to deal with disregarding, or setting aside a patient’s’ perceived need or to move onto a more pressing practice target during these personal discussions’ [24–26, 28–30, 40]. “I tend to deal with the problem patients come with first. And then if it’s appropriate to ask questions, you know, ticking the boxes, I will do that at the end of the consultation.” (p. 231) [40] “We spend a lot of time visiting... and yet frequency of home visits doesn’t get QOF points ... Caring, that’s what doctors do.” (p. 136) [43] Papers showed that GPs were more likely to exception report indicators they perceived as having relatively little systematic evaluation or that they were not proven to work. They felt the indicator was contrary to their role as a patient advocate and in their clinical judgement, not relevant to individual patient-centred care [46]. Patientcentredness was defended by professionals in ‘everyday practice’, given the relevance to patient care and the patient-doctor relationship [52, 55–58]. “…Well I think it has put a lot of strain on the partners and practices to get all the QOF points … I mean when it came to get all these points just to get more money, I think it’s put more strain on doctors and it has lost the … just normal care for ----- patients, taking them as a patient rather than as another … object to get points.” (p. 283) [23] “I think that the art of the job has declined and, I don’t know, the sense of feeling that you could be with people rather than be doing. It’s quite hard to define but there’s more to general practice than doing ... clinical things.” (p. 136) [43] The synthesis indicates that the QOF embodied an approach to achieving evidence-based medicine (EBM), yet we found no evidence in the papers that linked the compatibility of EBM with a more holistic approach to patient-centred care, as perceived by the professionals and as linked to achieving aspects of Professionalism. Structural and organisational changes QOF was viewed as increasing the responsibility of lead partners (doctors) in most areas of their practice. This included supervising the work of nursing colleagues, which was seen as an increase in their workloads [26, 28, 40, 44]. “There is an environment and ethos of increased surveillance and performance monitoring.” (p. 232) [40] “I suppose it feels like I’m being watched. It’s a bit like big brother – you’ve not ticked these boxes.” (p. 232) [40] For some, this has come at the expense of work life balance which manifested as an astonishment with the way their selected profession grasped such issues [24, 43, 44]. “My practice does not understand the concept [work life balance]. And I, we’ve two or three away days a year, I’m often talking about it. And they don’t understand. They’ll take me aside and ‘what do you mean?’ I just find that astonishing you know…, if you have a bereavement of this or that, you just get on with it basically and you don’t expect to be sick for anything… So I mean its just life I’ve chosen, it’s very busy but I do manage to stay sane through it.” (p. 54) [45] Salaried GPs carried less responsibility for QOF activity than the QOF leaders in areas such as surveillance of others, meeting targets on time, and for the business side of the practice [23]. “I think the balance of, of that is [partners] have a lot more responsibility...you have to take a lot more responsibility for the practice and more leadership. And I quite enjoy ... coming in doing the job and, and not having to worry about that so much. And you get paid more money but I think the balance of the hours you’d be spending and their stress of the job would probably be higher as a partner.” (p. 284) [23] Those who eventually wanted to succeed to GP principal status took greater responsibility for QOF activity from those who wanted to remain salaried [24, 40, 45]. “But sometimes you do feel that you are not really involved in decision making. That’s fine for some people, but for me, I do like a bit of control. So I think at the moment its fine, but I think eventually I would want part of the decision making process.” (p. 285) [23] We found that the QOF also impacted the role of nurses but not entirely in same way as it did their GP Colleagues [43, 44]. Nurses initially perceived the changes to their role to be beneficial, which led to professional progression (related to achievement values), however not to any greater authority or any increase in status, which for their GP colleagues were achieved through alignment with the QOF income. “I’m not comparing it [GP salary] to what the papers say they were walking off with, but (they got) financial rewards for a lot of the work that has been done by nurses.” (p. 714 ) [43] P4P schemes have been focused on certain medical professionals that make up the healthcare workforce, and the incentives were focused on rewarding those professionals. Our analysis indicates that the QOF work was distributed throughout primary care practice, involving nurses, managerial staff and healthcare assistants but without monetary reward for these groups [28, 43, 44] and this was experienced by other practice staff as an injustice in the reward system. Yet, the effect of income inequality on population health status continues to be described and the link between population health status and socioeconomic status has long been recognised [57] however this link was discounted by the scheme. Our analysis also showed that except for certain medical professionals, all other groups that made up the primary care staff adhered to the targets without the incentivised reward. As such monetary gain was not the only powerful determinant of employee motivation or positive returns in terms of the QOF performance and success. We also found that the QOF changes for nurses were experienced in isolation of ----- their self-interest and power values, or formal rank (specific to the Nursing discipline), inferring a feeling of continued inequity in primary care practice and healthcare systems. “Because the workload had increased particularly monitoring wise. We needed to do an awful lot more monitoring of the routine measures. So the combination of that, plus the fact that our nurse had done the diabetes course and asthma course and a prescribing course, we felt that she could move on to something a bit more senior and someone else [the new healthcare assistant (HCA)] could do the routine blood pressures and bloods.” (p. 56) [24] As a consequence of achievement and increase in workload, the papers in the synthesis revealed that there was an increased blurring of the boundaries with other medical tasks and between different practitioners [25, 28, 43, 44]. “I do in fact do most of the work for the contract and in many ways that’s not a good thing as it is supposed to be team work.” (p. 714) [43] “... We do the work, the doctor gets the rewards and it is up to him whether he decides to pass it on or not because he gets the global sum now. So that is a bit of a conflict with a lot of the nurses at the moment. So our role and responsibility has expanded but at the same time the wages are staying much the same.” (p. 714) [43] IT systems were seen as a beneficial tool to help professionals as a form of a reminder, to manage record and collect relevant patient illness related data. But on the other hand, it was a system that made visible the performance of professional work against what was increasingly experienced as ‘outsider implemented targets’. It was not perceived as well by the professionals [23, 40, 44, 45] as there was little scope for the professionals’ to retain personal beliefs or to include patient agendas during reviews [26, 29, 43, 44]. “The more templates that get introduced, it takes away the clinicians freedom and that sort of rapport that you can build with a patient is much more difficult when you have to go through set [depression scores] questions.” (p. 413) [42] Application of the findings to the Schwartz value theory and implications for professionalism In addition to identifying ten basic Values, the Values Theory explicates a structure of dynamic relations among them (see Fig. 1). One basis of the value structure is the fact that actions in pursuit of any Value have consequences that conflict with some Values but are congruent with others. Essentially, choosing an action alternative that promotes one Value (e.g., following template work—conformity) may literally contravene or violate a competing Value (disregarding a patients concerns—benevolence). When we think of our values, we think of what is important to us, each of us hold numerous values (e.g., achievement, security, benevolence) with varying degrees of importance [18]. Furthermore, those actions in pursuit of some values alone had practical, psychological, and social consequences for professionals. Participants in some papers stated that most of the internationally agreed attributes of Medical Professionalism were not perceived or described as being threatened by the introduction of pay for performance [42]. Contrariwise, the findings of the synthesis revealed that there was some discord experienced by practitioners with some aspects of Professionalism which we present in this section (See Table. 2). Triggered values: relinquishing control and retaining independence Complexity of both patient problems and health systems now requires professionals to work as an interrelated team within the newer hierarchies and hence a relinquishing of control in achieving QOF targets. Initially the issue of retaining control in making decisions in clinical practice was seen as a contentious issue. The concerns were especially regarding who, which or where the body of evidence that was influencing ‘everyday clinical decisions was originating from’ [24, 45]. Other concerns were about government regulation and its influence on the process of care and protecting the well fare of patients and their treatment [24, 45]. Schwartz argued that when Values are triggered, they become infused with feeling [16]. For instance, Schwartz posited four steps in the activation of personal norms that apply equally to basic values [17]. These steps include, awareness of need, awareness of viable actions, perceiving one-self as able to help and then triggering a sense of responsibility to become involved. The synthesis indicates that the introduction of QOF targets influenced behaviour of professionals and it was the operative feature of the targets that triggered the Value for independence linked to the welfare of patients and the care they received (self-transcendence value). Consequently, it was the tension experienced by GP’s in routine practice between their accountability and role requirements under the QOF conditions, which indicated a decrease and loss of professional autonomy. It is important to acknowledge that professional autonomy is recognised by the Royal College of Physicians as a core professional value ----- (Table. 2) [22]. Our analysis proposes that GPs further experienced self-restriction, hierarchical struggle, and outsider control due to the tension imposed by the QOFs influence on the development of indicators. In particular, the Values that were aligned to Professionalism, such as self-direction and stimulation seemingly were experienced as opposing the security, conformity and tradition values, supported by the QOF (Fig. 3). As a result, the restrictiveness of the self-direction Value may have led to the triggering of these conflicts. There were other aspects of the indicators where medical professionals themselves had limited influence (e.g. patient cooperation and access), which further challenged their confidence in achieving the QOF targets [26, 42, 46] causing concern. Incentivised conformity in driving the required actions Typically, people adapt their values to their circumstances [59] and they successively upgrade the importance they attribute to values they can readily attain and downgrade the importance of values whose pursuit is blocked [59]. When the QOF was first announced, primary care had been underfunded, there were large variations in quality between doctors, and general demoralisation within the primary care workforce [60]. Studies in our review suggest that QOF related behaviours raised the profile of general practice (achievement, power, status). This (already) set context may have also contributed to high QOF opt in rates (voluntary) for this P4P scheme in general practice. However, upgrading attainable values and downgrading thwarted values applies to most, but not to all values [55]. We found that Values that concern material well-being achievement, power and security were particularly aligned to the QOF. We also found evidence that when such Values were obstructed, their importance increased and when they were easily attained their importance dropped [61]. “Well it’s certainly improved my income. Probably increased my workload, not to the same degree as it increased my income. But I’m a bit worried that we’ve sold our soul to the devil to some degree, because they can change the goal posts later.” (p. 230) [40] The presence of the QOF was a requisite and binding, so despite having the choice to opt in, ‘no way out’ of QOF was experienced by those that were in specific QOF leadership roles [24]. Those GPs, for whom social order, justice, and helpfulness in the specific context of the QOF work were important values, would ideally be the target individuals and therefore most likely be motivated to pursue the incentivised goals in the context of this P4P scheme. This however, was experienced by others as confusing in relation to their role of the professional as a patient advocate. For example following a form of prescriptive QOF work was experienced as, taking away time to listen to patient concerns [29] which were perceived as participating in a form of ‘poor’ or ‘low value’ patient care impacting the patient-doctor relationship. RCP suggest this aspect of Professionalism requires training, practice, and mentorship to highlight such antagonisms in patient care [22]. Marcotte et al., propose physicians can and should embrace professionalism as the motivation for redesigning care. Payment reform incentives that align with their professional values should follow and encourage these efforts; that is, payment reform should not be the impetus for redesigning care [62]. Significance of cherished values; continuity of care, holism and the caring role of clinicians in primary care Values guide the selection or evaluation of actions, policies, people, and events. Therefore, medical professionals work in self-regulated disciplines, where the profession sets out the parameters of what is good or bad, justified or illegitimate, worth doing or avoiding, based on possible consequences for their cherished values [22] that are related to their profession. However, the impact of Values in everyday decisions is rarely conscious, power values can conflict with universalism and benevolence and these were evident in the accounts of professionals’ which resulted in high arousal to maintain professional behaviours that were linked to their role as patient partners and that were aligned especially to Professionalism (Fig. 5). “It distracts from the consultations and it can leave you know feeling a bit confused and perhaps as though that, the thing the patients regard as the problem hasn’t been addressed properly.” (p. 8) [26] The conflicts in Values or changes that were occurring would not have been at the forefront of every professional’s awareness, not until they had started to operate under the QOF conditions or for example when they experienced or became aware of a discontinuity of care for the patients in their daily practice. “In a sense that it’s still a patient presenting to a doctor with a problem, yes it is the same as it always was. The difference is that it’s more likely that the patient and the doctor won’t know each other.” (p.230) [40] This highlights the importance of intrinsic motivations [6, 23] for professionals in their day to day work, which if thwarted leads to deepening any individually held disappointment with their profession ----- (satisfaction, stimulation). Recent, GP career intention data has shown that morale had reduced over the past 2 years and intention to leave/retire in the next 2 years increased from 13% in the 2014 survey to 18% 2017 [51]. As a result the theme of personal congruence carried the message that the internal values of a doctor should match the external behaviour and actions [63]. We found that the QOF work was more amenable to the values under conservation and self-enhancement dimensions, and hence directly opposed to the values under self-transcendence and openness to change dimensions (see Fig. 4). As a result, practitioners who were self-directed and worked for the welfare of patients were constrained in their ability to use knowledge attained from previous interactions (patient agendas) with patients in guiding future consultations. This may have led professionals to view standardised care as a ‘box-ticking’ exercise, and at odds with their professional training and their caring role [30]. Holism and patient-centred care were significant values that were particularly vulnerable to QOF changes. “I thought that you were supposed to tailor this care to every individual patient and meet patient needs...I think it takes away patient, you know, centred care really...I don’t think people appreciate being phoned up all the time and reminding them to come in and things...rightly or wrongly [lead partner] strives for perfection and I think sometimes you have to acknowledge you don’t get perfection all the time and whenever you’re dealing with patients and people you’ll never get perfection anyway.” (p. 56) [45] Some of the papers, described the need of professionals to defend efforts to continue to deliver nonincentivised care as part of their professional role [25, 44, 45]. Initially, some GPs were apprehensive about the consequences of implementation of indicators in ‘everyday clinical practice’ [26, 29, 30]. Furthermore, there seemed to be insufficient governmental, organisational, administrative, executive, and managerial recognition of the link between the ‘doctors on the ground floor’ working in ‘everyday clinical practice’, and the consequences for ‘routine clinical practice’ and for the professional-patient relationships [24, 25, 45]. Acquiring a say in the development of indicators through negotiations between the BMA and the NHS was an important aspect for professionals, linked to freedom to practice ‘in patients’ best interest’ [6, 24, 42, 45]. “I’d like to see performance measures that really reflect the care.” (p. 553) [46] “...Some things are ...within the control of the providers, but some things really aren’t, even done ... with good intent.” (p. 553) [46] “...Often what happens with physicians is things are mandated to us and we don’t have any input in...the process of how things some to us.” (p. 553) [46] Valderas et al., recommends that person-centred care should be a guiding principle for the development of assessment frameworks and quality indicators. As peoplecentredness is a core value of health systems, which acknowledges that individual service users should be the key stakeholders and their values, goals and priorities should shape care delivery [64]. Structural & organisational changes: the trade-off between multiple values The synthesis showed that all practices had changed their styles of operation in response to the QOF [24, 25, 44, 65]. This involved an increase in the number of administrative staff, including those with responsibility for information technology (IT) [25, 65] and the new managerial stratum worked to align clinical activities to the wider organisational goals [24]. The findings from the synthesis also propose that the QOF targets that were aligned to the conservation and self-enhancement values of GPs, had led to extra income and sizable pay differentials at the practice level were the enabling factor that allowed for the vast organisational and structural changes that took place. These changes were described as a success (achievement) for practices and patients. “…it’s benefitting the patients, that they don’t get missed, they don’t slip through the net, they get their medicines reviewed, they get their blood tests, they’re kept on optimum treatment.” (p. 135) [44] Subsequently, the threat to status through competition (stimulation) was seen as a motivator [26]. “It does feel a bit like competition with other surgeries, I don’t know how others feel but I wouldn’t like to come last in our locality.” (p. 7) [26] Yet, those professionals who were motivated to remain self-directed and aligned their behaviours and attitudes to the welfare of patients, experienced restriction in their ability to use the knowledge attained from patient interactions to guide their future consultations. “So it’s made the two agendas a little bit clearer and I guess you’ve always had a health agenda and mine ----- is probably never been the same, but now that mine is encapsulated by QOF…it’s a bit more blatantly not the same. So I think there is an intrusion there and it’s not an entirely patient-led agenda, because you’ve got things that you want to do that you think are more important.” (p. 231) [40] Professionals were making the trade-offs among relevant opposing values based on the QOF targets and that these were guiding the attitudes and behaviours of health care providers in their practice. When Values are in conflict, practitioners will often attribute more importance to the achievement of one set of values at the expense of the others. This hierarchical relationship between values also distinguishes values from norms and attitudes that can be followed unfeelingly. Any attitude or behaviour typically has implications for more than one value, for example, a ‘tick box’ approach to medicine encouraged by P4P indicators promoted EBM and conformity values leading to success, achievement and status at the expense of self-direction, hedonism and stimulation values. Discussion This study involved a meta-synthesis of qualitative studies of provider views of the QOF program. We analysed the literature through the lens of Schwartz’s, Theory of Values as a theoretical framework and to contextualise our analysis we also used Professionalism as a conceptual lens. Using this theoretical framework, we found that QOF related work was experienced by providers as incongruent with their self-direction and benevolence values that are pivotal to professionalism as defined by the Royal College of Physicians [22]. This understanding is likely the result of the QOF being experienced as a mechanism of value activation for only certain values (see Table 2). Values affect behaviour only if they are activated [61]. Activation may or may not entail conscious thought about a value and much information processing occurs outside of awareness [61]. The more accessible a value (the more easily it comes to mind) and the more likely it will be activated and because more important values are more accessible, they relate more to behaviour [18, 66]. For policy and decision makers such insights are valuable in terms of designing P4P schemes. In a report on designing incentive payments for quality care the Conference Board of Canada identified three key guiding concepts – getting the right blend of incentives, alignment with health care goals, global experience and human motivation [67]. Also recognising the importance of values the NHS England review of QOF argued that the scheme needed to be to repositioned “… as a scheme which recognises and supports the professional values of GPs and their teams in the delivery of first contact, comprehensive, coordinated and person-centred care” [65]. Our analysis suggests that the pursuit of achievement Values in QOF related work was experienced as compatible with the pursuit of wealth, authority, success, and ambition values that were linked to seeking personal success for GPs. This was likely to reinforce and be supported by QOF actions that were aimed at enhancing GPs social position and status. This also included expanding practice activity, size and overall income, which may be considered as organisational success factors by some GPs. Values such as creativity, social justice, equality, benevolence were experienced as restricted as a result of the QOF targets. Accordingly, when Values are activated, they become infused with feeling both positive or negative [18]. Our synthesis has shown that definition of ‘high quality care’ must be accepted by general practitioners for it to be integrated into practice behaviour. If it is merely derived from an ‘outside regulation’ of clinical practice and assembled by an ‘outside agency’ it will not achieve enduring behaviour change [24, 25, 45, 65]. The direct involvement of providers in the definition of ‘high quality care’ could be one mechanism to balance the discord that was experienced with QOF work. Correspondingly, quality improvement initiatives that are constructed and implemented for the patients’ benefit should be compatible with both EBM and encompass a `patient-centred’ approach. Embedding concepts of high quality primary care, such as those highlighted by Mead and Bower which include a biopsychosocial perspective, `patient-as-person’, sharing power and responsibility, therapeutic alliance, and `doctor-as-person’ in quality improvement initiatives may alleviate some of the tensions that have created unease in general practice as a result of the QOF [68]. A recent systematic review has shown that four of Mead and Bower’s dimensions are still relevant today, and ‘coordinated care’ was a new dimension, reflecting increasing complexity of the health care system [69]. This will likely become more significant as integrated care is planned as a more efficient client-oriented health model [70]. The papers in our analysis described the caring role that encompassed softer values such as the pursuit of novelty, change and stimulation values was likely to be seen to undermine the safeguarding of older customs/ tradition values of medicine such as the more biomedical care model. Our analysis also demonstrates that the pursuit of traditional values (clinician-centred care, EBM, and templates) is essentially congruent with the pursuit of conformity values as both motivate actions of submission to external expectations (QOF targets). The Values Theory suggests that everyone experiences conflict between pursuing openness to change values or ----- conservation values and between pursuing selftranscendence or self-enhancement values. Conflicts between specific Values (e.g., power vs. universalism, tradition vs. hedonism) are also near-universal [18]. Values serve as standard or criteria and they tend to guide the selection or evaluation of actions, policies, and events. For example, individuals also decide what is good or bad, justified or illegitimate, worth doing or avoiding based on possible consequences for their cherished values [18]. Achieving some kind of balance in this now appears to have been crucial and the evidence suggests that embracing a more complimentary working between the two, with more focus on the combined efforts is more likely to drive successful complex initiatives. Historically, practices have been autonomous in managerial terms and GPs have been traditionally independently minded [71]. They possess a wide range of norms and values, many of which are desirable but some of which may not be suited to the changes required in complex health systems. For this reason, there are obvious tensions within this relationship with regards to the changes that are ‘softly required’ by health system managers [72]. The synthesis suggests allowing practitioners with competing high priority values to be part of quality improvement initiatives, to take on an influencer role within those initiatives, instead of being ‘responsive agents’ [73]. Initiatives need to consider and engage with concerns of professionals as changes occur in health systems, with timely consultation, piloting and prior to implementation [42]. Strengths and limitations Meta-ethnography does offer considerable potential for preserving the interpretive properties of primary data [74]. We acknowledge that the qualitative synthesis cannot be reduced to a set of mechanistic tasks, which raises issues of the transparency of the process [75] which we have tried to make transparent. The goal is to increase understanding, leading to greater explanatory effect, rather than to aggregate and merge findings in a kind of averaging process [76]. We did not have the added benefit of access to any raw data (including transcriptions, reflective notes, and author insight about the context of the studies) as some other meta synthesis have done [76]. Yet, Estabrooks and Field (1994) suggest that the recurrence of themes between compared studies adds to validity similarly to triangulation that is another technique, said to ensure soundness in analysis [77]. Pielstick (1998) understands this as using multiple studies and (meta-synthesis does this by definition) [78]. Undertaking a meta-synthesis is a demanding and laborious process, and would benefit from the development of suitable software [73]. However we feel this will help manage the large amounts of data that emerge from the papers but will not add anything to the process of analysis itself. Conclusion The QOF was instrumental in bringing fundamental changes to general practice organisations. Furthermore, these changes have endured and been embedded into general practice institutions, despite ongoing proposed changes to the QOF. As a mechanism for activating and triggering a select set of Values, the QOF is compatible with the pursuit of wealth, authority, success, ambition and achievement that have implications for Professionalism. In its implementation, QOF also created a ‘standardised success model’ for GPs to motivate and implement ‘actions of submission’ to achieve QOF targets. While QOF was aligned with traditional medical values, influenced by clinician centred care, EBM, and clinical guidelines; our analysis suggests that despite conforming to core medical values there were still some dilemmas regarding whether to pursue income and organisational goals above patient-centred practice. This analysis of the impact of QOF suggests that in order for quality improvement initiatives, such as P4P schemes to be endurable; they need to be compatible with provider values. P4P schemes need to be designed in order to integrate the personal and professional values that professionals’ find are essential to their practice. Professionals’ have shown that they are driven by their views, beliefs, and experiences, and not just by hierarchy and externally imposed constructs. Our review indicates that policy makers and health service planners need to carefully construct schemes in order to prioritise intrinsic professional values rather than rely on extrinsic motivators that show more limited alignment with Professionalism and its professional core values. Research on QOF has identified that use of performance targets has a limited impact on the quality of care and caused some internal conflicts during the process of carrying out the QOF work. In the UK the shift towards quality improvement approaches that are framed by national priorities and allow for professionals to design their improvement approach, may provide a way of harnessing values of professional autonomy and control as well as building on the motivation to develop patientcentred care. Moves to more network (groups of practices) based schemes may require further thinking as they will be a more complex context with potentially differing, and possibly competing, motivations between practices and practitioners. Our review of the QOF recommends valuable insights that provide those designing P4P systems. It also identifies the need for more qualitative research on the implementation of P4P schemes to fully understand their individual and organisational impact. Further research is also needed to more fully ----- understand how schemes can influence practitioners and support high quality care. In particular it is clear that context; in terms the of the wider organisational structure, payment systems and health system design, need to be more fully considered to fully understand the link between financial incentives, behaviour – both individual and organisational, and quality of care. Supplementary information [Supplementary information accompanies this paper at https://doi.org/10.](https://doi.org/10.1186/s12875-020-01208-8) [1186/s12875-020-01208-8.](https://doi.org/10.1186/s12875-020-01208-8) Additional file 1. Search Strategy Additional file 2. Table 4. Contextual Information for the 18 Published Papers Additional file 3. Table. 5 First and Second Order Constructs from the Published Papers Additional file 4. Table. 6 Identifying Third Order Constructs Additional file 5. Table. 7 Application of the Third Order Constructs to the Ten Motivational Values Abbreviations QOF: Quality and Outcomes Framework; P4P: Pay for Performance; LOA: Lines-of-argument synthesis; BMA: British Medical Association; RCP: Royal College of Physicians; NHS: National Health Service; QI: Quality Improvement; UK: United Kingdom; EBM: Evidence Based Medicine; GP: General Practitioner; ICT: Information and Communication Technology; NICE: Institute for Health and Care Excellence. Acknowledgements We would like to thank Professor Martin Roland, Emeritus Professor of Health Services Research- Fellow, Murray Edwards College, for some detailed comments on the relevant papers at the inclusion stage of our research. Authors’ contributions N.K designed and managed the review and wrote the first draft of the manuscript. N.K led the analysis of the data and led the writing. D.R assisted with the second level constructs in the analyses and writing of the early drafts. SP contributed to the analysis of the third order constructs and the writing of later drafts. The literature searches were carried out with the assistance of a specialist librarian M.M, with input from N.K and D.R. Authors N.K and S.P participated equally in the editing of this manuscript. Funding No funding was received for this study. Availability of data and materials Data was generated from the published papers that were included in this synthesis. The dataset used and/or analysed during the current study are available from the corresponding author on reasonable request. N.K, D.R and S.P had access to all relevant papers included, customised tables, and data necessary for verifying the integrity of the data and the accuracy of the analysis. All authors read and approved the final manuscript. Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests None declared. All authors received no personal or financial gain from carrying out this work. Author details 1Independent Researcher, Ontario, Canada. 2Faculty of Health Sciences, University of Ontario Institute of Technology, 2000 Simcoe Street North, Unit UA3000, Oshawa, ON L1H 7K4, Canada. [3]Ontario Shores Centre for Mental Health Sciences, 700 Gordon Street, Whitby, ON L1N 5S9, Canada. [4]Centre for Health Services Studies, University of Kent, Kent, Canterbury CT2 7NF, UK. Received: 24 July 2019 Accepted: 23 June 2020 References 1. Roland M, Guthrie B. Quality and Outcomes Framework: what have we [learnt?: BMJ. 2016 Aug 4;354:i4060. Available from: http://www.bmj.com/](http://www.bmj.com/lookup/doi/10.1136/bmj.i4060) [lookup/doi/10.1136/bmj.i4060 [cited 2018 22 Nov].](http://www.bmj.com/lookup/doi/10.1136/bmj.i4060) 2. Waqar S, Snow-Miller R, Weaver R. Building the Workforce - the New Deal for General Practice GP Induction and Refresher Scheme, 2015-2018. NHS [England; 2015. https://heeoe.hee.nhs.uk/sites/default/files/gp_induction_](https://heeoe.hee.nhs.uk/sites/default/files/gp_induction_and_refresher_scheme_2015-18.pdf) [and_refresher_scheme_2015-18.pdf.](https://heeoe.hee.nhs.uk/sites/default/files/gp_induction_and_refresher_scheme_2015-18.pdf) 3. Bostock N. Exclusive NHS England rejected QOF Suspension bacause of [opposition from GPs. GPonline; 2017. https://www.gponline.com/](https://www.gponline.com/exclusivenhs-england-rejected-qof-suspension-because-opposition-gps/article/1423168) [exclusivenhs-england-rejected-qof-suspension-because-opposition-gps/](https://www.gponline.com/exclusivenhs-england-rejected-qof-suspension-because-opposition-gps/article/1423168) [article/1423168.](https://www.gponline.com/exclusivenhs-england-rejected-qof-suspension-because-opposition-gps/article/1423168) 4. England N. Delivering the Forward View: NHS planning guidance [Internet]. [2016 [cited 2018 Nov 22]. Available from: https://www.england.nhs.uk/wp-](https://www.england.nhs.uk/wp-content/uploads/2015/12/planning-guid-16-17-20-21.pdf) [content/uploads/2015/12/planning-guid-16-17-20-21.pdf.](https://www.england.nhs.uk/wp-content/uploads/2015/12/planning-guid-16-17-20-21.pdf) 5. Forbes LJ, Marchand C, Doran T, Peckham S. The role of the Quality and Outcomes Framework in the care of long-term conditions: a systematic review. Br J Gen Pract [Internet]. 2017 Nov 25;67(664):e775–e784. Available [from: http://www.ncbi.nlm.nih.gov/pubmed/28947621 [cited 2018 5 Sep].](http://www.ncbi.nlm.nih.gov/pubmed/28947621) 6. NHS England. Report of the Review of the Quality and Outcomes [Framework in England [Internet]. 2018. Available from: https://www.](https://www.england.nhs.uk/wp-content/uploads/2018/07/quality-outcome-framework-report-of-the-review.pdf) [england.nhs.uk/wp-content/uploads/2018/07/quality-outcome-framework-](https://www.england.nhs.uk/wp-content/uploads/2018/07/quality-outcome-framework-report-of-the-review.pdf) [report-of-the-review.pdf [cited 2019 22 May].](https://www.england.nhs.uk/wp-content/uploads/2018/07/quality-outcome-framework-report-of-the-review.pdf) 7. Glenton, C., Lewin, S., Norris, S., & Norris S. 15 . Using evidence from qualitative research to develop WHO guidelines [Internet]. WHO handbook [for guideline development. 2014 [cited 2018 Sep 16]. Available from: http://](http://www.who.int/publications/guidelines/Chp15_May2016.pdf) [www.who.int/publications/guidelines/Chp15_May2016.pdf.](http://www.who.int/publications/guidelines/Chp15_May2016.pdf) 8. Roland M. Does pay-for-performance in primary care save lives? Lancet. [2016 Jul 16;388(10041):217–218. Available from: https://linkinghub.elsevier.](https://linkinghub.elsevier.com/retrieve/pii/S014067361600550X) [com/retrieve/pii/S014067361600550X [cited 2019 19 Jun].](https://linkinghub.elsevier.com/retrieve/pii/S014067361600550X) 9. McNeil R, Guirguis-Younger M, Dilley LB. Recommendations for improving the end-of-life care system for homeless populations: A qualitative study of the views of Canadian health and social services professionals. BMC Palliat Care. 2012. 10. Harrison MJ, Dusheiko M, Sutton M, Gravelle H, Doran T, Roland M. Effect of a national primary care pay for performance scheme on emergency hospital admissions for ambulatory care sensitive conditions: controlled [longitudinal study. BMJ. 2014 Nov 11;349:g6423. Available from: http://www.](http://www.ncbi.nlm.nih.gov/pubmed/25389120) [ncbi.nlm.nih.gov/pubmed/25389120 [cited 2019 19 Jun].](http://www.ncbi.nlm.nih.gov/pubmed/25389120) 11. McDonald RRM. Pay for performance in primary care in England and California: comparison of unintended consequences. Ann Fam Med. 2009; 7(2):121–7. 12. Marshall M, Harrison S. It’s about more than money: financial incentives and internal motivation. Qual Saf Heal Care. 2005;14(1):4–5. 13. Langlois E V., Tunçalp Ö, Norris SL, Askew I, Ghaffar A. Qualitative evidence to improve guidelines and health decision-making. Bull World Health [Organ. 2018;96(2):79-79A. Available from: http://www.who.int/entity/](http://www.who.int/entity/bulletin/volumes/96/2/17-206540.pdf) [bulletin/volumes/96/2/17-206540.pdf [cited 2018 16 Sep].](http://www.who.int/entity/bulletin/volumes/96/2/17-206540.pdf) 14. Barnett-Page E, Thomas J. Methods for the synthesis of qualitative research: [a critical review. BMC Med Res Methodol. 2009;9(1):59 Available from: http://](http://bmcmedresmethodol.biomedcentral.com/articles/10.1186/1471-2288-9-59) [bmcmedresmethodol.biomedcentral.com/articles/10.1186/1471-2288-9-59](http://bmcmedresmethodol.biomedcentral.com/articles/10.1186/1471-2288-9-59) [cited 2018 25 Jun]. 15. Noblit GW, Hare RD. Meta-Ethnography: Synthesizing Qualitative Studies (Qualitative Research Methods) [Internet]. Vol. 44, Counterpoints. Peter Lang [AG; 1988 [cited 2018 Oct 7]. 88 p. Available from: https://www.jstor.org/](https://www.jstor.org/stable/42975557) [stable/42975557.](https://www.jstor.org/stable/42975557) 16. Schwartz SH. Are there universal aspects in the structure and contents of human values? J Soc Issues. 1994;50(4):19–45. 17. Schwartz SH. An Overview of the Schwartz Theory of Basic Values. Psychol [Cult. 2012;2(1):12–13. Available from: http://scholarworks.gvsu.edu/orpc/](http://scholarworks.gvsu.edu/orpc/vol2/iss1/11) [vol2/iss1/11 [cited 2018 23 Nov].](http://scholarworks.gvsu.edu/orpc/vol2/iss1/11) ----- 18. Schwartz SH. Universals in the content and structure of values: Theoretical advances and empirical tests in 20 countries. In: Zanna M, editor. Advances in experimental social psychology. New York: Academic Press; 1992. p. 1–65. 19. Estabrooks CA, Field PA, Morse JM. Aggregating Qualitative Findings: An Approach to Theory Development. Qual Health Res. 1994;4(4):503–511. [Available from: http://journals.sagepub.com/doi/10.1177/1049732394004](http://journals.sagepub.com/doi/10.1177/104973239400400410) [00410 [cited 2019 21 Jan].](http://journals.sagepub.com/doi/10.1177/104973239400400410) 20. Schwartz SH. Universals in the content and structure of values: Theoretical advances and empirical tests in 20 countries. Adv Exp Soc Psychol. 1992; [25(C):1–65. Available from: https://plato.stanford.edu/entries/value-theory/](https://plato.stanford.edu/entries/value-theory/supplement2.html) [supplement2.html [cited 2020 [cited 2020 Mar 11].](https://plato.stanford.edu/entries/value-theory/supplement2.html) 21. Plochg T, Klazinga NS, Starfield B. Transforming medical professionalism to fit changing health needs. BMC Med. 2009;7(1):1–7. 22. Tweedie J, Hordern J, Dacre J. Advancing medical professionalism. Pepito J. [2018. 1–123 p. Available from: http://www.rcplondon.ac.uk [cited 2020 10 Mar].](http://www.rcplondon.ac.uk) 23. Cheraghi-Sohi S, McDonald R, Harrison S, Sanders C. Experience of contractual change in UK general practice: A qualitative study of salaried [GPs. Br J Gen Pract. 2012;62(597):e282–e287. Available from: http://bjgp.org/](http://bjgp.org/cgi/doi/10.3399/bjgp12X636128) [cgi/doi/10.3399/bjgp12X636128 [cited 2018 19 Apr].](http://bjgp.org/cgi/doi/10.3399/bjgp12X636128) 24. Cheraghi-Sohi S, Calnan M. Discretion or discretions? Delineating professional discretion: Thecase of English medical practice. Soc Sci Med. [2013;96:52–59. Available from: http://www.ncbi.nlm.nih.gov/pubmed/24034](http://www.ncbi.nlm.nih.gov/pubmed/24034951%20) [951 [cited 2018 19 Apr].](http://www.ncbi.nlm.nih.gov/pubmed/24034951%20) 25. Checkland K, Harrison S. The impact of the Quality and Outcomes Framework on practice organisation and service delivery: summary of evidence from two qualitative studies. Qual Prim Care. 2010;18(2):139–46 [Available from: http://www.ncbi.nlm.nih.gov/pubmed/20529476 [cited 2018](http://www.ncbi.nlm.nih.gov/pubmed/20529476) 19 Apr]. 26. Hackett J, Glidewell L, West R, Carder P, Doran T, Foy R. ‘Just another incentive scheme’: a qualitative interview study of a local pay-for-performance scheme [for primary care. BMC Fam Pract. 2014;15(1):168 Available from: http://www.](http://www.ncbi.nlm.nih.gov/pubmed/25344735) [ncbi.nlm.nih.gov/pubmed/25344735 [cited 2018 24 Jul].](http://www.ncbi.nlm.nih.gov/pubmed/25344735) 27. Gill PJ, Hislop J, Mant D, Harnden A. General practitioners’ views on quality markers for children in UK primary care: a qualitative study. BMC Fam Pract. [2012;13(1):92. Available from: http://bmcfampract.biomedcentral.com/](http://bmcfampract.biomedcentral.com/articles/10.1186/1471-2296-13-92) [articles/10.1186/1471-2296-13-92 [cited 2018 24 Jul].](http://bmcfampract.biomedcentral.com/articles/10.1186/1471-2296-13-92) 28. Maxwell M, Harris F, Hibberd C, Donaghy E, Pratt R, Williams C, et al. A qualitative study of primary care professionals’ views of case finding for depression in patients with diabetes or coronary heart disease in the UK. [BMC Fam Pract 2013;14(1):46. Available from: http://www.ncbi.nlm.nih.gov/](http://www.ncbi.nlm.nih.gov/pubmed/23557512) [pubmed/23557512 [cited 2018 24 Jul].](http://www.ncbi.nlm.nih.gov/pubmed/23557512) 29. Chew-Graham CA, Hunter C, Langer S, Stenhoff A, Drinkwater J, Guthrie EA, et al. How QOF is shaping primary care review consultations: a longitudinal [qualitative study. BMC Fam Pract. 2013;14(1):103. Available from: http://](http://www.ncbi.nlm.nih.gov/pubmed/23870537) [www.ncbi.nlm.nih.gov/pubmed/23870537 [cited 2016 27 Apr].](http://www.ncbi.nlm.nih.gov/pubmed/23870537) 30. Lester HE, Hannon KL, Campbell SM. Identifying unintended consequences of quality indicators: A qualitative study. BMJ Qual Saf. 2011;20(12):1057–61 [Available from: http://www.ncbi.nlm.nih.gov/pubmed/21693464 [cited 2018](http://www.ncbi.nlm.nih.gov/pubmed/21693464) 19 Apr]. 31. Fong W, Kwan YH, Yoon S, Phang JK, Thumboo J, Leung YY, et al. Assessment of medical professionalism: preliminary results of a qualitative study. BMC Med Educ. 2020 Jan 30;20(1):27. 32. Di Blasi Z, Harkness E, Ernst E, Georgiou A, Kleijnen J. Influence of context effects on health outcomes: A systematic review. Lancet. 2001;357(9258): 757–62. 33. Maxwell JA. Qualitative Research Design: An Interactive Approach: An Interactive Approach [Internet]. 2012 [cited 2020 May 11]. p. 218. Available [from: http://books.google.com/books?hl=en&lr=&id=DFZc28cayiUC&pgis=1.](http://books.google.com/books?hl=en&lr=&id=DFZc28cayiUC&pgis=1) 34. British Sociological Association. Criteria for the evaluation of qualitative research papers. Med Sociol News. 1996;22:34–7. 35. Coventry PA, Small N, Panagioti M, Adeyemi I, Bee P. Living with complexity; Marshalling resources: A systematic review and qualitative meta-synthesis of lived experience of mental and physical multimorbidity Multimorbidity in Primary Care Knowledge, attitudes, behaviors, education, [and communication. BMC Fam Pract. 2015;16(1):171 Available from: http://](http://www.biomedcentral.com/1471-2296/16/171) [www.biomedcentral.com/1471-2296/16/171 [cited 2020 25 Mar].](http://www.biomedcentral.com/1471-2296/16/171) 36. Campbell R, Pound P, Morgan M, Daker-White G, Britten N, Pill R, et al. Evaluating meta-ethnography: systematic analysis and synthesis of qualitative research. Health Technol Assess (Rockv). 2011;15(43):1–164. 37. Mitchell C, Dwyer R, Hagan T, Mathers N. Impact of the QOF and the NICE guideline in the diagnosis and management of depression: a qualitative [study. Br J Gen Pract. 2011;61(586):e279–e289. Available from: http://bjgp.](http://bjgp.org/cgi/doi/10.3399/bjgp11X572472) [org/cgi/doi/10.3399/bjgp11X572472 [cited 2018 19 Apr].](http://bjgp.org/cgi/doi/10.3399/bjgp11X572472) 38. Hannon KL, Lester HE, Campbell SM. Recording patient preferences for endof-life care as an incentivized quality indicator: What do general practice [staff think? Palliat Med. 2012;26(4):336–341. Available from: http://www.ncbi.](http://www.ncbi.nlm.nih.gov/pubmed/21680749) [nlm.nih.gov/pubmed/21680749 [cited 2018 24 Jul].](http://www.ncbi.nlm.nih.gov/pubmed/21680749) 39. Alderson SL, Russell AM, McLintock K, Potrata B, House A, Foy R. Incentivised case finding for depression in patients with chronic heart disease and diabetes in primary care: an ethnographic study. BMJ Open. 2014;4(8): [e005146–e005146. Available from: http://www.ncbi.nlm.nih.gov/](http://www.ncbi.nlm.nih.gov/pubmed/25138803) [pubmed/25138803 [cited 2018 24 Jul].](http://www.ncbi.nlm.nih.gov/pubmed/25138803) 40. Campbell SM, McDonald R, Lester H. The experience of pay for performance in english family practice: A qualitative study. Ann Fam Med. 2008;6(3):228– [34 Available from: http://www.ncbi.nlm.nih.gov/pubmed/18474885 [cited](http://www.ncbi.nlm.nih.gov/pubmed/18474885) 2018 22 Jul]. 41. Campbell S, Hannon K, Lester H. Exception reporting in the quality and outcomes framework: Views of practice staff - A qualitative study. Br J Gen [Pract. 2011;61(585):e183–e189. Available from: http://www.ncbi.nlm.nih.gov/](http://www.ncbi.nlm.nih.gov/pubmed/21439176) [pubmed/21439176 [cited 2018 24 Jul].](http://www.ncbi.nlm.nih.gov/pubmed/21439176) 42. Lester H, Matharu T, Mohammed MA, Lester D, Foskett-Tharby R. Implementation of pay for performance in primary care: A qualitative study 8 years after introduction. Br J Gen Pract. 2013;63(611):e408–15 Available from: [http://bjgp.org/lookup/doi/10.3399/bjgp13X668203 [cited 2018 19 Apr].](http://bjgp.org/lookup/doi/10.3399/bjgp13X668203) 43. McGregor W, Jabareen H, O’Donnell CA, Mercer SW, Watt GC. Impact of the 2004 GMS contract on practice nurses: a qualitative study. Br J Gen Pract. [2008;58(555):711–719. Available from: http://bjgp.org/cgi/doi/10.3399/](http://bjgp.org/cgi/doi/10.3399/bjgp08X342183) [bjgp08X342183 [cited 2018 19 Apr].](http://bjgp.org/cgi/doi/10.3399/bjgp08X342183) 44. Maisey S, Steel N, Marsh R, Gillam S, Fleetcroft R, Howe A. Effects of payment for performance in primary care: qualitative interview study. J Heal Serv Res Policy. 2008;13(3):133–9. 45. McDonald R, Harrison S, Checkland K. Incentives and control in primary health care: findings from English pay-for-performance case studies. J [Health Organ Manag. 2008;22(1):48–62 Available from: http://www.ncbi.nlm.](http://www.ncbi.nlm.nih.gov/pubmed/18488519) [nih.gov/pubmed/18488519 [cited 2018 19 Apr].](http://www.ncbi.nlm.nih.gov/pubmed/18488519) 46. Arbaje AI, Newcomer AR, Maynor KA, Duhaney RL, Eubank KJ, Carrese JA. Excellence in Transitional Care of Older Adults and Pay-for-Performance: Perspectives of Health Care Professionals. Jt Comm J Qual patient Saf. 2014; [40(12):550–551. Available from: http://www.ncbi.nlm.nih.gov/](http://www.ncbi.nlm.nih.gov/pubmed/26111380%20) [pubmed/26111380 [cited 2018 19 Apr].](http://www.ncbi.nlm.nih.gov/pubmed/26111380%20) 47. Hannon KL, Lester HE, Campbell SM. Patients’ views of pay for performance in primary care: A qualitative study. Br J Gen Pract. 2012;62(598):e322–8 Available [from: http://www.ncbi.nlm.nih.gov/pubmed/22546591 [cited 2018 1 may].](http://www.ncbi.nlm.nih.gov/pubmed/22546591) 48. Corti L, Backhouse G. Acquiring qualitative data for secondary analysis. Forum Qual Sozialforsch. 2005;. 49. GILLAM S, Steel N. QOF points: valuable to whom? [Internet]. Vol. 346, BMJ (Overseas and retired doctors ed.). 2013 [cited 2020 Jun 5]. Available from: [www.nice.org.uk/media/E19/EA/.](http://www.nice.org.uk/media/E19/EA/) 50. Heath I. Divided we fail: the Harveian oration 2011. Nurs Stand. 2011;12(20):16. 51. Owen K, Hopkins T, Shortland T, Dale J. GP retention in the UK: A worsening crisis. Findings from a cross-sectional survey. BMJ Open. 2019;1, 9(2). 52. Gillam S, Steel N. QOF points: valuable to whom? Source BMJ Br med J. 2013;. 53. Fleetcroft R, Steel N, Cookson R, Walker S, Howe A. Incentive payments are not related to expected health gain in the pay for performance scheme for UK primary care: cross-sectional analysis. BMC Health Serv Res. 2012;12(1):94. 54. Woolhandler S, Ariely D, Himmelstein DU. Why pay for performance may be incompatible with quality improvement. BMJ. 2012;345(7870). 55. Schwartz SH. Unit 2 Theoretical and Methodological Issues Subunit 1 Conceptual Issues in. Psychol Cult Artic [Internet]. 2012 [cited 2018 Nov 22]; [11:12–3. Available from: https://doi.org/10.9707/2307-0919.1116.](https://doi.org/10.9707/2307-0919.1116) 56. Lester H, Schmittdiel J, Selby J, Fireman B, Campbell S, Lee J, et al. The impact of removing financial incentives from clinical quality indicators: longitudinal analysis of four Kaiser Permanente indicators. BMJ. 2010;340: [c1898 Available from: http://www.ncbi.nlm.nih.gov/pubmed/20460330](http://www.ncbi.nlm.nih.gov/pubmed/20460330) [cited 2019 21 Jan]. 57. Durrheim D. Income inequality and health status: a nursing issue [Internet]. Vol. 25, AUSTRALIAN JOURNAL OF ADVANCED NURSING. 2007 [cited 2020 [May 12]. Available from: https://s3.amazonaws.com/academia.edu.](https://s3.amazonaws.com/academia.edu.documents/56474455/PROPUES_TPM.pdf?response-content-disposition=inline%3B) [documents/56474455/PROPUES_TPM.pdf?response-content-disposition=](https://s3.amazonaws.com/academia.edu.documents/56474455/PROPUES_TPM.pdf?response-content-disposition=inline%3B) [inline%3B filename%3DDiseno_de_un_plan_de_Mantenimiento_Prod.](https://s3.amazonaws.com/academia.edu.documents/56474455/PROPUES_TPM.pdf?response-content-disposition=inline%3B) pdf&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential= AKIAIWOWYYGZ2Y53UL3A%2F20200314%2Fus-e. ----- 58. Corti Bishop L. L. Strategies in Teaching Secondary Analysis of Qualitative Data. FQS (Forum Qual Soc Res. 2005;. 59. Schwartz SH, Bardi A. Influences of Adaptation to Communist Rule on Value Priorities in Eastern Europe. Polit Psychol. 1997;18(2):385–410. Available from: [http://doi.wiley.com/10.1111/0162-895X.00062 [cited 2018 28 Nov].](http://doi.wiley.com/10.1111/0162-895X.00062) 60. Lester H, Campbell S. Developing Quality and Outcomes Framework (QOF) indicators and the concept of “QOFability”. Qual Prim Care. 2010;18(2):103–9 [Available from: http://www.ncbi.nlm.nih.gov/pubmed/20529471 [cited 2019](http://www.ncbi.nlm.nih.gov/pubmed/20529471) 15 Apr]. 61. Inglehart R. Modernization and postmodernization : cultural, economic, and political change in 43 societies [Internet]. Princeton University Press; 1997 [[cited 2018 Nov 28]. 453 p. Available from: https://press.princeton.edu/](https://press.princeton.edu/titles/5981.html) [titles/5981.html.](https://press.princeton.edu/titles/5981.html) 62. Marcotte LM, Moriates C, Wolfson DB, Frankel RM. Professionalism as the bedrock of high-value care. Acad Med. 2019;1. 63. Wagner P, Hendrich J, Moseley G, Hudson V. Defining medical professionalism: [a qualitative study. Med Educ. 2007;41(3):288–94 Available from: http://doi.](http://doi.wiley.com/10.1111/j.1365-2929.2006.02695.x) [wiley.com/10.1111/j.1365-2929.2006.02695.x [cited 2020 5 Jun].](http://doi.wiley.com/10.1111/j.1365-2929.2006.02695.x) 64. Valderas JM, Gangannagaripalli J, Nolte E, Boyd CM, Roland M, SarriaSantamera A, et al. Quality of care assessment for people with [multimorbidity. J Intern Med. 2019;285(3):289–300 Available from: https://](https://onlinelibrary.wiley.com/doi/abs/10.1111/joim.12881) [onlinelibrary.wiley.com/doi/abs/10.1111/joim.12881 [cited 2020 25 Mar].](https://onlinelibrary.wiley.com/doi/abs/10.1111/joim.12881) 65. Marmot M. Report of the Review of the Quality and Outcomes Framework in England [Internet]. 2010 [cited 2019 Feb 20]. Available [from: https://www.england.nhs.uk/wp-content/uploads/2018/07/quality-](https://www.england.nhs.uk/wp-content/uploads/2018/07/quality-outcome-framework-report-of-the-review.pdf%0Ahttps:/www.england.nhs.uk/publication/report-of-the-review-of-the-quality-and-outcomes-framework-in-england) [outcome-framework-report-of-the-review.pdf%0Ahttps://www.england.](https://www.england.nhs.uk/wp-content/uploads/2018/07/quality-outcome-framework-report-of-the-review.pdf%0Ahttps:/www.england.nhs.uk/publication/report-of-the-review-of-the-quality-and-outcomes-framework-in-england) [nhs.uk/publication/report-of-the-review-of-the-quality-and-outcomes-](https://www.england.nhs.uk/wp-content/uploads/2018/07/quality-outcome-framework-report-of-the-review.pdf%0Ahttps:/www.england.nhs.uk/publication/report-of-the-review-of-the-quality-and-outcomes-framework-in-england) [framework-in-england.](https://www.england.nhs.uk/wp-content/uploads/2018/07/quality-outcome-framework-report-of-the-review.pdf%0Ahttps:/www.england.nhs.uk/publication/report-of-the-review-of-the-quality-and-outcomes-framework-in-england) 66. Bardi A. Relations of values to behavior in everyday situations. The Hebrew University; 2000. 67. Goldfarb D. Family Doctor Incentives: Getting closer to the Sweet Spot. The [Conference Board of Canada; 2014;22. https://www.newswire.ca/](https://www.newswire.ca/newsreleases/doctors-pay-should-be-based-on-the-right-blend-of-incentives-514340331.html) [newsreleases/doctors-pay-should-be-based-on-the-right-blend-of-](https://www.newswire.ca/newsreleases/doctors-pay-should-be-based-on-the-right-blend-of-incentives-514340331.html) [incentives-514340331.html.](https://www.newswire.ca/newsreleases/doctors-pay-should-be-based-on-the-right-blend-of-incentives-514340331.html) 68. Mead N, Bower P. Measuring patient-centredness: A comparison of three observation-based instruments. Patient Educ Couns 2000;. 69. Langberg EM, Dyhr L, Davidsen AS. Development of the concept of patientcentredness – A systematic review. Patient Educ Couns [Internet]. 2019 Feb [27 [cited 2019 Mar 18]; Available from: https://www.sciencedirect.com/](https://www.sciencedirect.com/science/article/pii/S073839911830733X) [science/article/pii/S073839911830733X.](https://www.sciencedirect.com/science/article/pii/S073839911830733X) 70. Tracy DK, Hanson K, Brown T, James AJB, Paulsen H, Mulliez Z, et al. Integrated care in mental health: Next steps after the NHS Long Term Plan. Vol. 214, British Journal of Psychiatry. Cambridge University Press; 2019. p. 315–7. 71. Marshall M, Sheaff R, Rogers A, Campbell S, Halliwell S, Pickard S, et al. A qualitative study of the cultural changes in primary care organisations needed to implement clinical governance. British Journal of General Practice. 2002. 72. Marshall M, Sheaff R, Rogers A, Campbell S, Halliwell S, Pickard S, et al. A qualitative study of the cultural changes in primary care organisations needed to implement clinical governance. Br J Gen Pract. 2002;52(481):641– [645. Available from: http://www.ncbi.nlm.nih.gov/pubmed/12171222%5](http://www.ncbi.nlm.nih.gov/pubmed/12171222/nhttp:/www.pubmedcentral.nih.gov/articlerender.fcgi?artid=PMC1314382) [Cnhttp://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=PMC1314382](http://www.ncbi.nlm.nih.gov/pubmed/12171222/nhttp:/www.pubmedcentral.nih.gov/articlerender.fcgi?artid=PMC1314382) [cited 2018 7 Oct]. 73. Rogers A, Pilgrim D. A sociology of mental health and illness [Internet]. Primary Care. Open University Press; 2005 [cited 2018 Oct 4]. 269 p. [Available from: http://psycnet.apa.org/psycinfo/1994-97819-000.](http://psycnet.apa.org/psycinfo/1994-97819-000) 74. Dixon-woods M, Agarwal S, Young B, Jones D, Sutton A. Integrative approaches to qualitative and quantitative evidence [Internet]. Vol. 181, [Health Development Agency. 2004 [cited 2018 Oct 2]. Available from: www.](http://www.hda.nhs.uk) [hda.nhs.uk.](http://www.hda.nhs.uk) 75. Khan N, Bower P, Rogers A. Guided self-help in primary care mental health: Meta-synthesis of qualitative studies of patient experience. Br J Psychiatry. [2007;191(3):206–11 Available from: http://www.ncbi.nlm.nih.gov/pubmed/1](http://www.ncbi.nlm.nih.gov/pubmed/17766759) [7766759 [cited 2016 27 Apr].](http://www.ncbi.nlm.nih.gov/pubmed/17766759) 76. Walsh D, Downe S. Meta-synthesis method for qualitative research: A [literature review. J Adv Nurs. 2005;50(2):204–11 Available from: http://www.](http://www.ncbi.nlm.nih.gov/pubmed/15788085) [ncbi.nlm.nih.gov/pubmed/15788085 [cited 2019 25 Feb].](http://www.ncbi.nlm.nih.gov/pubmed/15788085) 77. Protheroe J, Rogers A, Kennedy AP, Macdonald W, Lee V. Promoting patient engagement with self-management support information: A qualitative meta-synthesis of processes influencing uptake. Implement Sci. 2008;3(1):44 [Available from: http://www.ncbi.nlm.nih.gov/pubmed/18851743 [cited 2016](http://www.ncbi.nlm.nih.gov/pubmed/18851743) 27 Apr]. 78. Pielstick, C. D., The transforming leader: A meta-ethnographic analysis, [Community College Review, 1998:26(3):15–34 https://doi.org/10.1177/](https://doi.org/10.1177/009155219802600302) [009155219802600302.](https://doi.org/10.1177/009155219802600302) Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. -----
{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC7359468, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://bmcprimcare.biomedcentral.com/track/pdf/10.1186/s12875-020-01208-8" }
2,020
[ "JournalArticle" ]
true
2020-07-13T00:00:00
[ { "paperId": "faa45a320ea6e2a762277e492abe82bdce623b7d", "title": "Modernization and Postmodernization" }, { "paperId": "653cc66bb9cde633dda203de6aace8bb3715a848", "title": "Professionalism as the Bedrock of High-Value Care" }, { "paperId": "47e6a43c6562f2fb77e71a35c80f35ff808d1f4e", "title": "Assessment of medical professionalism: preliminary results of a qualitative study" }, { "paperId": "f0cfc9a8b2f34d9b27164aeeeeb384e522406dfa", "title": "Development of the concept of patient-centredness - A systematic review." }, { "paperId": "3ee294f2839fdefb0c1eeff269990881036c3710", "title": "Integrated care in mental health: next steps after the NHS Long Term Plan." }, { "paperId": "363aea77fc6105c862e22266de1e8a07c21ce12e", "title": "Quality of care assessment for people with multimorbidity" }, { "paperId": "407f053dc06ed935fccd60e0ff0249d097dcfaab", "title": "GP retention in the UK: a worsening crisis. Findings from a cross-sectional survey" }, { "paperId": "4667782d7abab0d982f7be947dc1a9428cbdb5c6", "title": "Qualitative evidence to improve guidelines and health decision-making" }, { "paperId": "55f17a6e9032267d91c1063c1274dd92066aa023", "title": "The role of the Quality and Outcomes Framework in the care of long-term conditions: a systematic review" }, { "paperId": "4834efca2fd183faf8433c0398fb7d731e31586a", "title": "Does pay-for-performance in primary care save lives?" }, { "paperId": "2dc7dc944b955539d0aee023934bc19059a64da2", "title": "Living with complexity; marshalling resources: a systematic review and qualitative meta-synthesis of lived experience of mental and physical multimorbidity" }, { "paperId": "9ace79e4e45a4346126b1375ca4798d7ea40e3c4", "title": "Excellence in Transitional Care of Older Adults and Pay-for-Performance: Perspectives of Health Care Professionals." }, { "paperId": "46337e1c1cc7b7437c90539a10b19ed7dace2564", "title": "Effect of a national primary care pay for performance scheme on emergency hospital admissions for ambulatory care sensitive conditions: controlled longitudinal study" }, { "paperId": "505dd5cb144738cc60458b42faec7600235ffd2a", "title": "‘Just another incentive scheme’: a qualitative interview study of a local pay-for-performance scheme for primary care" }, { "paperId": "782ba17210a7744bf620dda16e4cbd1084ef919f", "title": "Incentivised case finding for depression in patients with chronic heart disease and diabetes in primary care: an ethnographic study" }, { "paperId": "62cfc00c93f82a89beb1e5d133e651c68ed545ea", "title": "Discretion or discretions? Delineating professional discretion: the case of English medical practice." }, { "paperId": "e5988fdbd549355d0d73c6d4d19b812b4f0b061c", "title": "How QOF is shaping primary care review consultations: a longitudinal qualitative study" }, { "paperId": "d3245e4b9dd344a984390b3242119e0adb00af1a", "title": "Implementation of pay for performance in primary care: a qualitative study 8 years after introduction." }, { "paperId": "459e66e2d5c6907faa26fbe27f34223da6393b5e", "title": "A qualitative study of primary care professionals’ views of case finding for depression in patients with diabetes or coronary heart disease in the UK" }, { "paperId": "de1067944cbe3574469d641738995de886622ddb", "title": "An Overview of the Schwartz Theory of Basic Values" }, { "paperId": "6d1ca248e97573555f4ddf811158edff33c0d6cc", "title": "Recommendations for improving the end-of-life care system for homeless populations: A qualitative study of the views of Canadian health and social services professionals" }, { "paperId": "5822da792f03bc44d36cfb4ff4bd80d397a0b1dd", "title": "General practitioners’ views on quality markers for children in UK primary care: a qualitative study" }, { "paperId": "c36e74a4f5375c3ce54be43ec4d9b4663b758ffa", "title": "Why pay for performance may be incompatible with quality improvement" }, { "paperId": "93c126cac72473335be3bcff40a002d2bbbec20a", "title": "Recording patient preferences for end-of-life care as an incentivized quality indicator: What do general practice staff think?" }, { "paperId": "b1be1a503022981978a85c5a76df6a0b9b5f9a95", "title": "Patients' views of pay for performance in primary care: a qualitative study." }, { "paperId": "d64d378d9e4edfe87a86a47abf1b41edd65a6210", "title": "Incentive payments are not related to expected health gain in the pay for performance scheme for UK primary care: cross-sectional analysis" }, { "paperId": "2cb24dbb70315530f1f5746b8a32bee9028c0bf7", "title": "Experience of contractual change in UK general practice: a qualitative study of salaried GPs." }, { "paperId": "9bcd99fc26c861b6033f1f31da43c5ee39164821", "title": "Evaluating meta-ethnography: systematic analysis and synthesis of qualitative research." }, { "paperId": "788c334aa8b2bb5a2b956fa3708ed7cedfc34cc5", "title": "Living with complexity" }, { "paperId": "44be08b4f740f0e937fd0597edeea21c3f48c543", "title": "Identifying unintended consequences of quality indicators: a qualitative study" }, { "paperId": "7f68515edffc5bb8f60d2c16d6f4970fb5dd59a2", "title": "Impact of the QOF and the NICE guideline in the diagnosis and management of depression: a qualitative study." }, { "paperId": "99b9b37608977dc69da3cb46a6e6c5c8c9fea2d0", "title": "Exception reporting in the Quality and Outcomes Framework: views of practice staff - a qualitative study." }, { "paperId": "1ea28192d9c00f69393e15b9b846cc2b64a507e2", "title": "The impact of removing financial incentives from clinical quality indicators: longitudinal analysis of four Kaiser Permanente indicators" }, { "paperId": "f8641d2eb2fd50258595f4cf1ce22811aba7a5f3", "title": "Transforming medical professionalism to fit changing health needs" }, { "paperId": "38ff5157d201cd9cc6284e063d00bbc28d0e0898", "title": "Pay for Performance in Primary Care in England and California: Comparison of Unintended Consequences" }, { "paperId": "49fa7627db03cc6581dde1ddf0da8890be672ea6", "title": "The association of posttraumatic stress disorder and metabolic syndrome: a study of increased health risk in veterans" }, { "paperId": "804eb93c3975d489582f71519c491889a9125a79", "title": "Promoting patient engagement with self-management support information: a qualitative meta-synthesis of processes influencing uptake" }, { "paperId": "31eaa9d55217f96741055c1afa942eca8f5ea94e", "title": "Impact of the 2004 GMS contract on practice nurses: a qualitative study." }, { "paperId": "94532e3f6b414f215fe9176bbdac807f4002aa35", "title": "Effects of payment for performance in primary care: Qualitative interview study" }, { "paperId": "ebb7f2f7286932a8b7a7432340b8c651c3429583", "title": "The Experience of Pay for Performance in English Family Practice: A Qualitative Study" }, { "paperId": "a20ded5ef5dd3ed6f1d66f0d5bd5b6146548f395", "title": "Incentives and control in primary health care: findings from English pay-for-performance case studies." }, { "paperId": "7fae41a5db640ddfbfae636de23a7b98490f8bea", "title": "Income inequality and health status: a nursing issue" }, { "paperId": "146e328b5b63522fe0f6015e6637fea9126945bc", "title": "Guided self-help in primary care mental health" }, { "paperId": "e4e44dd89b8988863eecb15a98ab9b6b1d74de65", "title": "Defining medical professionalism: a qualitative study" }, { "paperId": "5a948b9fbf9c9f86fc24e1a010db3389f1923273", "title": "Acquiring Qualitative Data for Secondary Analysis" }, { "paperId": "4dc2a922b783dc7a090bad91c8c1f1b7799d223a", "title": "Meta-synthesis method for qualitative research: a literature review." }, { "paperId": "779246c4ed513451cbadff8eb5c273995e8828f7", "title": "Strategies in Teaching Secondary Analysis of Qualitative Data" }, { "paperId": "20010b4c4eea80a9dd99f98147e32c4976c18743", "title": "A qualitative study of the cultural changes in primary care organisations needed to implement clinical governance." }, { "paperId": "0baa6343fad5baa372d90552a583b8cef02306ab", "title": "Influence of context effects on health outcomes: a systematic review" }, { "paperId": "eb77a32d997fcb8a25ffa24bdd27ea5b9b61b917", "title": "Modernization and Postmodernization: Cultural, Economic and Political Change in 43 Societies" }, { "paperId": "79e3aacd2c5bc1ccafb6acaaeb2ab76d62e2af9d", "title": "The Transforming Leader: A Meta-Ethnographic Analysis" }, { "paperId": "409ce0841305791cf639a240eb7089b1b9950a42", "title": "Qualitative Research Design: An Interactive Approach" }, { "paperId": "6d441a5723b5708413789cad352bce68bf8e283a", "title": "Primary care" }, { "paperId": "d03566dc38b24426e3a63a2dc7b5af2a6fc377a7", "title": "Readers panel Spending money: How should the NHS spend the £300 million set aside to prevent a winter bed crisis? Our readers panel replies" }, { "paperId": "ba1b982bdea9f9f721fb5906aa15b3aacd17ff9c", "title": "Influences of Adaptation to Communist Rule on Value Priorities in Eastern Europe" }, { "paperId": "8c57841534b8102fcd351a22c06edf5c3da659f2", "title": "Aggregating Qualitative Findings: An Approach to Theory Development" }, { "paperId": null, "title": "a nursing issue [Internet" }, { "paperId": null, "title": "Development of the concept of patientcentredness – A systematic review" }, { "paperId": null, "title": "" }, { "paperId": null, "title": "Advancing medical professionalism" }, { "paperId": null, "title": "Exclusive NHS England rejected QOF Suspension bacause of opposition from GPs. GPonline" }, { "paperId": null, "title": "Exclusive NHS England rejected QOF Suspension bacause of opposition from GPs" }, { "paperId": "7bda6229597353802c281405e6b8e83ed5649a17", "title": "质量与结果框架:我们学到了什么 ? Quality and Outcomes Framework : what have we learnt ?" }, { "paperId": null, "title": "Delivering the Forward View: NHS planning guidance [Internet" }, { "paperId": null, "title": "Building the Workforce - the New Deal for General Practice GP Induction and Refresher Scheme, 2015-2018" }, { "paperId": null, "title": "Building the Workforce -the New Deal for General Practice GP Induction and Refresher Scheme" }, { "paperId": "bc6bcc9d7047c88ebf1500d5e0c177e075d4eed2", "title": "Criteria for the evaluation of qualitative research" }, { "paperId": null, "title": "Using evidence from qualitative research to develop WHO guidelines [Internet]. WHO handbook for guideline development" }, { "paperId": null, "title": "Family Doctor Incentives: Getting closer to the Sweet Spot" }, { "paperId": null, "title": "Family Doctor Incentives: Getting closer to the Sweet Spot. The Conference Board of Canada" }, { "paperId": null, "title": "QOF points: valuable to whom" }, { "paperId": null, "title": "QOF points: valuable to whom? [Internet" }, { "paperId": null, "title": "Overseas and retired doctors ed.)" }, { "paperId": null, "title": "QOF points: valuable to whom? Source" }, { "paperId": "edd7e294b0f0e0208ad57beb9765da8b14c20461", "title": "Unit 2 Theoretical and Methodological Issues Subunit 1 Conceptual Issues in Psychology and Culture Article 6 8-1-2002 Psychology of Morality" }, { "paperId": null, "title": "Divided we fail: the Harveian oration 2011" }, { "paperId": null, "title": "" }, { "paperId": null, "title": "" }, { "paperId": "17c11597f1d27be7e4cafc62b6c30682795f3f3c", "title": "The impact of the Quality and Outcomes Framework on practice organisation and service delivery: summary of evidence from two qualitative studies." }, { "paperId": "912877933b56af14cdf6c118652d9bf0b10255ae", "title": "Developing Quality and Outcomes Framework (QOF) indicators and the concept of 'QOFability'." }, { "paperId": null, "title": "" }, { "paperId": null, "title": "" }, { "paperId": null, "title": "Income inequality and health status: a nursing issue [Internet" }, { "paperId": "e5634ecd1809029680abe04ca53157cc0e5a3fa7", "title": "Health Development Agency" }, { "paperId": null, "title": "It’s about more than money: financial incentives and internal motivation" }, { "paperId": null, "title": "It's about more than money: financial incentives and internal motivation. Qual Saf Heal Care" }, { "paperId": null, "title": "L . Strategies in Teaching Secondary Analysis of Qualitative Data" }, { "paperId": "9929e90f4fc3c1e040eace8ad21b571f7f346729", "title": "Integrative approaches to qualitative and quantitative evidence" }, { "paperId": "34b9635d7779e219e9d60e0d3d33919ca9bc123c", "title": "Publisher's Note" }, { "paperId": "62f5c9a7d5e8ac3e87ea32081427874c44844e99", "title": "Measuring patient-centredness: a comparison of three observation-based instruments." }, { "paperId": null, "title": "Relations of values to behavior in everyday situations. The Hebrew University" }, { "paperId": null, "title": "Relations of values to behavior in everyday situations" }, { "paperId": "f8eadcbd6b684b86a964664925dbbdfde54754ee", "title": "A Sociology of Mental Health and Illness" }, { "paperId": "aae327c6e142861dacdca6a4b3345d095ffec4ce", "title": "Modernization and Postmodernization: Cultural, Economic, and Political Change in 43 Societies" }, { "paperId": null, "title": "British Sociological Association. Criteria for the evaluation of qualitative research papers" }, { "paperId": null, "title": "" }, { "paperId": "01315b84c630055a16e45f91b97509807680c70d", "title": "Are There Universal Aspects in the Structure and Contents of Human Values" }, { "paperId": "dc49e27d0ed890cd3ed2e80ca0b0107207f12a64", "title": "Universals in the Content and Structure of Values: Theoretical Advances and Empirical Tests in 20 Countries" }, { "paperId": null, "title": "" }, { "paperId": null, "title": "Meta-Ethnography: Synthesizing Qualitative Studies (Qualitative Research Methods) [Internet" }, { "paperId": "ccd8148a65f127e1d06c8d97a041ace14982b60f", "title": "The sociology of mental health and illness" }, { "paperId": "bb194b9ae6428d282d60b5952b1b428e0545ad61", "title": "Report of the Review of the Quality and Outcomes Framework in England" }, { "paperId": "fcb05f2c34cc55f46e44a4b389ad78d1dc98a5de", "title": "Bmc Medical Research Methodology Open Access Optimal Classifier Selection and Negative Bias in Error Rate Estimation: an Empirical Study on High-dimensional Prediction" }, { "paperId": null, "title": "Delivering the Forward View: NHS planning guidance" }, { "paperId": null, "title": "Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations" }, { "paperId": null, "title": "" } ]
25,156
en
[ { "category": "Psychology", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Psychology", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0132509cb91893022691151e825cdef22692df5c
[ "Psychology", "Medicine" ]
0.858548
Sensorimotor Rhythm Neurofeedback Enhances Golf Putting Performance.
0132509cb91893022691151e825cdef22692df5c
Journal of Sport & Exercise Psychology (JSEP)
[ { "authorId": "2004830021", "name": "M. Cheng" }, { "authorId": "117129350", "name": "Chung-Ju Huang" }, { "authorId": "3155854", "name": "Yu-Kai Chang" }, { "authorId": "21728617", "name": "D. Koester" }, { "authorId": "9500975", "name": "T. Schack" }, { "authorId": "79743957", "name": "T. Hung" } ]
{ "alternate_issns": null, "alternate_names": [ "J Sport Exerc Psychol (JSEP", "J Sport Exerc Psychol", "Journal of Sport & Exercise Psychology" ], "alternate_urls": null, "id": "de07b8c7-cfb1-4ac5-bf01-bce16955ff22", "issn": "0895-2779", "name": "Journal of Sport & Exercise Psychology (JSEP)", "type": "journal", "url": "http://www.humankinetics.com/JSEP/" }
null
**_Journal of Sport & Exercise Psychology, 2015, 37, 626 -636_** http://dx.doi.org/10.1123/jsep.2015-0166 © 2015 Human Kinetics, Inc. ORIGINAL RESEARCH # Sensorimotor Rhythm Neurofeedback Enhances Golf Putting Performance #### Ming-Yang Cheng,[1] Chung-Ju Huang,[2] Yu-Kai Chang,[3] Dirk Koester,[1] Thomas Schack,[1] and Tsung-Min Hung[4] 1Bielefeld University; 2University of Taipei; 3National Taiwan Sport University; 4National Taiwan Normal University Sensorimotor rhythm (SMR) activity has been related to automaticity during skilled action execution. However, few studies have bridged the causal link between SMR activity and sports performance. This study investigated the effect of SMR neurofeedback training (SMR NFT) on golf putting performance. We hypothesized that preelite golfers would exhibit enhanced putting performance after SMR NFT. Sixteen preelite golfers were recruited and randomly assigned into either an SMR or a control group. Participants were asked to perform putting while electroencephalogram (EEG) was recorded, both before and after intervention. Our results showed that the SMR group performed more accurately when putting and exhibited greater SMR power than the control group after 8 intervention sessions. This study concludes that SMR NFT is effective for increasing SMR during action preparation and for enhancing golf putting performance. Moreover, greater SMR activity might be an EEG signature of improved attention processing, which induces superior putting performance. **_Keywords: precision sports, attention, EEG, sensorimotor rhythm, automaticity_** The quality of mental regulation can differentiate superior from inferior performance in precision sports activities such as golf putting. In golf, the putt is considered one of the most important parts of the game, representing on average 43% of all shots taken during a single round (Pelz & Frank, 2000). From a technical perspective, putting is the simplest skill used in golf. However, mentally, putting is the most stressful and demanding activity in the game (Nicholls, 2007). The mental challenge of putting is reflected by previous psychophysiological studies showing complex brain processes during putting performance (Babiloni et al., 2008). Hence, the maintenance Ming-Yang Cheng is with Center of Excellence “Cognitive Interaction Technology” (CITEC), Bielefeld University, Bielefeld, Germany. Chung-Ju Huang is with the Graduate Institute of Sport Pedagogy, University of Taipei, Taipei City, Taiwan, Republic of China. Yu-Kai Chang is with the Graduate Institute of Athletics and Coaching Science, National Taiwan Sport University, Taoyuan County, Taiwan, Republic of China. Dirk Koester is with the Center of Excellence “Cognitive Interaction Technology” (CITEC), Bielefeld University, Bielefeld, Germany. Thomas Schack is with the Center of Excellence “Cognitive Interaction Technology” (CITEC), Bielefeld University, Bielefeld, Germany. Tsung-Min Hung is with the Department of Physical Education, National Taiwan Normal University, Taipei City, Taiwan, Republic of China. Address author correspon[dence to Tsung-Min Hung at ernesthungkimo@yahoo.com.tw.](mailto:ernesthungkimo@yahoo.com.tw) of a mental state conducive to skilled execution is critical for ideal precision sports performance. Superior performance in precision sports can be characterized as an automatic process as opposed to a controlled process, which is typically observed in less skilled performers (Fitts & Posner, 1967). An automatic process is by nature reflexive, whereas a controlled process is an intentionally initiated sequence of cognitive activity (Schneider & Shiffrin, 1977). Achieving automatic process in action execution is the primary goal of mastery (Logan, Hockley, & Lewandowsky, 1991). Differences between these two levels of cognitive processing are reflected at the neurophysiological level: participants who were in the automatic stage exhibited weaker activity of the bilateral cerebellum, presupplementary motor area, premotor cortex, parietal cortex, and prefrontal cortex compared with novices (Wu, Chan, & Hallett, 2008). In addition, the somatosensory cortex has been related to conscious perception of somatosensory stimuli (Nierhaus et al., 2015), such that lower activity in the somatosensory cortex might be a signature of reduced conscious involvement in movement execution, as is frequently observed in highly skilled performers. Although previous studies of the brain function underlying superior golf putting performance have provided insights into adaptive mental states and their cortical processes, few studies have examined the cortical processes that are more directly associated with somatosensory activity. For example, Babiloni ----- et al. (2008) demonstrated that successful putting was preceded by higher high-frequency alpha (10–12 Hz) event-related desynchronization over the frontal midline and the right primary sensorimotor area compared with unsuccessful putting performance. Similarly, studies found that reduced (Kao, Huang, & Hung, 2013) and stable (Chuang, Huang, & Hung, 2013) frontal midline theta power was the precursor of superior performance in precision sports. Since high-frequency alpha power in these cortical areas reflect only task-related attention (Klimesch, Doppelmayr, Pachinger, & Ripper, 1997) whereas frontal midline theta power indicates top-down sustained attention (Sauseng, Hoppe, Klimesch, Gerloff, & Hummel, 2007), these findings support the importance of specialized task-related attention on superior motor performance. However, the information encoded during automatic somatosensory processing during skilled precision sport performance remains unexamined as yet. Sensorimotor rhythm (SMR), the 12- to 15-Hz oscillation of the sensorimotor cortex, has shown promising as a link between adaptive mental states (e.g., automatic process-related attention) and skilled visuomotor performance. Sensorimotor rhythm is considered an indicator of cortical activation, which is inversely related to somatosensory processing (Mann, Sterman, & Kaiser, 1996). A recent study showed that skilled dart-throwing players demonstrated higher SMR power before dart release than novices in a dart-throwing task (Cheng et al., 2015). This result suggests that lower cognitive involvement in processing somatosensory information as reflected by higher SMR power is characteristic of skilled performance. Furthermore, several lines of studies pertaining to SMR power tuning for enhancing adaptive cortical processing in motor performance have shown promising results. Augmented SMR power resulting from neurofeedback training (NFT) has been identified as a relaxed focus state without somatosensory intervention (Gruzelier, Foks, Steffert, Chen, & Ros, 2014). Similarly, a reduced trait anxiety score and task-processing time during microsurgery were observed after augmented SMR NFT (Ros et al., 2009). Moreover, a facilitative sense of control, confidence, and feeling at-one with a role was demonstrated after augmented SMR NFT before acting performance (Gruzelier, Inoue, Smart, Steed, & Steffert, 2010). Thus, increased SMR activity implies the maintenance of a relaxed, focused state by reducing motor perception (e.g., somatosensory processing) by the sensorimotor cortex (Vernon et al., 2003). This interpretation is similar to the mental characteristics of peak performance in skilled athletes (Krane & Williams, 2006) and is in agreement with the concept of automaticity proposed by Fitts and Posner (1967). Hence, SMR power not only might be a sensitive indicator of the activity of sensorimotor cortex (Mann et al., 1996) but also shows potential for a performance-enhancing intervention. Although there is no direct evidence to support the effectiveness of SMR NFT on performance enhancement in precision sport, two lines of research lend support to its potential use in sports. First, previous studies have demonstrated the effectiveness of NFT on performance enhancement in precision sports. For example, Landers et al. (1991) demonstrated that [“]correct[”] NFT (i.e., augmented slow cortical potential at the left temporal lobe) led to superior performance, whereas [“]incorrect[”] NFT (i.e., augmented slow cortical potential at the right temporal lobe) impaired performance in skilled archers. Similarly, Kao, Huang, and Hung (2014) reported that NFT targeting to reduce the frontal midline theta resulted in improved performance in skilled golfers. These findings support the feasibility of tuning EEG to improve behavioral outcome in precision sports. The second line of evidence is the finding that SMR NFT has a beneficial effect on attention-related performance in various attentional tasks. For example, an increased P300b amplitude at frontal, central, and parietal sites during the auditory oddball task and reduced commission errors, and a reduction in reaction time variability during the Test of Variables of Attention (TOVA) was observed after augmented SMR NFT (Egner, Zech, & Gruzelier, 2004). These findings suggest that augmenting SMR power might improve attention-related processes by improving impulse control and the ability to integrate relevant environmental stimuli. Similarly, Ros et al. (2009) reported that a shorter operation time and reduced trait anxiety score were observed in surgeons following augmented SMR NFT, suggesting that augmented SMR enhanced the learning of a complex medical specialty by developing sustained attention and a relaxed attentional focus as well as increasing working memory (Vernon et al., 2003). Furthermore, Doppelmayr and Weber (Doppelmayr & Weber, 2011) revealed that augmented SMR NFT not only resulted in a significant SMR amplitude increase accompanied by a significant increase in reward threshold, but also facilitated the performance of spatialrotation, simple, and choice-reaction time tasks. These results indicate that visuospatial processing, semantic memory regulation, and the integration of relevant stimuli can be improved following augmented SMR NFT. Collectively, the benefits of augmented SMR NFT can be attributed to an improved regulation of somatosensory and sensorimotor pathways, which in turn leads to more efficient attention allocation (Kober et al., 2014) that results in an improved processing of task-relevant stimuli. To the best of our knowledge, no study has directly examined the effect of SMR NFT on precision sport performance. Thus, this study investigated the effect of SMR NFT on a golf putting task. We predicted that golfers would be able to increase SMR power before putting execution following augmented SMR NFT. More importantly, we predicted that increased SMR power improves putting performance as a result of augmented SMR NFT. ## Methods #### Participants Fourteen male and two female preelite and elite golfers were recruited (mean handicap = 0, _SD = 3.90)._ ----- Participants were matched based on performance history supplemented by the assessment of a professional coach and then randomly assigned into either an SMR neurofeedback group (SMR NFT) or a control group (seven male and one female for each group). The mean age of the SMR NFT and control group were 20.6 (1.59) and 22.3 (2.07), respectively. The years of experience in golf were 9.5 (2.67) for the SMR NFT group and 9.2 (1.83) for the control group. An independent t test showed no difference in age [t(14) = 1.895, _p = .079] or years of experience in golf [t(14) = 0.273, p_ = .789] between the two groups. None of the participants reported psychiatric and neurological disorders and had never been hospitalized for general brain damage. #### Procedures For the pretest and posttest, we used the same procedure to collect data. At pretest, after being informed of the general purpose of the study, all participants were asked to read and sign an informed consent form approved by our institutional review board. They were then given the opportunity to ask questions about the experiment. The participants were individually tested in a sound-proof indoor artificial golf green, where they were initially required to stand 3 m from a hole 10.8 cm in diameter to obtain an individual putting distance (Arns, Kleinnijenhuis, Fallahpour, & Breteler, 2008). Participants performed a series of 10 putts, which were scored as successfully holed or not holed. The percentage of successful putts in a series was determined after each series. This process was repeated until each participant achieved 50% accuracy. After the individual putting distance was determined, participants were fitted with a Lycra electrode cap (Neuroscan, Charlotte, NC, USA). After a 10-min warm-up, participants were first asked to undergo a resting EEG recording, including eye-closed and eye-opened conditions, while assuming a normal putting stance for 1 min each. Then, all participants performed golf putting tasks consisting of 40 self-paced putting trials in four separate recording blocks while EEGs were recorded. The participants performed the putting task in the standing position and were allowed to take a brief rest between each putt. They were also allowed to sit briefly after each block of 10 putts. The score was calculated based on the linear distance from the edge of hole to the edge of the ball (cm). Putting into the hole successfully was determined as score 0. Putting trials in which the ball was deflected by contacting the edge of the hole were excluded, and participants were asked to perform extra putting trials to complete the forty trials. The experiment lasted approximately 2 hr in total. After completing the pretest, all participants were scheduled to go through 8 sessions of neurofeedback training. Then the posttest, which was identical to the pretest, followed the neurofeedback intervention. #### Instrumentation **_Electroencephalography._** For the pretest and posttest, EEGs were recorded at 32 electrode sites (FP1, FP2, F7, F8, F3, F4, FZ, FT7, FT8, FC3, FC4, C3, C4, CZ, T3, T4, T5, T6, TP7, TP8, CP3, CP4, CPZ, A1, A2, P3, P4, PZ, O1, O2, OZ) corresponding to the International 10–10 system (Chatrian, Lettich, & Nelson, 1985). In addition, four electrodes were attached to acquire horizontal and vertical oculography (HEOL, HEOR, VEOU and VEOL). All sites were initially referenced to A1 and then rereferenced to linked ears offline. A frontal midline site (FPz) served as the ground. EEG data were collected and amplified using a Neuroscan Nuamps amplifier (Neuroscan, Charlotte, NC, USA) with a band-pass filter setting of 1–100 Hz and a 60-Hz notch filter. The EEG and EOG signals were sampled at 500 Hz and recorded online with NeuroScan 4.5 (Neuroscan, Charlotte, NC, USA) software installed on a Lenovo R400 laptop (Lenovo, Taipei City, R.O.C). Vertical and horizontal eye movement artifacts were recorded via bipolar electro-oculographic activity (EOG), in which vertical EOG was assessed by electrodes placed above and below the left eye (VEOU and VEOL), whereas horizontal EOG was assessed by electrodes located at the outer canthi (HEOL, HEOR). Impedance values for all electrode sites were maintained below 5 kΩ. An infrared ray sensor was set to detect the swing for each putt. Once the back swing movement was detected, an event mark was sent to the EEG data, which served as the time point for analyzing the EEG activity before putting. Twelve to fifteen hertz of Cz was extracted as the SMR (Babiloni et al., 2008). **_Neurofeedback._** Neurofeedback training was completed with a NeuroTek Peak Achievement Trainer (NeuroTek, Goshen, KY). The EEG data from the assessment were band-pass filtered using the BioReview software (NeuroTek, Goshen, KY). The active scalp electrode was placed at Cz for SMR training, with the reference placed on both mastoids. Signal was acquired at 256 Hz and then A/D converted and band filtered to extract the SMR (12–15 Hz). The amplitude of the SMR was transformed online into graphical feedback representations including the low-frequency audio-feedback tone by acoustic bass (No. 33) in the BioReview software. #### Neurofeedback Training Procedure Participants underwent an eight-session training program lasting 5 weeks. Each session was composed of neurofeedback training lasting from 30 to 45 min. On average, a total of 12 training trials were performed in a single session. Each training trial comprised 30 s. The total duration of a single session was approximately 30 min. The SMR NFT group aimed to increase absolute SMR amplitude over the designated threshold, which was individually determined by averaging 1.5 s of each participant’s successful putting trials during the pretest. To enhance the participants’ efficacy during NFT, a progressive adjustment of the training threshold difficulty was employed. The standard for adjusting the training threshold was based on the individualized standard deviation which derived from the SMR power of the final three 0.5-s time windows before putting during the pretest. When ----- participants’ SMR power was higher than the threshold, the acoustic bass sound was displayed. Participants were instructed to perform based on their own putting routine while receiving the auditory feedback. The successful training ratio, defined as the time spent above threshold divided by the total time of a single training trial (30 s), was reported to participants following every training trial. In the control group, the training protocol was similar to that used by Egner, Strawson, and Gruzelier (2002) to establish a mock feedback condition. This protocol was designed to prevent study participants from learning to regulate SMR by using the randomly prerecorded feedback tone during the training trials from SMR NFT group. The total length of this prerecorded mock feedback tone was 4 min that were derived from a randomly chosen participant in the SMR NFT group during the Session 1 training. Researchers played the mock feedback tone from a random starting point to guarantee a randomized feedback tone was received by participants in the control group. On average, a total of seven training trials were performed in a single session and the total duration of a single session was approximately 30 min. To evaluate the neurofeedback learning effect, the mean successful training ratio of each session was recorded and computed for subsequent analysis. To reduce the number of sessions necessary for statistical evaluation of the learning efficiency between the two groups, we combined two consecutive sessions into one section [e.g., Section 1 = (Session 1 + Session 2) / 2]. #### Data Reduction The EEG data reduction was conducted offline using the Scan 4.5 software (Neuroscan, Charlotte, NC, USA). EEG data were sampled 1.5 s before putting execution and were triggered by the event-related marker from infrared ray sensors. Trial preparation periods of less than 1.5 s were excluded to establish the common structure of artifact-free data across trials and participants. EOG correction (Semlitsch, Anderer, Schuster, & Presslich, 1986) was carried out on continuous EEG data to eliminate blink artifacts. EEG segments with amplitudes exceeding ±100 μV from baseline were excluded from subsequent analysis. After artifact-free EEG data were acquired, fast Fourier transforms were calculated at 50% overlap on 256-sample Hanning windows for all artifact-free segments to transform to spectral power (μV[2]). Sensorimotor rhythm power was computed as the mean of 12–15 Hz from Cz and then natural log transformed (Davidson, 1988). To compute a normalized EEG power for each golfer, the relative power was used, for which the ratio of power at 12–15 Hz to 1–30 Hz was computed (Niemarkt et al., 2011). #### Statistical Analyses The average putting score and standard deviation between the two groups was analyzed by a 2 (Group: SMR NFT, Control) × 2 (Test: pretest, posttest) ANOVA with repeated measures on the test factor. The difference score (posttest to pretest) for the relative power of SMR was subjected to a 2 (Group: SMR NFT, Control) × 3 [Time window: –1.5 to –1.0 s (T1), –1.0 to –0.5 s (T2), –0.5 to 0 s (T3)] ANOVA with repeated measures on the time window factor. In addition, we ran several control analyses to provide additional evidence to support our conclusions. The success of the training ratio was tested by a 2 (Group: SMR NFT, Control) × 4 (Training section: Section 1: sessions 1–2; Section 2: sessions 3–4; Section 3: sessions 5–6; Section 4: sessions 7–8) ANOVA with repeated measures on the training section. To characterize the within-session learning effect, we compared the successful training ratio of the first and last trials of each session across all eight sessions. A 2 (Group: SMR NFT, Control) × 8 (Session: session 1, 2, 3, 4, 5, 6, 7, 8) × 2 (Trial: first trial, last trial) three-way ANOVA with repeated measures on the session, and trial was used to examine this issue. To ensure control of neurofeedback in the SMR NFT group within the training program, we employed a one-way ANOVA with training section (Training section: Section 1: sessions 1–2; Section 2: sessions 3–4; Section 3: sessions 5–6; Section 4: sessions 7–8) as a variable to detect the threshold fluctuation within the four training sections. To examine the regional fluctuation of 12–15 Hz power before and after training, we carried out a 2 (Group: SMR NFT, Control) × 4 (Region: frontal, central, parietal, occipital) two-way ANOVA with repeated measures on the region. The examination of concurrent changes in neighboring frequency bands was conducted by analyzing the pre-to-post difference scores for theta (4–7 Hz), alpha (8–12 Hz), low beta (13–20 Hz), high beta (21–30 Hz), and broad beta (13–30 Hz) frequency bands with a 2 (Group: SMR NFT, Control) × 3 [Time window: –1.5 to –1.0 s (T1), –1.0 to –0.5 s (T2), –0.5 to 0 s (T3)] twoway ANOVA. Mauchly’s test was used to assess the validity of the ANOVA sphericity assumption whenever necessary. The degrees of freedom were corrected using the Greenhouse–Geisser procedure, and least significant difference analysis was used for post hoc comparisons (p < .05). The partial eta square was used to estimate the effect size, with values of .02, .12, and .26 suggesting relatively small, medium, and large effect sizes, respectively (Cohen, 1992). ## Results #### Putting Performance The mean distance of the SMR group in the pretest and posttest was 29.62 cm (5.59) and 16.59 cm (8.92), respectively. The control group distance was 20.17 cm (12.07) and 18.80 cm (5.58), respectively. An independent t test showed no difference in the mean distance in the pretest between two groups [t(14) = 2.008, p = .073, ----- η[2]p = .224]. The 2 (Group: SMR NFT, Control) × 2 (Test: pretest, posttest) mixed-model ANOVA revealed a significant interaction effect on putting performance [F(1, 14) = 5.029, p = .042, η[2]p = .264]. The SMR neurofeedback group exhibited a shorter distance from the hole in posttest than pretest [t(7) = 3.417, p = .011, η[2]p = .625]. No significant difference was observed for other comparisons. #### Putting Performance in Standard Deviation A marginal interaction effect was observed in the 2 (Group: SMR NFT, Control) × 2 (Test: pretest, posttest) ANOVA [F(1, 14) = 4.121, _p = .062,_ η[2]p = .227]. We did not observe an effect on Group factor [F(1, 14) = 0.136, p = .717, η[2]p = .010]. The SMR group exhibited a significantly lower SD in the posttest (16.11 cm) than in the pretest (24.70 cm) [t(7) = 4.408, p = .003, η[2]p = .735], whereas the control group showed no significant variation in SD (21.03 cm to 18.38 cm) [t(7) = 1.208, p = .266, η[2]p = .173]. #### SMR Relative Power The difference scores of the SMR group members for T1, T2, and T3 was 0.481 (0.588), 0.186 (0.378), and 0.040 (0.268), respectively. For the control group, the difference scores was –0.200 (0.424), –0.143 (0.440), and 0.009 (0.444), respectively. We compared the difference scores with a 2 (Group: SMR NFT, Control) × 3 [Time window: –1.5 to –1.0 s (T1), –1.0 to –0.5 s (T2), –0.5 to 0 s (T3)] two-way ANOVA and observed a marginally significant two-way interaction effect [F(2, 28) = 3.315, p = .051, η[2]p = .191]. To explore this marginal interaction effect and examine the training effect before and after NFT, a subsequent simple main effect analysis was performed and revealed a marginal Time effect [F(2, 14) = 3.470, _p = .060, η[2]p = .331] in the SMR NFT group. Post hoc_ analysis showed that the SMR power was significantly greater in T1 than in T3 [t(7) = 2.925, p = .022, η[2]p = .550]. No significant simple main effect was observed in the control group [F(2, 14) = .671, p = .567, η[2] = .141]. In addition, a simple main effect analysis revealed that the SMR NFT group exhibited a relatively higher SMR power than that of the control at T1 [t(14) = 2.657, p = .019, η[2]p = 335]. The significant group main effect revealed that the SMR NFT group had a higher SMR power than that of the control group [F(1, 14) = 4.665, p = .049, η[2]p = .250]. The difference scores between the two groups are depicted in Figure 1. #### Control Analyses **_Successful Training Ratio._** The overall mean of the golfers’ successful training ratio was 62.39 (8.88) % for the SMR training group and 22.27 (22.28) % for the control group. The 2 (Group: SMR NFT, Control) × 4 **Figure 1 — The difference scores of SMR relative power between the SMR NFT and control groups at T1 (–1.5 to –1.0 s), T2** (–1.0 to –0.5 s), and T3 (–0.5 to 0 s). ----- (Training section: Section 1: sessions 1–2; Section 2: sessions 3–4; Section 3: sessions 5–6; Section 4: sessions 7–8) ANOVA showed no interaction effect [F(3,42) = 0.694, p = .497, η[2]p = .047], but a significant group main effect was observed [F(1,14) = 22.188, p = .001, η[2]p = .613]. The SMR group showed a significantly higher percentage of successful training ratios than did the control group. Table 1 lists the successful training ratio for each group during the training sections. **_Within-Session Learning._** The results of NFT can be affected by day-to-day fluctuations in arousal level (Gruzelier et al., 2014). Thus, in addition to comparing the average successful training ratios of the eight sessions between these two groups, we compared the successful training ratios of the first and last trials of each session for all eight sessions between the two groups to determine whether participants in the NFT group improved within each training session. We hypothesized that the successful training ratio would be greater in the last trial than in the first trial for the SMR NFT group but not for control group. A 2 (Group: SMR NFT, Control) × 8 (Session: sessions 1, 2, 3, 4, 5, 6, 7, 8) × 2 (Trial: first trial, last trial) three-way ANOVA was employed to test this hypothesis. The result showed that although the 3-way interaction effect was not significant [F(7, 98) = 2.063, p = .082, η[2]p = .128], a 2 (Group: SMR NFT, Control) × 2 (Trial: first trial, last trial) interaction effect [F(1, 14) = 33.192, p = .001, η[2]p = .703] was revealed. Post hoc analysis was consistent with our prediction; only the SMR NFT group demonstrated a greater successful training ratio in the last trial (M = 77.65, SD = 7.84) than in the first trial (M = 50.58, SD = 10.65) for all sessions [t(7)= 8.344, p = .001, η[2]p = 909]. The control group did not show a significant difference between the first trial (M = 12.19, SD = 11.86) and last trial (M = 16.32, SD = 17.00) [t(7) = 1.784, p = .118, η[2]p = 313]. In addition, the SMR NFT group demonstrated a significantly higher training ratio on the first trial [t(7) = 6.810, p = .001, η[2]p = 768] and last trial [t(7) = 9.267, p = .001, η[2]p = .860] than did the control group (Figure 2). **_Threshold Increments Within SMR Training Sessions._** Although our control analyses provided supportive evidence for the learning progress made by the SMR NFT group, we further analyzed the change in threshold during each session of SMR NFT. In our study, threshold level was used as a difficulty index in the SMR NFT group, in which golfers were instructed to increase the SMR above designated level to meet our training demand. Thus, an improvement in the successful training ratio from the two previous control analysis was meaningful only when the threshold for each session was also examined. Previous studies evaluated the threshold variation within day-to-day sessions and suggested that the increased threshold could serve as a marker for improvement of the controllability due to neurofeedback training (Doppelmayr & Weber, 2011). Thus, we converted the eight training sessions into four sections as described in the methods section and examined the training threshold variation by employing an one-way ANOVA to examine the effect of Training section (Section 1: sessions 1–2; Section 2: sessions 3–4; Section 3: sessions 5–6; Section 4: sessions 7–8) in the SMR group. We hypothesized that the threshold value would increase after the first training section, which supports an improvement in controllability due to SMR neurofeedback training. The average training thresholds for sections one to four in the SMR NFT group were 5.862 (2.781), 7.636 (3.368), 8.214 (3.718), and 7.750 (3.816), respectively. As predicted, a significant difference was detected by the one-way ANOVA [F(3, 18) = 9.945, _p = .001,_ η[2]p = .624]. Post hoc analysis demonstrated that the training thresholds in the second, third, and fourth sections were significantly higher than that of the first section. **_Electrode Specificity._** Although the current study demonstrated that the relative SMR power of the SMR NFT group was significantly higher than that of the control group following SMR NFT, it remained unknown whether the greater 12–15 Hz EEG relative power after training was limited to the sensorimotor cortex or there was a spillover to other regions, such as the frontal, parietal and occipital cortices. Thus, we compared the difference scores at 12–15 Hz EEG relative power among Fz, Cz, Pz, and Oz between pre- and posttest sessions. Previous work has shown that the SMR originated in the centro-parietal region (Grosse-Wentrup, Schölkopf, & Hill, 2011). Thus, we hypothesized that the difference score of 12–15 Hz at Cz would be greater than that of the frontal and occipital regions for SMR group participants after training. A 2 (Group: SMR NFT, Control) × 4 (Region: Frontal, Central, Parietal, Occipital) two-way ANOVA between the two groups was performed to test this hypothesis. The difference scores at Fz, Cz, Pz, and Oz were 0.035 (0.200), 0.212 (0.178), 0.135 (0.298), and 0.003 (0.241), respectively, for the SMR NFT group. For the **Table 1 The Successful Training Ratios Between the SMR NFT** **and Control Groups Across the Four Training Sections** **(Every Two Consecutive Sessions Were Folded Resulting In Four Sections)** **Section 1** **Section 2** **Section 3** **Section 4** **Total** SMR 53.82 (19.71) 63.85 (12.53) 65.63 (9.52) 66.27 (17.91) 62.39 (5.08) Control 20.51 (24.11) 23.02 (26.31) 22.94 (21.58) 22.62 (19.61) 22.27 (1.09) _Note. The unit is the percentage of increasing time for successfully controlling SMR power._ ----- **Figure 2 — The mean successful training ratio for the first and last trial between the SMR NFT and control groups across the** eight training sessions. control group, the difference scores at Fz, Cz, Pz, and Oz were –0.056 (0.309), –0.438 (0.169), –0.150 (0.268), and –0.168 (0.640), respectively. This result yielded a marginally significant interaction effect [F(3, 42) = 2.680, p = .089, η[2]p = .161]. Because of the exploratory nature of this study, we conducted a follow-up analysis of this interaction effect. The independent _t tests of_ the four regions between the two groups showed that significance was only observed at a difference score of Cz [t(14) = 5.159, p = .001, η[2]p = 655], in which the SMR NFT group exhibited a significantly higher difference score than the control group. Moreover, oneway ANOVA of four regions in the SMR NFT group reached marginal significance [F(3, 21) = 2.644, p = .076, η[2]p = .274]. The follow-up pairwise t tests found that the difference score of Cz was higher than that of Fz [t(7) = 3.740, p = .007, η[2]p = 666] and Oz [t(7) = 2.530, _p = .039, η[2]p = .478]. These lines of evidence provide_ preliminary support for the electrode specificity of SMR NFT in this study. **_Frequency Specificity._** Previous studies have shown that neurofeedback training may generate concurrent changes in flanking frequency bands (Enriquez-Geppert et al., 2014). The aim of this analysis was to investigate whether SMR NFT resulted in a change in frequency bands close to SMR. We compared the relative power difference scores of theta (4–7 Hz), alpha (8–12 Hz), low beta (13–20 Hz), high beta (21–30 Hz), and broad beta (13–30 Hz) frequency bands before golf putting from pretest and posttest between the two groups. The 2 (Group: SMR NFT, Control) × 3 [Time window: –1.5 to –1.0 s (T1), –1.0 to –0.5 s (T2), –0.5 to 0 s (T3)] twoway ANOVA showed that neither interaction effects on theta power [F(2, 28) = 0.550, p = .583, η[2]p = .038], alpha power [F(2, 28) = 0.113, p = .802, η[2]p = .011], low beta power [F(2, 28) = 0.052, p = .949, η[2]p = .004], high beta power [F(2, 28) = 0.503, p = .496, η[2]p = .035], and broad beta band [F(2, 28) = 0.883, p = .425, η[2]p = .059] nor group main effects on theta power [F(1, 14) = 0.032, p = .860, η[2]p = .002], alpha power [F(1, 14) = 0.070, p = .795, η[2]p = .005], low beta power [F(1, 14) = 0.764, p = .397, η[2]p = .052], high beta power [F(1, 14) = 0.677, p = .424, η[2]p = .046], and broad beta power [F(1, 14) = 0.023, p = .881, η[2]p = .002] were observed. The difference scores among these five frequency bands are listed in Table 2. ## Discussion The aim of this study was to investigate the effect of SMR neurofeedback training on golf putting performance. Our results showed that golfers receiving SMR neurofeedback training demonstrated enhanced SMR activity during the final 1.5 s before golf putting, resulting in better putting performance compared with the control group. This finding lends preliminary support to the hypothesis that SMR NFT is effective for increasing SMR power, and leads to superior putting performance. ----- **Table 2 Difference Scores (%) of Relative Power for Theta, Alpha, Low Beta, High Beta,** **and Beta Frequency Bands in Three Time Windows Between the Two Groups, SMR and Control** **T1 (–1.5 to –1.0 s)** **T2 (–1.0 to –0.5 s)** **T3 (–0.5 to 0 s)** **Relative Power** **SMR** **Control** **SMR** **Control** **SMR** **Control** Theta .025 (.621) .338 (.493) –.234 (.172) –.186 (.528) .311 (1.071) .085 (.452) Alpha .006 (.134) .048 (.221) .052 (.177) .017 (.216) –.006 (.465) –.058 (.223) Low beta .035 (.258) .014 (.135) –.069 (.124) –.029 (.164) –.097 (.183) –.033 (.082) High beta .014 (.190) .015 (.109) –.046 (.085) .128 (.593) –.047 (.152) –.030 (.070) Beta .034 (.164) .053 (.010) –.064 (.094) –.029 (.077) –.050 (.137) –.082 (.112) Increased SMR power by NFT results in better visuomotor performance. For behavioral data, we observed that SMR neurofeedback training improved skilled golfers’ putting performance, as indicated by the reduced average distance from the hole and the variability of the score. No significant change in putting performance was observed in the control group. Previous studies have demonstrated that augmenting SMR by NFT improved visual motor performance (Ros et al., 2009) and increased self-rating scores of subjective flow state in dancers (Gruzelier et al., 2010). Furthermore, augmenting SMR by NFT was related to an improved attention-related mental state (Vernon et al., 2003) and memory performance (Hoedlmoser et al., 2008). In addition, converging lines of evidence support the effectiveness of NFT based on non-SMR variables enhancing performance in the sport domain (Arns et al., 2008; Gruzelier et al., 2010; Kao et al., 2014; Landers et al., 1991; Raymond, Sajid, Parkinson, & Gruzelier, 2005; Ring, Cooke, Kavussanu, McIntyre, & Masters, 2015). Nevertheless, the current study is the first, to our best knowledge, to use the SMR protocol to investigate the effectiveness of NFT on sport performance. Our results support the finding of the augmented SMR power which is linked with more adaptive fine-motor performance (Cheng et al., 2015) and extend the potential facilitation effects of SMR training to the sport domain. Less task-irrelevant interference of somatosensory and sensorimotor processing, as reflected in augmented SMR power after training, leads to improved putting performance. A previous study has indicated that participants in the automatic stage showed weaker activity in the presupplementary motor area, premotor cortex, parietal cortex, and prefrontal cortex compared with novices in a self-paced sequential finger movement task (Wu et al., 2008). A negative relationship between SMR power and sensorimotor activity has been suggested (Mann et al., 1996). The drop in sensorimotor activity, as reflected by increased SMR power, may indicate a greater adaptive task-related attention allocation that facilitates the execution of sport performance (Gruzelier et al., 2010). Increasing SMR power through NFT is also related to more efficient and modulated visuomotor performance (Gruzelier et al., 2010; Ros et al., 2009). These results suggest that augmenting SMR power led to an improved adjustment of somatosensory and sensorimotor pathways (Kober et al., 2014), which resulted in increased task-related attention toward specific tasks (Egner & Gruzelier, 2001). Moreover, previous studies have suggested that enhanced SMR power leads to a relatively higher flow state (Gruzelier et al., 2010) and calming mood (Gruzelier, 2014a). Based on the functional role of SMR, these findings imply that a reduction in sensorimotor activity may lessen the conscious processing involved in motor execution, which would lead to a more conceptual automatic process (Cheng et al., 2015). This interpretation is in line with converging evidence supporting a beneficial effect of augmented SMR on focusing and sustaining attention, working memory, and psychomotor skills (Egner & Gruzelier, 2001; Ros et al., 2009). Collectively, the superior golf putting performance observed in the present SMR NFT group might be the result of reduced somatosensory information processing before the back swing, which leads to refined golf putting performance. The interpretation that a reduction in conscious interference facilitates motor operation is in line with the concept of automatic processing proposed by Fitts and Posner (1967). However, given the relatively small sample size, future research should verify the causal relationship between augmented SMR power and finemotor performance. Reduced cortical activity in the sensorimotor area, as reflected by the higher power of 12–15 Hz, is sensitive to superior putting performance. First, the electrode specificity of SMR NFT was demonstrated. Although electrode specificity has been suggested to be an important step in support of the NFT training effect on the corresponding EEG component at a specific brain region (Gruzelier, 2014b), this is the first study in the area of NFT and sport performance to provide such preliminary evidence for the localized training effects. The lack of difference between Cz and Pz might suggest that this region is also part of a network associated with SMR activity in motor performance. This speculation is in line with the evidence that the parietal region is involved in processing visual-spatial information during motor performance (Del Percio et al., 2011). Second, frequency specificity was analyzed. One might argue that enhanced putting performance was caused by variation in another frequency band at the Cz ----- site, but this explanation is inconsistent with the lack of significant changes on difference scores in the theta, alpha, low beta and high beta frequency bands. These results suggest that it is primarily SMR power that accounts for the facilitating effect of SMR NFT on putting performance rather than other neighboring frequency bands. Our demonstration of electrode and frequency specificity strengthens the hypothesis that improved putting performance was the result of reduced sensorimotor activity before putting execution. The SMR NFT group improved the putting performance through the refined strategy for controlling the SMR power and reached the training goal as a result of the training program. First, our data showed that the SMR group demonstrated a higher successful training ratio than did the control group. Second, previous studies proposed that the training effect would emphasize daily training improvement (Gruzelier et al., 2014). In our control analysis, we compared the successful training ratio of the first and the last trial within eight sessions. A significantly higher successful training ratio for the last trial than for the first trial was observed, suggesting that golfers in the SMR NFT group learned the tuning strategy successfully after the initial trials and that the strategies were effective in the subsequent trials of the remaining sessions. This result lends support to the concept of neurofeedback trainability and further confirms the possibility of EEG tuning within a single training session (Kao et al., 2014; López-Larraz, Escolano, & Minguez, 2012). Furthermore, we found a significant threshold increase after the first session only in SMR NFT group, suggesting that our training protocol is facilitative to golfers. This evidence was in line with previous work in which the SMR amplitude increased above the daily adjusted threshold (Weber, Köberl, Frank, & Doppelmayr, 2011). We have several suggestions with regard to future neurofeedback studies. First, combining these studies with neuroimaging tools is necessary. Although we have provided evidence that the regulation of SMR power can enhance putting performance, this result would be benefit from the experiments conducted with high-spatialresolution neuroimaging tools, such as fMRI, to provide a more precise anatomical description of the NFT effect. Second, the phenomenological report of neurofeedback learning and its effects is often overlooked (Gruzelier, 2014b). A sophisticated measurement of subjective mental state, such as an in-depth questionnaire or scale, is needed to further elucidate the mental state associated with NFT (Gruzelier, 2014a). Third, the retention of learning driven by NFT must be examined. Thus far, this issue has received little attention, but it is critical from a practical viewpoint to determine how long the performance enhancement due to NFT lasts. Fourth, to explore the effect of SMR NFT on anticipative motor planning is needed. Future study should investigate the link between neurophysiological and cognitive processes by using the priming tests to further understand the neurocognitive architecture of golf performance. Last but not least, the changes in network dynamics after NFT should be further examined to fill the knowledge gap of cortical interaction caused by NFT. For example, the parietal and sensorimotor cortex networks are thought to be functionally relevant during motor performance (Baumeister et al., 2013). Our findings should be interpreted with caution due to the limitations of the study. First, the sample size was limited. Some of our statistical analyses reached only marginal significance, likely due to the small sample size. Furthermore, given the exploratory nature of the study, it is reasonable to speculate implications regarding the the marginally significant effects. Second, although the neurophysiological source of the SMR could not be precisely located due to limited spatial resolution by surface EEG, the finding of a marginally significant larger SMR difference score at the Cz site compared with the Fz and Oz sites as well as the finding that the largest magnitude of 12–15 Hz differences occurred at the Cz site rather than other frequency bands in the SMR group provide indirect evidence to support the impact of somatosensory activity on superior putting performance after SMR NFT. Third, putting is only one of many fundamental motor skills involved in golf performance. Our results may be difficult to generalize to other golf motor skills (e.g., the drive shot and tee shot). Future studies should, therefore, examine different skills involved in golf performance to determine the generalizability of the present findings. Fourth, the skill levels of the participants may impact the effect of NFT, and caution should be exercised when generalizing these findings to golfers at other skill levels. In conclusion, an eight-session SMR NFT exhibited a putting performance enhancement and increased SMR power in SMR NFT group compared with control group, suggesting that SMR NFT is an effective protocol for enhancing putting performance through fine-tuning somatosensory interference, as reflected by augmented SMR. **Acknowledgments** The work of Tsung-Min Hung was supported in part by the National Science Council (Taiwan) under grant NSC 98-2410H-003 -124 -MY2. ## References Arns, M., Kleinnijenhuis, M., Fallahpour, K., & Breteler, R. (2008). Golf Performance Enhancement and Real-Life Neurofeedback Training Using Personalized EventLocked EEG Profiles. _Journal of Neurotherapy,_ _11(4),_ [11–18. doi:10.1080/10874200802149656](http://dx.doi.org/10.1080/10874200802149656) Babiloni, C., Del Percio, C., Iacoboni, M., Infarinato, F., Lizio, R., Marzano, N., . . . Eusebi, F. (2008). Golf putt outcomes are predicted by sensorimotor cerebral EEG rhythms. _The Journal of Physiology,_ _[586(1), 131–139. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=17947315&dopt=Abstract)_ [doi:10.1113/jphysiol.2007.141630](http://dx.doi.org/10.1113/jphysiol.2007.141630) Baumeister, J., Von Detten, S., Van Niekerk, S.M., Schubert, M., Ageberg, E., & Louw, Q.A. (2013). Brain activity in predictive sensorimotor control for landings: An EEG pilot ----- study. International Journal of Sports Medicine, _34(12),_ [1106–1111. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=23740338&dopt=Abstract) [doi:10.1055/s-0033-1341437](http://dx.doi.org/10.1055/s-0033-1341437) Chatrian, G.E., Lettich, E., & Nelson, P.L. (1985). Ten percent electrode system for topographic studies of spontaneous and evoked EEG activity. The American Journal of EEG _Technology,_ _25, 83–92._ Cheng, M.Y., Hung, C.L., Huang, C.J., Chang, Y.K., Lo, L.C., Shen, C., & Hung, T.M. (2015). Expert-novice differences in SMR activity during dart throwing. Biological _Psychology,_ _[110, 212–218. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=26277266&dopt=Abstract)_ [doi:10.1016/j.bio-](http://dx.doi.org/10.1016/j.biopsycho.2015.08.003) [psycho.2015.08.003](http://dx.doi.org/10.1016/j.biopsycho.2015.08.003) Chuang, L.Y., Huang, C.J., & Hung, T.M. (2013). The differences in frontal midline theta power between successful and unsuccessful basketball free throws of elite basketball players. International Journal of Psychophysiology, _90(3),_ [321–328. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=24126125&dopt=Abstract) [doi:10.1016/j.ijpsycho.2013.10.002](http://dx.doi.org/10.1016/j.ijpsycho.2013.10.002) Cohen, J. (1992). A power primer. _Psychological Bul-_ _letin,_ _[112(1), 155–159. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=19565683&dopt=Abstract)_ [doi:10.1037/0033-](http://dx.doi.org/10.1037/0033-2909.112.1.155) [2909.112.1.155](http://dx.doi.org/10.1037/0033-2909.112.1.155) Davidson, R.J. (1988). EEG measures of cerebral asymmetry: conceptual and methodological issues. The International _Journal of Neuroscience,_ _[39(1-2), 71–89. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=3290140&dopt=Abstract)_ [doi:10.3109/00207458808985694](http://dx.doi.org/10.3109/00207458808985694) Del Percio, C., Iacoboni, M., Lizio, R., Marzano, N., Infarinato, F., Vecchio, F., . . . Babiloni, C. (2011). Functional coupling of parietal alpha rhythms is enhanced in athletes before visuomotor performance: a coherence electroencephalographic study. _Neuroscience,_ _[175, 198–211. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=21144884&dopt=Abstract)_ [doi:10.1016/j.neuroscience.2010.11.031](http://dx.doi.org/10.1016/j.neuroscience.2010.11.031) Doppelmayr, M., & Weber, E. (2011). Effects of SMR and theta/ beta neurofeedback on reaction times, spatial abilities, and creativity. Journal of Neurotherapy, _[15(2), 115–129. doi:](http://dx.doi.org/10.1080/10874208.2011.570689)_ [10.1080/10874208.2011.570689](http://dx.doi.org/10.1080/10874208.2011.570689) Egner, T., & Gruzelier, J. (2001). Learned self-regulation of EEG frequency components affects attention and eventrelated brain potentials in humans. Neuroreport, _12(18),_ [4155–4159. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=11742256&dopt=Abstract) [doi:10.1097/00001756-200112210-](http://dx.doi.org/10.1097/00001756-200112210-00058) [00058](http://dx.doi.org/10.1097/00001756-200112210-00058) Egner, T., Strawson, E., & Gruzelier, J. (2002). EEG signature and phenomenology of alpha/theta neurofeedback training versus mock feedback. _Applied Psycho-_ _physiology and Biofeedback,_ _[27(4), 261–270. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=12557453&dopt=Abstract)_ [doi:10.1023/A:1021063416558](http://dx.doi.org/10.1023/A:1021063416558) Egner, T., Zech, T.F., & Gruzelier, J. (2004). The effects of neurofeedback training on the spectral topography of the electroencephalogram. Clinical Neurophysiology, _115(11),_ [2452–2460. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=15465432&dopt=Abstract) [doi:10.1016/j.clinph.2004.05.033](http://dx.doi.org/10.1016/j.clinph.2004.05.033) Enriquez-Geppert, S., Huster, R.J., Scharfenort, R., Mokom, Z.N., Zimmermann, J., & Herrmann, C.S. (2014). Modulation of frontal-midline theta by neurofeedback. Biological _Psychology,_ _[95(1), 59–69. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=23499994&dopt=Abstract)_ [doi:10.1016/j.biopsy-](http://dx.doi.org/10.1016/j.biopsycho.2013.02.019) [cho.2013.02.019](http://dx.doi.org/10.1016/j.biopsycho.2013.02.019) Fitts, P.M., & Posner, M.I. (1967). Human Performance. Basic _[Concepts in Psychology. Retrieved from http://lccn.loc.](http://lccn.loc.gov/67011662)_ [gov/67011662](http://lccn.loc.gov/67011662) Grosse-Wentrup, M., Schölkopf, B., & Hill, J. (2011). Causal influence of gamma oscillations on the sensorimotor rhythm. _NeuroImage,_ _[56(2), 837–842. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=20451626&dopt=Abstract)_ [doi:10.1016/j.neuroimage.2010.04.265](http://dx.doi.org/10.1016/j.neuroimage.2010.04.265) Gruzelier, J. (2014a). Differential effects on mood of 12–15 (SMR) and 15–18 (beta1) Hz neurofeedback. International _Journal of Psychophysiology,_ _[93(1), 112–115. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=23357178&dopt=Abstract)_ [doi:10.1016/j.ijpsycho.2012.11.007](http://dx.doi.org/10.1016/j.ijpsycho.2012.11.007) Gruzelier, J. (2014b). EEG-neurofeedback for optimising performance. III: A review of methodological and theoretical considerations. _Neuroscience and Biobehavioral_ _Reviews,_ _[44, 159–182. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=24690579&dopt=Abstract)_ [doi:10.1016/j.neubio-](http://dx.doi.org/10.1016/j.neubiorev.2014.03.015) [rev.2014.03.015](http://dx.doi.org/10.1016/j.neubiorev.2014.03.015) Gruzelier, J., Foks, M., Steffert, T., Chen, M.J.L., & Ros, T. (2014). Beneficial outcome from EEG-neurofeedback on creative music performance, attention and well-being in school children. Biological Psychology, _95(1), 86–95._ [PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=23623825&dopt=Abstract) [doi:10.1016/j.biopsycho.2013.04.005](http://dx.doi.org/10.1016/j.biopsycho.2013.04.005) Gruzelier, J., Inoue, A., Smart, R., Steed, A., & Steffert, T. (2010). Acting performance and flow state enhanced with sensory-motor rhythm neurofeedback comparing ecologically valid immersive VR and training screen scenarios. Neuroscience Letters, _[480(2), 112–116. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=20542087&dopt=Abstract)_ [doi:10.1016/j.neulet.2010.06.019](http://dx.doi.org/10.1016/j.neulet.2010.06.019) Hoedlmoser, K., Pecherstorfer, T., Gruber, G., Anderer, P., Doppelmayr, M., Klimesch, W., & Schabus, M. (2008). Instrumental conditioning of human sensorimotor rhythm (12-15 Hz) and its impact on sleep as well as declarative learning. Sleep, _[31(10), 1401–1408. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=18853937&dopt=Abstract)_ Kao, S.C., Huang, C.J., & Hung, T.M. (2013). Frontal midline theta is a specific indicator of optimal attentional engagement during skilled putting performance. _Journal of_ _Sport & Exercise Psychology,_ _35(5), 470–478 Retrieved_ from [http://www.ncbi.nlm.nih.gov/pubmed/24197715.](http://www.ncbi.nlm.nih.gov/pubmed/24197715) [PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=24197715&dopt=Abstract) Kao, S.C., Huang, C.J., & Hung, T.M. (2014). Neurofeedback training reduces frontal midline theta and improves putting performance in expert golfers. _Journal of Applied_ _Sport Psychology,_ _26(3), 271–286._ [doi:10.1080/104132](http://dx.doi.org/10.1080/10413200.2013.855682) [00.2013.855682](http://dx.doi.org/10.1080/10413200.2013.855682) Klimesch, W., Doppelmayr, M., Pachinger, T., & Ripper, B. (1997). Brain oscillations and human memory: EEG correlates in the upper alpha and theta band. Neuroscience _Letters,_ _[238(1-2), 9–12. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=9464642&dopt=Abstract)_ [doi:10.1016/S0304-](http://dx.doi.org/10.1016/S0304-3940(97)00771-4) [3940(97)00771-4](http://dx.doi.org/10.1016/S0304-3940(97)00771-4) Kober, S.E., Witte, M., Stangl, M., Väljamäe, A., Neuper, C., & Wood, G. (2014). Shutting down sensorimotor interference unblocks the networks for stimulus processing: An SMR [neurofeedback training study. Clinical Neurophysiology.](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=24794517&dopt=Abstract) [PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=24794517&dopt=Abstract) Krane, V., & Williams, J.M. (2006). Psychological characteristics of peak performance. In J.M. Williams (Ed.), Applied _sport psychology: Personal growth to peak performance._ New York: McGraw-Hill. Landers, D.M., Petruzzello, S.J., Salazar, W., Crews, D.J., Kubitz, K.A., Gannon, T.L., & Han, M. (1991). The influence of electrocortical biofeedback on performance in pre-elite archers. Medicine and Science in Sports and Exer_cise,_ _[23(1), 123–129. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=1997806&dopt=Abstract)_ [doi:10.1249/00005768-](http://dx.doi.org/10.1249/00005768-199101000-00018) [199101000-00018](http://dx.doi.org/10.1249/00005768-199101000-00018) Logan, G.D., Hockley, W.E., & Lewandowsky, S. (1991). Automaticity and memory. In W.E. Hockley & S. Lewandowsky (Eds.), _Relating Theory and Data: Essays on Human_ ----- _Memory in Honor of Bennet B. Murdock (pp. 347–366)._ East Sussex, UK: Psychology Press. López-Larraz, E., Escolano, C., & Minguez, J. (2012). Upper alpha neurofeedback training over the motor cortex increases SMR desynchronization in motor tasks. _Pro-_ _ceedings of the Annual International Conference of the_ _IEEE Engineering in Medicine and Biology Society,_ _EMBS, 2012, 4635–4638._ [http://doi.org/doi:10.1109/](http://doi.org/) [EMBC.2012.6347000](http://dx.doi.org/10.1109/EMBC.2012.6347000) Mann, C.A., Sterman, M.B., & Kaiser, D.A. (1996). Suppression of EEG rhythmic frequencies during somato-motor and visuo-motor behavior. International Journal of Psy_chophysiology,_ _[23(1-2), 1–7. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=8880361&dopt=Abstract)_ [doi:10.1016/0167-](http://dx.doi.org/10.1016/0167-8760(96)00036-0) [8760(96)00036-0](http://dx.doi.org/10.1016/0167-8760(96)00036-0) Nicholls, A.R. (2007). A longitudinal phenomenological analysis of coping effectiveness among Scottish international adolescent golfers. _European Journal of Sport Science,_ _[7(3), 169–178. doi:10.1080/17461390701643034](http://dx.doi.org/10.1080/17461390701643034)_ Niemarkt, H.J., Jennekens, W., Pasman, J.W., Katgert, T., Van Pul, C., Gavilanes, A.W.D., . . . Andriessen, P. (2011). Maturational changes in automated EEG spectral power analysis in preterm infants. _Pediatric Research,_ _70(5),_ [529–534. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=21772227&dopt=Abstract) [doi:10.1203/PDR.0b013e31822d748b](http://dx.doi.org/10.1203/PDR.0b013e31822d748b) Nierhaus, T., Forschack, N., Piper, S.K., Holtze, S., Krause, T., Taskin, B., . . . Villringer, A. (2015). Imperceptible somatosensory stimulation alters sensorimotor background rhythm and connectivity. _The Journal of Neuroscience,_ _[35(15), 5917–5925. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=25878264&dopt=Abstract)_ [doi:10.1523/JNEURO-](http://dx.doi.org/10.1523/JNEUROSCI.3806-14.2015) [SCI.3806-14.2015](http://dx.doi.org/10.1523/JNEUROSCI.3806-14.2015) Pelz, D., & Frank, J.A. (2000). Dave Pelz’s putting bible: The _complete guide to mastering the green. New York, NY:_ Doubleday. Raymond, J., Sajid, I., Parkinson, L.A., & Gruzelier, J. (2005). Biofeedback and dance performance: A preliminary investigation. Applied Psychophysiology and Biofeedback, _[30(1), 65–73. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=15889586&dopt=Abstract)_ [doi:10.1007/s10484-005-2175-x](http://dx.doi.org/10.1007/s10484-005-2175-x) Ring, C., Cooke, A., Kavussanu, M., McIntyre, D., & Masters, R. (2015). Investigating the efficacy of neurofeedback training for expediting expertise and excellence in sport. _Psychology of Sport and Exercise,_ _16, 118–127._ [doi:10.1016/j.psychsport.2014.08.005](http://dx.doi.org/10.1016/j.psychsport.2014.08.005) Ros, T., Moseley, M.J., Bloom, P.A., Benjamin, L., Parkinson, L.A., & Gruzelier, J. (2009). Optimizing microsurgical skills with EEG neurofeedback. BMC Neuroscience, _10(1),_ [87. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=19630948&dopt=Abstract) [doi:10.1186/1471-2202-10-87](http://dx.doi.org/10.1186/1471-2202-10-87) Sauseng, P., Hoppe, J., Klimesch, W., Gerloff, C., & Hummel, F.C. (2007). Dissociation of sustained attention from central executive functions: Local activity and interregional connectivity in the theta range. _The European Journal_ _of Neuroscience,_ _[25(2), 587–593. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=17284201&dopt=Abstract)_ [doi:10.1111/](http://dx.doi.org/10.1111/j.1460-9568.2006.05286.x) [j.1460-9568.2006.05286.x](http://dx.doi.org/10.1111/j.1460-9568.2006.05286.x) Schneider, W., & Shiffrin, R.M. (1977). Controlled and automatic human information processing: I. Detection, search, and attention. Psychological Review, _84(1), 1–66._ [doi:10.1037/0033-295X.84.1.1](http://dx.doi.org/10.1037/0033-295X.84.1.1) Semlitsch, H.V., Anderer, P., Schuster, P., & Presslich, O. (1986). A solution for reliable and valid reduction of ocular artifacts, applied to the P300 ERP. _Psychophysiology,_ _[23(6), 695–703. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=3823345&dopt=Abstract)_ [doi:10.1111/j.1469-8986.1986.](http://dx.doi.org/10.1111/j.1469-8986.1986.tb00696.x) [tb00696.x](http://dx.doi.org/10.1111/j.1469-8986.1986.tb00696.x) Vernon, D., Egner, T., Cooper, N., Compton, T., Neilands, C., Sheri, A., & Gruzelier, J. (2003). The effect of training distinct neurofeedback protocols on aspects of cognitive performance. International Journal of Psycho_physiology,_ _[47(1), 75–85. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=12543448&dopt=Abstract)_ [doi:10.1016/S0167-](http://dx.doi.org/10.1016/S0167-8760(02)00091-0) [8760(02)00091-0](http://dx.doi.org/10.1016/S0167-8760(02)00091-0) Weber, E., Köberl, A., Frank, S., & Doppelmayr, M. (2011). Predicting successful learning of SMR neurofeedback in healthy participants: Methodological considerations. _Applied Psychophysiology and Biofeedback,_ _36(1), 37–45._ [PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=21053066&dopt=Abstract) [doi:10.1007/s10484-010-9142-x](http://dx.doi.org/10.1007/s10484-010-9142-x) Wu, T., Chan, P., & Hallett, M. (2008). Modifications of the interactions in the motor networks when a movement becomes automatic. The Journal of Physiology, _586(Pt 17),_ [4295–4304. PubMed](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=18617569&dopt=Abstract) [doi:10.1113/jphysiol.2008.153445](http://dx.doi.org/10.1113/jphysiol.2008.153445) _Manuscript submitted: June 30, 2015_ _Revision accepted: October 08, 2015_ -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1123/jsep.2015-0166?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1123/jsep.2015-0166, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "other-oa", "status": "GREEN", "url": "https://pub.uni-bielefeld.de/download/2901056/2901057/Cheng%20et%20al.%20-%202015%20-%20Sensorimotor%20rhythm%20neurofeedback%20enhances%20golf%20putting%20performance.pdf" }
2,015
[ "JournalArticle" ]
true
2015-12-01T00:00:00
[]
18,046
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Law", "source": "s2-fos-model" }, { "category": "Economics", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/013973151bd1db3552ea8d8278c6dece1bbdd352
[]
0.934056
How to issue a privacy-preserving central bank digital currency
013973151bd1db3552ea8d8278c6dece1bbdd352
Social Science Research Network
[ { "authorId": "3170736", "name": "Christian Grothoff" }, { "authorId": "145140500", "name": "T. Moser" } ]
{ "alternate_issns": null, "alternate_names": [ "SSRN, Social Science Research Network (SSRN) home page", "SSRN Electronic Journal", "Soc Sci Res Netw", "SSRN", "SSRN Home Page", "SSRN Electron J", "Social Science Electronic Publishing presents Social Science Research Network" ], "alternate_urls": [ "www.ssrn.com/", "https://fatcat.wiki/container/tol7woxlqjeg5bmzadeg6qrg3e", "https://www.wikidata.org/wiki/Q53949192", "www.ssrn.com/en", "http://www.ssrn.com/en/", "http://umlib.nl/ssrn", "umlib.nl/ssrn" ], "id": "75d7a8c1-d871-42db-a8e4-7cf5146fdb62", "issn": "1556-5068", "name": "Social Science Research Network", "type": "journal", "url": "http://www.ssrn.com/" }
JEL
### SUERF Policy Briefs ### No 114, June 2021 # How to issue a privacy-preserving central bank digital currency ## By Christian Grothoff and Thomas Moser[1] _JEL codes: E42, E51, E52, E58, G2._ _Keywords: Central Bank Digital Currency, privacy, blind signatures._ _Many central banks are currently investigating Central Bank Digital Currency (CBDC) and possible designs. A_ _recent survey conducted by the European Central Bank has found that both citizens and professionals consider_ _privacy the most important feature of a CBDC. We show how a central bank could issue a CBDC that would be_ _easily scalable and allow the preservation of a key feature of physical cash: transaction privacy. At the same_ _time the proposed design would meet regulatory requirements and thus offer an appropriate balance between_ _privacy and legal compliance._ ##### Introduction A central bank digital currency (CBDC) for the general public would be a new type of money issued by central banks, alongside banknotes and reserve accounts for selected financial market participants. Despite initial skepticism, the number of central banks investigating CBDC has grown steadily over the past three years. However, there is currently no consensus on how a CBDC should be designed and what features it should have. These questions are being intensively debated and researched. 1 Christian Grothoff, Bern University of Applied Sciences and Taler Systems SA. Thomas Moser, Swiss National Bank. ----- A recent survey conducted by the European Central Bank has found that both citizens and professionals consider privacy the most important feature of a digital Euro (ECB 2021). This may be surprising, but the fact that citizens place a high value on privacy is a consistent finding of many surveys. Skeptics sometimes counter that citizens express just the opposite in their behavior; they consistently choose convenience, speed, and financial savings over privacy. However, in doing so they are often not fully aware of the extent to which technological advances have improved the ability to track, aggregate, and disseminate personal information. They also often do not expect their data to be shared and used in a context other than the one in which they disclose the data (Nissenbaum 2010). Over the past decade, the public has become increasingly aware and concerned about the vast scale of data collected and stored by governments and corporations. Goldfarb and Tucker (2012) provide behavior-based evidence of increasing consumer privacy concerns. Payments data are particularly revealing, and a CBDC could potentially provide a great deal of data on citizens, making them vulnerable to malicious use for political or commercial purposes. We thus believe that a successful CBDC would need to provide credible transaction protections in order to gain broad public acceptance. Moreover, privacy is not just an individual value, it also has a social value. Privacy is essential for a free society and democracy. At the same time, a CBDC should not provide protection for illegal transactions and tax evasion. A privacypreserving CBDC must ensure legal compliance, particularly compliance with anti-money laundering (AML) and combating the financing of terrorism (CFT) regimes. It is thus crucial to find the right balance between privacy [and legal compliance. We believe that our proposal recently published as a Swiss National Bank Working Paper](https://www.snb.ch/n/mmr/reference/working_paper_2021_03/source/working_paper_2021_03.n.pdf) promises to do just that (Chaum et al. 2021). It builds on and improves the eCash technology (Chaum, 1983, and Chaum et al. 1990) and uses GNU Taler (Dold, 2019). Taler is part of the GNU project, which is a collaborative project for the development of “Free/Libre and Open Source Software” (FLOSS).[2] With FLOSS, all interested parties have access to the source code and the right to tailor the software to their needs. The patent-free, open standard protocol improves interoperability and competition among service providers. Instead of using proprietary secrets or hardware security modules, Taler exclusively uses cryptographic software with public specifications to provide privacy and security. ##### Privacy in payments: accounts versus tokens Payment systems can be account-based or token-based. In an account-based system, a payment is made by debiting the payer's account and crediting the payee's account. This implies that the transaction must be recorded and involved parties identified. In a token-based system, a payment is made by transferring a token that represents monetary value. The prime example is cash – coins or banknotes. Paying with cash means handing over a coin or banknote. There is no need to record the transfer or identify the parties involved, possession of the token is sufficient. However, the payee must be able to verify the token’s authenticity. It has been suggested that the distinction between account- and token-based systems is not applicable to digital currencies (Garratt et al. 2020). We believe that the distinction is also useful for digital currencies. The critical distinction is the information carried by the information asset. In an account-based system, the assets (accounts) are associated with transaction histories that include all of the credit and debit operations involving the accounts. In a token-based system, the assets (tokens) carry information about their value and the entity that issued the 2 For more information about GNU, see [https://www.gnu.org. GNU Taler is released free of charge under the GNU](https://www.gnu.org) Affero General Public License by the GNU Project. GNU projects popular among economists are the software packages «R» and “GNU Regression, Econometrics and Time-series Library” (GRETL). ----- token. The only possibility of attaining the transaction privacy property of cash, therefore, lies in token-based systems. We propose a token-based, software-only CBDC, where the CBDC token is issued and distributed just like banknotes. Consequently, we will simply refer to these CBDC tokens as “coins.” Customers withdraw coins by withdrawing money from their bank account; that is, they load coins onto their smartphone or computer and their bank debits their account for the corresponding amount. The proposed CBDC is genuine digital bearer instrument; it is stored locally on the computer or smartphone; there is no account or ledger involved. There is also no record linking the CBDC and the owner. Privacy is achieved with a cryptographic technique called blind signatures. Before the user interacts with the central bank to obtain a digitally signed coin, a blinding operation performed locally on the user's device hides the numeric value representing a coin from the central bank before requesting the signature. In GNU Taler, this numeric value is a public key, with the associated private key only being known to the owner of the coin. The coin derives its value from the central bank’s signature on the coin’s public key. The central bank makes the signature with its own private key. A merchant or payee can use the central bank’s corresponding “public key” to verify the central bank’s signature and thereby the authenticity of the CBDC. Because the blind signatures are carried out under the control of the users themselves, users do not have to trust the central bank or the commercial bank to safeguard their private spending history. The central bank only learns the total amounts of digital cash withdrawn and the total amount spent. Commercial banks learn how much digital cash their customers withdraw, but not how much an individual customer has spent or where they are spending it. Privacy in this design is thus not a question of confidentiality; it is cryptographically guaranteed. ##### The benefits of NOT using Distributed Ledger Technology (DLT) for CBDC Most central banks experiment with distributed ledger technology (DLT). DLT is an interesting design if no central party is available or desired: the purpose of a blockchain or DLT is to establish an immutable consensus across multiple parties. However, this is not required in the case for a retail CBDC issued by a trusted central bank. Distributing the central bank’s ledger merely increases transaction costs; it does not provide tangible benefits in a central bank deployment. A critical benefit of not using DLT is improved scalability. Our proposed scheme would be easily scalable and as cost-effective as modern RTGS systems currently used by central banks. GNU Taler can easily handle tens of thousands of transactions per second. The main cost of the system would be the secure storage of 1-10 kilobytes per transaction. Using Amazon Web Services pricing, experiments with an earlier prototype of GNU Taler showed that the cost of the system (memory, bandwidth, and computation) at scale would be less than USD 0.0001 per transaction. Furthermore, achieving privacy with DLT is a challenge, because DLT is essentially an account-based system. The only difference to a traditional account-based system is that the accounts are not kept in a central database but in a decentralized append-only database. Cryptographic privacy-enhancing technologies such as zero-knowledge proofs are possible but computationally demanding in a DLT-context, so that the high resource requirements make their use on mobile devices impractical. This does not apply to the Chaum-style blind signature protocol used in GNU Taler, which can be executed efficiently and quickly. ----- ##### How to prevent double-spending in a token-based system Money has value only when it is scarce, which means, among other things, that double-spending of a monetary asset is prevented. In a token-based system, one way to prevent double spending is to make it difficult to copy the token. This is the approach that central banks take with banknotes. With digital currencies, however, preventing copying is a challenge. Two potential technologies to prevent digital copying are unclonable functions and secure zones in hardware. However, physically unclonable functions cannot be exchanged over the Internet (eliminating the main use case of CBDCs), and security features in copy-prevention hardware have been repeatedly compromised. Our proposal, which consists only of software, does not even attempt to prevent token copying. Rather, double spending is prevented by the fact that each coin can be spent exactly once only. Once a coin has been spent, the number of the corresponding coin goes on a list of spent coins managed by the central bank. This list contains only the number of the spent coin but no transaction history. The coins also cannot be linked to the payers because they blinded the coins when the CBDC was withdrawn. When a payee receives a coin, the payee consults this list to see if the coin has already been spent before. If it has, the payment is rejected as invalid. Because our proposal requires online checks to prevent double-spending it does not enable offline payments. While this could be considered a disadvantage, Grothoff and Dold (2021) point out that any offline payment system has inherent and severe risks and thus its own drawbacks. Given that central banks do not intend to replace physical cash with CBDC, but rather to issue CBDC in addition to physical cash, physical cash can be used as the secure offline fallback in the event of power outages or cyber attacks. ##### Regulatory and policy consideration In the proposed scheme, central banks do not learn the identities of consumers or merchants or transaction amounts. Central banks only see when electronic coins are withdrawn and when they are redeemed. Commercial banks continue to provide crucial customer and merchant authentication and, in particular, remain the guardians of know-your customer information. Commercial banks observe when merchants receive funds and can limit the amount of CBDC per transaction that an individual merchant may receive, if required. Additionally, transactions are associated with the relevant customer contracts. The resulting income transparency enables the system to be compliant with the AML/CFT regulations. The proposed scheme thus offers one-sided privacy, allowing the buyer to remain anonymous while making the seller's incoming transactions and underlying contractual obligations available upon request by competent authorities. If unusual patterns of merchant income are detected, the commercial bank, tax authorities, or law enforcement can obtain and inspect the business contracts underlying the payments to determine whether the suspicious activity is nefarious. Overall, the system implements privacy-by-design and privacy-by default approaches (as required by the EU’s General Data Protection Regulation). Merchants do not inherently learn the identity of their customers, banks have only necessary insights into their own customers’ activities, and central banks are blissfully divorced from detailed knowledge of citizens’ activities. A potential financial stability concern often raised with retail CBDCs is banking sector disintermediation. While this would be a serious concern with an account-based CBDC, it should be less of a concern with a token-based CBDC. Hoarding a token-based CBDC entails risks of theft or loss similar to those of hoarding cash. However, should hoarding or massive conversions of money from bank deposits to CBDC become a problem, the proposed ----- design would give central banks several options, including imposing per-account withdrawal limits or negative interest rates. Imposing limits could also be a requirement of the AML/CFT regime. While GNU Taler by design allows its users to transact any amount in any currency, legislation could impose an enforceable ceiling on individual transactions, requiring merchants that receive transactions exceeding the transaction limit to determine the identity of the buyer. However, since there are no accounts, it would not be possible to impose holding limits. But this is a good thing. Technologically enforced restrictions on holding or receiving CBDC should be avoided anyway, as such restrictions would result in failures where users are unable to execute transactions despite sufficient liquidity. With the proposed design, central banks, commerce and citizens could reap the full benefits of the digital economy. The efficiency and cost effectiveness, along with the improved consumer usability that comes from shifting from authentication to authorization, make this system likely the first to support the long-envisioned goal of online micropayments. In addition, the use of coins to cryptographically sign electronic contracts would enable the use of smart contracts. This could lead to the emergence of entirely new applications for payment systems. A recently designed extension for GNU Taler integrates privacy-preserving age verification that allows legal guardians to impose age restrictions on digital purchases made with coins given to wards. Merchants would only learn that the buyer meets the age requirement for the goods sold, while the identity and exact age of the child would remain private. This is just one example of how central banks could use this system to issue programmable money. ∎ ----- ##### About the authors **_Christian Grothoff_** _is a Professor for Computer Network Security at the Bern University of Applied Sciences,_ _researching future Internet architectures. His research interests include compilers, programming languages,_ _software engineering, networking, security and privacy. Before, he was leading the Décentralisé research team at_ _INRIA and an Emmy Noeter research group leader at TU Munich. He earned his PhD in computer science from UCLA,_ _an M.S. in computer science from Purdue University, and a Diplom in mathematics from the University of Wuppertal._ _He is an Ashoka Fellow and co-founder of Taler Systems SA and Anastasis SARL. He also served as an expert court_ _witness, and has reported on technology and national security as a freelance journalist._ **_Thomas Moser_** _is an Alternate Member of the Governing Board of the Swiss National Bank. Before joining the Swiss_ _National Bank, he was an Executive Director at the International Monetary Fund (IMF), and earlier in his career an_ _Economist at the Swiss Institute for Business Cycle Research (KOF) at the Swiss Federal Institute of Technology_ _(ETH), Zurich. Thomas Moser is also a member of the Managing Committee of the Swiss Institute of Banking and_ _Finance at the University of St. Gallen, a Member of the Board of Directors of Orell Füssli Ltd., and a Member of the_ _Advisory Board of the NZZ Swiss International Finance Forum. Thomas Moser holds a Master and a Doctorate in_ _Economics from the University of Zurich._ **SUERF is a network association of** central bankers and regulators, academics, and practitioners in the financial sector. The focus of the association is on the analysis, discussion and understanding of financial markets and institutions, the monetary economy, the conduct of regulation, supervision and monetary policy. SUERF’s events and publications provide a unique European network for the analysis and discussion of these and related issues. **SUERF Policy Briefs (SPBs)** serve to promote SUERF Members' economic views and research findings as well as economic policy-oriented analyses. They address topical issues and propose solutions to current economic and financial challenges. SPBs serve to increase the international visibility of SUERF Members' analyses and research. The views expressed are those of the author(s) and not necessarily those of the institution(s) the author(s) is/are affiliated with. All rights reserved. **Editorial Board** Ernest Gnan Frank Lierman David T. Llewellyn Donato Masciandaro Natacha Valla SUERF Secretariat c/o OeNB Otto-Wagner-Platz 3 A-1090 Vienna, Austria Phone: +43-1-40420-7206 www.suerf.org • suerf@oenb.at -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.2139/ssrn.3965050?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.2139/ssrn.3965050, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://arbor.bfh.ch/15002/1/suerf.pdf" }
2,021
[]
true
null
[]
3,944
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01398bbd4ecc589fecef134641d1016fdc1c044c
[ "Computer Science", "Medicine" ]
0.83742
Design and Implementation of High-Performance ECC Processor with Unified Point Addition on Twisted Edwards Curve
01398bbd4ecc589fecef134641d1016fdc1c044c
Italian National Conference on Sensors
[ { "authorId": "46977722", "name": "Md. Mainul Islam" }, { "authorId": "144778068", "name": "Md. Selim Hossain" }, { "authorId": "82670319", "name": "Moh. Khalid Hasan" }, { "authorId": "153087089", "name": "M. Shahjalal" }, { "authorId": "1816029", "name": "Y. Jang" } ]
{ "alternate_issns": null, "alternate_names": [ "SENSORS", "IEEE Sens", "Ital National Conf Sens", "IEEE Sensors", "Sensors" ], "alternate_urls": [ "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-142001", "http://www.mdpi.com/journal/sensors", "https://www.mdpi.com/journal/sensors" ], "id": "3dbf084c-ef47-4b74-9919-047b40704538", "issn": "1424-8220", "name": "Italian National Conference on Sensors", "type": "conference", "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-142001" }
With the swift evolution of wireless technologies, the demand for the Internet of Things (IoT) security is rising immensely. Elliptic curve cryptography (ECC) provides an attractive solution to fulfill this demand. In recent years, Edwards curves have gained widespread acceptance in digital signatures and ECC due to their faster group operations and higher resistance against side-channel attacks (SCAs) than that of the Weierstrass form of elliptic curves. In this paper, we propose a high-speed, low-area, simple power analysis (SPA)-resistant field-programmable gate array (FPGA) implementation of ECC processor with unified point addition on a twisted Edwards curve, namely Edwards25519. Efficient hardware architectures for modular multiplication, modular inversion, unified point addition, and elliptic curve point multiplication (ECPM) are proposed. To reduce the computational complexity of ECPM, the ECPM scheme is designed in projective coordinates instead of affine coordinates. The proposed ECC processor performs 256-bit point multiplication over a prime field in 198,715 clock cycles and takes 1.9 ms with a throughput of 134.5 kbps, occupying only 6543 slices on Xilinx Virtex-7 FPGA platform. It supports high-speed public-key generation using fewer hardware resources without compromising the security level, which is a challenging requirement for IoT security.
# sensors _Article_ ## Design and Implementation of High-Performance ECC Processor with Unified Point Addition on Twisted Edwards Curve **Md. Mainul Islam** **[1]** **, Md. Selim Hossain** **[2]** **, Moh. Khalid Hasan** **[1]** **, Md. Shahjalal** **[1]** **and Yeong Min Jang** **[1,]*** 1 Department of Electronics Engineering, Kookmin University, Seoul 02707, Korea; mainul.islam@ieee.org (M.M.I.); khalidrahman45@ieee.org (M.K.H.); mdshahjalal26@ieee.org (M.S.) 2 Department of Electrical and Electronic Engineering, Khulna University of Engineering & Technology (KUET), Khulna 9203, Bangladesh; selim@eee.kuet.ac.bd ***** Correspondence: yjang@kookmin.ac.kr; Tel.:+82-02-910-5068 Received: 23 August 2020; Accepted: 4 September 2020; Published: 10 September 2020 [����������](http://www.mdpi.com/1424-8220/20/18/5148?type=check_update&version=1) **�������** **Abstract: With the swift evolution of wireless technologies, the demand for the Internet of Things** (IoT) security is rising immensely. Elliptic curve cryptography (ECC) provides an attractive solution to fulfill this demand. In recent years, Edwards curves have gained widespread acceptance in digital signatures and ECC due to their faster group operations and higher resistance against side-channel attacks (SCAs) than that of the Weierstrass form of elliptic curves. In this paper, we propose a high-speed, low-area, simple power analysis (SPA)-resistant field-programmable gate array (FPGA) implementation of ECC processor with unified point addition on a twisted Edwards curve, namely Edwards25519. Efficient hardware architectures for modular multiplication, modular inversion, unified point addition, and elliptic curve point multiplication (ECPM) are proposed. To reduce the computational complexity of ECPM, the ECPM scheme is designed in projective coordinates instead of affine coordinates. The proposed ECC processor performs 256-bit point multiplication over a prime field in 198,715 clock cycles and takes 1.9 ms with a throughput of 134.5 kbps, occupying only 6543 slices on Xilinx Virtex-7 FPGA platform. It supports high-speed public-key generation using fewer hardware resources without compromising the security level, which is a challenging requirement for IoT security. **Keywords: elliptic curve cryptography (ECC); elliptic curve point multiplication (ECPM); twisted** Edwards curve; unified point addition; simple power analysis (SPA) attacks **1. Introduction** The Internet of Things (IoT) refers a global network, where billions of devices are connected through the Internet and share data with each other. Since most of these devices have constrained resources, data are usually stored in the cloud, where people can continuously upload and download data from anywhere via the Internet [1]. Security concerns arise as data owners have no control over the data management in the cloud-computing environment. The importance of data security and the limited resources of IoT devices motivate us to install lightweight cryptographic schemes that can satisfy the security, low-energy, and low-memory requirements of the existing IoT applications. Elliptic curve cryptography (ECC), a public-key cryptography (PKC), has become a promising approach to the IoT security, smart card security, and digital signatures as it provides high levels of security with smaller key sizes. Compared with traditional Rivest–Shamir–Adleman (RSA) algorithm, ECC provides an equal level of security but with a shorter key length [2–4]. ECC can be implemented with low hardware resource usage and low energy consumption without degrading its security ----- _Sensors 2020, 20, 5148_ 2 of 19 level. Owing to low hardware use, it is well suited for the security of low-power, low-memory, and resource-constrained IoT devices. ECC implemented in a small chip can provide high-speed data encryption and decryption facilities. In addition, it prevents unauthorized devices from gaining access to wireless sensor networks (WSNs) by providing a key agreement protocol for the wireless sensor nodes connected to the IoT infrastructures in the networks [5–8]. An elliptic curve cryptosystem would be one of the best candidates to meet the privacy and security challenges emerged in radio-frequency identification (RFID) technologies [9–11]. Presently, ECC-based untraceable RFID authentication protocols are used in smart healthcare environments to enhance medical data security [12–14]. Elliptic curve-based digital signature schemes such as elliptic curve digital signature algorithm (ECDSA) [2] and Edwards curve digital signature algorithm (EdDSA) [15,16] are adopted in wireless body area networks (WBANs) to fulfill the security requirements for real-time health data (e.g., blood pressure, heart rate, and pulse) management [17–19]. Modern security protocols such as transport layer security (TLS) and datagram transport layer security (DTLS) deploy these signature schemes for the energy efficient mutual authentication of the servers and clients in IoT platforms [20–22]. An ECC hierarchy is equipped with four consecutive levels as shown in Figure 1. The first level contains finite field arithmetic, such as addition, subtraction, multiplication, squaring, and inversion, which can be performed in both the Galois binary field GF(2[n]) and Galois prime field GF(p). The second level incorporates elliptic curve group operations, such as point addition (PA) and point doubling (PD). In the third level, elliptic curve point multiplication (ECPM) is accomplished by combining the elliptic curve group operations in a sequential manner. The top level includes ECC protocols such as ECDSA and EdDSA. The central and the most time-consuming operation in an elliptic curve-based cryptographic system is ECPM. The principle of ECPM can be specified as Q = kP, where P is a base point on an elliptic curve, k is a nonzero positive integer, and Q is another point on the curve [23]. Q and _k are considered to be public key and private key, respectively, and P is regarded as the public-key_ generator. The retrieval of k knowing the points P and Q is known as elliptic curve discrete logarithm problem (ECDLP) [2] that measures the security strength of the ECPM operation and finds out the weaknesses of the system. The easiest technique to accomplish ECPM is the binary/double-and-add (DAA) algorithm [2] that requires fewer hardware resources compared with other available methods. Therefore, ECC schemes adopting the DAA-based ECPM are suited for IoT applications because of their lower hardware resource requirements and lower power consumption. The major disadvantage of the DAA method is that the DAA-based ECPM is vulnerable to simple power analysis (SPA) attacks [24,25] unless it uses unified point operations. **Figure 1. Hierarchy of elliptic curve cryptography.** Edwards curves, a family of elliptic curves, are gaining enormous attention among security researchers because of their simplicity and high resistance against SCAs [26]. ECPM on Edwards curves is faster and more secure than that on the Weierstrass form of elliptic curves [27,28]. Edwards curves ----- _Sensors 2020, 20, 5148_ 3 of 19 have the advantage of providing strongly unified addition formulas [28], which cover both PA and PD. Separate hardware architectures for PA and PD are not required to perform ECPM. Moreover, unified PA prevents probable SPA attacks by making the secret key indistinguishable from power tracing. When ECPM adopts the same module for PA and PD, the binary bit pattern of the secret key cannot be retrieved by SPA. The twisted Edwards curves are a generalization of Edwards curves [29], which are mainly used in the digital signature scheme EdDSA. One of the most compatible twisted Edwards curves in digital signature systems is Edwards25519, which is the Edwards form of the elliptic curve Curve25519 [23,30]. In modern times, Edwards25519 curve is used for a high-speed, high-security digital signature scheme called Ed25519 [15,16]. ECPM using unified twisted Edwards curve not only provides high resistance against SPA but also it reduces the area of ECC processors. ECC can be accomplished with both hardware and software approaches. Although the software implementation is simple and cost-effective, it cannot provide high-speed computation as the hardware implementation can. Indeed, the hardware implementation of ECC with limited resources is a highly challenging task because low hardware use leads to a lower computational speed. In this point of view, Edwards curves are more effective than classical elliptic curves as they can be implemented on a smaller area with higher processing speed. Most of the hardware implementations of ECC reported in the literature are based on the Weierstrass form of elliptic curves. Few hardware implementations based on twisted Edwards curves over GF(p) have been reported. Baldwin et al. [31] first documented hardware implementation of a reconfigurable 192-bit ECC processor adopting twisted Edwards curve over GF(p). They provide a comparison between the FPGA implementation of an elliptic curve-based point multiplication and that of a twisted Edwards curve for different number of arithmetic logic units (ALUs) operated in parallel, which shows the Edwards curve as more efficient. Additionally, the twisted Edwards curve point operations are compared with the unified version of these operations. Although the unified version shows little bit worse performance, it provides a higher resistance against SPA. Liu et al. [21] present a computable endomorphism on twisted Edwards curves to boost the speed of ECDSA verification process. They provide area-efficient hardware architecture for signature verification with its FPGA implementation. Application specific integrated circuit (ASIC) implementation of the architecture is also provided for low-cost applications. The implementation results show that the design reduces approximately 50% of the number of PD operations required. Parallel architectures for ECPM on extended twisted Edwards are proposed by Abdulrahman et al. [32]. The authors present a new radix-8 ECPM algorithm to cope with SCAs and speed up computations. However, no hardware implementation of these architectures is reported. In this paper, a lightweight FPGA-based hardware implementation of ECC over GF(p) is proposed for IoT appliances. The major contributions of this paper are summarized as follows: - An efficient radix-4 interleaved modular multiplier is proposed to perform 256-bit modular multiplication over a prime field. - A novel hardware architecture for strongly unified PA on the Edwards25519 curve is proposed. - An efficient ECPM scheme is proposed to perform high-speed point multiplication on the Edwards25519 curve. The same module is used for PA and PD to prevent probable SPA attacks. The area required by the scheme is significantly lower than other available designs for ECPM. - ECPM is performed in projective coordinates to avoid the most expensive (in terms of computational complexity) modular division operation. In addition, a projective-to-affine (P2A) converter is proposed to transform the projective output into its affine form. This type of transformation reduces the computation time additionally required for the modular division operation performed in affine coordinate-based PA. - An ECC processor is designed by combining the ECPM scheme and the P2A converter in such a manner as to reduce the number of modular inversion operations required. The area-delay product of the proposed ECC processor is considerably small that ensures a better performance of our processor. ----- _Sensors 2020, 20, 5148_ 4 of 19 The rest of this paper is organized as follows: Section 2 presents the mathematical background of the twisted Edwards curve and unified PA formula. Section 3 presents the proposed hardware architectures for field operations (modular multiplication and modular inversion), unified PA, ECPM, and ECC processor. Section 4 presents the implementation results of the proposed designs. Section 5 shows a performance comparison of our proposed ECC processor with other related processors. Finally, Section 6 concludes this research study. **2. Mathematical Background** This section presents the twisted Edwards curve with its affine and projective representations as well as the unified PA formula for the curve. _2.1. Twisted Edwards Curve_ The affine representation of a twisted Edwards curve over a prime field Fp with not characteristic 2 is given by the equation [23,29]: _ta,d : ax[2]_ + y[2] = 1 + dx[2]y[2], (1) where a, b ∈ Fp \ {0, 1} with a ̸= d. When a = 1, the curve is called untwisted Edwards curve or, formally, Edwards curve. In the case of a = 1, the curve will be _−_ _td : −x[2]_ + y[2] = 1 + dx[2]y[2]. (2) when a = 1, d = 121665/121666, and p = 2[255] 19, the curve is called Edwards25519 that is the _−_ _−_ _−_ Edwards form of the elliptic curve Curve25519 [23]. In a projective or Jacobian coordinate system, each point (x, y) on ta,d is represented by a triplet form (X, Y, Z). The affine point P(x, y) corresponds to the projective point P(X = x, Y = y, Z = 1). The projective point P(X, Y, Z) corresponds to the affine point P(x = X/Z, y = Y/Z) with Z = 0. _̸_ The projective representation of the curve ta,d is given by the equation [23,29]: _Ta,d : (aX[2]_ + Y[2])Z[2] = Z[4] + dX[2]Y[2]. (3) The projective form of the curve td is given by the equation: _Td : (−X[2]_ + Y[2])Z[2] = Z[4] + dX[2]Y[2]. (4) _2.2. Unified Point-Addition Formula_ PA on the curve Td in projective coordinates is given by the equation: _P1(X1, Y1, Z1) + P2(X2, Y2, Z2) = P3(X3, Y3, Z3)._ (5) where P1 and P2 are two points on the curve and P3 is the resultant point. The unified PA formula [29] for Td can be given as follows: _A = X1X2, B = Y1Y2, C = Z1Z2, D = X1Y2,_ _E = X2Y1, F = AB, G = A + B,_ _H = C[2], I = D + E, J = dF, K = CG,_ _L = IC, M = H + J, N = H_ _J,_ _−_ _X3 = LN, Y3 = MK, Z3 = MN._ (6) The above formula is applicable for both PA and PD. PD can be performed considering that the points P1 and P2 are identical. ----- _Sensors 2020, 20, 5148_ 5 of 19 **3. Proposed Hardware Architectures** This Section presents the proposed hardware architectures for ECC operations and the final ECC processor. _3.1. Modular Multiplication_ Modular multiplication is the most important arithmetic operation of an ECC processor. The speed and occupied area of the processor entirely depend on it. Although a radix-2 multiplier consumes less hardware resources compared to higher radix (e.g., radix-4 and radix-8) multipliers [33], it is not compatible for high-speed multiplication because of its high latency. To reduce the latency, an efficient radix-4 interleaved modular multiplication algorithm is proposed as demonstrated in Algorithm 1. It requires n/2 + 1 clock cycles (CCs) to multiply two n-bit integers A and B over the prime field GF(p), where p is an n-bit prime number. Figure 2 illustrates the proposed modular multiplier based on this algorithm. **Algorithm 1 Proposed Radix-4 Interleaved Modular Multiplication** **Input : A = ∑i[n]=[−]0[1]** _[a][i][2][i][,][ B][ =][ ∑]i[n]=[−]0[1]_ _[b][i][2][i][;][ a][i][,][ b][i][ ∈{][0, 1][}]_ **Output : C = (A** _B) mod p_ _·_ 1: C ← 0; 2: T ← _B||"01";_ 3: while T(n 1 downto 0) = 0 do _−_ _̸_ 4: _D ←_ 4C; 5: **if T(n + 1 downto n) = "01" then** 6: _E_ _D + A;_ _←_ 7: **else if T(n + 1 downto n) = "10" then** 8: _E_ _D + 2A;_ _←_ 9: **else if T(n + 1 downto n) = "11" then** 10: _E_ _D + 3A;_ _←_ 11: **else** 12: _E ←_ _D;_ 13: **end if;** 14: _C ←_ _E mod p;_ 15: _T ←_ _T(n −_ 1 downto 0)||"00"; \\left shift operation 16: end while; 17: return C; Modular multiplication is obtained by performing iterative addition of its interim partial products reducing to modulo p. A shift-left register “Reg T” is used to perform left to right bitwise multiplication and for a synthesizable loop operation. T[(n + 1) : 2] is precomputed as the multiplier B and T[1 : 0] is precomputed as “01”. These two extra bits are added at the rightmost position of the register T to determine the appropriate end of the loop in the case of b0 = 0. At the beginning of each iteration, accumulator C is quadrupled and computed as D. For the bitwise multiplication, A, 2A, and 3A are separately added to D. MUX1 is used to select one of the four outputs D, D + A, D + 2A, and D + 3A as E based on the three bits T[(n + 1) : n]. If Tn+1 and Tn both are zero, D remains unchanged and E becomes D. At the end of each iteration, E is reduced to modulo p and T is shifted to the left by 2 bits. The modulo operation (E mod p) is performed by subtracting the prime numbers p to (j 1)p from E, _−_ where E is always less than jp; (j = 3, 4, 5...). In this module, (E mod p) is obtained by subtracting the prime numbers p to 6p from E as E is always less than 7p. These subtractions are executed using the 2’s complement method. MUX2 selects one of the seven outputs E, E − _p, E −_ 2p, E − 3p, E − 4p, E − 5p, and E − 6p as C for the next iteration based on the comparisons E ≥ _p, E ≥_ 2p, E ≥ 3p, E ≥ 4p, E ≥ 5p, and E 6p. These comparisons are obtained by checking the three bits E[(n + 1) : (n 1)]. After n/2 _≥_ _−_ ----- _Sensors 2020, 20, 5148_ 6 of 19 number of iterations, B, as well as T[(n 1) : 0], is shifted to zero value and the execution is stopped. _−_ The final content of the register “Reg C” is the modular multiplication of A and B. **Figure 2. Proposed modular multiplier.** A total of n/2 + 1 CCs are required to perform the modular multiplication operation, where n/2 CCs correspond to n/2 number of iterations and one extra CC is required for the initialization. To perform modular squaring, the inputs A and B are taken as identical. _3.2. Modular Inversion_ Modular inversion is the costliest (in terms of the hardware resource requirements) arithmetic operation in finite fields. In affine representations, PA and PD require modular inversion operation to perform modular division. In this study, although our ECC processor is designed in projective coordinates, modular inversion is required for P2A conversion. Algorithm 2 [2] demonstrates the binary modular inversion for the P2A conversion module proposed in this paper. The hardware architecture of this module is depicted in Figure 3. **Figure 3. Proposed hardware architecture for modular inversion.** ----- _Sensors 2020, 20, 5148_ 7 of 19 **Algorithm 2 Binary Modular Inversion [2]** **Input : B = ∑i[n]=[−]0[1]** _[b][i][2][i][;][ b][i][ ∈{][0, 1][}]_ **Output : C = B[−][1]** _mod p_ 1: C ← 0, q ← _B, r ←_ _p, s ←_ 1, t ← 0; 2: while q = 1 do _̸_ 3: **while q(0) = 0 do** 4: _q ←_ _q/2;_ 5: **if s(0) = 0 then** 6: _s ←_ _s/2;_ 7: **else** 8: _s_ (s + p)/2; _←_ 9: **end if;** 10: **end while;** 11: **while r(0) = 0 do** 12: _r ←_ _r/2;_ 13: **if t(0) = 0 then** 14: _t ←_ _t/2;_ 15: **else** 16: _t_ (t + p)/2; _←_ 17: **end if;** 18: **end while;** 19: **if q > r then** 20: _q ←_ _q −_ _r;_ 21: **if s > t then** 22: _s ←_ _s −_ _t;_ 23: **else** 24: _s_ _s + p_ _t;_ _←_ _−_ 25: **end if;** 26: **else** 27: _r ←_ _r −_ _q;_ 28: **if t > s then** 29: _t ←_ _t −_ _s;_ 30: **else** 31: _t_ _t + p_ _s;_ _←_ _−_ 32: **end if;** 33: **end if;** 34: end while; 35: return s mod p; The contents of the registers “Reg Q”, “Reg R”, “Reg S”, and “Reg T” are updated in every iteration. Five multiplexers such as MUX1, MUX2, MUX3, MUX4, and MUX5 are used to select corresponding outputs, satisfying different conditions by their select lines. In the case of q being even, MUX1 selects q/2 and MUX3 selects s/2 if s is even or (s + p)/2 if s is odd. In the case of q being odd and greater than r, MUX1 selects q _r and MUX3 selects s_ _t if s > t or s + p_ _t if s < t._ _−_ _−_ _−_ The comparisons q > r and s > t are obtained by checking the sign bits of the subtractions q _r and_ _−_ _s −_ _t, respectively. If q is odd and less than r, q and r both remain unchanged. Similarly, MUX2 selects_ one of the three outputs r, r/2, and r _q based on the conditions r(0) = 0 and r > q. MUX4 selects_ _−_ one of the five outputs t, t/2, (t + p)/2, t _s, and t + p_ _s based on the conditions r(0) = 0, t(0) = 0,_ _−_ _−_ _r > q, and t > s. MUX5 is used to select the final result as (s mod p) if q = 1. In this regard, q is_ subtracted by 2 to check whether q < 2 at the end of each iteration. When the sign bit of the subtraction _q_ 2 is 1, (s mod p) is stored in the register “Reg C”, which is the modular inversion of B. _−_ ----- _Sensors 2020, 20, 5148_ 8 of 19 In this architecture, on average n + n/4 CCs are required to perform the modular inversion operation, where n number of iterations are to reduce the n-bit variable q to 1 in a regular manner and additional n/4 number of iterations are for such uncertain case as q being odd. The clock cycles required for the modular inversion operation may vary from our estimation depending on the binary bit pattern of B. _3.3. Unified Point Addition_ Unified PA is required to perform both PA and PD by the same module so as to prevent possible SPA attacks in ECPM. The proposed hardware architecture for the unified PA formula described in (6) is depicted in Figure 4. The architecture includes 12 multiplications, 1 squaring, 3 additions, and 1 subtraction, which can be denoted as (12M+1S+4A). The proposed design consists of four consecutive levels, in which the arithmetic modules are connected in a sequential manner. The modules are arranged in horizontally parallel among the levels to achieve the shortest data path. The whole architecture is efficiently balanced to reduce the area required. Start signals are used to start the arithmetic operations and Done signals are used to confirm the end of the operations. The Done signals of the modules at each level are considered to be the Start signals of the modules at its subsequent level. AND blocks are used to synchronize the horizontal modules in time (e.g., if the Done signals d1, d2, d3, d4, and d5 all be 1, the Start signal s1 will be 1; otherwise, it will be 0). The modular multiplier and the squarer require n/2 + 1 CCs to perform modular multiplication and squaring. Modular addition and subtraction are completed in only one CC. The level that contains any multiplication or squaring operation requires n/2 + 1 CCs and the level that contains no multiplication or squaring requires one CC to jump to the next level. In this design, a total of 2n + 5 CCs are required to complete the unified PA operation. **Figure 4. Proposed hardware architecture for unified PA.** ----- _Sensors 2020, 20, 5148_ 9 of 19 _3.4. Elliptic Curve Point Multiplication_ ECPM is the ultimate operation of an ECC processor. It multiplies a point on an elliptic curve with a scalar. The execution time of ECC schemes is dominated by ECPM. Let P(X, Y, Z) be a point on the curve Td, k be a scalar that is considered to be secret key. A public key Q(X, Y, Z) is generated from the known base point P and the secret key k by performing ECPM as follows: _Q = kP,_ (7) where Q is also a point on the curve. It can be obtained by adding P to itself k − 1 times such as _Q = P + P + ....... + P._ � �� � _k−1 times_ (8) If k is expressible as a power of 2, Q can be obtained by doubling P on itself log2k times such as _Q = ...2(2(2(P)))._ � �� � _log2k times_ (9) In the binary/ DAA method, ECPM is performed by a combination of PD and PA following the binary bit pattern of the secret key as shown in Algorithm 3. In this algorithm, separate modules are required to perform PA and PD. The power consumption of the two separate modules are different. Monitoring these two power levels by SPA, the bit pattern of k can be retrieved as shown in Figure 5. Moreover, k can be assumed by timing analysis; hence, ECPM based on this algorithm is vulnerable to SPA attacks. To cope with SPA, Algorithm 3 is modified to Algorithm 4, where PD is replaced by unified PA. According to this algorithm, power is only consumed for PA with a fixed power consumption, which is independent of the bit pattern of k as shown in Figure 6. Since the power consumption is the same across all the iterations, this algorithm is free from SPA. Figure 7 illustrates the proposed hardware architecture for ECPM based on Algorithm 4. Two point-addition blocks PA1, PA2 and three multiplexers MUX1, MUX2, MUX3 are used in this processor. Initially, Q1 is precomputed as _P. PA1 adds the point Q1 to itself and the output Q2 goes to the input of PA2. Identical inputs are_ inserted in PA1 to perform PD by means of PA. One of the two inputs of PA2 is the output of PA1 and the other one is P or 0. If ki = 1, PA2 adds the point P to the point Q2 and the output Q3 goes to the input of the PA1 via the register Rg. On the contrary, if ki = 0, PA2 remains idle and the output of PA1 directly goes to its input via Rg. MUX1 is used to select the ith bit of k by log2l number of select lines, where l is the bit length of k. Based on ki, MUX2 selects P or 0 as one of the two inputs of PA2; MUX3 selects Q2 or Q3 as the input Q1 for the subsequent iteration. **Algorithm 3 DAA ECPM without Unified PA [2]** **Input : P(X, Y, Z), k = ∑i[l]=[−]0[1]** _[k][i][2][i][;][ k][i][ ∈{][0, 1][}][,][ k][l][−][1][ =][ 1]_ **Output : Q(X, Y, Z)** 1: Q ← _P;_ 2: for i from l − 2 to 0 do 3: _Q ←_ 2Q; _\\PD_ 4: **if ki = 1 then** 5: _Q_ _Q + P;_ PA _←_ _\\_ 6: **end if;** 7: end for; 8: return Q; ----- _Sensors 2020, 20, 5148_ 10 of 19 **Algorithm 4 Proposed Unified PA-based ECPM** **Input : P(X, Y, Z), k = ∑i[l]=[−]0[1]** _[k][i][2][i][;][ k][i][ ∈{][0, 1][}][,][ k][l][−][1][ =][ 1]_ **Output : Q(X, Y, Z)** 1: Q1 ← _P;_ 2: for i from l − 2 to 0 do 3: _Q2 ←_ _Q1 + Q1;_ _\\PA_ 4: **if ki = 1 then** 5: _Q3 ←_ _Q2 + P;_ _\\PA_ 6: _Q1 ←_ _Q3;_ 7: **else** 8: _Q1 ←_ _Q2;_ 9: **end if** 10: end for 11: return Q1; **Figure 5. Simplistic representation of SPA in conventional DAA ECPM.** **Time** **Figure 6. Simplistic representation of SPA in proposed unified PA-based ECPM.** **Figure 7. Proposed hardware architecture for ECPM.** ----- _Sensors 2020, 20, 5148_ 11 of 19 For the l-bit k, the register stores kP as the final result after l − 1 number of iterations. The average CCs required to perform the ECPM can be calculated as ECPMCC = (l − 1) × (PA1CC + RgCC) + l/2 × PA2CC = (l 1) (2n + 5 + 1) + (l/2) (2n + 5) (10) _−_ _×_ _×_ = 3nl 2n + 8.5l 6. _−_ _−_ For l = n, ECPMCC = 3n[2] + 6.5n − 6. (11) PA1 and Rg remain active in every iteration, whereas PA2 goes idle in the case of ki = 0. In every iteration, a total 2n + 6 CCs are spent by PA1 and Rg. Additional 2n + 5 CCs are spent by PA2 if ki = 1. On average, l(n + 2.5) CCs are spent by PA2 across the ECPM. For the n-bit k, the latency of the ECPM is approximately 3n[2] + 6.5n 6 CCs. This latency may vary depending on the bit pattern of the key; _−_ it increases with the number of 1 and decreases with the number 0 present in the bit pattern. In this study, an average case is considered. This means that the key has equal number of 1 and 0 in its bit pattern, although this is not always the case. _3.5. Proposed ECC Processor_ A time-area-efficient ECC processor is designed for public-key generation using the proposed projective coordinate-based ECPM along with a P2A converter as shown in Figure 8. This processor will generate a public key from a private key and a base point on Td. Initially, the affine base point P(x, y) is transformed into its projective form such as P(X, Y, Z) by an affine-to-projective (A2P) converter. The public key Q(X, Y, Z) is obtained by performing ECPM of the projective point P(X, Y, Z) with the secret key k. Finally, Q(X, Y, Z) is transformed into its affine form such as Q(x, y) by the P2A converter. For the P2A conversion, Z is inverted by the proposed modular inversion module and separately multiplied by X and Y. The latency required by the processor to process the ECPM operation along with the coordinate conversions is 3n[2] + 8.25n 5 CCs, which is the total sum of the latency of ECPM, _−_ modular inversion, and modular multiplication. **Figure 8. Proposed ECC processor for public-key generation.** ----- _Sensors 2020, 20, 5148_ 12 of 19 **4. Implementation Results** The proposed ECC processor was programmed in VHDL and implemented using the Xilinx ISE 14.7 Design Suite software. Xilinx ISim simulator was used to simulate the ECC operations. The simulation results were verified by the Maple 18 software. Synthesizing, mapping, placing, and routing of the proposed ECC modules were performed on Xilinx Virtex-7 and Virtex-6 FPGA platforms, separately. The details of these FPGA platforms and settings are as follows: - Platform 1: Virtex-7 (XC7VX690T) - Platform 2: Virtex-6 (XC6VHX380T) - Design Goal: Balanced - Design Strategy: Xilinx Default - Optimization Goal: Speed - Optimization Effort: Normal The implementation results of the proposed ECC modules are summarized in Table 1. On Platform 1, all the modules run at a maximum frequency of 104.39 MHz. The proposed ECC processor occupies 6543 slices (25,898 LUTs) and generates a public key from a given 256-bit private key in 1.9 ms with a throughput of 134.5 kbps. On Platform 2, the modules operate at a maximum frequency of 93.23 MHz. The numbers of slices and LUTs used by the processor are 6579 and 25,968, respectively, the delay of the public-key generation is 2.13 ms, and the throughput is 120.1 kbps. **Table 1.** Implementation results of the proposed ECC modules on different FPGA platforms over Fp-256. **Area** **Operation** **Platform** **CCs** **(Slices)** **Area** **(LUTs)** **Maximum** **Frequency (MHz)** **Time** **(µs)** **Throughput** **(Mbps)** Modular multiplication Modular inversion Virtex-7 129 416 1451 104.39 1.24 207.2 Virtex-6 129 420 1460 93.23 1.38 185 Virtex-7 320 1197 4155 110.65 2.89 83.5 Virtex-6 320 1209 4156 97.94 3.27 74.6 Virtex-7 517 4159 15,594 104.39 4.95 51.7 Unified PA Virtex-6 517 4292 15,593 93.23 5.55 46.2 ECPM (projective) Public-key generation Virtex-7 198,266 5457 21,194 104.39 1899 134.8 10[−][3] _×_ Virtex-6 198,266 5541 21,224 93.23 2126 120.4 10[−][3] _×_ Virtex-7 198,715 6543 25,898 104.39 1903 134.5 10[−][3] _×_ Virtex-6 198,715 6579 25,968 93.23 2131 120.1 10[−][3] _×_ The performance of the ECC modules on the Virtex-6 FPGA platform is a little bit worse compared to the Virtex-7 FPGA platform in terms of speed. However, the area use of the different modules on these platforms are almost the same. It must be noted that no digital signal processing (DSP) slice is used to implement our processor. Although DSP slices increase processing speed, they increase processor’s cost as well. **5. Performance Comparison** Several hardware implementations of ECC have been reported in [34–53], where some authors aimed to minimize the area use while others tried to reduce the computation time. Achieving a higher processing speed with low-area use is technically challenging. We tried to maintain a balance ----- _Sensors 2020, 20, 5148_ 13 of 19 between area and time as they are two important performance criteria of a cryptographic processor. A performance comparison of our proposed ECC processor with other related designs is presented in Table 2. **Table 2.** Performance comparison of the proposed ECC processor with other related designs over Fp-256. **Area** **Frequency** **Design** **Platform** **CCs** **(Slices)** **(MHz)** **Time** **(ms)** **Throughput** **Area × Time** **(kbps)** Ours (a) Virtex-7 6.5k 198.7 104.39 1.9 134.49 _[a]_ 12.35 Ours (b) Virtex-6 6.6k 198.7 93.23 2.13 120.12 _[a]_ 14.05 [34] Virtex-7 24.2k, 2.8k DSPs 215.9 72.9 2.96 1816.2 71.63 [35] Kintex-7 11.3k 397.3 121.5 3.27 78.28 36.95 [36] Virtex-6 65.6k 153.2 327 0.47 546.42 _[a]_ 30.83 [37] Virtex-5 8.7k 361.6 160 2.26 113.27 _[a]_ 19.66 [38] Virtex-5 10.2k 442.2 66.7 6.63 38.61 _[a]_ 67.63 [39] Virtex-4 12k 459.9 36.5 12.6 20.32 _[a]_ 151.20 [40] Virtex-4 9.4k, 14 DSPs 610 20.44 29.84 8.58 _[a]_ 280.50 [41] Virtex-4 35.7k 207.1 70 2.96 86.53 _[a]_ 105.67 [42] Virtex-4 13.2k 200 40 5 51 66.00 [43] Virtex-4 20.6k 191.6 49 3.91 65.47 80.55 [44] Virtex-4 20.1k 331.1 43 7.7 33.25 _[a]_ 154.77 [45] Virtex-4 20.8k, 32 DSPs 414 60 6.9 37.10 _[a]_ 143.52 [46] Virtex-4 7k, 8 DSPs 993.7 182 5.46 46.88 _[a]_ 38.22 [47] Spartan-3 27.6k 708 40 17.7 14.46 _[a]_ 488.52 [48] Virtex-II Pro 12k 337.7 36 9.38 27.29 _[a]_ 112.56 [49] Virtex-II Pro 8.3k 163.2 37 4.41 58.04 _[a]_ 36.60 [50] Virtex-II Pro 15.8k, 256 DSPs 151.4 39.5 3.86 66.74 60.98 [51] Virtex-II Pro 41.6k 252.1 94.7 2.66 96.17 _[a]_ 110.66 [52] Virtex-E 16.4k 156.8 39.7 3.95 64.82 _[a]_ 64.78 [53] Virtex-E 14.2k 118.3 34.7 3.41 75.09 _[a]_ 48.42 _a Estimated by the authors of this paper as Throughput = (Maximum frequency ÷ CCs)× 256._ The residue number system (RNS)-based ECC design reported in [34] provides a higher throughput (1816.2 kbps) by performing ECPM on 21 keys in parallel. Conventional DAA method is adopted for ECPM, where PA and PD are executed by separate modules carrying high risk of SPA attacks. On Virtex-7 FPGA, the design consumes 96,867 LUTs (approx. 24,216 slices) with 2799 additional DSP slices. Although the throughput of this design is higher than that of our design, it costs 3.7 times more hardware resources. The novelty of this design is that it processes 21 keys simultaneously, which prevents template-based attacks by increasing the computation complexity. In [35], the authors propose a high-performance ECC processor with its ASIC and FPGA implementations. A novel hardware architecture for combined PA-PD operation in Jacobian ----- _Sensors 2020, 20, 5148_ 14 of 19 coordinates is proposed to achieve high-speed ECPM with low hardware use. On Kintex-7 FPGA, the processor separately designed in affine and Jacobian coordinates performs ECPM in 4.7 ms and 3.27 ms, occupying 9.3k and 11.3k slices, respectively. Our processor implemented on 7-series FPGA is 1.72 times faster and costs 1.73 times less slices as compared with this processor designed in Jacobian coordinates. The throughput of our design is 1.76 times higher. A high-speed ECC processor is proposed in [36] providing redundant signed digit (RSD)-based carry free modular arithmetic. The processor performs high-speed ECPM with a higher throughput. However, it occupies 10 times more slices on Virtex-6 FPGA than our processor. Although RSD representation offers fast computation, it consumes a vast amount of hardware resources, which makes processor bulky and hence not suited for low-power IoT devices. The high-speed RSD-based modular multiplier proposed in this paper performs single multiplication in only 0.34 µs, consuming 22k LUTs. In comparison with this multiplier, our proposed modular multiplier performs single multiplication in 1.45 µs and consumes only 1.3k LUTs with almost 4 times better efficiency in terms of area-time (AT) product. The RSD-based ECC processors reported in [37,38] present comprehensive pipelining technique for Karatsuba–Ofman multiplication to achieve high throughput. Our processor has smaller AT product compared with these processors. Liu et al. [39] propose a hardware-software approach for flexible duel-field ECC processor with its ASIC and FPGA implementations. The traditional DAA method for ECPM is replaced by the double-and-add-always (DAAA) method to protect the processor from SPA attacks. Although the DAAA method for ECPM provides high resistance against SPA, it increases the computational complexity and hence reduces the frequency and throughput. In addition, it consumes more power than the conventional DAA method as PA and PD are performed in every iteration. Our processor is protected against SPA attacks by implementing the cost-effective DAA algorithm with unified PA. When compared to our processor, the main advantage of this processor is that it is flexible and reconfigurable over different field orders. In addition, it can perform ECPM over both GF(2[n]) and GF(p), whereas our processor performs ECPM over GF(p) only. Hu et al. [40] propose an SPA-resistant ECC design over GF(p), providing its ASIC and FPGA implementations. The design uses 9370 slices with 14 additional DSP slices on Virtex-4 FPGA. Despite employing additional DSP slices, the speed of this design is considerably low. It takes 29.84 ms with a frequency of 20.44 MHz to perform single ECPM over a 256-bit prime field. The advantage of this design that makes it well suited for embedded applications is its reconfigurable computing capability. A low latency ECPM design is proposed in [41] exploiting parallel multiplication over GF(p). Protection against timing and SPA attacks is provided by using the DAAA method for ECPM. The latency of this design is 3n[2] + 37n + 4n CCs, whereas the latency of our design is 3n[2] + 8.25n 5 CCs. Therefore, the computational complexity of ECPM in this design is higher _−_ than that in our design. The radix-4 parallel interleaved modular multiplier proposed in this paper performs multiplication in 0.79 µs, consuming 6.3k LUTs. Four multiplier units are operated in parallel to speed up the multiplication process. The main feature of this design is its capability to perform ECPM over GF(p) with any arbitrary value of p less than or equal to 256 bits in size. The design reported in [42] exploits the Montgomery ladder algorithm for SPA-resistant ECPM. Although the Montgomery ladder algorithm offers lower latency ECPM and higher resistance against SPA than the general DAA method [23], it deals with around 50% additional PA operations that results in a higher power consumption. Hence, the DAA method is more efficient than the Montgomery ladder technique in terms of energy consumption. The advantage of this design is that it supports any prime number p ≤ 256-bit. In [43], the authors present a high-performance hardware design for ECPM adopting non-adjacent form (NAF) method. Although NAF method has the advantage of reducing the latency of ECPM, the computational complexity and its vulnerability to SCAs are high in this method. Moreover, additional point subtraction operation is required for NAF scalar multiplication. Like the designs reported in [40,41], this design is programmable for any prime p ≤ 256-bit. Parallel crypto design is proposed in [44] using the DAAA method to perform SCA-resistant ECPM over different ----- _Sensors 2020, 20, 5148_ 15 of 19 field orders. The design is represented in affine coordinates, where PA and PD require modular division operations. Modular division is the most time-consuming arithmetic operation in finite fields. Therefore, this design is not convenient for high-speed computation. However, it provides high resistance against timing and SPA attacks by parallel computation of PA and PD. Ananyi et al. [45] propose a flexible hardware ECC processor that supports five National Institute of Standard and Technology (NIST) recommended prime curves. They provide a comparison between the binary and NAF ECPM over all five NIST prime fields such as p192, p224, p256, p384, and p521, where the NAF ECPM is found to be more time-efficient. Their processor consumes 20,793 slices (31,946 LUTs) with 32 additional DSP blocks on Virtex-4 FPGA and performs the binary ECPM in 6.9 ms and the NAF ECPM in 6.1 ms over p256. The modular inverter designed in this paper operates at a frequency of 58.6 MHz costing 10,921 slices with 32 DSP blocks, whereas our modular inverter implemented on Virtex-7 FPGA runs at 110.65 MHz consuming 1197 slices without any DSP block. A scalable ECC processor developed by Loi et al. [46] can perform ECPM on five NIST suggested prime curves such as P-192, P-224, P-256, P-384, and P-521 without hardware reconfiguring. On Virtex-4 FPGA, this processor performs ECPM in 5.46 ms, occupying 7020 slices along with 8 additional DSP slices. Despite using DSP slices, the computational speeds of the processors reported in [45,46] are low. The main significance of these processors is that they are flexible over the five NIST prime fields and hence they can be programmed to perform ECPM for variable prime numbers ranging from 192 to 521 bits in size without being architecturally reconfigured. The processors reported in [47–53], are implemented on some backdated FPGA platforms, which are now obsolete. Performance comparison in terms of AT product is shown in Figure 9. The AT product of our design is lower than that of the other designs tabulated in Table 2. Figure 10 shows performance comparison in terms of throughput per slice. The per slice throughput of our design is higher than that of the other designs except [34]. The RNS-based design reported in [34] provides a higher throughput by performing ECPM on 21 keys concurrently. Our processor’s low value of AT product and high value of throughput ensure a better performance in IoT platforms. However, a fair comparison is not possible because the compared processors are implemented on different FPGA platforms. Our proposed ECC processor is implemented only on the Virtex-7 and Virtex-6 FPGAs because the number of input/output blocks (IOBs) is limited in earlier FPGAs. Furthermore, the earlier FPGAs such as Virtex-II-Pro, Virtex-4, and Virtex-5 are not compatible with low-power devices because of their high power consumption. 500 450 400 350 300 250 200 150 100 50 0 (a) (b) [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] ### Reference **Figure 9. Performance comparison in terms of AT product.** ----- _Sensors 2020, 20, 5148_ 16 of 19 70 Spartan-3 Virtex-E 60 Virtex-II-Pro Virtex-4 Virtex-5 50 Virtex-6 Virtex-7 40 Kintex-7 30 20 10 0 |Col1|Col2|Col3|Col4|Spartan-3 Virtex-E Virtex-II-Pro Virtex-4 Virtex-5 Virtex-6 Virtex-7 Kintex-7|Spartan-3 Virtex-E Virtex-II-Pro Virtex-4 Virtex-5 Virtex-6 Virtex-7 Kintex-7| |---|---|---|---|---|---| ||||||| ||||||| ||||||| ||||||| ||||||| ||||||| ||||||| ||||||| ||||||| ||||||| ||||||| ||||||| ||||||| (a) (b) [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] ### Reference **Figure 10. Performance comparison in terms of throughput per slice.** **6. Conclusions** In this paper, a high-performance ECC processor has been proposed exploiting unified PA on Edwards25519 curve to perform SPA-resistant point multiplication. An efficient ECPM module has been designed in projective coordinates, which supports 256-bit point multiplication over a prime field. Unified PA is adopted for the ECPM module to provide strong protection against SPA attacks and reduce the area required by an additional PD module. To perform high-speed modular multiplication, an efficient radix-4 interleaved modular multiplier has been proposed. The proposed ECC processor performs fast point multiplication with a considerably lower area use, providing high resistance against SPA. Because of its less hardware resource requirements and high computation speed, it is well suited for resource-constrained IoT devices. Since it provides a faster ECPM that is a rising demand of elliptic curve-based digital signature schemes, it could be manipulated in Bitcoin-like cryptocurrencies for high-speed digital signature generation and verification, which would reduce latency in transaction confirmation. Based on the overall performance analyses, it can be concluded that the proposed ECC processor could be a good choice for the IoT security as well as the emerging technology “Blockchain”. **Author Contributions: All the authors contributed to this paper. Specifically, M.M.I. and M.S.H. proposed the** idea and programmed the design; M.M.I. analyzed and verified the implementation results and wrote the paper; M.K.H. and M.S. reviewed and edited the paper; and Y.M.J. supervised the work and provided funding support. All authors have read and agreed to the published version of the manuscript. **Funding: This research was supported by the Ministry of Science and ICT (MSIT), Korea, under the Information** Technology Research Center (ITRC) support program (IITP-2018-0-01396) supervised by the Institute for Information & communication Technology Promotion (IITP). **Acknowledgments: As the first author of this paper, I would like to thank my beloved mother and brother who** always supported me in every success and failure in my life. **Conflicts of Interest: This manuscript has not been published elsewhere and is not under consideration by** another journal. We have approved the manuscript and agree with submission in Sensors. There are no conflict of interest to declare. **References** 1. Ding, S.; Li, C. ; Li, H. A novel efficient pairing-free CP-ABE based on elliptic curve cryptography for IoT. _[IEEE Access 2018, 6, 27336–27345. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2018.2836350)_ 2. Hankerson, D.; Menezes, A.; Vanstone, S. Guide to Elliptic Curve Cryptography; Springe: New York, NY, USA, 2004. ----- _Sensors 2020, 20, 5148_ 17 of 19 3. ElGamal, T. A public key cryptosystem and a signature scheme based on discrete logarithms. IEEE Trans. _[Inf. Theory 1985, 31, 469–472. [CrossRef]](http://dx.doi.org/10.1109/TIT.1985.1057074)_ 4. Rivest, R.L.; Shamir, A.; Adleman, L. A method for obtaining digital signatures and public-key cryptosystems. _[Commun. ACM 1978, 21, 120–126. [CrossRef]](http://dx.doi.org/10.1145/359340.359342)_ 5. [Diffie, W.; Hellman, M. New directions in cryptography. IEEE Trans. Inf. Theory 1976, 22, 644–654. [CrossRef]](http://dx.doi.org/10.1109/TIT.1976.1055638) 6. Liu, Z.; Huang, X.; Hu, Z.; Khan, M.K.; Seo, H.; Zhou, L. On emerging family of elliptic curves to secure [internet of things: ECC comes of age. IEEE Trans. Dependable Secur. Comput. 2017, 14, 237–248. [CrossRef]](http://dx.doi.org/10.1109/TDSC.2016.2577022) 7. Challa, S.; Wazid, M.; Das, A.K.; Kumar, N.; Reddy, A.G.; Yoon, E.-J.; Yoo, K.-Y. Secure signature-based authenticated key establishment scheme for future IoT applications. _IEEE Access 2017, 5, 3028–3043._ [[CrossRef]](http://dx.doi.org/10.1109/ACCESS.2017.2676119) 8. Lara-Nino, C.A.; Diaz-Perez, A.; Morales-Sandoval, M. Elliptic curve lightweight cryptography: A survey. _[IEEE Access 2018, 6, 72514–72550. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2018.2881444)_ 9. Lee, Y.; Sakiyama, K.; Batina, L. Elliptic-curve-based security processor for RFID. IEEE Trans. Comput. 2008, _[57, 1514–1527. [CrossRef]](http://dx.doi.org/10.1109/TC.2008.148)_ 10. Liao, Y.; Hsiao, C. A secure ECC-based RFID authentication scheme integrated with ID-verifier transfer [protocol. Ad Hoc Netw. 2014, 18, 133–146. [CrossRef]](http://dx.doi.org/10.1016/j.adhoc.2013.02.004) 11. Chou, J. An efficient mutual authentication RFID scheme based on elliptic curve cryptography. J. Supercomput. **[2014, 70, 75–94. [CrossRef]](http://dx.doi.org/10.1007/s11227-013-1073-x)** 12. Zhang, Z.; Qi, Q. An efficient RFID authentication protocol to enhance patient medication safety using [elliptic curve cryptography. J. Med. Syst. 2014, 38, 5. [CrossRef] [PubMed]](http://dx.doi.org/10.1007/s10916-014-0047-8) 13. Zhao, Z. A secure RFID authentication protocol for healthcare environments using elliptic curve [cryptosystem. J. Med. Syst. 2014, 38, 5. [CrossRef] [PubMed]](http://dx.doi.org/10.1007/s10916-014-0046-9) 14. He, D.; Zeadally, S. An analysis of RFID authentication schemes for internet of things in healthcare [environment using elliptic curve cryptography. IEEE Internet Things J. 2015, 2, 72–83. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2014.2360121) 15. Bernstein, D.J.; Duif, N.; Lange, T.; Schwabe, P.; Yang, B.Y. High-speed high-security signatures. _[J. Cryptogr. Eng. 2012, 2, 77–89. [CrossRef]](http://dx.doi.org/10.1007/s13389-012-0027-1)_ 16. Liusvaara, I.; Josefsson, S. Edwards Curve Digital Signature Algorithm (EdDSA). In Internet-Draft: [Draft-irtf-cfrg-eddsa-05, Internet Engineering Task Force, 2017. Available online: https://tools.ietf.org/](https://tools.ietf.org/html/rfc8032) [html/rfc8032 (accessed on 1 January 2017).](https://tools.ietf.org/html/rfc8032) 17. Liu, J.; Zhang, Z.; Chen, X.; Kwak, K. Certificateless remote anonymous authentication schemes for wireless [body area networks. IEEE Trans. Parallel Distrib. Syst. 2014, 25, 332–342. [CrossRef]](http://dx.doi.org/10.1109/TPDS.2013.145) 18. He, D.; Zeadally, S.; Kumar, N.; Lee, J.-H. Anonymous authentication for wireless body area networks with [provable security. IEEE Syst. J. 2017, 11, 2590–2601. [CrossRef]](http://dx.doi.org/10.1109/JSYST.2016.2544805) 19. Saeed, M.E.S.; Liu, Q.-Y.; Tian, G.; Gao, B.; Li, F. Remote authentication schemes for wireless body area [networks based on the Internet of Things. IEEE Internet Things J. 2018, 5, 4926–4944. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2018.2876133) 20. Keoh, S.L.; Kumar, S.S.; Tschofenig, H. Securing the Internet of Things: A standardization perspective. _[IEEE Internet Things J. 2014, 1, 265–275. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2014.2323395)_ 21. Liu, Z.; Grosschadl, J.; Hu, Z.; Jarvinen, K.; Wang, H.; Verbauwhede, I. Elliptic curve cryptography with efficiently computable endomorphisms and its hardware implementations for the Internet of Things. _[IEEE Trans. Comput. 2017, 66, 773–785. [CrossRef]](http://dx.doi.org/10.1109/TC.2016.2623609)_ 22. Banerjee, U.; Wright, A.; Juvekar, C.; Waller, M.; Arvind, A.; Chandrakasan, A.P. An Energy-Efficient Reconfigurable DTLS Cryptographic Engine for Securing Internet-of-Things Applications. IEEE J. Comput. **[2019, 54, 2339–2352. [CrossRef]](http://dx.doi.org/10.1109/JSSC.2019.2915203)** 23. Islam, M.M.; Hossain, M.S.; Hasan, M.K.; Shahjalal, M.; Jang, Y.M. FPGA implementation of high-speed area-efficient processor for elliptic curve point multiplication over prime field. _IEEE Access 2019, 7,_ [178811–178826. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2958491) 24. Brier, E.; Joye, M. Weierstraß elliptic curves and side-channel attacks. In Public Key Cryptography (LNCS); Springer: Heidelberg, Germany, 2002; Volume 2274, pp. 335–345. 25. Joye, M. Elliptic curves and side-channel analysis. ST J. Syst. Res. 2003, 4, 17–21. 26. [Edward, H.M. A normal form for elliptic curves. Bull. Am. Math. Soc. 2007, 44, 393–422. [CrossRef]](http://dx.doi.org/10.1090/S0273-0979-07-01153-6) 27. Bernstein, D.J.; Lange, T. Faster addition and doubling on elliptic curves. In Proceedings of the Advances in _Cryptology (LNCS); Springer: Heidelberg, Germany, 2007; Volume 4833, pp. 29–50._ ----- _Sensors 2020, 20, 5148_ 18 of 19 28. Hisil, H.; Wong, K.K.H.; Carter G.; Dawson, E. Twisted edwards curves revisited. In Proceedings of the _Advances in Cryptology (LNCS); Springer: Heidelberg, Germany, 2008; Volume 5350, pp. 326–343._ 29. Bernstein, D.J.; Birkner, P.; Lange, T.; Peters, C. Twisted edwards curves. In Proceedings of the Advances in _Cryptology (LNCS); Springer: Heidelberg, Germany, 2008; Volume 5023, pp. 389–405._ 30. Bernstein, D.J. Curve25519: New Diffie-Hellman speed records. In Proceedings of the Public Key Cryptography _(LNCS); Springer: Heidelberg, Germany, 2006; Volume 3958, pp. 207–228._ 31. Baldwin, B.; Moloney, R.; Byrne, A.; McGuire, G.; Marnane, W.P. A hardware analysis of twisted Edwards curves for an elliptic curve cryptosystem. In Proceedings of the Reconfigurable Computing: Architectures Tools _and Applications (LNCS); Springer: Heidelberg, Germany, 2009; Volume 5453, pp. 355–361._ 32. Abdulrahman, E.A.H.; Masoleh, A.R. New regular radix-8 scheme for elliptic curve scalar multiplication [without pre-computation. IEEE Trans. Comput. 2015, 64, 438–451. [CrossRef]](http://dx.doi.org/10.1109/TC.2013.213) 33. Islam, M.M.; Hossain, M.S.; Shahjalal, M.; Hasan, M.K.; Jang, Y.M. Area-time efficient hardware implementation of modular multiplication for elliptic curve cryptography. IEEE Access 2020, 8, 73898–73906. [[CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.2988379) 34. Asif, S.; Hossain, M.S.; Kong, Y. High-throughput multi-key elliptic curve cryptosystem based on residue [number system. IET Comput. Digit. Tech. 2017, 11, 165–172. [CrossRef]](http://dx.doi.org/10.1049/iet-cdt.2016.0141) 35. Hossain, M.S.; Kong, Y.; Saeedi, E.; Vayalil, N. High-performance elliptic curve cryptography processor over [NIST prime fields. IET Comput. Digit. Tech. 2016, 11, 33–42. [CrossRef]](http://dx.doi.org/10.1049/iet-cdt.2016.0033) 36. Shah, Y.A.; Javeed, K.; Azmat, S.; Wang, X. Redundant signed digit based high-speed elliptic curve [cryptographic processor. J. Circuits Syst. Comput. 2018, 28, 1950081. [CrossRef]](http://dx.doi.org/10.1142/S0218126619500816) 37. Marzouqi, H. ; Al-Qutayri, M.; Salah, K.; Schinianakis, D.; Stouraitis, T. A high-speed FPGA implementation [of an RSD-based ECC processor. IEEE Trans. Very Large Scale Integr. (VLSI) Syst. 2016, 24, 151–164. [CrossRef]](http://dx.doi.org/10.1109/TVLSI.2015.2391274) 38. Marzouqi, H.; Al-Qutayri, M.; Salah, K. An FPGA implementation of NIST 256 prime field ECC processor. In Proceedings of the IEEE International Conference on Electronics, Circuits, and Systems (ICECS), Abu Dhabi, UAE, 8–11 December 2013; pp. 493–496. 39. Liu, Z.; Liu, D.; Zou, X. An efficient and flexible hardware implementation of the dual-field elliptic curve [cryptographic processor. IEEE Trans. Ind. Electron. 2017, 64, 2353–2362. [CrossRef]](http://dx.doi.org/10.1109/TIE.2016.2625241) 40. Hu, X.; Zheng, X.; Zhang, S.; Cai, S.; Xiong, X. A low hardware consumption elliptic curve cryptographic [architecture over GF(p) in embedded application. Electronics 2018, 7, 104. [CrossRef]](http://dx.doi.org/10.3390/electronics7070104) 41. Javeed, K.; Wang, X. Low latency flexible FPGA implementation of point multiplication on elliptic curves [over GF(p). Int. J. Circuit Theory Appl. 2016, 45, 214–228. [CrossRef]](http://dx.doi.org/10.1002/cta.2295) 42. Javeed, K.; Wang, X. FPGA based high-speed SPA-resistant elliptic curve scalar multiplier architecture. Int. J. _[Reconfigurable Comput. 2016, 2016, 1–10. [CrossRef]](http://dx.doi.org/10.1155/2016/6371403)_ 43. Javeed, K.; Wang, X.; Scott, M. High performance hardware support for elliptic curve cryptography over [general prime field. Microprocess. Microsyst. 2017, 51, 331–342. [CrossRef]](http://dx.doi.org/10.1016/j.micpro.2016.12.005) 44. Ghosh, S.; Alam, M.; Chowdhury, D.R.; Gupta, I.S. Parallel crypto-devices for GF(p) elliptic curve [multiplication resistant against side-channel attacks. Comput. Electr. Eng. 2009, 35, 329–338. [CrossRef]](http://dx.doi.org/10.1016/j.compeleceng.2008.06.009) 45. Ananyi, K.; Alrimeih, H.; Rakhmatov, D. Flexible hardware processor for elliptic curve cryptography over [NIST prime fields. IEEE Trans. Very Large Scale Integr. (VLSI) Syst. 2009, 17, 1099–1112. [CrossRef]](http://dx.doi.org/10.1109/TVLSI.2009.2019415) 46. Loi, K.C.C.; Ko, S.B. Scalable elliptic curve cryptosystem FPGA processor for NIST prime curves. IEEE Trans. _[Very Large Scale Integr. (VLSI) Syst. 2015, 23, 2753–27. [CrossRef]](http://dx.doi.org/10.1109/TVLSI.2014.2375640)_ 47. Sakiyama, K.; Mentas, N.; Batina, L.; Preneel, B.; Verbauwhede, I. Reconfigurable modular arithmetic logic unit for high-performance public-key cryptosystems. In Proceedings of the Reconfigurable Computing: _Architectures and Applications (LNCS); Springer: Heidelberg, Germany, 2006; Volume 3985, pp. 347–357._ 48. Ghosh, S.; Mukhopadhyay, D.; Roychowdhury, D. Petrel: Power and timing attack resistant elliptic curve scalar multiplier based on programmable GF(p) arithmetic unit. IEEE Trans. Circuits Syst. I-Regul. Pap. 2011, _[58, 1798–1812. [CrossRef]](http://dx.doi.org/10.1109/TCSI.2010.2103190)_ 49. Lee, J.-W.; Chung, S.C.; Chang, H.C.; Lee, C.Y. Efficient power-analysis-resistant dual-field elliptic curve cryptographic processor using heterogeneous dual-processing-element architecture. IEEE Trans. Very Large _[Scale Integr. (VLSI) Syst. 2014, 22, 49–61. [CrossRef]](http://dx.doi.org/10.1109/TVLSI.2013.2237930)_ 50. Mcivor, C.J.; Mcloone, M.; Mccanny, J.V. Hardware elliptic curve cryptographic processor over GF(p). _[IEEE Trans. Circuits Syst. I-Fundam. Theor. Appl. 2006, 53, 1946–1957. [CrossRef]](http://dx.doi.org/10.1109/TCSI.2006.880184)_ ----- _Sensors 2020, 20, 5148_ 19 of 19 51. Lai, J.Y.; Huang, C.-T. High-throughput cost-effective dual-field processors and the design framework for elliptic curve cryptography. IEEE Trans. Very Large Scale Integr. (VLSI) Syst. 2008, 16, 1567–1580. 52. Schinianakis, D.M.; Fournaris, A.P.L.; Michail, H.E.; Kakarountas, A.P.; Stouraitis, T. An RNS implementation [of an Fp elliptic curve point multiplier. IEEE Tran. Circuits Syst. I-Regul. Pap. 2009, 56, 1202–1213. [CrossRef]](http://dx.doi.org/10.1109/TCSI.2008.2008507) 53. Esmaeildoust, M.; Schinianakis, D.; Javashi, H.; Stouraitis, T.; Navi, K. Efficient RNS implementation of elliptic curve point multiplication over GF(p). IEEE Trans. Very Large Scale Integr. (VLSI) Syst. 2013, 21, [1545–1549. [CrossRef]](http://dx.doi.org/10.1109/TVLSI.2012.2210916) © 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution [(CC BY) license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/.) -----
{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC7571177, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/1424-8220/20/18/5148/pdf?version=1599703165" }
2,020
[ "JournalArticle" ]
true
2020-09-01T00:00:00
[ { "paperId": "716702e14dbe6b4d6e09af07061cb9c33fc6f847", "title": "A High-Speed FPGA Implementation of an RSD-Based ECC Processor" }, { "paperId": "db2de3062b4b056f9b881b20abd60f2e8b34d171", "title": "FPGA Implementation of High-Speed Area-Efficient Processor for Elliptic Curve Point Multiplication Over Prime Field" }, { "paperId": "83b1a49c0e8dbe9dd062e65077761fd02f4a13be", "title": "An Energy-Efficient Reconfigurable DTLS Cryptographic Engine for Securing Internet-of-Things Applications" }, { "paperId": "f6b33f8c49227a6e71c6f6688d3c0172901054bc", "title": "Redundant-Signed-Digit-Based High Speed Elliptic Curve Cryptographic Processor" }, { "paperId": "4adc45abf706cb05da2a596cba8dc79e26b48285", "title": "Elliptic Curve Lightweight Cryptography: A Survey" }, { "paperId": "9e480c26aec4d5ede4cc50abd0ceeade28b8e7ae", "title": "Remote Authentication Schemes for Wireless Body Area Networks Based on the Internet of Things" }, { "paperId": "c0a0830b27e1fbf48690176cd8c86ca30a1394ec", "title": "A Low Hardware Consumption Elliptic Curve Cryptographic Architecture over GF(p) in Embedded Application" }, { "paperId": "b15b766569553f9cf55eb2af7c54cc7d20763682", "title": "High-throughput multi-key elliptic curve cryptosystem based on residue number system" }, { "paperId": "904826e8b0f0ad6abe046ebbee57e4c22b9487fd", "title": "High performance hardware support for elliptic curve cryptography over general prime field" }, { "paperId": "b9ea0de9f99b0e7c61403fd6591ad08d388da0e5", "title": "On Emerging Family of Elliptic Curves to Secure Internet of Things: ECC Comes of Age" }, { "paperId": "8a063e91e3e563a63c2df23910860cf33e8bf8b6", "title": "Elliptic Curve Cryptography with Efficiently Computable Endomorphisms and Its Hardware Implementations for the Internet of Things" }, { "paperId": "e5ed11fcb1b41d952d7e099b60c3d0bc9a90c2c7", "title": "An Efficient and Flexible Hardware Implementation of the Dual-Field Elliptic Curve Cryptographic Processor" }, { "paperId": "e53d92bd1f2ddf080fd6088b32703225c04ff64f", "title": "Secure Signature-Based Authenticated Key Establishment Scheme for Future IoT Applications" }, { "paperId": "f32d4448498e05aac45c0d2dcdd23db80e61cc42", "title": "Low latency flexible FPGA implementation of point multiplication on elliptic curves over GF(p)" }, { "paperId": "c12a1fc6572b5cf2352bad74b7b6894a6aa00820", "title": "FPGA Based High Speed SPA Resistant Elliptic Curve Scalar Multiplier Architecture" }, { "paperId": "96578f7284c3d1898b9d822f8c7b82509ca39616", "title": "An Analysis of RFID Authentication Schemes for Internet of Things in Healthcare Environment Using Elliptic Curve Cryptography" }, { "paperId": "55ba62233d96b80e72f237858b1afa91e3827fe9", "title": "New Regular Radix-8 Scheme for Elliptic Curve Scalar Multiplication without Pre-Computation" }, { "paperId": "76d39926a82e57d4ad5596027245fdd1ea186414", "title": "Scalable Elliptic Curve Cryptosystem FPGA Processor for NIST Prime Curves" }, { "paperId": "a1c9b7885bdf5b2ba4f569982207ccf361d7765b", "title": "A secure ECC-based RFID authentication scheme integrated with ID-verifier transfer protocol" }, { "paperId": "c00c9d300d41246574e586176ec7a7b7224a4d1a", "title": "Securing the Internet of Things: A Standardization Perspective" }, { "paperId": "28121dee3de03f92d93ca4e6a5fb016e2de0bfb0", "title": "A Secure RFID Authentication Protocol for Healthcare Environments Using Elliptic Curve Cryptosystem" }, { "paperId": "363f8798b2eb393e393404e53a2c3311246e11b3", "title": "An Efficient RFID Authentication Protocol to Enhance Patient Medication Safety Using Elliptic Curve Cryptography" }, { "paperId": "591cc520dedf349550a274083797fd4435ba5a0f", "title": "Certificateless Remote Anonymous Authentication Schemes for WirelessBody Area Networks" }, { "paperId": "84b05e48ef50bab129b9170f5906616a8fcec917", "title": "An efficient mutual authentication RFID scheme based on elliptic curve cryptography" }, { "paperId": "47a42a726dc33098e1b44b220a9e5ca50913dc01", "title": "An FPGA implementation of NIST 256 prime field ECC processor" }, { "paperId": "b9ae6599edb62a858baaa0674e741509132245f3", "title": "Efficient RNS Implementation of Elliptic Curve Point Multiplication Over ${\\rm GF}(p)$" }, { "paperId": "218b4099a98a7c015ff5d9ad15af61de53be8fd6", "title": "High-speed high-security signatures" }, { "paperId": "45eda384030511576961120aeb85d393fbce15c8", "title": "Petrel: Power and Timing Attack Resistant Elliptic Curve Scalar Multiplier Based on Programmable ${\\rm GF}(p)$ Arithmetic Unit" }, { "paperId": "22204f2c9331275fdca046a6ca071580bdaa237b", "title": "Flexible Hardware Processor for Elliptic Curve Cryptography Over NIST Prime Fields" }, { "paperId": "7e9c64b2ef9b7ea35400cd1c7320475a612f9a0c", "title": "An RNS Implementation of an $F_{p}$ Elliptic Curve Point Multiplier" }, { "paperId": "b150b027371bca7e821aaa3abdfe23ff62237022", "title": "A Hardware Analysis of Twisted Edwards Curves for an Elliptic Curve Cryptosystem" }, { "paperId": "e14584885c368290bf057676d28323c94cda1421", "title": "Parallel crypto-devices for GF(p) elliptic curve multiplication resistant against side channel attacks" }, { "paperId": "3866633f00639437b8c776659b46f7884a8aa0d0", "title": "Twisted Edwards Curves Revisited" }, { "paperId": "a7697f50a24574d741deae1cf733fe8664db4e4e", "title": "Elliptic-Curve-Based Security Processor for RFID" }, { "paperId": "4dec5540a6d58d5ed694ba8820fd498d98839e0b", "title": "Elixir: High-Throughput Cost-Effective Dual-Field Processors and the Design Framework for Elliptic Curve Cryptography" }, { "paperId": "80b388f07313e609a0fbd7dadfbadc69ae3b653e", "title": "Protection" }, { "paperId": "6bd787497fb26ba4633f1e0ffc06d9b814234e6a", "title": "Twisted Edwards Curves" }, { "paperId": "7d366a6f74f2046ebddaf9baf664f2715dad92f9", "title": "Faster Addition and Doubling on Elliptic Curves" }, { "paperId": "93471564302ad9ac9fbe64e345e9888b50e096a5", "title": "A normal form for elliptic curves" }, { "paperId": "42938ecab8f3d7c5448098c016b99fa43d69502e", "title": "Hardware Elliptic Curve Cryptographic Processor Over$rm GF(p)$" }, { "paperId": "48febafd4e799a576235e75e6d1a9049426f5235", "title": "Curve25519: New Diffie-Hellman Speed Records" }, { "paperId": "eb7d79719f621bbb9fff85b08472de18a9ccf090", "title": "Reconfigurable Modular Arithmetic Logic Unit for High-Performance Public-Key Cryptosystems" }, { "paperId": "aba3474052dff485ca23c7262b078d110570b065", "title": "Weierstraß Elliptic Curves and Side-Channel Attacks" }, { "paperId": "bc533d2f27381d81d8e0cd3f445c54556e938816", "title": "The State of Elliptic Curve Cryptography" }, { "paperId": "ba624ccbb66c93f57a811695ef377419484243e0", "title": "New Directions in Cryptography" }, { "paperId": "48518658d5120d73518baabe74c549b1c50447b7", "title": "Area-Time Efficient Hardware Implementation of Modular Multiplication for Elliptic Curve Cryptography" }, { "paperId": "61cc545915203331c8bb67b603b3d20bc1dfefee", "title": "Cryptanalysis and Improvement of Anonymous Authentication for Wireless Body Area Networks with Provable Security" }, { "paperId": "dab5f9d4105074d2d587e40c6919715f8449a2d1", "title": "A Novel Efficient Pairing-Free CP-ABE Based on Elliptic Curve Cryptography for IoT" }, { "paperId": "1d152337a4b3126bf22d1648e64deea3481acde2", "title": "Edwards-Curve Digital Signature Algorithm (EdDSA)" }, { "paperId": "f3558c00bcbdddbcb7967b4318247ac2ab34d1a6", "title": "High-performance elliptic curve cryptography processor over NIST prime fields" }, { "paperId": "ae4ac89d4a347e5830ab4078f99f96d414b3947b", "title": "Efficient Power-Analysis-Resistant Dual-Field Elliptic Curve Cryptographic Processor Using Heterogeneous Dual-Processing-Element Architecture" }, { "paperId": "7fc91d3684f1ab63b97d125161daf57af60f2ad9", "title": "Elliptic Curves and Side-Channel Analysis" }, { "paperId": "a1cd437a924849d19e0713f042e45e79dc8b95a1", "title": "A public key cyryptosystem and signature scheme based on discrete logarithms" }, { "paperId": null, "title": "This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license" }, { "paperId": null, "title": "Draft-irtf-cfrg-eddsa-05, Internet Engineering Task Force, 2017" } ]
17,516
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/013d1b03a0f7dfb6675f903e52da4f3ae809f719
[ "Computer Science" ]
0.879323
Model for Simulation of Heterogeneous High-Performance Computing Environments
013d1b03a0f7dfb6675f903e52da4f3ae809f719
International Conference on High Performance Computing for Computational Science
[ { "authorId": "145921064", "name": "R. Mello" }, { "authorId": "1966773", "name": "Luciano José Senger" } ]
{ "alternate_issns": null, "alternate_names": [ "High Performance Computing for Computational Science (Vector and Parallel Processing)", "High Perform Comput Comput Sci (vector Parallel Process", "Int Conf High Perform Comput Comput Sci", "VECPAR" ], "alternate_urls": null, "id": "1eed7a6b-9e3b-431b-a1a1-c61ba1a420d6", "issn": null, "name": "International Conference on High Performance Computing for Computational Science", "type": "conference", "url": "https://link.springer.com/conference/vecpar" }
null
# Model for Simulation of Heterogeneous High-Performance Computing Environments Rodrigo Fernandes de Mello[1] and Luciano José Senger[2][⋆] 1 Universidade de São Paulo – Departamento de Computação Instituto de Ciências Matematicas e de Computação Av. Trabalhador Saocarlense, 400 Caixa Postal 668 CEP 13560-970 São Carlos, SP, Brazil mello@icmc.usp.br 2 Universidade Estadual de Ponta Grossa – Departamento de Informatica Av. Carlos Cavalcanti, 4748 CEP 84030-900 Ponta Grossa, PR, Brazil ljsenger@icmc.usp.br **Abstract. This paper proposes a new model to predict the process execution be-** havior on heterogeneous multicomputing environments. This model considers the process execution costs such as processing, hard disk acessing, message transmitting and memory allocation. A simulator of this model was developed which help to predict the execution behavior of processes on distributed environments under different scheduling techniques. Besides the simulator, it was developed a suite of benchmark tools in order to parameterize the proposed model with data collected from real environments. Experiments were conduced to evaluate the proposed model which used a parallel application executing on a heterogeneous system. The obtained results show the model ability to predict the actual system performance, providing an useful model for developing and evaluating techniques for scheduling and resource allocation over heterogeneous and distributed systems. ## 1 Introduction The evaluation of a computing system allows the analysis of its technical and economic feasibility, safety, performance and correct execution of processes. In order to evaluate a system, techniques that estimate its behavior on different situations are used. Such techniques provide numerical results which allow the comparison among different solutions for the same problem [1]. The evaluation of a computing system may use elementary or indirect techniques. The elementary ones are directly applied over the system, so it is necessary to have it previously implemented. The indirect ones allow the system evaluation before its implementation, what is relevant at the project phase [2–6]. The indirect techniques use mathematic models to represent the behavior of the main system components. Such models should be as similar as possible to the real problems, generating results for a good evaluation without being necessary to implement ⋆ The authors thank to William Voorsluys for improving the source code of the benchmark memo and the fundings from Capes and Fapesp Brazilian Foundations (under the process number 04/02411-9). ----- them [6]. Several models have been proposed for the evaluation of the execution time and the process delay. They consider the CPU consumption, the performance slowdown due to the use of the virtual memory [7] and the time spent with messages transmitted through the communication network [8]. Amir et al. [7] have proposed a method for job assignment and reassignment on cluster computing. This method uses a queuing network model to represent the slowdown caused by virtual memory usage. In such model the static memory m(j) used by the process is known. This model defines the load of each computer in accordance with the equation 1, where: L(t, i) is the load of computer i at the instant t; lc(t, i) is the CPU occupation; lw(t, i) is the amount of main memory used; rw(i) is maximum capacity of the main memory; β is the slowdown factor due to the use of virtual memory. Such factor increases the process response time, what consequently reflects in a lower final performance. This work attempts to minimize the slowdown by means of scheduling operations. L(t, i) = � lc(t, i) if lw(t, i) ≤ rw(i) (1) lc(t, i) ∗ β otherwise Mello et al. [9] have proposed improvements to the slowdown model by Amir et _al. [7]. This work includes new parameters which allow a better modelling of process_ slowdown. Such parameters are the capacity of CPU and memory, throughput for reading and writing on hard disk and delays generated by the use of the communication network. However, this model presents similar limitations to the work by Amir et al. [7], as it does not offer any resource to model, through equations, the delay caused by the use of virtual memory (represented in equation 1 by the parameter β), nor consider other delays of the process execution time generated by: message transmission, hard disk access and other input/output operations. The modeling of message transmission delays is covered by other works [8,10]. Culler et al. [8] have proposed the LogP model to quantify the overhead and the network communication latency among processes. The overhead and latency cause delays among processes which communicate. This model is composed of the following parameters: L which represents the high latency limit or delay incurred in transmitting a message containing a word (or a small number of words) from the source computer to a destination; o represents the overhead which is the time spent by processor to prepare a message for sending or receiving; g is the minimum time interval between consecutive message transmittion (sending or receiving); P is the number of processors. The LogP model assumes a finite capacity network with the maximum transmission defined by L/g messages. Sivasubramaniam [10] used the LogP model to propose a framework to quantify the overhead of parallel applications. In such framework are considered aspects such as the processing capacity and the communication system usage. This framework joins efforts of actual experiments and simulations to refine and define analytic models. The major limitation of this work is that it does not present a complete case study. The LogP model can be aggregated to the model by Amir et al. [7] and Mello et _al. [9], permitting to evaluate the process execution time and slowdowns considering_ the resources of CPU, memory and transmitted messages on the network. Although ----- unifying the models, they are still incomplete because do not consider the spatial and message generation probability distributions. Motivated by such limitations, some studies have been proposed [11,12]. Chodnekar et al. [11] have presented a study to characterize the probability distribution of messages on communication systems. In such work, the 1D-FFT and IS [13], Cholesky and Nbody [14], Maxflow [15], 3D-FFT and MG [16] parallel applications are evaluated executing on real environments. In the experiments, some informations have been captured such as the message sending and receiving moments, size of messages and destination. These informations were analyzed through statistic tools, and the spatial and message generation probability distributions obtained. The spatial distribution defines the frequency each process communicates with others. The message generation distribution defines the probability that each process sends messages to others. They have concluded that the most usual message generation probability distribution for parallel applications are the exponential, hyperexponential and Weibull. It has also been concluded that the spatial distribution is not uniform and there are different traffic patterns during the applications’ execution. In the most part of applications there is a process which receives and sends a large number of messages to the remainder processes (like a master for PVM – Parallel Virtual Machine – and MPI – Message Passing Interface – applications). The work also presents some features about message volume distribution, but there is not a precise analysis about the message size, overhead and latency. Vetter and Mueller [12] have studied the communication behavior of scientific applications using MPI (Message Passing Interface). This study quantifies the average volume of transmitted messages and their size. It has been concluded that in peer-topeer systems 99% of the transmitted messages vary from 4 to 16384 bytes. In collective calls this number varies from 2 to 256 bytes. This was combined with the studies on spatial and message generation distributions by Chodnekar et al. in [11] and to the LogP model [8] which allow the identification of overhead and communication latency in computing systems. By unifying these studies to the previously described slowdown models it is possible to evaluate the process behavior considering CPU, virtual memory and message transmittion. However, it is not possible to model voluntary delays in the execution of processes (generated by sleep calls) and accesses to hard disks. Motivated by the unification of the previously presented models, the aggregation of the applications’ voluntary delays and hard disk access, this paper presents the UniMPP (Unified Modeling for Predicting Performance) model. This model unifies the CPU consumption considered in the models by Amir et al. [7] and Mello et al. [9], the time spent to transmit messages modeled by Culler et al. [8] and Sivasubramaniam [13], the message volume and the spatial and message generation probability distributions by Chodnekar et al. [11], and Vetter and Mueller [12]. Experiments confirmed that this model can be used to predict the behavior of process execution on heterogeneous environments, once it generates the process response times very similar to the observed on real executions. This model was implemented in a simulator which is parameterized with system configurations (CPUs, main and virtual memories, hard disk thoughput and network capacity) and receives processes for execution. Distribution functions are used to char ----- acterize the process CPU, memory, hard disk and network occupations. The simulator also generates new processes according to a probability distribution function, allowing to evaluate different scheduling and load balancing policies without needing the real execution. As presented before, the simulator needs to be parameterized with the actual system configurations. For this purpose, a suite of benchmark tools was developed to collect informations such as the capacity of CPUs in MIPS (millions of instructions per second), the main and virtual memory behavior under a progressive occupation (this generates delay functions), the hard disk throughput in reading and writing operations (in MBytes per second) and the network delay (considering the overhead and latency in seconds). The main contribution of this work is the UniMPP model which can be used with the simulator allied to the benchmark tools to predict the process execution time on heterogeneous environments. The simulator is prepared to receive new scheduling and load balancing policies and evaluate them using different workload models [17]. This paper is divided into the following sections: 2 The model; 3 Parameterization; 4 Model Validation; 5 Conclusions and References. ## 2 The Model Motivated by the unification of the virtual memory slowdown models [7,9], by the models of delays in process execution caused by messages transmission [8, 10], by studies about spatial and message generation probability distributions [11], by the slowdown caused in main and virtual memory ccupation, by the definition of voluntary delay and access to hard disks, the UniMPP (Unified Modeling for Predicting Performance) model has been designed. These models are presented in the previous section. Unifying the ideas of each model and adding voluntary delays and hard disk access, we have defined a new model to predict the execution behavior of processes running on heterogeneous computers. By using this model, researchers can evalutate different techniques such as scheduling and load balancing without being necessary to run an application on an real environment. In this model, a process pj arrives at the system, following a probability distribution function, at the instant aj. Such process is started by the computer ci. Each computer maintains its queue qi,t of processes at the instant t. In this model, every computer ci is composed of the sextuple {pci, mmi, vmi, dri, dwi, loi}, where: pci is the total computing capacity of each computer measured in instructions per unit of time; mmi is the total main memory; vmi is the total virtual memory capacity; dri is the hard disk reading throughput; dwi is the hard disk writing throughput; loi is the time spend in sending and receiving messages. In the UniMPP, each process is represented by the sextuple {mpj, smj, pdfdmj, pdfdrj, pdfdwj, pdfnetj}, where: mpj represents the processing consumption; smj is the amount of static memory allocated by the process; pdfdmj is the probability distribution function used to represent the dynamic memory occupation; pdfdrj is the probability distribution function used to represent the hard disk reading; pdfdwj is the probability distribution function used to represent the hard disk writing; pdfnetj is the ----- probability distribution function used to represent the sending and receiving operations on communication system. Having formally defined computers and processes, equations were defined to obtain the process response time and delay. The first equation (equation 2) presents the response time (T Epj,ci) of a process pj being executed in a computer ci, where the total computing capacity pci of ci and the processing consumption of pj should be represented by the same metric, such as MI (millions of instructions when the capacity of processors was obtained in Mips – Millions of instructions per second) or MF (millions of float-point instructions when the capacity of processors was obtained in Mflops – Millions of float-point instructions per second). T Epj,ci = [mp][j] (2) pci The equation 2 presents a calculation method for the execution time of a process under ideal conditions, in which there is no competition nor delays caused by the memory and input/output usage. The work by Amir et al. [7] presents a more adequate equation in which, from the moment that the virtual memory starts to be used, there is a delay in the process execution. These authors use a constant delay in their equations. However, by using the benchmark tools described in section 3, it was observed that there are limitations in their model, since the performance slowdown is linear during the main memory usage and exponential from the moment the virtual memory starts to be used. T EMpj,ci = T Epj,ci (1 + α) (3) ∗ The Amir’s performance model does not consider this linear performance slowdown caused by the use of the main memory and considers a constant factor for the performance slowdown caused by the use of the virtual memory when, in fact, this slowdown is exponential. The UniMPP models the process performance slowdown generated by the use of main and virtual memories, by the equation 3, where α represents a percentage obtained from a delay function and T E is presented in equation 2. This delay function is generated by a benchmark tool (section 3) where in the x axis is the mem− ory occupation up to use all the virtual memory and in the y axis is the α value (the − slowdown imposed in the process execution by the memory occupation). A model which considers the process execution slowdown caused by the use of main and virtual memories become more adequate, however, it does not allow the precise quantification of the total execution time of processes which perform input and output operations to the hard disk. For this reason, experiments have been conduced and equations developed to measure the delays generated by accesses to hard disk. The equation 4 models the process delay generated by reading operations from hard disk, where: nr represents the number of reading accesses; bsize represents the data buffer size; dri represents the throughput capacity for reading accesses from hard disk; and wtdrk represents the waiting time for using the resource. SLDRpj,ci = nr � k=1 bsizek + wtdrk (4) dri ----- The hard disk writing delay is defined by equation 5, where: nw represents the number of writing accesses; bsize represents the data buffer size to be written; dwi is the throughput capacity for writing accesses in hard disk; wtdw is the waiting time for using the resource. SLDWpj,ci = nw � k=1 bsizek + wtdwk (5) dwi In addition to the delays caused by memory usage and input/output to hard disks, there are delays generated by sending and receiving messages on communication systems. Such delays vary according to the network bandwidth, latency and overhead of communication protocols [18–20]. The protocol latency involves the transmission time on communication system, which vary in accordance with the message size and control messages generated by the protocol [18–20]. The protocol overhead is the time involved for packing and unpacking messages for transmission. This time also varies according to the messages size [18–20]. The delay for sending and receiving messages is defined by equation 6, where: nm represents the number of sent and received messages; θs,k, described in equation 7 is the time used for sending and receiving messages on communication system, not considering the wait for resources; and wtnk represents the wait time, the queue time, to send or receive a message, when the resource is busy. The components of equation 7 are: os,k overhead, which when multiplied by two allows the quantification of packing time (by the sender) and the unpacking time (by the receiver) of a message; and ls,k is the latency to transmit a message. SLNpj,ci = nm � θs,k + wtnk (6) k=1 θs,k = 2 ∗ os,k + ls,k (7) Aiming the unification of all previously described delay models, it is proposed the equation 8, which allows the definition of the response time (the prediction of this time in a real enviroment) of a process pj in a computer ci, where: lz is the process voluntary delay generated by the system calls sleep. In the case of load transference (that is, process migration) the communication channels may modify their behaviors and perform a higher or lower number of input/output operations (a process migrating to a computer where there are others which it communicates, reduces the latency and overhead because does not use the communication system, although it can overload the CPU). The equation 9 is the response time of a process pj transferred among n computers. SLpj = SLpj,ci = T EMpj,ci + SLDRpj,ci + SLDWpj,ci + SLNpj,ci + lz (8) SLpj = n � SLpj,ck (9) k=1 ----- The UniMPP model unifies the concepts from models by Amir et al. [7], Mello et _al. [9] and Culler et al. [8] and extends them by adding voluntary delay equations and_ the time for reading and writing accesses to hard disks. In addition, based on experiments, this work proposes new equations to define the main and virtual memory slowdown. By these equations, it was observed that the slowdown is linear when using the main memory, and exponential using the virtual. Such experiments were carried though using the benchmark tools from section 3. This model allows studies of scheduling, load balancing algorithms and prediction of process response times on heterogeneous environments. The proposed model has been implemented in a simulator, named SchedSim [3], which allows other researchers to conduct related studies. Such simulator is implemented in Java language and uses the object oriented concepts that simplify its extension and functionality additions. The simulator is parameterized with system configurations (CPUs, main and virtual memories, hard disk thoughput and network capacity) and receives processes for execution. It generates new processes according to a probability distribution function, allowing to evaluate different scheduling and load balancing policies without needing the real execution. ## 3 Parameterization In order to parameterize the SchedSim simulator using real environment characteristics, a suite of benchmark tools[4] was developed. These tools measure the capacity of CPU, reading and writing hard disk throughput and the message transmission delays. Such tools evaluate these characteristics until they reach a minimum sample size based on the central limit theorem, allowing to apply statistical summary measures such as confidence interval, standard deviation and average [21]. This suite is composed by the following tools: 1. mips: it measures the capacity of a processor, in millions of instructions per second. This tool uses a bench() function implemented by Kerrigan [22]; 2. memo: it creates child processes until all main and virtual memories are filled up, measuring the delays of the context switches among processes. The child processes only allocate the memory and then sleep for some seconds, thus it does not consider the processor usage; 3. discio: it measures the average writing throughput (buffered and unbuffered) and the average reading throughput in local storage devices (hard disks) or remote storage devices (via network file systems); 4. net - it is composed of two applications, a customer and a server, which allow the evaluation of the time spent to send and receive messages over communication networks (based on the equation 7). 3 Source code available at http://www.icmc.usp.br/˜mello/outr.html 4 Benchmark – source code available at http://www.icmc.usp.br/˜mello/outr.html ----- ## 4 Validation In order to validate the proposed model, executions of a parallel application developed in PVM (Parallel Virtual Machine) [23] in a scenario composed of two homogeneous computers have been considered. This adopted application is composed of a master and worker processes. The master process launches one worker on each computer and defines three parameters: the problem size, that is, the number of mathematic operations executed to solve an integral (eq. 10) defined between two points a and b using the trapezium rule [24,25], the number of bytes that will be transferred over the network and recorded in the hard disk. These workers are composed of four stages: message receiving, processing, writing into the hard disk and message sending. The message exchange happens between master and worker at the beginning and at the end of the workers’ execution. The workers are instrumented to account the time consumed in operations. � b 2 sin x + e[x] (10) ∗ a Scenario details are presented on the table 1 and they have been obtained with the benchmark suite. A message size of 32 bytes has been considered for the benchmark _net. The table 2 presents the slowdown equations generated by using main and virtual_ memories, respectively, on the computers c1 and c2. Such equations have been obtained through the experiments with the benchmark memo. The linear format of the equations is used when the main memory is not completely filled up, for instance, in the case of computers c1 and c2 not exceed 1 Gbyte of its memory capacity. After exceeding such limit, the virtual memory is used and the delay is represented by the exponential funtion. **Table 1. System details** **Resource** c1 c2 CPU (Mips) 1145.86 1148.65 Main memory (Mbytes) 1Gbyte 1Gbyte Virtual memory (Mbytes) 1Gbyte 1Gbyte Disk writing throughput (MBytes/seg) 65.55 66.56 Disk reading throughput (MBytes/seg) 76.28 75.21 Overhead + Latency (seconds) 0.000040 The experiment results are presented in the table 3. It may be observed that the error among the curves is low, close to zero. Ten experiments have been conduced for different numbers of applications, each one composed of two workers executing on two computers. Such experiment was used to saturate the capacity of all computing resources of the environment. The figure 1 shows the experiment and simulation results. ----- **Table 2. Memory slowdown functions for computers c1 and c2** **Memory** **Regression** **Equation** R[2] Main memory Linear y = 0.0012x 0.0065 0.991 − Main and Virtual memory Exponential y = 0.0938 e[0][.][0039][x] 0.8898 ∗ **Table 3. Simulation results for computers c1 and c2** **Processes Actual Average Predicted Error (%)** 10 151.40 149.51 0.012 20 301.05 293.47 0.025 30 447.70 437.46 0.022 40 578.29 573.58 0.008 50 730.84 714.92 0.021 60 856.76 862.52 0.006 70 1002.10 1012.17 0.009 80 1147.44 1165.24 0.015 90 1245.40 1318.37 0.055 100 1396.80 1471.88 0.051 1600 1400 1200 1000 800 600 400 200 0 10 20 30 40 50 60 70 80 90 100 Number of Processes **Fig. 1. Actual and predicted average response times for computers c1 and c2** The simulation obtained results show the model ability to reproduce the real system behavior. It is important to notice the increasing of the prediction errors when the system runs a number of processes between 90 and 100. The real executions, using 90 and 100 processes, overloaded the computers and some processes were killed by the PVM system. The premature stopping of processes ----- (at about 5 processes where killed) decreases the computer’s load, justifiyng the model prediction error. The simulator was used aiming to predict the system behavior considering a number of processes greater than the number of processes executed by PVM. After experiments in an homogeneous system, a new environment composed of heterogeneous computers were parameterized using the benchmark tools. In this environment, it was executed the same application, which computes an integral function between two points using the trapezium rule. The features of the heterogeneous computers are presented in the table 4. **Table 4. System details** **Resource** c3 c4 CPU (Mips) 927.55 1600.40 Main memory (Mbytes) 256 512 Virtual memory (Mbytes) 400 512 Disk write throughput (MBytes/seg) 47.64 15.99 Disk read throughput (MBytes/seg) 41.34 32.55 Overhead + Latency (seconds) 0.000056924 The tables 5 and 6 present the slowdown equations, obtained by the memo benchmarking, considering the main and virtual memory usage. **Table 5. Memory slowdown functions for computer c3** **Memory** **Regression** **Equation** R[2] Main memory Linear y = 0.0018x 0.0007 0.9998 − Main and Virtual memory Exponential y = 0.7335 e[0][.][0097][x] 0.8856 ∗ **Table 6. Memory slowdown functions for computer c4** **Memory** **Regression** **Equation** R[2] Main memory Linear y = 0.0018x 0.0035 0.9821 − Main and Virtual memory Exponential y = 0.0924 e[0][.][0095][x] 0.8912 ∗ The experiment results are presented in the table 7. The error values obtained comparing the simulated and the actual execution time values are close to 0, allowing to confirm the model ability in predicting real executions. The figure 2 shows the experiment and simulation results. ----- **Table 7. Simulation results for computers c3 and c4** **Processes Actual Average** **Predicted Error (%)** 10 153.29 152.38 0.0059 20 306.63 304.66 0.0064 30 457.93 457.46 0.0010 40 593.66 610.78 0.0280 50 760.02 764.65 0.0060 60 892.29 918.97 0.0290 70 1040.21 1074.18 0.0316 80 1188.14 1230.75 0.0346 90 1333.70 1388.14 0.0392 100 1488.97 1572.22 0.0529 1600 1400 1200 1000 800 600 400 200 0 10 20 30 40 50 60 70 80 90 100 Number of Processes **Fig. 2. Actual and predicted average response times for computers c3 and c4** When a number at about 60 processes are running, some problems were observed, due to PVM process management. It was observed that using some computers with less processing power, PVM started to kill processes earlier, when running more than 60 processes. These problems explain the difference between the actual and the simulated time values and the increasing in predicting errors. The experiments presented in this section validate the model used by the simulator. The model and the simulator is able to predict the behavior of a real and dynamic system, modelling distinct parallel applications which solve problems from different areas, such as: aeronautics, fluid dynamics and geoprocessing. Thus, the system behavior can be predicted earlier, in project phase, minimizing the development costs. ----- ## 5 Conclusions Several models have been proposed to measure the response time of processes in computing systems [7,9]. Such models have presented some contributions, considering that the virtual memory occupation causes delays in process executions [7,9], as well as delays generated by the message transmissions on communication systems [8, 10]. Nevertheless, such models do not unify all possible delays of a process execution. Motivated by such limitations, this work has presented a new unified model to predict the applications’ execution running on heterogeneous distributed envionments. This model considers the process execution time in accordance with the processing, accesses to hard disk, message transmissions on communication networks, main and virtual memory slowdowns. This work has contributed by modeling the delays in reading and writing accesses to hard disks and presenting a new technique which uses equations to represent the delays generated by the main and virtual memory usage. This has complemented studies by Amir et al. [7] and Mello et al. [9], which consider a constant delay. In addition it was developed a simulator of the proposed model which can be used to predict the execution of applications on heterogeneous multicomputing environments. Such simulator has been developed considering extensions such as the design of new scheduling and load balancing policies. This simulator is licensed under GNU/GPL which allows its broad use by the researchers interested in developing and evaluating resource allocation techniques. In order to complement this simulator and allow its parameterization using real environment information, a suite of benchmark tools was developed and is also available under the GNU/GPL license. In order to validate the simulator, a parallel application was implemented, simulated and executed on a real environment. It was observed that the percentage error obtained between the actual and the predicted execution times was lower than 1%, what confirms the accuracy of the proposed model to predict the application execution on heterogeneous multicomputing environments. ## References 1. de Mello, R.F.: Proposta e Avaliacão de Desempenho de um Algoritmo de Balanceamento de Carga para Ambientes Distribuídos Heterogêneos Escaláveis. PhD thesis, SEL-EESC-USP (2003) 2. et. al, E.L.: Quantitative System Performance: Computer System Analysis Using Queueing Networks Models. Prentice Hall (1984) 3. et. al, P.B.: A Guide to Simulation. Spring-Verlag (1987) 4. Kleinrock, L.: Queueing Systems - Volume II: Computer Applications. John Wiley & Sons (1976) 5. Lavenberg, S.S.: Computer Performance Modeling Handbook. Academic Press (1983) 6. Jain, R.: The Art of Computer Systems Performance Analysis: Techniques for Experimental Design, Measurements, Simulation and Modeling. John Wiley & Sons (1991) 7. Amir, Y.: An opportunity cost approach for job assignment in a scalable computing cluster. IEEE Transactions on Parallel and Distributed Systems 11(7) (2000) 760–768 ----- 8. Culler, D.E., Karp, R.M., Patterson, D.A., Sahay, A., Schauser, K.E., Santos, E., Subramonian, R., von Eicken, T.: LogP: Towards a realistic model of parallel computation. In: Principles Practice of Parallel Programming. (1993) 1–12 9. et. al, R.F.M.: Analysis on the significant information to update the tables on occupation of resources by using a peer-to-peer protocol. In: 16th Annual International Symposium on High Performance Computing Systems and Applications, Moncton, New-Brunswick, Canada (2002) 10. Sivasubramaniam, A.: Execution-driven simulators for parallel systems design. In: Winter Simulation Conference. (1997) 1021–1028 11. et. al, S.C.: Towards a communication characterization methodology for parallel applications. In: Proceedings of the 3rd IEEE Symposium on High-Performance Computer Architecture (HPCA ’97), IEEE Computer Society (1997) 310 12. Vetter, J.S., Mueller, F.: Communication characteristics of large-scale scientific applications for contemporary cluster architectures. J. Parallel Distrib. Comput. 63(9) (2003) 853–865 13. Sivasubramaniam, A., Singla, A., Ramachandran, U., Venkateswaran, H.: An approach to scalability study of shared memory parallel systems. In: Measurement and Modeling of Computer Systems. (1994) 171–180 14. Singh, J.P., Weber, W., Gupta, A.: Splash: Stanford parallel applications for shared-memory. Technical report (1991) 15. Anderson, R.J., Setubal, J.C.: On the parallel implementation of goldberg’s maximum flow algorithm. In: Proceedings of the fourth annual ACM symposium on Parallel algorithms and architectures, San Diego, California, United States, ACM Press (1992) 168–177 16. Bailey, D.H., Barszcz, E., Barton, J.T., Browning, D.S., Carter, R.L., Dagum, D., Fatoohi, R.A., Frederickson, P.O., Lasinski, T.A., Schreiber, R.S., Simon, H.D., Venkatakrishnan, V., Weeratunga, S.K.: The NAS Parallel Benchmarks. The International Journal of Supercomputer Applications 5(3) (1991) 63–73 17. Feitelson, D.G., Rudolph, L., Schwiegelshohn, U., Sevcik, K.C., Wong, P.: Theory and Practice in Parallel Job Scheduling. In: Job Scheduling Strategies for Parallel Processing. Volume 1291. Springer (1997) 1–34 Lect. Notes Comput. Sci. vol. 1291. 18. Chiola, G., Ciaccio, G.: A performance-oriented operating system approach to fast communications in a cluster of personal computers. In: In Proc. 1998 International Conference on Parallel and Distributed Processing, Techniques and Applications (PDPTA’98). Volume 1., Las Vegas, Nevada (1998) 259–266 19. Chiola, G., Ciaccio, G.: (Gamma: Architecture, programming interface and preliminary benchmarking) 20. Chiola, G., Ciaccio, G.: Gamma: a low cost network of workstations based on active messages. In: Proc. Euromicro PDP’97, London, UK, January 1997, IEEE Computer Society (1997) 21. W.C.Shefler: Statistics: Concepts and Applications. The Benjamin/Cummings (1988) 22. Kerrigan, T.: Tscp benchmark (2004) 23. Beguelin, A., Gueist, A., Dongarra, J., Jiang, W., Manchek, R., Sunderam, V.: PVM: Parallel Virtual Machine: User’s Guide and tutorial for Networked Parallel Computing. MIT Press (1994) 24. Pacheco, P.S.: Parallel Programming with MPI. Morgan Kaufmann Publichers (1997) 25. Burden, R.L., Faires, J.D.: Análise Numérica. Thomson (2001) -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-540-71351-7_9?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-540-71351-7_9, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "http://vecpar.fe.up.pt/2006/programme/papers/5.pdf" }
2,006
[ "JournalArticle" ]
true
2006-06-10T00:00:00
[]
8,167
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/013f3dbde05a9be858502c6cc5e85ea5ebae5ab6
[ "Computer Science" ]
0.879892
Collective hybrid intelligence: towards a conceptual framework
013f3dbde05a9be858502c6cc5e85ea5ebae5ab6
International Journal of Crowd Science
[ { "authorId": "145660977", "name": "Morteza Moradi" }, { "authorId": "2056161260", "name": "Mohammad Moradi" }, { "authorId": "2065706784", "name": "Farhad Bayat" }, { "authorId": "2130483840", "name": "Adel Nadjaran Toosi" } ]
{ "alternate_issns": null, "alternate_names": [ "Int J Crowd Sci" ], "alternate_urls": [ "http://www.emeraldinsight.com/loi/ijcs" ], "id": "1c3fd765-9228-4d18-ba06-e2927f0d352d", "issn": "2398-7294", "name": "International Journal of Crowd Science", "type": "journal", "url": "http://www.emeraldgrouppublishing.com/services/publishing/ijcs/index.htm" }
PurposeHuman or machine, which one is more intelligent and powerful for performing computing and processing tasks? Over the years, researchers and scientists have spent significant amounts of money and effort to answer this question. Nonetheless, despite some outstanding achievements, replacing humans in the intellectual tasks is not yet a reality. Instead, to compensate for the weakness of machines in some (mostly cognitive) tasks, the idea of putting human in the loop has been introduced and widely accepted. In this paper, the notion of collective hybrid intelligence as a new computing framework and comprehensive.Design/methodology/approachAccording to the extensive acceptance and efficiency of crowdsourcing, hybrid intelligence and distributed computing concepts, the authors have come up with the (complementary) idea of collective hybrid intelligence. In this regard, besides providing a brief review of the efforts made in the related contexts, conceptual foundations and building blocks of the proposed framework are delineated. Moreover, some discussion on architectural and realization issues are presented.FindingsThe paper describes the conceptual architecture, workflow and schematic representation of a new hybrid computing concept. Moreover, by introducing three sample scenarios, its benefits, requirements, practical roadmap and architectural notes are explained.Originality/valueThe major contribution of this work is introducing the conceptual foundations to combine and integrate collective intelligence of humans and machines to achieve higher efficiency and (computing) performance. To the best of the authors’ knowledge, this the first study in which such a blessing integration is considered. Therefore, it is believed that the proposed computing concept could inspire researchers toward realizing such unprecedented possibilities in practical and theoretical contexts.
## IJCS 3,2 198 Received 26 March 2019 Revised 3 June 2019 Accepted 11 July 2019 International Journal of Crowd Science Vol. 3 No. 2, 2019 pp. 198-220 EmeraldPublishingLimited 2398-7294 [DOI 10.1108/IJCS-03-2019-0012](http://dx.doi.org/10.1108/IJCS-03-2019-0012) www.emeraldinsight.com/2398-7294.htm # Collective hybrid intelligence: towards a conceptual framework ## Morteza Moradi ### Department of Electrical Engineering, University of Zanjan, Zanjan, Iran ## Mohammad Moradi ### Young Researchers and Elite Club, Qazvin, Islamic Republic of Iran ## Farhad Bayat ### Department of Electrical Engineering, University of Zanjan, Zanjan, Iran, and ## Adel Nadjaran Toosi ### Faculty of Information Technology, Monash University, Melbourne, Australia Abstract Purpose – Human or machine, which one is more intelligent and powerful for performing computing and processing tasks? Over the years, researchers and scientists have spent significant amounts of money and effort to answer this question. Nonetheless, despite some outstanding achievements, replacing humans in the intellectual tasks is not yet a reality. Instead, to compensate for the weakness of machines in some (mostly cognitive) tasks, the idea of putting human in the loop has been introduced and widely accepted. In this paper, the notion of collective hybrid intelligence as a new computing framework and comprehensive. Design/methodology/approach – According to the extensive acceptance and efficiency of crowdsourcing, hybrid intelligence and distributed computing concepts, the authors have come up with the (complementary) idea of collective hybrid intelligence. In this regard, besides providing a brief review of the efforts made in the related contexts, conceptual foundations and building blocks of the proposed framework are delineated. Moreover, some discussion on architectural and realization issues are presented. Findings – The paper describes the conceptual architecture, workflow and schematic representation of a new hybrid computing concept. Moreover, by introducing three sample scenarios, its benefits, requirements, practical roadmap and architectural notes are explained. Originality/value – The major contribution of this work is introducing the conceptual foundations to combine and integrate collective intelligence of humans and machines to achieve higher efficiency and (computing) performance. To the best of the authors’ knowledge, this the first study in which such a blessing integration is considered. Therefore, it is believed that the proposed computing concept could inspire researchers toward realizing such unprecedented possibilities in practical and theoretical contexts. Keywords Crowdsourcing, Human computation, Autonomous control, Collective machine intelligence, Human–machine collaboration, Hybrid intelligence Paper type Conceptual paper © Morteza Moradi, Mohammad Moradi, Farhad Bayat and Adel Nadjaran Toosi. Published in International Journal of Crowd Science. Published by Emerald Publishing Limited. This article is published under the Creative Commons Attribution (CC BY 4.0) licence. Anyone may reproduce, distribute, translate and create derivative works of this article (for both commercial and non-commercial purposes), subject to full attribution to the original publication and authors. The full terms of this licence may be seen at http://creativecommons.org/licences/by/4.0/legalcode ----- 1. Introduction The concept of computation has evolved over the years with respect to real-world requirements and technological advancements (Mahoney, 1988; Copeland, 2000). In this regard, many computing paradigms have been introduced so far, such as Kephart and Chess (2003), Bargiela and Pedrycz (2016); and Shi et al. (2016). In addition to the infrastructural necessities of any computing process, an old dream in this context is the realization of full autonomy in computing, decision making and similar intellectual processes. Achieving this level of automation, in essence, needs to add intelligence to the process in some way. In other words, to be able to come up with (super) human-level decisions, an autonomous (computing/control) system should be equipped with adequate infrastructural facilities, computing power and intelligence (Feigenbaum, 2003; Nilsson, 2005; Cassimatis, 2006). Nowadays, thanks to the availability of powerful hardware, advanced processing components, inexpensive data storage equipment, sophisticated algorithms and so on, the major challenge in achieving such dreamy machines is the lack of sufficient human-level intelligence. Although many efforts have been spent in this direction (Decker, 2000; Hibbard, 2001; Zadeh, 2008; Bundy, 2017), replacing human intelligence by machines’ has not yet been realized literally. On the other side, leveraging humans’ brainpower to improve machines’ performance has become an efficient approach during recent years (Weyer et al., 2015; Ofli et al., 2016; Chang et al., 2017). Therefore, one may think that instead of trying to build machines to take the place of humans, it would be better to establish a foundation to facilitate joint work of humans and machines to tackle large-scale problems. Although hybrid intelligence paradigm introduces some opportunities to take benefits of human and machine intelligence (Huang et al., 2017), lack of a reference model/general architecture to adhere to its principles causes some non-uniformity. Moreover, adhering to this approach may not warrant taking advantages of available possibilities. On the other side, volunteer computing (Beberg et al., 2009) as an interesting and working idea mainly focuses on leveraging computing resources of the participants, e.g. their PCs and browsers (Fabisiak and Danilecki, 2017). One can apparently observe that despite the huge available opportunities to synthesize various capabilities of humans and machines, absence of a comprehensive approach to make the most of them is an obvious drawback. In other words, any framework/mechanism which could integrate intelligence and computational resources of human agents and machine entities in different levels could come up with the best of both worlds. In this respect, with the aim of studying previous efforts and current status of similar researches, a brief overview is conducted. Then, to take the efficiency of such human–machine cooperation and collaboration to an unprecedented level, the conceptual architecture of a new evolutionary computing/automation framework, entitled collective hybrid intelligence (CHI), is proposed and its related issues and considerations are discussed in detail. According to the current findings and achievements as the building blocks of the introduced solution, it is expected that the proposed concept could extend borders of the researches in the field to increase efficacy of human–machine synergy in performing computing tasks. The rest of this paper is organized as follows. At first, an overview of the context and intention of the paper is provided in Section 2. The background and preliminary concepts are briefly overviewed in section 3. The concept of Collective Hybrid Intelligence, its fundamentals, benefits, challenges and realization models are discussed in Section 4. Finally, to clearly describe and discuss how typical systems of this kind (that is constructed based on the proposed framework of CHI) may work in different application domains, three example scenarios are delineated in Section 5. ## Collective hybrid intelligence 199 ----- ## IJCS 3,2 200 2. Big picture Undoubtedly, computers – i.e. smart/intelligent machines – are among the most important and influential inventions of the modern era. Their ever-increasing capabilities in handling a wide variety of computational problems have made computers the artificial superheroes of all times. Over the years and with thanks to the outstanding progress in hardware technology, computing paradigms, machine learning and artificial intelligence, the machines have received an overestimated (and even exaggerated) applause. Affected by science-fiction stories and movies, the public though may be concerned of an early domination of machines over human race. In this regard, defeating the world chess champion by a computer (i.e. IBM’s Deep Blue) in 1997[1] and beating a professional Go player by DeepMind’s AlphaGo in 2015[2] were convincing evidences for robophobics to conclude that machines finally win over humans and they will be coronated in the near future. Despite many advancements, the truth is that even latest machines are not jack of all trades and there are many battlefields in which humans can defeat a billion bucks machine[3]. In other words, when it comes to cognitive and intelligent tasks, current machines are not stronger than humans at all (for some example, see Fleuret et al., 2011; Stabinger et al., 2016; Dodge and Karam, 2017). Such facts have driven the research community to rethink the computational paradigms by putting humans in the loop. In addition to compensate the machine’s weaknesses in some ways, human agents could provide human-level training data for machine learning purposes (Zhong et al., 2015; Yang et al., 2018). Because of effectiveness of such cooperation, the (mostly fictional) war between humans and machines has turned into a synergistic collaboration. However, this is not the final destination for the long journey of achieving super intelligence and computational capabilities. The authors believe that the last step before realization of super human intelligence (or artificial super intelligence) is to make the most of current neglected potentiality that humans and machines can present in a cooperative way. In the rest of the paper, roles of both parties as the building blocks of a new comprehensive computational concept, entitled Collective Hybrid Intelligence, are investigated. As concluding remark, throughout the paper the term machine refers to any non-human and intelligent entity including computers, programs, robots, etc. 3. Background 3.1 Collective human intelligence Human is an integral part of any computing process; however, over the years his role, position and responsibilities have been changed and evolved. User, operator, supervisor and collaborator are main categories that could reflect humans’ role in such processes (Folds, 2016), “For thousands of years, humans’ intelligence, problem solving and reasoning abilities presented numerous game-changing ideas and inventions to make the life easier (Sarathy, 2018). Nonetheless, handling sophisticated and complicated situations and issues needed something more than a genius or intelligent decision-maker. Such a fact probably was sparked the motivation to establish the first councils and organized group decisionmaking bureaus (Burnstein and Berbaum, 1983; Maoz, 1990; Zanakis et al., 2003; Buchanan and O’Connell, 2006). In the age of computers, for years humans were mostly consumers while a minority group of supervisors were in charge of keeping the machines up and running. In fact, those days can pessimistically be referred to as human-independent computing or machine-driven computing era. Fortunately, many things have changed forever by introduction of crowdsourcing concept (Howe, 2006). The underlying idea of this revolutionary paradigm ----- was taking advantages of humans’ collective abilities and efforts to provide more efficient performance. Thanks to its potentials, the initial concept has been soon after widely accepted and evolved into a working decision making and problem-solving strategy (Brabham, 2008; Guazzini et al., 2015; Yu et al., 2018). Although the idea was not an essentially new one[4]; its formulation and attitudes towards leveraging wisdom of crowds and collective human intelligence to cope with problems have made it a popular approach. Based upon the preliminary idea, several computing concepts such as human computation (Von Ahn, 2008), social computing (Wang et al., 2007) and community intelligence (Luo et al., 2009) have been introduced. Within the recent decade, putting the human in the loop of computing, decision-making (Chiu et al., 2014), ideation (Huang et al., 2014; Schemmann et al., 2016) and similar processes have gained momentum so that one can witness a wide variety of application domains that taking benefits of humans’ intelligence and problem-solving potentials. Nonetheless, there is not any serious intention to completely replace machines with humans because this is impossible at all. Instead, the major goal of human-based computation is to compensate machines’ deficiency in performing some specific tasks and processes including cognitive and intelligence-intensive ones (Wightman, 2010; Quinn and Bederson, 2011). For example, outsourcing image labeling tasks to the people can provide more accurate efficient and in some cases less-expensive results than relying on machines (Nowak and Rüger, 2010). In other words, when it comes to the situation in which human-level intelligence is needed, regarding the current machines’ state, recruiting human participant is the silver bullet. Further, one can expect more insightful and elaborated answers through involving experts in the form of expert crowdsourcing (Retelny et al., 2014) (Figure 1). Such benefits, by the way, will not come without cost because employment and management of a remarkable number of users in crowdsourcing projects can be a pain in the neck. Therefore, there is need for elaborated and reliable infrastructure, managerial supervision and workflows. The good news in this context is that availability of technological support and platforms such as Amazon Mechanical Turk (AMT)[5], TurkPrime (Litman et al., 2017) and Figure-Eight[6] (formerly Crowdflower) have made conducting a crowdsourcing campaign as simple as posting a blog. 3.2 Collective machine intelligence Speaking about artificial intelligence, one of the first things will prompt in the mind is sciencefiction movies. Despite the remarkable advancements in the field (Dai and Weld, 2011; ## Collective hybrid intelligence 201 Figure 1. Simplified schematic of CHI workflow ----- ## IJCS 3,2 202 Pan, 2016; Makridakis, 2017; Lu et al., 2018; Li et al., 2018) and predictions concerned about future of AI (Del Prado, 2015; Müller and Bostrom, 2016; Russell, 2017), there is a long unpaved way to the age of predomination of machines which are capable of controlling everything. Therefore, one should not be concerned of becoming slave or even agent of an artificial entity in the near future. Things are far different in the real world and (perhaps) the major issue in the field is how to make the most of machines to be more useful and efficient. From a general point of view, machine intelligence can be interpreted as capabilities of machines in handling and performing computational and processing tasks as well as decision making in a more accurate, accelerated and effective way than humans. Needless to say that coming up with a universal and comprehensive definition of machine intelligence is a controversial and interdisciplinary issue and out of scope of this paper. Anyway, following studies can provide some useful information in this regard (Hernández-Orallo and Minaya-Collado, 1998; Bien, et al., 2002; Legg and Hutter, 2007; Dobrev, 2012). As mentioned earlier, however, in some cases – including cognitive tasks – machines could not even present human-level performance (Fleuret et al., 2011; Stabinger et al., 2016; Dodge and Karam, 2017); there are many scenarios (such as huge computation, high-volume data analysis, real-time knowledge-based decision making and so on) that may not be realized without help of them. Such outstanding achievements are owing to many years of research and development in machine learning and artificial intelligence as well as advancements in hardware technology and communication/computation infrastructures. All these facilities and progresses, though, could not quench humans’ thirst of creating comprehensive and polymath machines. The ultimate intention in the field is to realize the idea of universal AI (Everitt and Hutter, 2018) or Artificial General Intelligence (Gurkaynak et al., 2016) rather than case-specific ones, e.g. Artificial Narrow Intelligence (Gurkaynak et al., 2016). Achieving such level of autonomy and intelligence, of course, is not practically impossible; however a great deal of (multidimensional) intelligence and resources are needed. Looking for such an ambitious vision asserts that the days of kingdom of independent and single-dimension artificial intelligence are gone (or will be gone soon) (Wiedermann, 2012; Yampolskiy, 2015; Miailhe and Hodes, 2017). This ongoing revolution borrowed the idea from humans who could think and operate more effectively when being organized in the form of a crowd (Bonabeau, 2009; Leimeister, 2010). The adoption of the concept of collective human intelligence in the context of machines known as collective machine intelligence (Halmes, 2013), wisdom of artificial crowds (Yampolskiy and El-Barkouky, 2011), collective robot intelligence (Kube and Zhang, 1992), etc. (Figure 2). Regardless of differences in nomenclature and (even) details, the goal is almost a similar and identical one: aggregation and integration of independent (homogeneous/heterogeneous) machines’ intelligence, power and resources to produce more effective and efficient outputs. Seems to be partially similar to swarm intelligence (Kennedy, 2006), cluster computing (Sadashiv and Kumar, 2011) and so on, collective machine intelligence (CMI) is a comprehensive and multipurpose concept aimed at taking advantages of (almost) every aspects of a single machine to improve the team performance. Moreover, in such multi-agent systems the ultimate intention is facilitating collaborative learning, knowledge, experience and resource sharing (Gifford, 2009). Clearly, the core concept of CMI is synergy and all-out cooperation. One of the very early well-experienced realization of the concept is SETI@home project in which millions of computers all over the world contributed in search for the extraterrestrial intelligence through analyzing radio ----- ## Collective hybrid intelligence 203 Figure 2. Simplified schematic of CMI workflow signals (Anderson, et al., 2002). Although the major goal of the project was compensating the lack of adequate processing resources rather establishing a platform to aggregate independent machine’s intelligence; it could be an inspirational case study to prove the applicability of such a strategy. Further, several remarkable research works have been conducted to empirically study the efficiency of teaming up machines to benefit more of their aggregated utilization, such as projects reported in (Chien et al., 2003; Larson et al., 2009; Pedreira and Grigoras, 2017). Of course, there is still a notable challenge that, e.g. a cluster of powerful machines may face severe difficulties to handle it, namely lack of human-level, cognitive intelligence. 3.3 Hybrid intelligence The major untouchable difference between humans and most powerful artificial intelligence is the humanity. Thinking, understanding, learning, recognizing and judging like what humans do are the essential barriers that no artificial human-made creature (i.e. machine) could yet overcome them[7][8][9]. Regarding this fact, behind every successful machine, there is a least one human that is in charge of supervising, training or collaborating with it (Folds, 2016). Emphasizing on the intellectual aspects of such constructive symbiosis, it is referred to as hybrid intelligence (Kamar, 2016). Taking a closer look at the literature reveals there are cases in which the term (hybrid intelligence) was used to point out to other concepts, especially collective machine intelligence, e.g. research conducted in (Deng et al., 2012). In other words, in those instances applying various machine learning algorithms to perform same task in a more efficient way interpreted as leveraging hybrid intelligence. Such an appellation, by the way, may not be completely wrong and irrelevant; though, according to the aforementioned concepts and principles, the term collective machine intelligence can better reflect the underlying concept of interest. ----- ## IJCS 3,2 204 Figure 3. Simplified schematic of hybrid intelligence workflow Whether clearly stated or not, when it comes to supporting machine learning algorithms with human intelligence (usually in the form of crowdsourcing), the hybrid intelligence is leveraged (Vaughan, 2017; Nushi et al., 2018; Klumpp et al., 2019) (Figure 3). One can witness best practices of following this strategy in the field of robotics (Chang, et al., 2017) and particularly for human-robot interaction purposes (Breazeal et al., 2013). Such an approach – at the simplest scenario- can be simulated by training an image processing algorithm with human-labeled images (data sets) (Vaughan, 2017). Among various advantages of incorporating human intelligence in the machine learning workflow (Barbier et al., 2012; Vaughan, 2017; Verhulst, 2018), the followings can be enumerated: � simplifying problems and making them machine-understandable; � compensating machines’ weaknesses and inefficiency, especially for cognitive tasks; � facilitating and optimizing learning process; and � saving costs and time. Mapping general problems into computational ones and making them machine-readable and –understandable are of hard-to-tackle challenges. Equipping machines with general intelligence – if possible at this time- may not be economical in every case and demands a great deal of efforts and resources with no guarantee of being efficient. Specifically, when it comes to cognitive and human-specific issues, machines face extremely sophisticated challenges. Therefore, taking advantages of humans’ intelligence and problem solving power could be considered as the silver bullet. In spite of many advantages hybrid intelligence can present, there is also room for further improvement by mobilizing all the possibilities for great, unprecedented breakthroughs. ----- 3.4 Discussion (Are these enough?) To be or not to be? To answer this question about the need for another intelligence-oriented computing concept, the first and foremost is evaluation of the current state progress and challenges. From a high level perspective, computing tasks and processes – based on the contextual and intrinsic requirements- can be categorized into two major classes: intelligence-intensive and resource-intensive. The former refers to the tasks that require some type of cognitive-based judgments, intelligent decision-making, computational intelligence and similar soft (and mostly human-specific) abilities (Maleszka and Nguyen, 2015; Chen and Shen, 2019). On the other side, the latter ones are of time- and powerconsuming tasks which introduce dealing with large amount of data (Liu et al., 2015; Jonathan et al., 2017) and high computational and processing requirements (Ilyashenko et al., 2017; 2019; Singh et al., 2019). Natural language processing, semantic-based processing, concept understanding and interpretation are some general intelligence-intensive tasks, while multi-dimensional information processing, big data analysis, high volume communication control and management are among resource-intensive challenges. Notwithstanding the wide variety of real-world needs and requirements, numerous computational processes with different levels of complexity could be introduced. Therefore, to efficiently handle such situations, the most appropriate computing concept should be used. As an overview on the previously mentioned concepts, their features are summarized and compared in the following table (Table I). As noted in the Table I, there are some essential issues with current computational paradigms such as scalability and insufficiency to deal with complicated, hybrid tasks that require both enormous intelligence and resources. For example, assume a series of very large-scale semantic and cognitive image and video processing tasks that should provide real-time outputs as well as presenting reliable continuous performance. As we know, none of the described computational solutions could properly cope with these challenges and being satisfied with the current available solutions is, in fact, a case of any port in a storm. In this regard, it seems necessary to take advantages of current infrastructures and facilities in a novel arrangement for dealing with ever-growing computational requirements. 4. A new human–machine cooperation framework The availability of human participants, computing resources and software platforms as building blocks of any computational process have facilitated ambitious perspectives. Clearly, we are facing an unprecedented presence and distribution of tangled intelligence and computing power that have partially been overlooked and remained unused. At the lowest level, a very large, active and interested community of intelligent participants who equipped with the state-of-the-art smartphones are yet to be recruited. Strategy Context Major challenges Major drawbacks CHI Intelligence- User management, incentive Scalability, non-real time response, limited intensive tasks mechanism design types of tasks CMI Resource- Implementation, cooperation Lack of standard interaction modality, lack intensive tasks management, task allocation of human intelligence, availability issues Hybrid (Mostly) Human–machine interaction, Scalability, machine-dependent intelligence intelligence- synchronization performance intensive tasks ## Collective hybrid intelligence 205 Table I. Summarization of computing paradigms ----- ## IJCS 3,2 206 Mobile data mining (Stahl et al., 2010) as well as location-based computing (Karimi, 2004), further, have leveraged such smart entities as the most eligible candidates to take part in computational processes of all kind (Vij and Aggarwal, 2018; Zhao et al., 2019). On the other hand, distributed, ubiquitous and cloud computing paradigms, high-speed network connection and communication as well as similar technological facilities have provided a fertile land of opportunities to tame the groundbreaking possibilities. Therefore, not as a completely mold-breaking concept but as a complementary and evolutionary one, Collective Hybrid Intelligence (CHI) has everything to be realized. Defined as a framework for “integration and convergence of (intelligent and nonintelligent) capabilities of humans and machines in an organized and structured way to perform a (series of) specific (intelligence- and resource-intensive) computing tasks,” CHI can be considered as a comprehensive, multipurpose and scalable concept. The notion of collective hybrid intelligence, in addition to intelligence-intensive processes, can also be extended to any human–machine cooperative tasks. Basically, besides sharing the intelligence, the agents can collaborate for, e.g. data collection, testing, validation, ideation and any process that needs a remarkable amount of cooperative efforts. The CHI, principally, is an umbrella term to describe various ways of leveraging human– machine cooperation and collaboration to come up with solutions for highly complicated and sophisticated problems. In other words, this study is aimed to put forward a brand new vision for enabling humans and machines (in a bilateral way) to establish some type of super-collaboration. According to the concept, every single entity with sufficient capabilities and qualifications can be a nominee (i.e. potential contributor) to participate in a computational process. In this regard, in the presence of appropriate utilization mechanisms, e.g. computing platforms and portals, various computational and processing tasks of interest can be performed in (almost) everywhere and at every time (Figure 4). Owing to wide range of possible situations, requirements and computational problems, the proposed framework is presented at the conceptual level. Doing so, in addition to make it flexible so as to be able to fit various needs, implementation of different instances in different contexts will be facilitated. Therefore, the architectural notes in the following sections present a high-level view of the framework and its fundamentals (i.e. general organization of CHI) not a specific implementation of that. Besides proposing a modern computing perspective, CHI is greatly related to the concepts discussed in the previous section. Such relationships are illustrated in Figure 5. 4.1 Architectural notes From a general point of view, the conceptual architecture of a typical realization of CHI-based systems can be depicted as in Figure 6. According to this conceptual representation, any practical realization needs a complicated and multi-level implementation. Specifically, some mechanisms are required for distributed task management, result aggregation, integration and validation. The general workflow of such a system can be described as follows. After specifying the goal [i.e. problem(s) to be solved] and decomposing it into subtasks, the active agents will be identified/selected based on some criteria. Then, the task management component firstly analyzes the (ordered) task to determine its requirements, including primary resources, priority, estimated completion time, etc. Then, the appropriate available resources will be specified for performing the task in an efficient way. Decomposition of the initial task into several subtasks for distributing them over the computing network is the next step. Such a partitioning was based on the type of tasks and ----- ## Collective hybrid intelligence 207 Figure 4. Simplified schematic of CHI workflow Figure 5. Relationships between CHI and related concepts available resources. For example, managing a data-intensive task is far different from a time-dependent one. Finally, the subtasks will be assigned to the selected agents. Moreover, the task management component is in charge of aggregating and integrating the results, i.e. agent-generated responses. The agent management component maintains a complete and ----- ## IJCS 3,2 208 Figure 6. Conceptual architecture of a CHI-based system continuously updating profile (list) for all the available agents and their processing and computational capabilities. The agents will be prioritized based on some major factors, such as availability, active resources and (quality of) performance history. Those information plays a vital role in assigning tasks to the agents. Generally, two main scenarios can be considered for the task assignment process. First, the tasks will be presented in a task pool, then the volunteer agents in an auctionlike process and based on their capabilities, resources and also problem requirements will take responsibility of performing those tasks. In the second approach, those agents in the ready queue that match the requirements (such as being in an appropriate geographical location, having a specific resource, etc.) specified by the task coordinator; will be selected to perform the tasks. Then, the tasks will be performed by the participants and the outputs will be returned to the cloud-based server. Finally, the gathered results will be integrated and validated so that they become usable for the intended goal(s) (Figure 7). To demonstrate how such an approach may be benefited, three example scenarios are described in the section 6. According to the aforementioned workflow, as a high-level viewpoint, such a system should be shaped over a cloud-based infrastructure to support huge communication and computing processes. To manage the computing procedures, including task management and integration, a distributed computing platform should be leveraged as a middleware. ----- However, handling such possibly huge computing processes may face with many difficulties; thanks to the emerging fog (Bonomi et al., 2012) and edge computing (Shi et al., 2016) concepts, they can be managed efficiently. As illustrated in the layered architecture (Figure 8), on the top of the stack, a web service is in charge of providing participant agents with appropriate interface – similar to existing crowdsourcing platforms- so that they could perform assigned tasks. One important aspect of adhering to the CHI principles is leveraging maximum benefits of distributed computing. Specifically, thanks to flourishing of mobile crowdsourcing and data mining; location-based intelligence and computing are pervasively available. Moreover, thanks to ubiquitous smart devices spread globally, including smartphones, gadgets, laptops, closed-circuit cameras, PCs and state-of-the-art game consoles, we are witnessing a highly distributed, untamed computing potentialities. To capture such diverse dynamics, there are needs to well-organized and purposeful mechanisms and platforms. As the inspirational practical examples of how humans’ power ## Collective hybrid intelligence 209 Figure 7. General internal workflow of CHI Figure 8. Layered architectural representation of CHI ----- ## IJCS 3,2 210 Figure 9. Homogeneous realization of CHI could be used and converged, general- and specific-purpose crowdsourcing platforms, such as (Willis et al., 2017; Peer et al., 2017), are worth studying. In addition to take advantages of current crowdsourcing systems, there may be need to design customized systems to fit the case-specific requirements of computational processes. From another point of view, establishing reliable mechanisms to organize machines’ participation and joint work is an essential requirement. In this regard, development of platforms through which machines could interact and collaborate with each other put forward priceless benefits. Previous efforts of this kind such as Robot-specific social networks (Wang et al., 2012) and social internet of things (SIoT) (Atzori et al., 2012) are great sources of inspiration, by the way. 4.2 Realization models Based upon the proposed framework, machines, as passive entities, are thought to be in charge of providing computational power and processing infrastructure. Therefore, a PC, laptop, supercomputer and even a smartphone or a large network of computers can be regarded as an independent/hybrid agent in the process. From another viewpoint, the human agent besides his traditional roles (user or supervisor) can present a cooperative and interactive character to assist machines in a broad range from collecting training data sets to perform more complicated tasks, such as result validation and verification. Moreover, decision-making on how to distribute tasks between humans and machines is another important and determining consideration. Such a decision affects the bilateral human– machine cooperation as well as resource management. For example, inefficient separation of an intelligence-intensive task between agents may result in wasting times of machines for what those are not very powerful in and imposing complex and heavy computations (that take too long to complete) on humans. To avoid such flaws in realization of the CHI, two general task separation models are presented. The first one is a homogeneous model in which the tasks will be presented to the machines and humans in a distinctive manner. Then the results produced by each group will be collected and integrated. In the final stage, both results generated by the machine and human will be combined to produce the expected output (Figure 9). As a heterogeneous solution, the second model is based on using direct human–machine collaboration in the form of hybrid intelligence from the very early steps (Figure 10). ----- ## Collective hybrid intelligence 211 Figure 10. Heterogeneous realization of CHI As mentioned earlier, such a separation of tasks and duties comes in handy for managing available resources, costs, completion time and accuracy as well as striking a balance between efficiency and complexity. This is mainly because, not all tasks are appropriate for all agents and not all problems can be solved in an identical way. The first model, in essence, is the appropriate choice for the mostly resources-intensive tasks or those ones in which requirements and different aspects of tasks are clearly distinctive and separable. In such a situation, this kind of organization can drastically resolve unnecessary complexities. Accordingly, intrinsically hybrid and complicated processes are better to be organized based on the second realization model. 4.3 Discussion Generally, crowdsourcing-based and distributed processes introduce some intrinsic challenges and difficulties. Consequently, when it comes to synthesize these processes in an organized and cooperative workflow, facing unexampled and incidental challenges are inevitable. As a matter of fact, in spite of its presumed efficiency and applicability, the major challenge CHI struggles with is a cost-effective and reliable implementation. However, the authors are working to come up with such a solution, it seems there are needs more efforts and time to that point. In this respect, to cope with such issues, some essential considerations [including general (1-4), human-centric (5-7) and machine-centric (7, 8) ones] should be taken into account as follows. 4.3.1 Problem formulation. CHI is basically a high-level solution when the problem is a multidimensional, computationally expensive and usually large-scale one. Such a problem, on its own, addresses several intrinsic complexities that may affect the effectiveness of the process. Therefore, there is need to a preliminary analysis step for specifying different aspects of the problem, the category it belongs to, required resources and so on. Such a preevaluation provides necessary information to map the problem to the appropriate realization approach. As the matter of fact, the heart of a system constructed based on the proposed concept is efficient separation of duties (tasks) among the participants and this largely depends on the problem formulation process. 4.3.2 Distribution management. The distribution of tasks among agents and managing them is one of the most important and critical issues. Owing to intrinsic heterogeneity of the participant agents in the process, managing and coordinating them so as to result in ----- ## IJCS 3,2 212 providing most efficient and possible performance is of the highest importance. Analyzing performance log records, real-time agent management facilities as well as continuous monitoring and efficiency assessment are among the major considerations in this regard. 4.3.3 Interaction facilitation. The communication among various agents involved in the process and their interaction with control/management unit are other essential issues that should be taken into account. In addition to demand for (possibly) some new communication protocols, there is an essential need to an interface (agent interaction modality), e.g. a task management system such as Amazon Mechanical Turk, through which agents can interact with the system, perform the assigned tasks and submit the results. 4.3.4 Availability management. Although the availability issue is a well-studied topic for distributed systems (Kondo et al., 2008; Rawat et al., 2016); dealing with similar problems in the context of the proposed concept is way different and more challenging. Specifically, there should be several strategies for the cases in which human participants refuse to complete tasks in the scheduled time. Such problems are particularly associated with voluntary participation. The case will be more critical if the unavailability occurs in hybrid (heterogeneous) processes by each of the participant parties. 4.3.5 Participation engagement. In the context of crowdsourcing, attracting participation is an influential and challenging issues. Because relying on volunteer participants could not guarantee the desired performance in most of cases (Mao et al., 2013; Baruch et al., 2016); some strict, foolproof and reliable engagement strategies are needed. According to the best practices (Pilz and Gewald, 2013; Khoi et al., 2018), monetary incentives can be convincing for most of humans. So, when it comes to recruiting professional (expert) crowdworkers, higher costs (and even other incentives) may be imposed. Further, using non-human agents (i.e. machines) is even more difficult and troublesome. A probably working suggestion is establishing a cloud-based market in the reverse direction through which individuals could sell their own machines’ capabilities by enrolling in available computational processes. Then, they will be paid per completed tasks. 4.3.6 Quality assurance. One of the most important concerns in human-mediated processes in general and crowdsourcing in particular is the quality (i.e. accuracy and preciseness) of performance (e.g. submitted results). Despite efforts have been made to cope with this issue (Daniel et al., 2018), its unfavorable consequences can be severe in complicated and multidimensional projects. As an example, low quality labels in a crowdsourced image annotation process address very limited negative effects in contrast with inaccurate evaluation of a machine learning model. In addition to considering strict criteria for crowdworker recruitment, monitoring participants’ performance and adhering to rigorous task assignment standards are some practical steps to ensure the quality of the completed tasks. 4.3.7 Adversarial intentions. Untruthful workers and those with adversarial intentions in mind (Difallah et al., 2012; Steinhardt et al., 2016) can threaten any crowdsourcing process. Hence, trust management (Yu et al., 2012; Feng et al., 2017) plays a key role in participant recruitment and task assignment processes to deal with inaccurate and wrong submissions or even organized attacks aimed at affecting the process. Because there are situations in which some private information can be revealed (Boutsis and Kalogeraki, 2016), relying on untrusted workers may result in privacy breach and violation. Therefore, the needs for identifying malicious participants (both humans and machines), neutralizing wrongdoings and preserving privacy (for information and even participants (Kajino et al., 2014) are a must. 4.3.8 Machine inefficiency. Owing to differences in hosting systems’ configuration, implementation, initial training data and so on, the efficiency of (even same) machine ----- learning algorithms may vary case by case. For this reason, various machines introduce various levels of efficiency for different problems. In this regard, there should be some mechanisms to manage such unbalanced capabilities and performance – specifically in the case of hybrid collaboration- to make the computational process as reliable as possible. 5. Example scenarios Explaining the operation of a system that works based on the proposed concept, three motivating example scenarios are presented in this section. Applications of CHI are not limited to these cases; however, they could be regarded as inspirational instances to generalize the underlying concepts. 5.1 Collective hybrid intelligence for computing tasks In this example, the given goal is to recognize similar images from a large data set and annotating them to obtain appropriate results. To participate in this location-independent (and mostly intelligence-intensive) task, there are no specific criteria for human agents but their position in the task allocation queue. On the other side, being equipped with Open CV machine vision library is the specified criterion for the machines. Then, such machines will be selected from the ready queue to be a participant. Though, there are various methods for assigning tasks to the workers (agents), “In the context of this example, the tasks are divided into two groups: Resource-intensive and cognitive ones. Thanks to the development in the field of machine vision and image processing, finding similar images, in general, is not a difficult task. Therefore, these relatively time-consuming tasks that do not need high level of cognitive ability will be assigned to the machines. Moreover, machines are in charge of performing initial automatic annotation. To guarantee the accuracy and efficiency of annotations, for a specific image or a set of images that convergence rate, similarity of classification and annotation are less than a determined threshold, the results will be assigned to humans for further considerations. Moreover, the output of humans’ efforts, after analysis, may be leveraged as a gold standard to evaluate machines’ performance. Also, such human-generated data can be used to train machines. 5.2 Collective hybrid intelligence for autonomous urban vehicles control One of the most important issues in controlling autonomous vehicle is need for an accurate, up-to-date and comprehensive map or some advanced peripherals to provide environmental information in real-time, (Vochin et al., 2018; Bayat et al., 2018) and references therein. In this example, the application of CHI in providing such a specialized map is considered. Doing so, in one side, human agents should collect information from different streets of the city including rush hour situations, the safest paths, detours in various times and conditions. Moreover, their own experiences and recommendations for navigation in such situations are of the high importance. On the other side, traffic cameras and other urban monitoring sensors provide specialized machines (i.e. specific-purpose computers) with some real-world information on different situations of the city. Alongside with satellite and global maps information, such machines which leverage advanced algorithms can come up with some navigation patterns for the autonomous vehicles. Finally, fusing these two types of intelligence – that could be gathered asynchronously – can be used for predictive control of such vehicles within different streets of a crowded city in different times. 5.3 Collective hybrid intelligence for human–robot cooperative surgery Human-robot cooperative surgery is another context that adhering to collective hybrid intelligence principles may improve its workflow and performance. As an imaginary ## Collective hybrid intelligence 213 ----- ## IJCS 3,2 214 scenario, the CHI can facilitate a complex operation as follows: depending on the case, the previous experiences and information are gathered from experts. Such invaluable data will feed the automatic robotic arm(s) with the necessary information. In the case of any unprecedented issues or exceptions, if the (expert) system could not find any reliable solution (recommendation), the experts who are monitoring the operation will present their ideas (suggestions) based on the situation and machine’s feedback. Then, the integrated responses will be sent to the robot as the collective advice. Needless to say that, in this case, all the mentioned processes should be performed in real-time. 6. Conclusion In this paper the notion and general concept of CHI as a new complementary computing and automation concept is proposed. The main idea behind the Collective Hybrid intelligence is leveraging humans and machines’ capabilities in a new manner to maximize the efficiency of human–machine cooperation and collaboration. The major building blocks of the presented framework are some well-experienced and successful approaches, namely distributed computing, collective human intelligence, human computing, hybrid intelligence and collective machine intelligence. To support the introduced idea, its different realization models, the conceptual architecture and workflow are delineated and discussed. The authors anticipate that this concept can provide unprecedented functionality and performance for human–machinecooperated processing and computing procedures in the near future. Meanwhile, it is emphasized that the proposed idea in this paper is in its early stages and there are still several unanswered questions and challenges yet to be resolved. Specifically, the implementation of a real-world system based on the presented framework is future work of the authors. Notes [1. www.wired.com/2017/05/what-deep-blue-tells-us-about-ai-in-2017/](http://www.wired.com/2017/05/what-deep-blue-tells-us-about-ai-in-2017/) [2. www.scientificamerican.com/article/how-the-computer-beat-the-go-master/?redirect=1](http://www.scientificamerican.com/article/how-the-computer-beat-the-go-master/?redirect=1) [3. www.dailymail.co.uk/sciencetech/article-6695515/Human-debate-champion-defeats-IBMs-smartest-](http://www.dailymail.co.uk/sciencetech/article-6695515/Human-debate-champion-defeats-IBMs-smartest-AI-powered-machine.html) [AI-powered-machine.html](http://www.dailymail.co.uk/sciencetech/article-6695515/Human-debate-champion-defeats-IBMs-smartest-AI-powered-machine.html) [4. www.crowdsource.com/blog/2013/08/the-long-history-of-crowdsourcing-and-why-youre-just-now-](http://www.crowdsource.com/blog/2013/08/the-long-history-of-crowdsourcing-and-why-youre-just-now-hearing-about-it/) [hearing-about-it/](http://www.crowdsource.com/blog/2013/08/the-long-history-of-crowdsourcing-and-why-youre-just-now-hearing-about-it/) [5. www.mturk.com/](http://www.mturk.com/) [6. www.figure-eight.com/](http://www.figure-eight.com/) [7. www.theguardian.com/technology/2019/mar/28/can-we-stop-robots-outsmarting-humanity-artificial-](http://www.theguardian.com/technology/2019/mar/28/can-we-stop-robots-outsmarting-humanity-artificial-intelligence-singularity) [intelligence-singularity](http://www.theguardian.com/technology/2019/mar/28/can-we-stop-robots-outsmarting-humanity-artificial-intelligence-singularity) [8. https://medium.com/@lancengym/3-simple-reasons-why-ai-will-not-rule-man-yet-22d8069d8321](https://medium.com//3-simple-reasons-why-ai-will-not-rule-man-yet-22d8069d8321) [9. https://thenextweb.com/syndication/2019/01/02/ai-is-incredibly-smart-but-it-will-never-match-human-](https://thenextweb.com/syndication/2019/01/02/ai-is-incredibly-smart-but-it-will-never-match-human-creativity/) [creativity/](https://thenextweb.com/syndication/2019/01/02/ai-is-incredibly-smart-but-it-will-never-match-human-creativity/) References Anderson, D.P., Cobb, J., Korpela, E., Lebofsky, M. and Werthimer, D. (2002), “SETI@ home: an experiment in public-resource computing”, Communications of the Acm, Vol. 45 No. 11, pp. 56-61. Atzori, L., Iera, A., Morabito, G. and Nitti, M. (2012), “The social internet of things (siot)–when social networks meet the internet of things: concept, architecture and network characterization”, Computer Networks, Vol. 56 No. 16, pp. 3594-3608. ----- Barbier, G., Zafarani, R., Gao, H., Fung, G. and Liu, H. (2012), “Maximizing benefits from crowdsourced data”, Computational and Mathematical Organization Theory, Vol. 18 No. 3, pp. 257-279. Bargiela, A. and Pedrycz, W. (2016), “Granular computing”, in Handbook on Computational Intelligence, Vol. 1, Fuzzy Logic, Systems, Artificial Neural Networks, and Learning Systems, pp. 43-66. Baruch, A., May, A. and Yu, D. (2016), “The motivations, enablers and barriers for voluntary participation in an online crowdsourcing platform”, Computers in Human Behavior, Vol. 64, pp. 923-931. Bayat, F., Najafinia, S. and Aliyari, M. (2018), “Mobile robots path planning: electrostatic potential field approach”, Expert Systems with Applications, Vol. 100, pp. 68-78. Beberg, A.L., Ensign, D.L., Jayachandran, G., Khaliq, S. and Pande, V.S. (2009), “Folding@ home: lessons from eight years of volunteer distributed computing”, Proceedings of International Parallel and Distributed Processing Symposium, Rome, Italy, pp. 1-8. Bien, Z., Bang, W.C., Kim, D.Y. and Han, J.S. (2002), “Machine intelligence quotient: its measurements and applications”, Fuzzy Sets and Systems, Vol. 127 No. 1, pp. 3-16. Bonabeau, E. (2009), “Decisions 2.0: the power of collective intelligence”, MIT Sloan Management Review, Article 45, Vol. 50 No. No. 2. Bonomi, F., Milito, R., Zhu, J. and Addepalli, S. (2012), “Fog computing and its role in the internet of things”, Proceedings of the first edition of the MCC workshop on Mobile Cloud Computing, ACM; pp. 13-16. Boutsis, I. and Kalogeraki, V. (2016), ““Location privacy for crowdsourcing applications”, Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing, ACM, pp. 694-705. Brabham, D.C. (2008), “Crowdsourcing as a model for problem solving: an introduction and cases”, Convergence: The International Journal of Research into New Media Technologies, Vol. 14 No. 1, pp. 75-90. Breazeal, C., DePalma, N., Orkin, J., Chernova, S. and Jung, M. (2013), “Crowdsourcing human-robot interaction: new methods and system evaluation in a public environment”, Journal of HumanRobot Interaction, Vol. 2 No. 1, pp. 82-111. Buchanan, L. and O’Connell, A. (2006), “A brief history of decision making”, Harvard Business Review, Vol. 84 No. 1, pp. 32-40. Bundy, A. (2017), “Smart machines are not a threat to humanity”, Communications of the Acm, Vol. 60 No. 2, pp. 40-42. Burnstein, E. and Berbaum, M.L. (1983), “Stages in group decision making: the decomposition of historical narratives”, Political Psychology, Vol. 4 No. 3, pp. 531-561. Cassimatis, N.L. (2006), “A cognitive substrate for achieving human-level intelligence”, AI Magazine, Vol. 27 No. 2, pp. 45-45. Chang, J.C., Amershi, S. and Kamar, E. (2017), “Revolt: collaborative crowdsourcing for labeling machine learning datasets”, Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, ACM, pp. 2334-2346. Chen, S.C.Y. and Shen, M.C. (2019), “The fourth industrial revolution and the development of artificial intelligence”, Contemporary Issues in International Political Economy, Palgrave Macmillan, Singapore, pp. 333-346. Chien, A., Calder, B., Elbert, S. and Bhatia, K. (2003), “Entropia: architecture and performance of an enterprise desktop grid system”, Journal of Parallel and Distributed Computing, Vol. 63 No. 5, pp. 597-610. Chiu, C.M., Liang, T.P. and Turban, E. (2014), “What can crowdsourcing do for decision support?”, Decision Support Systems, Vol. 65, pp. 40-49. Copeland, B.J. (2000), The Modern History of Computing, Zalta, E.N. (Ed.), Winter 2017 Edition The [Stanford Encyclopedia of Philosophy, available at: https://plato.stanford.edu/archives/win2017/](https://plato.stanford.edu/archives/win2017/entries/computing-history/) [entries/computing-history/](https://plato.stanford.edu/archives/win2017/entries/computing-history/) Dai, P. and Weld, D.S. (2011), “Artificial intelligence for artificial intelligence”, Proceedings of TwentyFifth AAAI Conference on Artificial Intelligence. ## Collective hybrid intelligence 215 ----- ## IJCS 3,2 216 Daniel, F., Kucherbaev, P., Cappiello, C., Benatallah, B. and Allahbakhsh, M. (2018), “Quality control in crowdsourcing: a survey of quality attributes, assessment techniques, and assurance actions”, ACM Computing Surveys, Article 7, Vol. 51 No. 1. Decker, M. (2000), “Replacing human beings by robots. How to tackle that perspective by technology assessment”, in Vision Assessment: Shaping Technology in 21st Century Society, Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 149-166. Del Prado, G.M. (2015), “18 Artificial intelligence researchers reveal the profound changes coming to [our lives”, Business Insider, available at: www.businessinsider.com/researchers-predictions-](http://www.businessinsider.com/researchers-predictions-future-artificial-intelligence-2015-10) [future-artificial-intelligence-2015-10 (accessed 10 September 2018).](http://www.businessinsider.com/researchers-predictions-future-artificial-intelligence-2015-10) Deng, W., Chen, R., Gao, J., Song, Y. and Xu, J. (2012), “A novel parallel hybrid intelligence optimization algorithm for a function approximation problem”, Computers and Mathematics with Applications, Vol. 63 No. 1, pp. 325-336. Difallah, D.E., Demartini, G. and Cudré-Mauroux, P. (2012), “Mechanical cheat: spamming schemes and adversarial techniques on crowdsourcing platforms”, Proceedings of the First International Workshop on Crowdsourcing Web Search, pp. 26-30. Dobrev, D. (2012), A Definition of Artificial Intelligence”, arXiv preprint arXiv:1210.1568. Dodge, S. and Karam, L. (2017), “A study and comparison of human and deep learning recognition performance under visual distortions”, Proceedings of the 26th International Conference on Computer Communication and Networks (ICCCN), IEEE, pp. 1-7. Everitt, T. and Hutter, M. (2018), “Universal artificial intelligence”, In Foundations of Trusted Autonomy, Springer, Cham; pp. 15-46. Fabisiak, T. and Danilecki, A. (2017), “Browser-based harnessing of voluntary computational power”, Foundations of Computing and Decision Sciences, Vol. 42 No. 1, pp. 3-42. Feigenbaum, E.A. (2003), “Some challenges and grand challenges for computational intelligence”, Journal of the ACM ( Acm), Vol. 50 No. 1, pp. 32-40. Feng, W., Yan, Z., Zhang, H., Zeng, K., Xiao, Y. and Hou, Y.T. (2017), “A survey on security, privacy, and trust in mobile crowdsourcing”, IEEE Internet of Things Journal, Vol. 5 No. 4, pp. 2971-2992. Fleuret, F., Li, T., Dubout, C., Wampler, E.K., Yantis, S. and Geman, D. (2011), “Comparing machines and humans on a visual categorization test”, Proceedings of the National Academy of Sciences, Vol. 108 No. 43, pp. 17621-17625. Folds, D.J. (2016), “Human executive control of autonomous systems: a conceptual framework”, Proceedings of IEEE International Symposium on Systems Engineering (ISSE), pp. 1-5. Gifford, C.M. (2009), “Collective machine learning: team learning and classification in multi-agent systems”, Doctoral dissertation, University of Kansas. Guazzini, A., Vilone, D., Donati, C., Nardi, A. and Levnaji�c, Z. (2015), “Modeling crowdsourcing as collective problem solving”, Scientific reports, Vol. 5, Article number: 16557. Gurkaynak, G., Yilmaz, I. and Haksever, G. (2016), “Stifling artificial intelligence: human perils”, Computer Law and Security Review, Vol. 32 No. 5, pp. 749-758. Halmes, M. (2013), “Measurements of collective machine intelligence”, arXiv preprint arXiv:1306.6649. Hernández-Orallo, J. and Minaya-Collado, N. (1998), “A formal definition of intelligence based on an intensional variant of algorithmic complexity”, Proceedings of International Symposium of Engineering of Intelligent Systems (EIS98), pp. 146-163. Hibbard, B. (2001), “Super-intelligent machines”, ACM SIGGRAPH Computer Graphics, Vol. 35 No. 1, pp. 11-13. Howe, J. (2006), “The rise of crowdsourcing”, Wired Magazine, Vol. 14 No. 6, pp. 1-4. Huang, Y., Vir Singh, P. and Srinivasan, K. (2014), “Crowdsourcing new product ideas under consumer learning”, Management Science, Vol. 60 No. 9, pp. 2138-2159. Huang, F.Y., Wang, K., An, Y. and Lasecki, W.S. (2017), “Towards hybrid intelligence for robotics”, Proceedings of The 5th Edition of the Collective Intelligence Conference. ----- Ilyashenko, A.S., Lukashin, A.A., Zaborovsky, V.S. and Lukashin, A.A. (2017), “Algorithms for planning resource-intensive computing tasks in a hybrid supercomputer environment for simulating the characteristics of a quantum rotation sensor and performing engineering calculations”, Automatic Control and Computer Sciences, Vol. 51 No. 6, pp. 426-434. Jonathan, A., Ryden, M., Oh, K., Chandra, A. and Weissman, J. (2017), “Nebula: distributed edge cloud for data intensive computing”, IEEE Transactions on Parallel and Distributed Systems, Vol. 28 No. 11, pp. 3229-3242. Kajino, H., Arai, H. and Kashima, H. (2014), “Preserving worker privacy in crowdsourcing”, Data Mining and Knowledge Discovery, Vol. 28 Nos 5/6, pp. 1314-1335. Kamar, E. (2016), “Directions in hybrid intelligence: complementing AI systems with human intelligence”, Proceedings of IJCAI, pp. 4070-4073. Karimi, H.A. (Ed.) (2004), “Telegeoinformatics: Location-Based Computing and Services”, CRC Press. Kennedy, J. (2006), “Swarm intelligence”, in Handbook of Nature-Inspired and Innovative Computing, Springer, pp. 187-219. Kephart, J.O. and Chess, D.M. (2003), “The vision of autonomic computing”, Computer, Vol. 1, pp. 41-50. Khoi, N., Casteleyn, S., Moradi, M. and Pebesma, E. (2018), “Do monetary incentives influence users’ behavior in participatory sensing?”, Sensors, Vol. 18 No. 5, Article 1426. Klumpp, M., Hesenius, M., Meyer, O., Ruiner, C. and Gruhn, V. (2019), “Production logistics and humancomputer interaction—state-of-the-art, challenges and requirements for the future”, The [International Journal of Advanced Manufacturing Technology, doi: 10.1007/s00170-019-03785-0.](http://dx.doi.org/10.1007/s00170-019-03785-0) Kondo, D., Andrzejak, A. and Anderson, D.P. (2008), “On correlated availability in internet-distributed systems”, Proceedings of the 2008 9th IEEE/ACM International Conference on Grid Computing, IEEE, pp. 276-283. Kube, C.R. and Zhang, H. (1992), “Collective robotic intelligence”, Proceedings of the Second International Conference on Simulation of Adaptive Behavior, pp. 460-468. Larson, S.M. Snow, C.D. Shirts, M. and Pande, V.S. (2009), “Folding@ Home and Genome@ Home: Using distributed computing to tackle previously intractable problems in computational biology”, arXiv preprint arXiv:0901.0866. Legg, S. and Hutter, M. (2007), “Universal intelligence: a definition of machine intelligence”, Minds and Machines, Vol. 17 No. 4, pp. 391-444. Leimeister, J.M. (2010), “Collective intelligence”, Business and Information Systems Engineering, Vol. 2 No. 4, pp. 245-248. Li, J., Pan, Z., Xu, J., Liang, B., Chen, Y. and Ji, W. (2018), “Quality-time-complexity universal intelligence measurement”, International Journal of Crowd Science, Vol. 2 No. 2, pp. 99-107. Litman, L., Robinson, J. and Abberbock, T. (2017), “TurkPrime. com: a versatile crowdsourcing data acquisition platform for the behavioral sciences”, Behavior Research Methods, Vol. 49 No. 2, pp. 433-442. Liu, J., Pacitti, E., Valduriez, P. and Mattoso, M. (2015), “A survey of data-intensive scientific workflow management”, Journal of Grid Computing, Vol. 13 No. 4, pp. 457-493. Lu, H., Li, Y., Chen, M., Kim, H. and Serikawa, S. (2018), “Brain intelligence: go beyond artificial intelligence”, Mobile Networks and Applications, Vol. 23 No. 2, pp. 368-375. Luo, S., Xia, H., Yoshida, T. and Wang, Z. (2009), “Toward collective intelligence of online communities: a primitive conceptual model”, Journal of Systems Science and Systems Engineering, Vol. 18 No. 2, pp. 203-221. Mahoney, M.S. (1988), “The history of computing in the history of technology”, Ieee Annals of the History of Computing, Vol. 10 No. 2, pp. 113-125. Makridakis, S. (2017), “The forthcoming artificial intelligence (AI) revolution: its impact on society and firms”, Futures, Vol. 90, pp. 46-60. ## Collective hybrid intelligence 217 ----- ## IJCS 3,2 218 Maleszka, M. and Nguyen, N.T. (2015), “Integration computing and collective intelligence”, Expert Systems with Applications, Vol. 42 No. 1, pp. 332-340. Mao, A., Kamar, E., Chen, Y., Horvitz, E., Schwamb, M.E., Lintott, C.J. and Smith, A.M. (2013), “Volunteering versus work for pay: incentives and tradeoffs in crowdsourcing”, Proceedings of the First AAAI Conference on Human Computation and Crowdsourcing. Maoz, Z. (1990), “Framing the national interest: the manipulation of foreign policy decisions in group settings”, World Politics, Vol. 43 No. 1, pp. 77-110. Miailhe, N. and Hodes, C. (2017), “The third age of artificial intelligence. Field actions science reports”, The Journal of Field Actions, (Special Issue), Vol. 17, pp. 6-11. Müller, V.C. and Bostrom, N. (2016), “Future progress in artificial intelligence: a survey of expert opinion”, In Fundamental Issues of Artificial Intelligence, Springer, Cham, pp. 555-572. Nilsson, N.J. (2005), “Human-level artificial intelligence? Be serious!”, AI Magazine, Vol. 26 No. 4, pp. 68-68. Nowak, S. and Rüger, S. (2010), “How reliable are annotations via crowdsourcing: a study about interannotator agreement for multi-label image annotation”, Proceedings of the international conference on Multimedia Information Retrieval, ACM, pp. 557-566. Nushi, B., Kamar, E. and Horvitz, E. (2018), “Towards accountable AI: hybrid human-machine analyses for characterizing system failure”, Proceedings of the Sixth AAAI Conference on Human Computation and Crowdsourcing. Ofli, F., Meier, P., Imran, M., Castillo, C., Tuia, D., Rey, N., Briant, J., Millet, P., Reinhard, F., Parkan, M. and Joost, S. (2016), “Combining human computing and machine learning to make sense of big (aerial) data for disaster response”, Big Data, Vol. 4 No. 1, pp. 47-59. Pan, Y. (2016), “Heading toward artificial intelligence 2.0”, Engineering, Vol. 2 No. 4, pp. 409-413. Pedreira, M.M. and Grigoras, C. (2017), “Scalable Global Grid catalogue for LHC Run3 and beyond”, arXiv preprint arXiv:1704.05272. Peer, E., Brandimarte, L., Samat, S. and Acquisti, A. (2017), “Beyond the Turk: alternative platforms for crowdsourcing behavioral research”, Journal of Experimental Social Psychology, Vol. 70, pp. 153-163. Pilz, D. and Gewald, H. (2013), “Does money matter? Motivational factors for participation in paid-and non-profit-crowdsourcing communities”, Wirtschaftsinformatik, Vol. 37, pp. 73-82. Quinn, A.J. and Bederson, B.B. (2011), “Human computation: a survey and taxonomy of a growing field”, Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, ACM, pp. 1403-1412. Rawat, A.S., Papailiopoulos, D.S., Dimakis, A.G. and Vishwanath, S. (2016), “Locality and availability in distributed storage”, IEEE Transactions on Information Theory, Vol. 62 No. 8, pp. 4481-4493. Retelny, D., Robaszkiewicz, S., To, A., Lasecki, W.S., Patel, J., Rahmati, N., Doshi, T., Valentine, M. and Bernstein, M.S. (2014), “Expert crowdsourcing with flash teams”, Proceedings of the 27th annual ACM symposium on User Interface Software and Technology, ACM, pp. 75-85. Russell, S. (2017), “Artificial intelligence: the future is superintelligent”, Nature, Vol. 548 No. 7669, pp. 520-522. Sadashiv, N. and Kumar, S.D. (2011), “Cluster, grid and cloud computing: a detailed comparison”, Proceedings of the 6th International Conference on Computer Science and Education (ICCSE), IEEE, pp. 477-482. Sarathy, V. (2018), “Real world problem-solving”, Frontiers in Human Neuroscience, Vol. 12PMC6028615. Schemmann, B., Herrmann, A.M., Chappin, M.M. and Heimeriks, G.J. (2016), “Crowdsourcing ideas: involving ordinary users in the ideation phase of new product development”, Research Policy, Vol. 45 No. 6, pp. 1145-1154. Shi, W., Cao, J., Zhang, Q., Li, Y. and Xu, L. (2016), “Edge computing: vision and challenges”, IEEE Internet of Things Journal, Vol. 3 No. 5, pp. 637-646. ----- Singh, S.P., Nayyar, A., Kaur, H. and Singla, A. (2019), “Dynamic task scheduling using balanced VM allocation policy for fog computing platforms”, Scalable Computing: Practice and Experience, Vol. 20 No. 2, pp. 433-456. Stabinger, S., Rodríguez-Sánchez, A. and Piater, J. (2016), “25 Years of CNNs: Can we compare to human abstraction capabilities?”, Proceedings of International Conference on Artificial Neural Networks, Springer, Cham, pp. 380-387. Stahl, F., Gaber, M.M., Bramer, M. and Philip, S.Y. (2010), “Pocket data mining: towards collaborative data mining in mobile computing environments”, Proceedings of 22nd IEEE International Conference on Tools with Artificial Intelligence (ICTAI), IEEE, Vol. 2, pp. 323-330. Steinhardt, J., Valiant, G. and Charikar, M. (2016), “Avoiding imposters and delinquents: adversarial crowdsourcing and peer prediction”, In Advances in Neural Information Processing Systems, pp. 4439-4447. Vaughan, J.W. (2017), “Making better use of the crowd: how crowdsourcing can advance machine learning research”, Journal of Machine Learning Research, Vol. 18, pp. 1-46. Verhulst, S.G. (2018), “Where and when AI and CI meet: exploring the intersection of artificial and collective intelligence towards the goal of innovating how we govern”, AI and Society, Vol. 33 No. 2, pp. 293-297. Vij, D. and Aggarwal, N. (2018), “Smartphone based traffic state detection using acoustic analysis and crowdsourcing”, Applied Acoustics, Vol. 138, pp. 80-91. Vochin, M., Zoican, S. and Borcoci, E. (2018), “Intelligent vehicle navigation system with assistance and alerting capabilities”, Concurrency and Computation: Practice and Experience, p. e4402, [available at: https://doi.org/10.1002/cpe.4402](https://doi.org/10.1002/cpe.4402) Von Ahn, L. (2008), “Human computation”, Proceedings of the 2008 IEEE 24th International Conference on Data Engineering, pp. 1-2. Wang, F.Y., Carley, K.M., Zeng, D. and Mao, W. (2007), “Social computing: from social informatics to social intelligence”, IEEE Intelligent Systems, Vol. 22 No. 2. Wang, W., Johnston, B. and Williams, M.A. (2012), “Social networking for robots to share knowledge, skills and know-how”, Proceedings of International Conference on Social Robotics, Springer, Berlin, Heidelberg, pp. 418-427. Weyer, J., Fink, R.D. and Adelt, F. (2015), “Human–machine cooperation in smart cars: an empirical investigation of the loss-of-control thesis”, Safety Science, Vol. 72, pp. 199-208. Wiedermann, J. (2012), “Is there something beyond AI? Frequently emerging, but seldom answered questions about artificial Super-Intelligence”, Proceedings of the International Conference Beyond AI, pp. 76-86. Wightman, D. (2010), “Crowdsourcing human-based computation”, Proceedings of the 6th Nordic Conference on Human-Computer Interaction: Extending Boundaries, ACM, pp. 551-560. Willis, C.G., Law, E., Williams, A.C., Franzone, B.F., Bernardos, R., Bruno, L., Hopkins, C., Schorn, C., Weber, E., Park, D.S. and Davis, C.C. (2017), “CrowdCurio: an online crowdsourcing platform to facilitate climate change studies using herbarium specimens”, New Phytologist, Vol. 215 No. 1, pp. 479-488. Yampolskiy, R.V. (2015), “On the limits of recursively self-improving AGI”, Proceedings of International Conference on Artificial General Intelligence, Springer, Cham, pp. 394-403. Yampolskiy, R.V. and El-Barkouky, A. (2011), “Wisdom of artificial crowds algorithm for solving NP-hard problems”, International Journal of Bio-Inspired Computation, Vol. 3 No. 6, pp. 358-369. Yang, J., Drake, T., Damianou, A. and Maarek, Y. (2018), “Leveraging crowdsourcing data for deep active learning an application: learning intents in Alexa”, Proceedings of the 2018 World Wide Web Conference on World Wide Web, pp. 23-32. Yu, C., Chai, Y. and Liu, Y. (2018), “Literature review on collective intelligence: a crowd science perspective”, International Journal of Crowd Science, Vol. 2 No. 1, pp. 64-73. ## Collective hybrid intelligence 219 ----- ## IJCS 3,2 220 Yu, H., Shen, Z., Miao, C. and An, B. (2012), “Challenges and opportunities for trust management in crowdsourcing”, Proceedings of the IEEE/WIC/ACM International Joint Conferences on Web Intelligence and Intelligent Agent Technology, IEEE, Vol. 2, pp. 486-493. Zadeh, L.A. (2008), “Toward human level machine intelligence-is it achievable? The need for a paradigm shift”, IEEE Computational Intelligence Magazine, Vol. 3 No. 3. Zanakis, S.H., Theofanides, S., Kontaratos, A.N. and Tassios, T.P. (2003), “Ancient Greeks’ practices and contributions in public and entrepreneurship decision making”, Interfaces, Vol. 33 No. 6, pp. 72-88. Zhao, X., Li, J., Han, R., Xie, B. and Ou, J. (2019), “GroundEye: a mobile crowdsourcing structure seismic response monitoring system based on smartphone”, in Health Monitoring of Structural and Biological Systems XIII, Vol. 10972, International Society for Optics and Photonics, available at: [https://doi.org/10.1117/12.2514905](https://doi.org/10.1117/12.2514905) Zhong, J., Tang, K. and Zhou, Z.H. (2015), “Active learning from crowds with unsure option”, Proceedings of International Joint Conferences on Artificial Intelligence, pp. 1061-1068. Further reading Moret-Bonillo, V. (2018), “Emerging technologies in artificial intelligence: quantum rule-based systems”, Progress in Artificial Intelligence, Vol. 7 No. 2, pp. 155-166. Corresponding author [Farhad Bayat can be contacted at: bayat.farhad@znu.ac.ir](mailto:bayat.farhad@znu.ac.ir) For instructions on how to order reprints of this article, please visit our website: www.emeraldgrouppublishing.com/licensing/reprints.htm Or contact us for further details: permissions@emeraldinsight.com -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1108/IJCS-03-2019-0012?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1108/IJCS-03-2019-0012, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.emerald.com/insight/content/doi/10.1108/IJCS-03-2019-0012/full/pdf?title=collective-hybrid-intelligence-towards-a-conceptual-framework" }
2,019
[ "JournalArticle", "Review" ]
true
2019-08-30T00:00:00
[ { "paperId": "9fc93a7ae7fabaabe1923a5914a29761dead1551", "title": "Future Progress in Artificial Intelligence: A Survey of Expert Opinion" }, { "paperId": "cd20a344091737e018ae1066f9d8476ec64dba73", "title": "Intelligent vehicle navigation system with assistance and alerting capabilities" }, { "paperId": "5249d9d8fd75488f14d65725bad51d1c8647cf3c", "title": "Production logistics and human-computer interaction—state-of-the-art, challenges and requirements for the future" }, { "paperId": "3528a3dd6c4ac1f9d154fcc1b5f4149ce470ffd1", "title": "Dynamic Task Scheduling using Balanced VM Allocation Policy for Fog Computing Platforms" }, { "paperId": "122466eb0c138e4774ab595a0dc4780847242150", "title": "GroundEye: a mobile crowdsourcing structure seismic response monitoring system based on smartphone" }, { "paperId": "258b3190707d25118208c8a1f2c56c75d063e67d", "title": "Smartphone based traffic state detection using acoustic analysis and crowdsourcing" }, { "paperId": "4690255ca7d4679a3fa10eaa93fb0c1e7c41b669", "title": "Real World Problem-Solving" }, { "paperId": "c280c95c17194de83e5234cca3586cd567ecd47c", "title": "Towards Accountable AI: Hybrid Human-Machine Analyses for Characterizing System Failure" }, { "paperId": "99f4c9115d4e086aa00515057ead867b7698e68a", "title": "Quality-time-complexity universal intelligence measurement" }, { "paperId": "79d576150b5d2da3d83d26e5c7c231ba6d4ea8a6", "title": "A Survey on Security, Privacy, and Trust in Mobile Crowdsourcing" }, { "paperId": "63fcc106f9a4053a79a1c2c1fd34211d827d1194", "title": "Mobile robots path planning: Electrostatic potential field approach" }, { "paperId": "baa9c67319222b4e7a74fb3130a24552a2d66c89", "title": "Do Monetary Incentives Influence Users’ Behavior in Participatory Sensing?" }, { "paperId": "2c08f8190d5dcd0278ce5f94ac39d0462c794508", "title": "Literature review on collective intelligence: a crowd science perspective" }, { "paperId": "0bf3f2c880333a8e7fe1ca430ab47ab349cb1021", "title": "Leveraging Crowdsourcing Data for Deep Active Learning An Application: Learning Intents in Alexa" }, { "paperId": "10b416132cf8c19c35ef3abb0acefb0352f48ad8", "title": "Where and when AI and CI meet: exploring the intersection of artificial and collective intelligence towards the goal of innovating how we govern" }, { "paperId": "48fb8ee6a786e13c3be21b4cd441b22dbc434f16", "title": "Emerging technologies in artificial intelligence: quantum rule-based systems" }, { "paperId": "72fdae43bb36c2470d04710e7192e4fe1093d14d", "title": "The Third Age of Artificial Intelligence" }, { "paperId": "6c0a71998932322d4d1658f0f1a4393e43e41352", "title": "Algorithms for planning Resource-Intensive computing tasks in a hybrid supercomputer environment for simulating the characteristics of a quantum rotation sensor and performing engineering calculations" }, { "paperId": "5fbf7af3b44172a1b071c8422ff7c5c87c83f066", "title": "Nebula: Distributed Edge Cloud for Data Intensive Computing" }, { "paperId": "fc0678e48ccc540c37c8cc3ceae0c9c935e60645", "title": "Artificial intelligence: The future is superintelligent" }, { "paperId": "9de3e1ee313a42d48bc340cad0c302570951fa67", "title": "CrowdCurio: an online crowdsourcing platform to facilitate climate change studies using herbarium specimens." }, { "paperId": "32754d496768685cfcc7fbdb323d87db2b33a09c", "title": "Brain Intelligence: Go beyond Artificial Intelligence" }, { "paperId": "2c21a89a91e6920ab9ac9a40af05055b865d5666", "title": "The Forthcoming Artificial Intelligence (AI) Revolution: Its Impact on Society and Firms" }, { "paperId": "89b4111f14cdf342188f96d3962581fd0afa042f", "title": "A Study and Comparison of Human and Deep Learning Recognition Performance under Visual Distortions" }, { "paperId": "05f39431776d2b8858132b2118b68a6a68bf4f14", "title": "Revolt: Collaborative Crowdsourcing for Labeling Machine Learning Datasets" }, { "paperId": "4c9f06e5e4eb93f16a739821be472364af66b427", "title": "Scalable global grid catalogue for Run3 and beyond" }, { "paperId": "a06707bf0d51d5a8c3980c7395b525b419729915", "title": "Browser-based Harnessing of Voluntary Computational Power" }, { "paperId": "e2847253f50ad7da6fc93de8eaa64c28c679978d", "title": "Smart machines are not a threat to humanity" }, { "paperId": "668bfea0aaa0565f58aa76d44d700e8a3a576f52", "title": "Heading toward Artificial Intelligence 2.0" }, { "paperId": "308e362e7893e8aa8a676e02a4068364a1bb23ff", "title": "The motivations, enablers and barriers for voluntary participation in an online crowdsourcing platform" }, { "paperId": "cff5cc26544f7ac057da867d7335ce3c4b9af461", "title": "Stifling artificial intelligence: Human perils" }, { "paperId": "c5824c3dcb42962cc1ece4ec4d2347afb6b06f8e", "title": "Human executive control of autonomous systems: A conceptual framework" }, { "paperId": "4cedc20e0414991487aced16ea068a6bb28b35f0", "title": "Location privacy for crowdsourcing applications" }, { "paperId": "71c4c255c1fbca2b25bd668abf9312e18b9ef1cf", "title": "25 Years of CNNs: Can We Compare to Human Abstraction Capabilities?" }, { "paperId": "545f142a246db9e65c65ec81f619d3ef093cca64", "title": "Directions in Hybrid Intelligence: Complementing AI Systems with Human Intelligence" }, { "paperId": "94ddef73ba921ed9e087537c97854fbf1f36aa2e", "title": "Crowdsourcing ideas: Involving ordinary users in the ideation phase of new product development" }, { "paperId": "e2d78d2800910060696c3a376043b88b92aa5665", "title": "Avoiding Imposters and Delinquents: Adversarial Crowdsourcing and Peer Prediction" }, { "paperId": "e3a442aa24e5df7e6b2a25e21e75c4c325f9eedf", "title": "Edge Computing: Vision and Challenges" }, { "paperId": "ca9f3a04a1b6524105f51676abdd198b118a7d18", "title": "Beyond the Turk: Alternative Platforms for Crowdsourcing Behavioral Research" }, { "paperId": "2f35b305b6c56046b631b0cdb6d2f08e4ee577a7", "title": "TurkPrime.com: A versatile crowdsourcing data acquisition platform for the behavioral sciences" }, { "paperId": "f8cd9e3da23d01ee7487be29ca251b6ad20889af", "title": "Combining Human Computing and Machine Learning to Make Sense of Big (Aerial) Data for Disaster Response" }, { "paperId": "9ffc2368a492669f1c330e3d787c04878d9c8ab7", "title": "Harder, Better, Faster, Stronger Convergence Rates for Least-Squares Regression" }, { "paperId": "068eb1a12766a5e8e408e650a3e1d4b4ec7ab42d", "title": "Active Learning from Crowds with Unsure Option" }, { "paperId": "b2012ac7cd3d78d4c49b20fab53f7fd4b6b63b50", "title": "On the Limits of Recursively Self-Improving AGI" }, { "paperId": "c8fdfab5996752372e21e7961c78556788551373", "title": "Modeling crowdsourcing as collective problem solving" }, { "paperId": "9ab91fdd78695cd11e5f4f9dfc0152bc48705d29", "title": "A Survey of Data-Intensive Scientific Workflow Management" }, { "paperId": "c1657703fc9da202679cd2cde91f47e3b41c211a", "title": "Human–Machine Cooperation in Smart Cars: An Empirical Investigation of the Loss-of-Control Thesis" }, { "paperId": "ff48b661bd15089c1e0bd58bd9b708234113d35e", "title": "Expert crowdsourcing with flash teams" }, { "paperId": "91fd67a6644cfb8745831dd9f72a4156210d4926", "title": "Crowdsourcing New Product Ideas Under Consumer Learning" }, { "paperId": "155e48e2159d08c0f810f418cb9f70f5f4f8ce8e", "title": "What can crowdsourcing do for decision support?" }, { "paperId": "191026ff5ced0363dd9396f6253724a1a8ec8822", "title": "Preserving worker privacy in crowdsourcing" }, { "paperId": "be29c6a7e23aebc1672bc5c47731d14b59944404", "title": "Locality and Availability in Distributed Storage" }, { "paperId": "7c56e633b6c6f1ab16684a65dad75755a9f47b9e", "title": "Volunteering Versus Work for Pay: Incentives and Tradeoffs in Crowdsourcing" }, { "paperId": "455b6d27c195d6f2308e31b47b2d41a9d491c5bf", "title": "Measurements of collective machine intelligence" }, { "paperId": "9561c635a6bc27c4940aca854958cac26e8717f0", "title": "Crowdsourcing human-robot interaction" }, { "paperId": "d1fd2c8d8c921658af8de93c6903abe449705dd2", "title": "Challenges and Opportunities for Trust Management in Crowdsourcing" }, { "paperId": "9cfe5fe748e0900dbd99d8d9195e317c5d4e5b4c", "title": "The Social Internet of Things (SIoT) - When social networks meet the Internet of Things: Concept, architecture and network characterization" }, { "paperId": "6b10723dbe6d37c836284945b3607dc740eea107", "title": "Social Networking for Robots to Share Knowledge, Skills and Know-How" }, { "paperId": "d4d9a80708d592ab2e48cff3508791b4e142fe52", "title": "Super-intelligent machines" }, { "paperId": "407d88cd832ba8ae80a325d7074fec8d0ada6422", "title": "Maximizing benefits from crowdsourced data" }, { "paperId": "2f47de50306aef67e17d124d86a4df1a768476cd", "title": "Wisdom of artificial crowds algorithm for solving NP-hard problems" }, { "paperId": "2ef26af206f7ebbbd90a7d68d77a7639c8dfc6b6", "title": "Comparing machines and humans on a visual categorization test" }, { "paperId": "6019408d642776763fd0d90c17ea3ddeb41f09a0", "title": "Cluster, grid and cloud computing: A detailed comparison" }, { "paperId": "77d899375144a418dea1f6a7bb9abacdb58e4c37", "title": "Human computation: a survey and taxonomy of a growing field" }, { "paperId": "1cc44baf36bb40a99084a7df9468e86b6f5e2691", "title": "Pocket Data Mining: Towards Collaborative Data Mining in Mobile Computing Environments" }, { "paperId": "69e9b4d8033545d390bc6bb5a912c34523d1fdcc", "title": "Crowdsourcing human-based computation" }, { "paperId": "4e227d4919caa1780ec2cd67e834430037e294bb", "title": "How reliable are annotations via crowdsourcing: a study about inter-annotator agreement for multi-label image annotation" }, { "paperId": "1c37285eb427f9d5bae45dc0624cb5ee78abe827", "title": "Folding@home: Lessons from eight years of volunteer distributed computing" }, { "paperId": "f537128e57f798c2345c162830a45200a22861e4", "title": "Human computation" }, { "paperId": "ed0497d39c66b2a3494e005fc8704a71bbaa9101", "title": "Toward collective intelligence of online communities: A primitive conceptual model" }, { "paperId": "090fda5b623ee2ed7cf2492857cf2c991ee7c8ad", "title": "Folding@Home and Genome@Home: Using distributed computing to tackle previously intractable problem" }, { "paperId": "e395797903f5a48c0e268a50b6c23f32861d9d30", "title": "On correlated availability in Internet-distributed systems" }, { "paperId": "c51b6605c2d2ae8e2d6d47ed75da8efa3dc1ad2b", "title": "Crowdsourcing as a Model for Problem Solving" }, { "paperId": "8e8ec502208f29ee9f78ded19226578e027ecd16", "title": "Universal Intelligence: A Definition of Machine Intelligence" }, { "paperId": "8b2f16ecdfd564f31a1ac1a899ed4f5a8828beff", "title": "A Cognitive Substrate for Achieving Human-Level Intelligence" }, { "paperId": "d27523090504a707c982ae6045755d928e2f59f7", "title": "Human-Level Artificial Intelligence? Be Serious!" }, { "paperId": "30a64bdf778b8f561af9ae589e822c2c800920b1", "title": "Universal artificial intelligence" }, { "paperId": "5549a758987bbb87efd228637876146a6741498e", "title": "Telegeoinformatics: Location-based Computing and Services" }, { "paperId": "5915a6f9e0df52c0dafde6a01f3b117fd0675d06", "title": "Ancient Greeks' Practices and Contributions in Public and Entrepreneurship Decision Making" }, { "paperId": "fad235dd7655c2ff1a51eb90a4e7089f71a44f41", "title": "Newsvendors Tackle the Newsvendor Problem" }, { "paperId": "9bfd3642299420fb8a874d543e2d9f138f527a69", "title": "Entropia: architecture and performance of an enterprise desktop grid system" }, { "paperId": "6663809bf91d253972e2f4408d58c399b42151c1", "title": "SETI@home: an experiment in public-resource computing" }, { "paperId": "8916ec41cab019e9023ccae12525861ef557ff8d", "title": "Virtual humans for validating maintenance procedures" }, { "paperId": "5482018d2109665b0b090d505c44e2790dcb66f6", "title": "Machine intelligence quotient: its measurements and applications" }, { "paperId": "2eaa54e926a95e2fd0c0edfe6dbb921fc426faf3", "title": "Framing the National Interest: The Manipulation of Foreign Policy Decisions in Group Settings" }, { "paperId": "af2a1f1d18ba7e5476cad0b8550da175072655c0", "title": "The History of Computing in the History of Technology" }, { "paperId": "6816899edd04b958f2d5968f26c0abeb9d3e254e", "title": "Stages in Group Decision Making: The Decomposition of Historical Narratives" }, { "paperId": "69b0a222fde51a052c2613f83b4aa98fad00c814", "title": "The Fourth Industrial Revolution and the Development of Artificial Intelligence" }, { "paperId": null, "title": "Quality control in crowdsourcing: a survey of quality attributes, assessment techniques, and assurance actions" }, { "paperId": "7e6ce5bd878f24839ba38fe8fb91f30aab3cd863", "title": "Making Better Use of the Crowd: How Crowdsourcing Can Advance Machine Learning Research" }, { "paperId": null, "title": "Towards hybrid intelligence for robotics" }, { "paperId": null, "title": "“ 18 Arti fi cial intelligence researchers reveal the profound changes coming to our lives ”" }, { "paperId": "ceb2dfed3921ced660d65a0a022e48bd4858d8b0", "title": "Collective Intelligence" }, { "paperId": "14c724d303202086dc4805f7545b7433c18d5f30", "title": "Does Money Matter? Motivational Factors for Participation in Paid- and Non-Profit-Crowdsourcing Communities" }, { "paperId": "135975122db060b3a425a86ef8b7ddf74a9aa8c7", "title": "Integration Computing and Collective Intelligence" }, { "paperId": "275dc12c439e06deec7b9694ec0220cae22a9093", "title": "A novel parallel hybrid intelligence optimization algorithm for a function approximation problem" }, { "paperId": "80a560f8e3a6bba85051ff8a418ed80f5cabd33f", "title": "Mechanical Cheat: Spamming Schemes and Adversarial Techniques on Crowdsourcing Platforms" }, { "paperId": "918a6d57f779b8f8269bc0d1489dcd014afb8120", "title": "Swarm Intelligence" }, { "paperId": null, "title": "Fog computing and its role in the internet of things" }, { "paperId": null, "title": "Is there something beyond AI? Frequently emerging, but seldom answered questions about artificial Super-Intelligence" }, { "paperId": "63eb4c86a1a881d269436f90ac02b679865c9fe7", "title": "Collective Machine Learning: Team Learning and Classification in Multi-Agent Systems" }, { "paperId": "65f0033ef461d287978987e6f54bf35f5102396c", "title": "Decisions 2.0: the power of collective intelligence" }, { "paperId": "d77f8db2b1ec89205da707ebf28ed1dc3366b86f", "title": "Granular Computing" }, { "paperId": "fc3b768e023bc756a352a530faf8d1ef8924e445", "title": "A brief history of decision making." }, { "paperId": "5185c810d1851dae0de7aee97d51dc59bbd1d2a9", "title": "Human-Level Artificial Intelligence ? Be Serious !" }, { "paperId": "f20efbc02ac331848c1e48b5dfb165f34b1c1bfe", "title": "The Modern History of Computing" }, { "paperId": null, "title": "“ The rise of crowdsourcing ”" }, { "paperId": "ee7384f216fe8272f9677b03147b986574e53372", "title": "Some challenges and grand challenges for computational intelligence" }, { "paperId": "6e51d8e92729366721de4b886ce3c26b124c9e58", "title": "A Formal Definition of Intelligence Based on an Intensional Variant of Algorithmic Complexity" }, { "paperId": "bff65e9da04d89e9248feb89110b1cc5fa1d99ca", "title": "The Vision of Autonomic Computing" }, { "paperId": "19f540f91131113a321f95af296b4cc4e1e32297", "title": "Replacing Human Beings by Robots. How to Tackle that Perspective by Technology Assessment" }, { "paperId": "dd4a7f3a5373bea8329d09c2d53ca6fa2dae74ba", "title": "Collective robotic intelligence" }, { "paperId": "ff11ca3ecafbd7690c96f58d6399725a825da6fc", "title": "Past and Present" }, { "paperId": "ce7e6aebb9d6174452d083f3cbbf10fb8a3ddcd3", "title": "August 2008 | Ieee Computational Intelligence Magazine 11 toward Human Level Machine Intelligence— Is It Achievable? the Need for a Paradigm Shift" }, { "paperId": null, "title": "www.theguardian.com/technology/2019/mar/28/can-we-stop-robots-outsmarting-humanity-arti fi cial-intelligence-singularity" }, { "paperId": null, "title": "www.crowdsource.com/blog/2013/08/the-long-history-of-crowdsourcing-and-why-youre-just-now-hearing-about-it/" }, { "paperId": null, "title": "www.dailymail.co.uk/sciencetech/article-6695515/Human-debate-champion-defeats-IBMs-smartest-AI-powered-machine.html" } ]
17,759
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01407e4317cba968ae41a1d029b154d8e50a0f08
[]
0.913731
Next Generation Middleware Technology for Mobile Computing
01407e4317cba968ae41a1d029b154d8e50a0f08
[ { "authorId": "66339021", "name": "B. Darsana" }, { "authorId": "2343710208", "name": "Karabi Konar" }, { "authorId": "2247386153", "name": "Sr. Lecturer" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
## International Journal of Computer and Communication International Journal of Computer and Communication Technology Technology ##### Volume 2 Issue 2 Article 1 April 2011 ## Next Generation Middleware Technology for Mobile Computing Next Generation Middleware Technology for Mobile Computing ##### B. Darsana Sr. Lecturer, Dept of ISE, The Oxford College of Engineering, Bangalore – 560068, Karnataka, darsana@gmail.com ##### Karabi Konar Sr. Lecturer, Dept of ISE, The Oxford College of Engineering, Bangalore – 560068, Karnataka, karabi@gmail.com [Follow this and additional works at: https://www.interscience.in/ijcct](https://www.interscience.in/ijcct?utm_source=www.interscience.in%2Fijcct%2Fvol2%2Fiss2%2F1&utm_medium=PDF&utm_campaign=PDFCoverPages) ##### Recommended Citation Recommended Citation Darsana, B. and Konar, Karabi (2011) "Next Generation Middleware Technology for Mobile Computing," International Journal of Computer and Communication Technology: Vol. 2 : Iss. 2, Article 1. DOI: 10.47893/IJCCT.2011.1074 [Available at: https://www.interscience.in/ijcct/vol2/iss2/1](https://www.interscience.in/ijcct/vol2/iss2/1?utm_source=www.interscience.in%2Fijcct%2Fvol2%2Fiss2%2F1&utm_medium=PDF&utm_campaign=PDFCoverPages) This Article is brought to you for free and open access by the Interscience Journals at Interscience Research Network. It has been accepted for inclusion in International Journal of Computer and Communication Technology by an authorized editor of Interscience Research Network. For more information, please contact [sritampatnaik@gmail.com.](mailto:sritampatnaik@gmail.com) ----- # Next Generation Middleware Technology for Mobile Computing #### B. Darsana, Karabi Konar Sr. Lecturer, Dept of ISE, The Oxford College of Engineering, Bangalore – 560068, Karnataka **Abstract-Current advances in portable devices, wireless** **_technologies, and distributed systems have created a_** **_mobile computing environment that is characterized by_** **_a large scale of dynamism. Diversities in network_** **_connectivity,_** **_platform_** **_capability,_** **_and_** **_resource_** **_availability can significantly affect the application_** **_performance. Traditional middleware systems are not_** **_prepared to offer proper support for addressing the_** **_dynamic aspects of mobile systems. Modern distributed_** **_applications need a middleware that is capable of_** **_adapting to environment changes and that supports the_** **_required level of quality of service._** **_This paper represents the experience of several research_** **_projects related to next generation middleware systems._** **_We first indicate the major challenges in mobile_** **_computing systems and try to identify the main_** **_requirements for mobile middleware systems. The_** **_different categories of mobile middleware technologies_** **_are reviewed and their strength and weakness are_** **_analyzed._** #### Key Words: dynamism, platform capability, quality of service, resource availability, network connectivity. #### 1. Introduction The availability of lightweight, portable computers and wireless technologies has created a new class of applications called mobile applications. These applications often run on scarce resource platforms such as personal digital assistants, notebooks, and mobile phones, each of which have limited CPU power, memory, and battery life. They are usually connected to wireless links, which are characterized by lower bandwidths, higher error rates, and more frequent disconnections. Most distributed applications and services were designed with the assumption that the terminals were powerful, stationary and connected to fixed networks. Conventional middleware technologies thus have focused on masking out the problems of heterogeneity and distribution to facilitate the development of distributed systems. They allow the application developers to focus on application functionality rather than on dealing explicitly with distribution issues. Under the highly variable computing environment conditions that characterize mobile platforms, it is believed that existing traditional middleware systems are not capable of providing adequate support for the mobile wireless computing environment. There is a great demand for designing modern middleware systems that can support new requirements imposed by mobility. This paper provides a most relevant mobile middleware systems and goals that still need to be achieved. #### 2. Mobile Architectural Requirements Middleware is an enabling layer of software that resides between the application program and the networked layer of heterogeneous platforms and protocols. It decouples applications from any dependencies on the plumbing layer that consists of heterogeneous operating systems, hardware platforms and communication protocols Middleware plays a vital role in hiding the complexity of distributed applications. These applications typically operate in an environment that may include heterogeneous computer architectures, operating systems, network protocols, and databases. It is unpleasant for an application developer to deal with such heterogeneous “plumbing”. Middleware’s primary role is to conceal this complexity from developers by deploying an isolated layer of APIs[6]. This layer bridges the gap between International Journal of Computer and Communication Technology (IJCCT) ISSN: 2231 0371 Vol 2 Iss 2 ----- application program and platform dependency. Middleware is defined as follows by Linthicum. ##### 2.1 The Limitations of Mobile Computing There are at least three common factors that affect the design of the middleware infrastructure required for mobile computing: mobile devices, network connection, and mobility, which vary from one to another in term of resource availability. Devices like laptops can offer fast CPUs and large amount of RAM and disk space while others like pocket PCs and phones usually have scarce resources. Hence, middleware should be designed to achieve optimal resource utilization. Network connections in mobile scenarios is characterized by limited bandwidth, high error rate, higher cost, and frequent disconnections due to power limitations, available spectrum, and mobility. Due to these limitations, conventional middleware technologies designed for fixed distributed systems are not prepared to support mobile systems. They target a static execution platform where the host location is fixed, the network bandwidth does not fluctuate, and services are well defined. We next identify a number of important requirements that must be provided by middleware for mobile computing. ##### 2.2 Analyzing the Requirements for Mobile Computing During the system lifetime, the application behavior may need to be altered due to dynamic changes in infrastructure facilities, such as the availability of particular services. Dynamic reconfiguration is thus required and can be achieved by adding a new behavior or changing an existing one at system runtime. Dynamic changes in system behavior and operating context at runtime can trigger re-evaluation and reallocation of resources. Middleware supporting dynamic reconfiguration needs to detect changes in available resources and either reallocate resources, or notify the application to adapt to the changes. Adaptability[9] is also part of the new requirements that allows applications to run efficiently and predictably under a broader range of conditions. Through adaptation a system can adapt its behavior instead of providing a uniform interface in all situations. The middleware needs to monitor the resource supply/demand, compute adaptation decisions, and notify applications about changes. Asynchronous interaction tackles the problems of high latency and disconnected operations that can arise with other interaction models. A client using asynchronous communication primitives issues a request and continues operating and then collects the result at any appropriate time. The client and server components do not need to be running concurrently to communicate with each other. A client may issue a request for a service, disconnect from the network, and collect the result later on. This type of interaction style reduces the network bandwidth consumption, achieves decoupling of client and server, and elevates system scalability. Context-awareness is an important requirement to build an effective and efficient adaptive system. The context of a mobile unit is usually determined by its current location which, in turn, defines the environment where the computation associated with the unit is performed. The context may include device characteristics, user’s activities, services, as well as other resources of the system. Context-awareness is used by several systems; however, few systems sense execution context other than location. The system performance can be increased when execution context is disclosed to the upper layer that assists middleware in making the right decision. Lightweight middleware needs to be considered when constructing middleware for mobile computing. Current middleware platforms like CORBA[7] are too heavy to run on devices with limited resources. By default, they contain a wide range of optional features and all possible functionalities, many of which will be unused by most applications. For example, invoking a method on a remote object involves only client side functionality and either Dynamic or Static Invocation Interface. Most of the existing ORB implementations provide either a single or two separate libraries for the client and server sides that contains all functionality. This forces the client program to be glued with the entire functionality without having a choice to select a specific subset of this functionality. #### 3. Mobile Middleware Technologies This section sheds some light on the different types of mobile middleware technologies. We start by introducing a classification that allows us to contrast and evaluate the different categories. Among the middleware systems we reviewed, we have identified four categories of middleware. Each category aims to support at least one of the above requirements imposed by mobility. These categories are reflective middleware, tuple space, contextaware middleware, and event-based middleware, each of which attempts to address the previous requirements using different approaches. The following table illustrates how various requirements are met by the different categories. ----- Table: 3.1. Requirements Vs Categories Tuple Context Event Requirements Reflective Space Aware Based Synchronous/ connection X X based Asynchronous/ connectionless X X ReX configuration Adaptation X X Awareness X X Light weight X The above table shows the relation between Requirements and Categories. For synchronous /connection based Reflective and Context Aware is applicable. For asynchronous/connection Event based and Tuple space is applicable. For Light weight requirement Event based is the best approach. For Awareness and Adaption Reflective and context Aware meet the requirements. ##### 3.1 Reflective Middleware The reflection technique was initially used in the field of programming languages to support the design of more open and extensible languages. Reflection is also applied in other fields including operating systems and more recently distributed systems. The principle of reflection enables a program to access, reason about and change its own behavior. Smith defined the concept of reflection in the following quote: “In as much as a computational process can be constructed to reason about an external world in virtue of comprising an ingredient process (interpreter) formally manipulating representations of that world, so too a computational process could be made to reason about itself in virtue of comprising an ingredient process (interpreter) formally manipulating representations of its own operations and structures”. A reflective system consists of two levels referred to as meta-level and base-level[11]. The former performs computation on the objects residing in the lower levels. The latter performs computation on the application domain entities. The reflection approach supports the inspection and adaptation of the underlying implementation (the base-level) at run time. A reflective system provides a meta-object protocol (meta-interface) to define the services available at the meta-level. The meta-level can be accessed via a concept of reification. Reification means exposing some hidden aspect of the internal representation and hence they can be accessed by the application (the base-level). The implementation openness offers a straightforward mechanism to insert some behavior to monitor or alter the internal behavior of the platform. This enables the application to be in charge of inspecting and adapting the middleware behavior based on its own needs. Thus, a lightweight middleware with a minimal set of functionality is achieved to run on mobile systems. The main motivation of this approach is to make the middleware more adaptable to its environment and better able to cope with changes. Examples of middleware systems that adopted the concept of reflection are OpenCorba, Open-ORB(Object request Broaker), DynamicTAO, FlexiNet, and Globe. ##### 3.2 Tuple Space Middleware Communication in a wireless environment is characterized by frequent disconnections and limited bandwidth. Communication models such as message passing, RPC, or RMI[6] all have the drawback of tight coupling. This means that a sender has to know the exact identity and address of a receiver. Also, the sender has to wait for the receiver to be ready for exchanging information (synchronization paradigm). In distributed open systems this tends to be too restrictive. A decoupled and opportunistic style of computing is thus required. Computing is expected to proceed even in the presence of disconnection and to exploit connectivity whenever it becomes available. One solution is the concept of tuple space, which was initially introduced by Gelernter in as part of the Linda coordination language. Tuple Space systems[10] have proved their ability for facilitating communication in wireless settings. In general, a tuple space is a globally shared, associatively addressed memory space that is used by processes to communicate. A tuple space system can be realized as a repository of tuples, which are basically a vector of typed values or fields. Client processes create tuples and place them in the tuple space using a write operation. Also, they can concurrently access tuples using read or take operations. Most tuple space systems support both versions of the tuple retrieval operations, blocking and non-blocking. A template, which is similar to a tuple, is used to match the contents of tuples in the tuple space during the retrieval operations. A template matches a tuple if they have an equal number of fields and each template field matches the corresponding tuple field. This form of communication fits well in mobile setting where logical and physical mobility is involved. |Requirements|Reflective|Tuple Space|Context Aware|Event Based| |---|---|---|---|---| |Synchronous/ connection based|X||X|| |Asynchronous/ connectionless||X||X| |Re- configuration|X|||| |Adaptation|X||X|| |Awareness|X||X|| |Light weight||||X| ----- The Tuple space communication model, such as the one used in Linda,provides great flexibility for modeling concurrent process. This approach has also been extended with distributed tuple space. ##### 3.3 Context-Aware Middleware Mobile systems run in an extremely dynamic environment. The execution context changes frequently due to the user’s mobility. Mobile hosts often roam around different areas, and services that are available before disconnecting may not be available after reconnecting. Also, the bandwidth and connectivity quality may quickly alter based on the mobile host movements and their locations. The application developers cannot predict all the possible execution contexts that allow the application to know how to react in every scenario. The middleware has to expose the context information to the application to make it aware of the dynamic changes in execution environment. The application then instructs the middleware on how to adapt its own behavior in order to achieve the best quality of service. Many research groups gave special attention in particular to location awareness. For example, location information was exploited to provide travelers directional guidance, to discover neighboring services, and to broadcast messages to users in a specific area. Most location-aware systems depend on the underlying network operating system to obtain location information and generate a suitable format to be used by the system. The heterogeneity of coordination information is not supported and hence different positioning systems are required to deal with different sensor technologies, such as the Global Positioning System (GPS) outdoors, and infrared and radio frequency indoors. MobiPADS is a middleware system for mobile environment. The principal entity is Mobilets, Which are entities that provide a service, and which can be migrated between different MobiPADS environment. ##### 3.4 Event-Based Middleware Invocation-based middleware systems such as CORBA(Common object request Broaker Architecture) or Java (Remote Method Invocation)[7] are useful abstractions for building distributed systems. The communication model for these platforms is based on a request/reply pattern: an object remains passive until a principle performs an operation on it. This kind of model is adequate for a local area network (LAN) with a small number of clients and servers, but it does not scale well to large networks like the Internet. The main reason is that the request/reply model only supports one-to-one communication and imposes a tight coupling between the involved participants because of the synchronous paradigm. This model is also unsuitable for unreliable and dynamic environment. The event-based communication paradigm is a possible alternative for dealing with large-scale systems. Event notification is the basic communication paradigm that is used by event-based middleware systems. Events contain data that describes a request or message. They are propagated from the sending components to the receiver components. In order to receive events, clients (subscribers) have to express (subscribe) their interest in receiving particular events. Once clients have subscribed, servers (publishers) publish events, which will be sent to all interested subscribers. This paradigm hence offers a decoupled, many-tomany communication model between clients and servers. Asynchronous notification of events is also supported naturally. There are several examples of middleware based on the event-based systems, but not limited to, Hermes, CEA, STEAM, JEDI and ToPSS. ##### 3.5 Other Middleware Solutions There are many other middleware solutions that have been proposed particularly to target mobility aspects. Unpredictable disconnections are one of the major mobility issues that have been addressed by several systems. Systems like Coda, its successor Odyssey], Bayou, and xmiddle have used data replication to increase data availability to mobile users. This allows users access to replicas and to continue their tasks whenever the disconnection operations take place. Each system uses different mechanisms to guarantee the ultimate consistency among the replicas. These mechanisms include the support for discovery of inconsistent data as well as data reconciliation. Services discovery is another well-know problem introduced by user mobility. In a static environment, new services can be easily discovered by asking service providers to register with a well-known location service. In a mobile computing environment, the situation is different since mobile hosts often roam around various areas. Services that were present before disconnecting from the network may not exist after reconnecting. Jini and Ninja Service Discovery Service (SDS) are examples of systems that support dynamic service discovery, Bayou is ----- the system which support s disconnected operations and Jini is the system which support discovery of services. #### 4. Analysis of next-generation middleware This section summarizes the previous discussion on next-generation middleware with an emphasis on lessons learned from investigating the proposed solutions presented in the previous section. We particularly aim to highlight in which extent these solutions are suitable for mobile settings. It is a major challenge to solve all problems of mobile distributed systems. This is true due to the high degree of dynamism in mobile environments. Current middleware platforms like CORBA cannot successfully run in such an environment. Hence, there is an urgent need for new solutions that support particular application requirements such as dynamic reconfiguration, context-awareness, and adaptation. We believe that the reflective approach provides a solid base for building next generation middleware platforms and overcomes the limitations of the current middleware technologies. More specifically, the architecture follows a white box philosophy that provides principled and comprehensive access to internal details. It can also decrease problems of maintaining integrity since each object/interface is attached to a single meta-object at a time. Therefore, any modification to a meta-object can only affect a single object. Some reflective systems support higher level of reflection since they can add or remove methods from objects and classes dynamically and even alter the class of an object at run time. In contrast, others concentrate on a simpler reflective paradigm to achieve a better performance. Their reflective mechanisms are not part of the usual flow of control and only invoked when required. Reflective middleware like FlexiNet and DynamicTAO are built around the concept of object-oriented and component frameworks respectively. Component Frameworks (CFs) were initially defined by Szyperski as “collection of rules and interfaces that govern the interaction of a set of components plugged into them” There are several advantages of using CFs over the object-oriented approach. The uses of CFs are not limited to a particular programming language and there is no inheritance relation between components and framework. Hence, components and CFs can be developed independently, distributed in binary form, and combined at run time. We have noticed that the issue of consistent dynamic reconfiguration is still under research. There is some work in this area that has focused on developing reconfiguration models and algorithms that enforce well defined consistency rules while minimizing system disturbance. Performance is another issue that remains a matter for further investigation. All of the reflective systems presented previously impose a heavy computational load that would cause significant performance degradation on mobile devices. Tuple-space systems exploit the decoupled nature of tuple spaces for supporting disconnected operations in a natural manner. By default they offer an asynchronous interaction paradigm that appears to be more appropriate for dealing with intermittent connection of mobile devices, as is often the case when a server is not in reach or a mobile client requires to voluntary disconnect to save battery and bandwidth. By using a tuple-space approach, we can decouple the client and server components in time and space. In other words, they do not need to be connected at the same time and in the same place. Tuple-space systems support the concept of a space of spaces that offers the ability to join objects into appropriate spaces for ease of access. This opens up the possibility of constructing a dynamic super space environment to allow participating spaces to join or leave at arbitrary time. The ability to use multiple spaces will elevate the overall throughput of the system. Throughout our study, we have noticed that JaveSpaces and TSpaces typically require at least 60Mbytes of RAM. This is not affordable by most handheld devices available on the market nowadays. Context-Aware systems provide mobile applications with the necessary knowledge about the execution context in order to allow applications to adapt to dynamic changes in mobile host and network condition. The execution context includes but is not limited to: mobile user location, mobile device characteristics, network condition, and user activity (i.e., driving or sitting in a room). The context information is typically disclosed in a convenient format to the applications that instruct the middleware system to apply a certain adaptation policy. To our knowledge, most context-aware applications are only focusing on a user’s location while other things of interest are also mobile and changing. We believe that a reflective approach may improve the development of context-aware services and applications. In general, a reflective system provides mobile applications with context information that they need to optimize middleware and their own behaviors. One reflection solution has suggested the use of metadata and reflection to support context-aware applications. Traditional, invocation-based middleware like CORBA follow a request/reply communication style, which does not scale well to large networks like the Internet. Event-based paradigms present an interesting style that supports the development of large-scale distributed systems. In such a system, clients first announce their ----- interest in receiving specific events and then servers broadcast events to all interested clients. Hence, the event-based model achieves a highly decoupled system and many-to-many interaction style between clients and servers. We believe that not a lot of work has managed to merge the publish/subscribe communication approach with event-based middleware systems. Most existing systems do not combine traditional middleware functionality (i.e., security, QoS, transactions, reliability, access control, etc.) with the event-based paradigm. We feel that event-based middleware can be more successful if such functionality is provided in the future. Event-based systems also do not integrate well with object-oriented programming languages due to the major mismatch between the concept of objects and events. Events are viewed as untyped collection of data (attribute/value pairs) whereas current programming languages only support typed objects. Hence, events should support data typing in order to be treated as objects. In addition, the developers are responsible for handling the low-level event transmission issues. Current publish/subscribe systems are restricted to certain application scenarios such as instant messaging and stock quote dissemination. This indicates that such systems are not designed as general middleware platforms. From this discussion, we can realize that until this moment there is no middleware system that can fully support the requirements for mobile applications. Several solutions have considered one aspect or another; however, the door for further research is still wide open. ### 5. Conclusion The proliferation and development of wireless technologies and portable appliances have paved the way for a new computing paradigm called mobile computing. Mobile computing software is expected to operate in environments that are highly dynamic with respect to resource availability and network connectivity. Traditional middleware products, like CORBA and Java RMI, are based on the assumptions that applications in distributed systems will run in a static environment; hence, they fail to provide the appropriate support for mobile applications. This gives a strong incentive to many researchers to develop modern middleware that supports and facilitates the implementation of mobile applications. We discussed the state-of-the-art of middleware for mobile computing. We presented common characteristics and a set of requirements for mobile computing middleware, which allows us to better understand the relationship between the existing bodies of work on nextgeneration middleware. We explained the reasons behind the failure of traditional middleware systems for supporting mobile settings. We also identified, illustrated, and comparatively discussed four middleware classes: reflective middleware, tuple space, context-aware middleware, and event-based middleware. Beside these four categories, a pool of other middleware solutions has been developed to address specific mobility issues. However, none of these middleware systems support all the requirements. We concluded each category with a simple qualitative evaluation and made a number of observations related to some issues that need further investigations. #### 6. References [1] Gordon S. Blair, Geoff Coulson, Anders Andersen, Lynne Blair, Michael Clarke, F´abio Costa, Hector Duran Limon, Tom Fitzpatrick, Lee Johnston, Rui Moreira, Nikos Parlavantzas, and Katia Saikoski, “The Design and Implementation of Open ORB 2”, IEEE Distributed Systems Online, 2(6), 2009. [2] R. Meier and V. Cahill, "STEAM: Event-Based Middleware for Wireless Ad Hoc Networks", in Proceedings of the International Workshop on Distributed Event-Based Systems (ICDCS/DEBS'02). Vienna, Austria, 2009, pp. 639-644. [3] Antonio Carzaniga and Alexander L. Wolf, “ContentBased Networking: A New Communication Infrastructure”, In NSF Workshop on an Infrastructure for Mobile and Wireless Systems, Scottsdale, AZ, October 2008. [4] Antony Rowstron and Peter Druschel, “Pastry: Scalable, Decentralized Object Location and Routing for Large-scale Peer-to-Peer Systems”, In Proc. of Middleware 2009, November 2009. [5] G. Ashayer, H. K. Y. Leung, and H.-A. Jacobsen, “Predicate Matching and Subscription Matching in Publish/Subscribe Systems,” in Proceedings of the Workshop on Distributed Event-based Systems, 22nd International Conference on Distributed Computing Systems, (Vienna, Austria), IEEE Computer Society Press, July 2008. [6]International Journal of Ad Hoc and Ubiquitous Computing (2007) Volume: 2, Issue: 4, Publisher: Inderscience Publishers, Pages: 263 ISSN: 17438225 ----- [7] The impact of research on middleware technology Volume 41, Issue 1 (January 2007 Pages: 89 - 112 Year of Publication: 2007 ISSN:01635980 [8]W. Andreas. HyperDesk's response to the ORB RFP. OMG TC Document 91.1.6, Object Management Group, 492 Old Connecticut Path, Framingham, MA, USA, Jan. 1991. [9] T. Andrews, F. Curbera, H. Dholakia, Y. Goland, J. Klein, F. Leymann, K. Liu, D. Roller, D. Smith, S. Thatte, I. Trickovic, and S. Weerawarana ANSA. [10] The Advanced Network Systems Architecture (ANSA). Reference manual, Architecture Project Management, Castle Hill, Cambridge, UK, 1989. [11] H. E. Bal. _The Shared Data Object Model as a_ _Paradigm for Programming Distributed Systems. PhD_ thesis, Dept. of Computer Science, Vrije Universiteit Amsterdam, The Netherlands, 1989. International Journal of Computer and Communication Technology (IJCCT) ISSN: 2231 0371 Vol 2 Iss 2 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.47893/ijcct.2011.1074?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.47893/ijcct.2011.1074, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "CLOSED", "url": "" }
2,010
[ "Review" ]
false
null
[]
6,842
en
[ { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01416385e99636f6fdff4d317f449a3df426cd4e
[]
0.861302
STAFL: Staleness-Tolerant Asynchronous Federated Learning on Non-iid Dataset
01416385e99636f6fdff4d317f449a3df426cd4e
Electronics
[ { "authorId": "2075371711", "name": "Feng Zhu" }, { "authorId": "2051697771", "name": "Jiangshan Hao" }, { "authorId": "2144174054", "name": "Zhong Chen" }, { "authorId": "2109811281", "name": "Yanchao Zhao" }, { "authorId": "2146711040", "name": "Bing Chen" }, { "authorId": "2248421", "name": "Xiaoyang Tan" } ]
{ "alternate_issns": [ "2079-9292", "0883-4989" ], "alternate_names": null, "alternate_urls": [ "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-247562", "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-247562", "https://www.mdpi.com/journal/electronics" ], "id": "ccd8e532-73c6-414f-bc91-271bbb2933e2", "issn": "1450-5843", "name": "Electronics", "type": "journal", "url": "http://www.electronics.etfbl.net/" }
With the development of the Internet of Things, edge computing applications are paying more and more attention to privacy and real-time. Federated learning, a promising machine learning method that can protect user privacy, has begun to be widely studied. However, traditional synchronous federated learning methods are easily affected by stragglers, and non-independent and identically distributed data sets will also reduce the convergence speed. In this paper, we propose an asynchronous federated learning method, STAFL, where users can upload their updates at any time and the server will immediately aggregate the updates and return the latest global model. Secondly, STAFL will judge the user’s data distribution according to the user’s update and dynamically change the aggregation parameters according to the user’s network weight and staleness to minimize the impact of non-independent and identically distributed data sets on asynchronous updates. The experimental results show that our method performs better on non-independent and identically distributed data sets than existing methods.
# electronics _Article_ ## STAFL: Staleness-Tolerant Asynchronous Federated Learning on Non-iid Dataset **Feng Zhu** **[1,2], Jiangshan Hao** **[1,]*** **, Zhong Chen** **[1,2], Yanchao Zhao** **[1,]*, Bing Chen** **[1]** **and Xiaoyang Tan** **[1]** 1 Collage of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China; smc@nuaa.edu.cn (F.Z.); 736531683@nuaa.edu.cn (Z.C.); cb_china@nuaa.edu.cn (B.C.); x.tan@nuaa.edu.cn (X.T.) 2 Nanjing Research Institute of Electronics Engineering, Nanjing 210007, China ***** Correspondence: jiangshanhao@nuaa.edu.cn (J.H.); yczhao@nuaa.edu.cn (Y.Z.) **Abstract: With the development of the Internet of Things, edge computing applications are paying** more and more attention to privacy and real-time. Federated learning, a promising machine learning method that can protect user privacy, has begun to be widely studied. However, traditional synchronous federated learning methods are easily affected by stragglers, and non-independent and identically distributed data sets will also reduce the convergence speed. In this paper, we propose an asynchronous federated learning method, STAFL, where users can upload their updates at any time and the server will immediately aggregate the updates and return the latest global model. Secondly, STAFL will judge the user’s data distribution according to the user’s update and dynamically change the aggregation parameters according to the user’s network weight and staleness to minimize the impact of non-independent and identically distributed data sets on asynchronous updates. The experimental results show that our method performs better on non-independent and identically distributed data sets than existing methods. [����������](https://www.mdpi.com/article/10.3390/electronics11030314?type=check_update&version=1) **�������** **Citation: Zhu, F.; Hao, J.; Chen, Z.;** Zhao, Y.; Chen, B.; Tan, X. STAFL: Staleness-Tolerant Asynchronous Federated Learning on Non-iid Dataset. Electronics 2022, 11, 314. [https://doi.org/10.3390/](https://doi.org/10.3390/electronics11030314) [electronics11030314](https://doi.org/10.3390/electronics11030314) Academic Editor: Claus Pahl Received: 22 December 2021 Accepted: 17 January 2022 Published: 20 January 2022 **Publisher’s Note: MDPI stays neutral** with regard to jurisdictional claims in published maps and institutional affil iations. **Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **Keywords: federated learning; edge computing; weight divergence; non-iid** **1. Introduction** Mobile phones, wearable devices, and autonomous vehicles are just a few of the modern distributed networks that are generating a wealth of data each day. Due to the growing computational power of these devices, coupled with concerns over transmitting private information, it is increasingly attractive to store data locally and push network computation to the edge. The concept of edge computing is not a new one. Indeed, computing simple queries across distributed, low-powered devices is a decades-old area of research that has been explored under the purview of query processing in sensor networks, computing at the edge, and fog computing [1]. Recent works have also considered training machine learning models centrally but serving and storing them locally; for example, this is a common approach in mobile user modeling and personalization [2]. As the storage and computational capabilities of the devices within edge computing grow, it is possible to leverage enhanced local resources on each device. However, privacy concerns over transmitting raw data require user-generated data to remain on local devices. This has led to a growing interest in federated learning, which explores training statistical models directly on remote devices. In federated learning, data will not appear in other places except the data source, and each edge device cooperates to train a shared global model. The existing research on federated learning is primarily about synchronous communication. Researchers use the method of client selection to reduce the negative impact of stragglers on the global model in federated learning. However, synchronous communication will waste computing resources, since different devices have different computing ----- _Electronics 2022, 11, 314_ 2 of 15 capabilities. In addition, the server can not efficiently use the data in all s due to only a small number of s participating in the model training of each global round. Compared with synchronous communication, asynchronous federated learning will increase the burden on the server and consume a large amount of communication resources; however, it can significantly improve training efficiency. Training efficiency is crucial in Internet of Things tasks since they increasingly focus on real-time. Most of the existing asynchronous algorithms are semi-asynchronous strategies. These methods allow s to upload updates independently; however, they still need to synchronize their information. Another challenge of federated learning is the heterogeneity of data. Since federated learning does not allow data to be transmitted in the communication link, data heterogeneity cannot simply resolve by traditional methods such as scheduling data. The convergence speed of the asynchronous federated learning mechanism will be severely affected if each user has the same aggregation parameter. Moreover, as a type of distributed machine learning, more than 90% of the information exchange in federated learning is redundant. Compressing updates sent from the can reduce the consumption of communication resources in asynchronous federated learning and reduce the overall convergence time of federated learning. In this paper, we design a staleness-tolerate asynchronous federated learning method, STAFL. STAFL can asynchronously receive and aggregate updates from the and reduce the impact of stragglers on the global model by adding penalty parameters. Secondly, STAFL will adjust the aggregation parameters according to the heterogeneity of the user’s data and dynamically change the aggregation factor for each epoch by maintaining a local model parameter list. Finally, this paper uses the method of bit quantization. We compress the updates uploaded by the user and the global information sent by the server. Specifically, the contributions of this paper are as follows: - We have designed a federated learning architecture for asynchronous communication, STAFL. After the user completes the local iteration, it can update the local model information at any time. The server immediately aggregates the update and delivers the latest global model to the user when receiving the update information, thereby reducing the waste of computing resources; - We use the weight divergence of the local model to group users and maintain a list of users’ updated information on the server-side. The longer the list, the more users are considered when aggregating. Different aggregation weights are assigned based on the arrival time of the user update, the amount of data of the user, and the group to which the user belongs to reduce the negative impact of non-independent and identically distributed data sets on asynchronous aggregation; - We conducted many experiments to prove the effectiveness of the proposed method and used the communication compression method to reduce the communication cost of asynchronous federated learning. The experimental results show that STAFL has a significant advantage in convergence speed compared with other methods. The organizational structure of our paper is as follows: in Section 1, we introduce the unique challenges and difficulties of asynchronous federated learning, and propose corresponding solutions. Section 2 describes the related work of this paper. We will introduce the design ideas and a detailed description of the STAFL system in Section 3. In order to verify the effectiveness of the proposed method, we show the performance of STAFL in different scenarios in Section 4. Finally, Section 5 summarizes the main work of this paper. **2. Related Work** In recent years, federated learning has received widespread attention as an efficient cooperative machine learning method. Federated learning can address the concerns of a response time requirement, battery life constraint, bandwidth cost-saving, and data safety and privacy [3–6]. Guo et al. [7] explore the security issues of federated learning and how to design efficient, fast, and verifiable aggregation methods in federated learning. As IoT applications have higher and higher requirements for real-time performance, asynchronous ----- _Electronics 2022, 11, 314_ 3 of 15 federated learning has begun to receive attention as a method that can speed up convergence and reduce wasted computing resources. Although asynchronous federated learning can improve real-time performance, some challenges still need to be resolved [8,9]. In [10], Lu et al. designed an asynchronous federated learning scheme for resource sharing on the Internet of Vehicles and solve many challenges in the Internet of Vehicles scenario with very ingenious methods. In [11], the author proposes an age-aware communication strategy that realizes federated learning through wireless networks by jointly considering the parameters on user devices and the staleness of heterogeneous functions. However, these schemes pay more attention to security and privacy in federated learning, sacrifice part of the training efficiency, and ignore the negative impact of non-IID datasets on the shared model. Damaskinos et al. [12] designed an online system that can be used as middleware on Android OS and machine learning applications to solve the staleness problem. Wu et al. proposed an asynchronous aggregated federated learning architecture [13]. The author designed a bypass to store user updates arriving later in the global epoch and used the user selection method to improve training efficiency. In [14], Chai et al. proposed the use of a tier to distinguish users arriving at different times in a global epoch, aggregate each tier, and analyze convergence. However, these methods are semi-asynchronous communication and do not consider completely asynchronous scenarios, therefore staleness will still affect the system. Statistical heterogeneity is also one of the problems that need to be solved in federated learning and attracts the attention of researchers [15,16]. The baseline FL algorithm FedAvg [17] is known to suffer from instability and convergence issues in heterogeneous settings related to device variability or non-identically distributed data [18]. Since McMahan et al. [17] proposed a benchmark synchronization federated learning data aggregation method, many researchers have explored communication methods, communication architectures, and user selection methods to make machine learning algorithms more efficient. Zhao et al. [19] proposed a method to ensure accuracy when the data are not independent and identically distributed. Sahu et al. [20] improved the algorithm of FedAvg. Chen et al. proposed a federated learning communication framework and joint learning method based on wireless networks [21]. Regarding the combination of non-independent identical distribution and asynchronous communication, Chen et al. [22] proposed a novel asynchronous federated learning algorithm and conducted a theoretical analysis of non-iid scenarios and asynchronous problems. **3. System Design** According to what we proposed in the previous section, this paper improves the existing work as follows. We believe that our method is better than existing methods in federated learning based on asynchronous communication. The symbols used in this paper are shown in Table 1. We assume that, in federated learning, users can upload updates independently and are not restricted by server selection. The central server has sufficient resources to encourage the user to participate in federated learning. In addition, the server has sufficient computing capacity to perform asynchronous aggregation operations without consuming a large amount of time. We consider that there are n users participating in the federated learning. The training data in each user’s device are subject to non-independent and identical distribution. That is, each user will have missing labels. The missing data will pull the weight of local updates in the wrong direction, and users need to cooperate in offsetting the negative impact of missing labels. Each local dataset will not be leaked to any other users during the entire training process. ----- _Electronics 2022, 11, 314_ 4 of 15 **Table 1. Notation and Parameters.** **Notation/Term** **Description** Dataset _D_ _Tg_ Global epoch _α_ staleness hyper-parameter _a_ staleness penalty parameter _S(t)_ staleness penalty function _E(m)_ data distribution payoff _η_ learning rate = [1, 2, . . ., m, . . . ] set of users _M_ model list _L_ _n/β_ length of L _nc_ total data of group c **wm** model weight of user m _3.1. Staleness Tolerant Model_ Compared with synchronous communication, asynchronous federated learning requires additional costs to reduce the staleness caused by straggles. Since there is no clear concept of the global epoch, the server needs to perform global aggregation every time it receives an update uploaded by the user. We assume that α is a hyper-parameter that controls the weight of each new user update. Therefore, for each user’s update on the server-side, there are the following aggregation methods: **wm,** if l = 0; _al1+b_ **[w][m][ + (][1][ −]** _al1+b_ [)][w][old][,] if 0 < l ≤ _n;_ _αwm + (1 −_ _α)wold,_ _otherwise._ **wnew =**    (1) where wold is the global weight information of the last aggregation; l is the total number of times the server aggregates user updates. Whenever the server aggregates a user update, the value of l is increased by one. The time at l denotes as tl, and wm is the update information of user i at time tl. As shown in Figure 1, if the user i’s update is the first update that arrives at the server (w1[1][), the server will send global weight to both users at the] same time (t1) after the next user’s update arrives, because aggregation for only one user is a useless operation. After each update arrives at the server, the server will aggregate this information except at time t0. The term 1 _al+b_ [in Equation (][1][) is the gradually decreasing aggregation parameter for] newly arrived users’ update. When l < n, the previous round of global aggregation wold does not represent the global data distribution. We call wold at this time the prototype global weight. As prototype aggregates more and more user updates, the difference between it and centralized updates will become smaller, and it will be significantly better than the update of a certain user. Therefore, we designed a linearly decreasing weight for aggregation. As the number of user updates aggregated by prototype increases, the aggregate weight of new arrival user updates will be smaller. In this paper, we set a = _n−2_ 2 [and][ b][ =][ 2]n[n]−[−]2[8] [. That] is, the weight will slowly drop from the original 1/2 to 1/4. For prototype, we do not set any penalty terms because the server needs to aggregate more updates at this time. ----- _Electronics 2022, 11, 314_ 5 of 15 **Local** **Model List** client1 client2 client3 client4 client5 1 1 _[T]1_ … … … _wmkm_ _[T]2_ _w4_ _wk_ |ww 01k11|Col2|Col3|T0 0 w 11w1 2w1 4 w|Col5|T1 0 w1w1 w1w2 1 2 4 1 w t|T0 1 …w1 4w 12w31 w|T1 1 …… …w mkm| |---|---|---|---|---|---|---|---| |||w1w1 1 2|||||| |w11 w12 w12 w31 w1 4|t t 0 1||1 t 2||||| ||||||||| ||||||||| ||||||||| ||||||||| ||||||||| **Server** _w51_ user device cloud server _w11_ update informationnormal _w31_ update informationstaleness _w51_ update informationmissing _S t( )4_ staleness penalty |w51|Col2| |---|---| **Figure 1. The flow of STAFL between the cloud server and user device.** When the server aggregates enough updates (l > n), we think that the global model can represent all users’ data and will impose corresponding penalties for stragglers at this time (reflected in α in Equation (1)). Before introducing α, we first introduce the concept of the global epoch of our method. There is a local update counter at the user-side, and the counter’s value will increase by one every time the user uploads an update. When the server receives the updates of the next round from more than 10% of users (i.e., when w[2]1 in Figure 1 arrives at the server), the server considers that it has entered the next global epoch. At this time, if the server receives the previous round of user updates, it will impose corresponding staleness penalties. As shown in Figure 1, when the first update information of client 4 arrives at the server, the server has already entered the second global epoch, and the server will impose staleness penalties on client 4’s update. The staleness penalties term _S(t) will make the aggregation term α smaller. However, α will not only be affected by S(t)._ The difference in user distribution will also affect the size of α. This will be explained in detail in the following subsection. The server considers that the user’s update information has been lost when it still does not receive a user’s update after two global epochs and will send the latest global parameters to this user. Second, the server will not aggregate updates two rounds away from the global epoch because this information will seriously slow down the convergence speed of the global model. _3.2. Weight Divergence Discussion_ Since data cannot be transmitted in the federated learning scenario, it is necessary to find a method to improve convergence efficiency without data scheduling. The server can adjust the appropriate aggregation parameters if it learns the user’s data distribution so that the weights of the global model quickly move closer to the weights of the centralized update. However, due to privacy considerations, users are unlikely to tell the server the number of each label in their local dataset. This would mean that the server cannot accurately obtain the user’s data distribution. Therefore, we need to use the information updated by users to obtain data distribution divergence between users. We first define the concept of the weight difference in this paper. That is the sum of the weight divergence of all layers of the same neural network. Recall the calculation equation of the user’s local iteration: **wm[k][m][+][1]** = w[k]m[m] _m_ [,][ D][m][)][,] (2) _[−]_ _[η][∇][F][(][w][k][m]_ where η is the local learning rate, F is the loss function, and Dm is the dataset owned by user m. It can be seen from Equation (2) that, when all users have the same neural network model, the data owned by the users are the main factor that affects the neural network weight of user m. Intuitively speaking, if the weight divergence in the models owned by two users is smaller, the data distribution is more similar. According to this insight, the ----- _Electronics 2022, 11, 314_ 6 of 15 server can assign different aggregation parameters to the user’s update based on the weight divergence of the update to the user. Intuitively, we need to impose penalties on the users whose local data distribution is more different from the global to ensure that these users’ data will not negatively affect the global model. However, this is not the case. The situation that weight uploaded by the user is quite different from the global model often indicates that the global model owned by the server has not fully aggregated this user’s “knowledge”. Each user’s data are helpful for machine learning tasks and can represent user preferences or other characteristics. Therefore, we give specific compensation to users with large weight divergence so that the server can adapt to the local models of these users more quickly. In this paper, we use E(m) to denote the data distribution payoff. Our initial method is based on the following insight: the server may not have the information carried by the arriving user update if this user update is very different from the global model. At this time, the server will allocate a more considerable aggregation parameter α to the user, which is called the data distribution reward in this paper, even if the user’s update is stale. However, under the asynchronous and parallel federated learning framework, one concern regarding such a method is that the excessive weight difference may be caused not only by uneven data distribution but by potential attackers sending the model trained by wrong or low-quality data. In addition, even if the data owned by each user are of a certain quality, the server cannot determine the divergence between the user’s local weight and the global weight is caused by whether the difference in data distribution is due to staleness. The former is easy to distinguish, since the weight of malicious users will not gradually decrease with the progress of aggregation, and the difference from the global model will always be very high. However, solving the latter server requires additional information. Recall the global aggregation formulation (1) of federated learning. After performing the first round of aggregation, the server has obtained the first round of updated information for most users. We also use experiments to verify the feasibility of the weight difference to measure the difference in data distribution. Figure 2 shows the effect of different distributions of data on the MNIST data set on the user’s model parameters. We can see that the weight divergence of users with the same data distribution will be slight. The data tags owned by user2 and user24 are both 1 and 5, and the number of each data tag is the same. It can be seen from the figure that the weight divergence between the two users is minimal. The data held by user38 are a fuzzy picture generated by GAN, therefore it is very different from all users. _3.3. Aggregation Parameter Settings_ The server will group users according to the user’s updates through the updates aggregated in the first few global epochs. Users with similar weights will be classified together, and the server will consider these users to have the same data distribution. Therefore, our aggregation parameter α consists of two parts: data distribution reward and staleness penalty. The amount of data will affect the time of local iteration. Some users have a small amount of data, and the user’s update may reach the server earlier in each global epoch. Similarly to the FedAvg algorithm, the global aggregation parameters will be affected by the amount of data the user has, that is, _[|D]D[m][|]_ [, where][ D][ is the total number of the training data.] Besides, when a straggler update arrives at the server, the server will add an exponential penalty term e[−][a][(][t][m][−][T][g][0][)] to the update information, where a is a parameter that controls the staleness penalty. The larger a is, the greater the penalty for stale updates. tm is the time when the user arrives at the server, and Tg[0] [is the start time of the latest global epoch. The] staleness penalty is a necessary setting for asynchronous aggregation. Next, we will mainly introduce the influence of data distribution on aggregation parameters. ----- _Electronics 2022, 11, 314_ 7 of 15 **Figure 2. Weight difference between users.** When the user’s update of m arrives at the server, the server will determine whether the user’s update is a delayed update according to the previously described method. When the number of local iterations km of the user m is the same as the global iteration counter _Tg, the staleness penalty parameter a is always 0. This is because most users’ updates can_ arrive in a short time. Imposing staleness penalties for these users is unnecessary and will have a negative impact on the convergence speed. As shown in Figure 1, the server maintains a model list (denoted as ) with a length _L_ of _[n]_ _β_ [.][ β][ is a parameter similar to the “client selection” in synchronous communication,] which controls the range of the aggregation model in a global epoch. That is, the latest _n_ _β_ [users’ updated information will be considered during each global aggregation. The] server dynamically maintains a topology of user model weight distribution based on this information. The server groups different users according to the weight divergence when the user uploads the update for the first time, and the distribution of users in each group is roughly the same (users 2, 29, 24, and 25 in Figure 2). As illustrated in Figure 2, when a user update arrives at the server, the server will calculate the weight divergence dm between the update wm and the previous global model **wold according to the stored local model list. Furhter, it will find the mean value of the** distance between wold and all weights in the model list L, that is, Mean(L) (the black dotted line in the figure). Obviously, when the user’s arrival update is in the outer circle, the global model is more susceptible to the user’s influence. Accordingly, the server will take corresponding reward and punishment measures based on the difference between the arrival update weight and the global model. Specifically, if ∥wm − **wold∥≤** _Mean(L), the_ payoff of the data distribution in the aggregate weight E(m) = 1. When ∥wm − **wold∥≥** ----- _Electronics 2022, 11, 314_ 8 of 15 _Mean(_ ), the server will determine which group the user belongs to. If the model-list _L_ stored on the server has aggregated the data of the group with more than [1] _β_ [, then the] server will add a penalty term to the user’s update. Otherwise, a reward will be given. In conclusion, we have: � 1, if dm < dM; _E(m) =_ (3) _β_ [1][−]nc[n][c] [,] if dm ≥ _dM,_ where dM = Mean(L), and nc is the total amount of data of user group c assigned to user _m by the server. We call the user m that satisfies dm < dM as the “central” user, otherwise_ they are the “edge” user. As shown in Figure 3, when the arriving user is an “edge” user, the server will determine whether enough updates of the data distribution have been aggregated. If the server has aggregated _[n][c]_ _β_ [data of the group, then the server will impose a] penalty for the update (the red cross in the figure). Otherwise, the corresponding reward will be given (orange cross in the picture). If we use α[∗] to denote the aggregate parameters of all users in the model list L, and α[∗][m] is the aggregate parameter of the user m’s update information in model list, then, when l > n, Equation (1) can be rewritten as: _L_ **wnew =** len(L) ### ∑ _m=0_ _α[∗][m]_ if l > n. (4) sum(α[∗]) _[L][[][m][]][,]_ where function sum() is the sum of all items in the aggregate parameter list α[∗]. The detailed process of the STAFL method is described in Algorithm 1. Aggregated update Unaggregated user updates _dm_ Different case of updates Global gradient after last aggregation _dm_ (b) give weight divergence payoff (a) no weight divergence payoff **label 2** **label 4** **Label** **3** **label 1** (c) different case for weight divergence payoff **Figure 3. The impact of data distribution on user weight.** ----- _Electronics 2022, 11, 314_ 9 of 15 **Algorithm 1 Staleness-tolerate asynchronous federated learning.** **Input: user update information wm, the amount of data nm owned by the user m, staleness** function S(l), data distribution payoff E(m), User m update counter km, global epoch counter Tg, aggregation counter l, model-list L. **Output: finalized global model** At server side: **while User local update w[k]m[m]** [arrives at the server][ do] **if user m is straggler then** compute staleness penalty S(l) = exp(a(tm − _Tg[0][))]_ **end if** **if l > n then** compute data distribution payoff E(m) accroding to Equation (3) compute α = _[|D]|D|[m][|]_ _[S][(][l][)][E][(][m][)]_ **end if** perform aggregation using Equation (1) update model-list _L_ **if meet the accuracy requirements then** break; **end if** **end while** At client side: **for each user m do** update local model accroding Equation (2) **end for** _3.4. Model List Update and Weight Divergence Computation_ As mentioned above, for the server to judge the aggregated data distribution of the global model, we need to maintain a model list on the server-side. Although storing user update information will consume additional storage costs, our experimental results show that even holding a very short model list can significantly improve the negative impact of the non-IID dataset on the global model. The longer the model list, the more accurately the server can estimate the data distribution required for the global model. When a user update arrives at the server, the server immediately aggregates the update and stores the user’s weight in the model list. We will delete the oldest user update stored in the model list if the model list is full. In this way, we maintained a fixed-length sequence of user-local model weights that represents the latest global model. We use the Euclidean distance between the weights of the two models to represent the difference in data distribution for users. When a user update arrives at the server, the server immediately computes the Euclidean distance dm between that update and the global model. In addition, it should be noted that, since the computational complexity of calculating the Euclidean distance of the neural network weight will increase exponentially with the number of users participating in the training, we will only group users based on their first-round update information. The local model information uploaded by the user for the first time will only be affected by its local training dataset and will not be affected by the weight of other user models. Therefore, the weight divergence between users with different data distributions in the first epoch will be the largest, thus facilitating server grouping. ----- _Electronics 2022, 11, 314_ 10 of 15 **User** de‐ **10** **1** local **2** train quantization quantization weight **3** **9** **Server** de‐ **5** local **6** model **7** global **8** quantization quantization weight aggregation weight **Figure 4. Weight quantification steps.** _3.5. Reduce Communication Overhead_ Since many users participate in asynchronous communication, it is necessary to use the communication compression mechanism to reduce the communication cost. In this paper, we use the method of communication quantization to reduce the weight accuracy of the network from float (32 bits) to 4 bits. The overall scheme of communication compression is shown in Figure 4. First, each user trains their model locally. After the training is completed, the user quantizes the model weight and sends the quantized content to the server. After receiving the quantized model weight, the server performs a de-quantization operation to restore the original update of the user. Subsequently, the server uses the weight obtained by de-quantization to perform model aggregation to obtain a new global model. Then, it quantizes the model and sends it to all users participating in the training. The user then starts the next local epoch based on the new global model obtained by de-quantization. **4. Experiment Evaluations** To make sure that our experiments are closer to the Internet of Things, we integrated the training of federal learning into the actual environment. We assume that 50 users participate in the training process of federated learning. We also assume that all users are evenly distributed. In other words, the distances between all users, the server, and the transmission delay consumed are equal. The network condition is basically stable, and there is no massive packet loss. However, our method can tolerate a certain amount of data loss and is robust to non-large amounts of data loss. In the comparative experiment, the waiting time for each global epoch is the longest time consumed by the user in that global epoch minus the shortest time. In order to simulate the heterogeneity of devices in federated learning, we set different computing capacities for different devices (that is, we reduce the computing capacities of a device by reducing the frequency) and set a different amount of data for all users. _4.1. Data Settings_ Datasets: we use MNIST, FEMNIST, and CIFAR-10 datasets in this experiment to prove the effectiveness of our proposed method. The MNIST dataset comprises 60,000 training samples and 10,000 testing samples. The image is a fixed size (28 × 28 pixels) with a value of 0 to 1. Each image is flattened and converted into a one-dimensional numpy array of 784 (28 28) features. The CIFAR-10 data set contains 60,000 32 32 color images which _×_ _×_ comprise 50,000 training images and 10,000 test images. These samples are divided into five training batches and one test batch. Each batch has 10,000 samples. The test batch contains 1000 randomly selected images from each category. Training batches contain the remaining images; however, some may contain more images from one category than another. The five training sets contain exactly 5000 images from each class. The Federated Extended |User|Col2|Col3| |---|---|---| |de‐ 10 1 local 2 train quantization quantization weight 3 9 Server||| |de‐ 5 local 6 model 7 global 8 quantization quantization weight aggregation weight||| |||| de‐ **5** local **6** model **7** global **8** quantization quantization weight aggregation weight **9** ----- _Electronics 2022, 11, 314_ 11 of 15 MNIST dataset (FEMNIST) is the extended MNIST dataset based on the writer of the digit and character. Non-independent and identically distributed setting: in order to simulate a nonindependent and identically distributed experimental environment, each user in the experiment randomly selects a part from different categories of the training data set, that is, each user does not completely own all the categories of the training data set, and the number of data in each user is different. In particular, we will replace the data set of one of the users with fuzzy data (noise is added to the image or an image data set without complete GAN training). _4.2. Experimental Results of Model List and Weight Divergence_ We first focus on the performance of our method under different data distributions. In order to simulate different degrees of non-independent and identically distributed data sets, we divide the experimental scenarios into two types. In the first type of nonindependent and identically distributed data set, most users have five or more data labels, and a small number of users only have data with one or two labels. Under this scenario, the data are not extremely non-independent and identically distributed. User updates that arrive continuously within a specific time period can cover all data labels when performing federated learning under this data distribution. In the second non-independent and identically distributed case, each user has at most two data labels. In this case, if the global aggregation operation is performed, the global model will be severely affected by the “non-independent and identically distributed”, and the accuracy jitter will be fairly obvious. As shown in Figures 5 and 6, we compare STAFL with the traditional asynchronous method (i.e., heuristic method, the aggregation parameter α is large at the beginning of training, and, as the global aggregation progresses, the updated aggregation parameter α assigned to new users update will be smaller). In addition, in order to verify the influence of the weight difference grouping on the convergence speed, we use STAFL-1 to represent the accuracy change in the model when we only perform the staleness penalty without the data distribution payoff. We use STAFL-2 to describe the situation where we impose the staleness penalty on stragglers and add the data distribution payoff to “edge” users. When the user’s training data has one or two labels, as shown in Figure 6, the test accuracy of _STAFL-2 improved the fastest. Compared with the heuristic method, only aggregating the_ model list can reduce the impact of the previous stale user update on the global model. _L_ The staleness penalty can also reduce the degree of accuracy jitter. As shown in Figure 5, when most users’ data have more than five labels, the weight divergence between users is not particularly obvious. Therefore it is difficult for the server to divide users into multiple groups. In this case, the effect of adding data distribution payoff to each user is not very obvious. However, it still has a certain outcome. (a) (b) **Figure 5. STAFL performance under scenario 1. (a) Test accuracy of scenario 1. (b) Test loss of** scenario 1. ----- _Electronics 2022, 11, 314_ 12 of 15 (a) (b) **Figure 6. STAFL performance under scenario 2. (a) Test accuracy of scenario 2. (b) Test loss of** scenario 2. _4.3. Model List Length Discussion_ As mentioned in Section 3, the parameter β controls the range of user updates considered during server aggregation. Compared with the traditional asynchronous aggregation method, aggregating the continuously updated model list can reduce the stale updates _L_ that the server aggregated a long time ago. The global model will only be affected by the latest few user updates. Correspondingly, the length of the model list L will affect the scope of server aggregation. Similar to the client selection of synchronous federated learning, the larger the β, the more user updates in a global aggregation, and the greater the possibility of covering all data distributions. Correspondingly, the length of the model list will affect the scope of server ag_L_ gregation. Similar to the client selection in synchronous federated learning, the larger the β, the more user updates in a global aggregation, and the greater the possibility of covering all data distributions. However, in asynchronous federated learning, the length of is not as long as possible. First of all, storing a user’s model requires storage space, _L_ especially a complex neural network, which requires keeping a large amount of model weight information. Secondly, the longer the model list, the more obvious the impact of stale updates during aggregation. Further, a one-time aggregation of some users’ updates is enough for the global model to learn the corresponding knowledge for the server. We explore the impact of different β on the global model in the FEMNIST data set. We use the final aggregated model in each global epoch as the evaluation standard. As can be seen from Figure 7, because the aggregation parameters can compensate for the weight divergence, when the length of the model list is within a certain range, the difference in the accuracy of the overall model is not very obvious. (a) (b) **Figure 7. Different β on FEMNIST dataset. (a) Test accuracy. (b) Test loss.** ----- _Electronics 2022, 11, 314_ 13 of 15 _4.4. Comparison of STAFL with Other Methods_ In this subsection, we compare the performance of STAFL with existing methods. These methods are: - **FedAvg [17]: one of the most traditional synchronous aggregation methods of feder-** ated learning randomly selects a part of users to participate in training in each global epoch and simply discards the stragglers; - **FedProx [23]: FedProx is an improvement over FedAvg. Compared to FedAvg, Fed-** Prox still aggregates updates from some stragglers. In addition, similar to STAFL, it also uses Euclidean distance to improve the model’s performance on non-IID datasets; - **ASO [24]: ASO is a novel asynchronous federated learning algorithm that adaptively** trades off convergence speed and accuracy based on staleness. We evaluate the performance of different methods on the FEMNIST and CIFAR-10 datasets. We assume that the data are extremely non-IID and that 50% of users are stragglers. Table 2 shows the evaluation results. We highlight the optimal performance in different situations in bold. As can be seen from the table, our proposed method does not perform well at the beginning of training. The reason for this is that STAFL needs to calculate the weight divergence between users at the beginning of training, as opposed to other schemes that only perform weighted averages or focus only on staleness. As mentioned above, when a large number of users participates in training, the server needs a lot of time to calculate the Euclidean distance between them, even if the server can perform many calculations in parallel. As training progresses, STAFL can dynamically set aggregation parameters based on the model list and previous grouping results, resulting in better performance than other methods. In contrast, the synchronous aggregation strategy of FedAvg and FedProx cannot maintain the original convergence efficiency when there are a lot of stragglers. The convergence effect of the ASO is also unsatisfactory when the data distribution is extremely heterogeneous. **Table 2. Comparison with other methods.** **Dataset** **Accuracy at 90 s** **Accuracy at 180 s** **Accuracy at 300 s** FEMNIST 0.39 0.51 0.70 **FedAvg** CIFAR-10 0.27 0.49 0.64 FEMNIST 0.37 0.58 0.75 **FedProx** CIFAR-10 **0.39** 0.47 0.69 FEMNIST **0.45** 0.49 0.71 **ASO** CIFAR-10 0.32 0.54 0.67 FEMNIST 0.41 **0.62** **0.81** **STAFL** CIFAR 0.24 **0.56** **0.72** _4.5. Communication Cost Comparison_ In order to reduce the communication overhead of asynchronous federated learning, we adopt the communication compression method described in Section 3. We set the number of users to 2, 3, 4, and 5 to verify the time spent before and after communication compression and the total time required to reach 90% accuracy. We distributed the data to the corresponding number of Raspberry Pis for experimentation. It can be seen from Figure 8a that, after compression, the communication time has been reduced by 10%. Compressing the transmission data does not increase the convergence time. As shown in Figure 8b, the time required for the global model to reach 90% accuracy is reduced. Figure 8c shows the accuracy comparison chart before and after communication compression on the MNIST data set. Since there are fewer users participating in the training, the influence of non-independent and identical distribution of data on convergence is not very obvious. ----- _Electronics 2022, 11, 314_ 14 of 15 (a) (b) (c) **Figure 8. Comparison before and after communication compression. (a) comparison of communica-** tion time. (b) comparison of training time. (c) comparison of test accuracy. **5. Conclusions** In this paper, we design a federated learning system architecture for asynchronous communication, called STAFL. In STAFL, users can upload local updates at any time. The server can use the information stored in a model list to determine whether enough information about a certain data distribution has been aggregated in the global model so that it can better impose corresponding penalties or rewards on arriving updates. We also use communication compression to reduce the communication cost caused by asynchronous aggregation. Compared with other methods, our method focuses more on the effect of data heterogeneity on the global model. STAFL can control the negative impact of non-IID datasets on the convergence rate with the least cost. The experimental results show that our method has a significant improvement in convergence efficiency compared with other methods. In future work, we will focus on model performance on other complex datasets and focus on the points where STAFL can be improved, such as obtaining prior knowledge of data distribution more efficiently and conducting theoretical research. **Author Contributions: Funding acquisition, F.Z., Y.Z. and Z.C.; Methodology, F.Z., J.H., Y.Z. and** Z.C.; Validation, X.T.; Visualization, B.C.; Writing—original draft, J.H.; Writing—review and editing, Y.Z. All authors have read and agreed to the published version of the manuscript. **Funding: This work was supported in part by the National Key Research and Development Program** of China under Grant 2019YFB2102000 and in part by the National Natural Science Foundation of China under Grant(No. 62172215) and in part by the Natural Science Foundation of Jiangsu Province(No. BK20200067), in part by the A3 Foresight Program of NSFC (Grant No. 62061146002). **Conflicts of Interest: The authors declare no conflicts of interest.** **Abbreviations** The following abbreviations are used in this manuscript: FL Federated Learning iid independent and identically distributed non-iid non-independent and identically distributed IoT Internet of Things **References** 1. Bonomi, F.; Milito, R.; Zhu, J.; Addepalli, S. Fog computing and its role in the internet of things. In Proceedings of the First Edition of the MCC Workshop on Mobile Cloud Computing, Helsinki, Finland, 17 August 2012; pp. 13–16. 2. Kuflik, T.; Kay, J.; Kummerfeld, B. Challenges and solutions of ubiquitous user modeling. In Ubiquitous Display Environments; Springer: Berlin/Heidelberg, Germany, 2012; pp. 7–30. 3. Shi, W.; Cao, J.; Zhang, Q.; Li, Y.; Xu, L. Edge computing: Vision and challenges. _IEEE Internet Things J. 2016, 3, 637–646._ [[CrossRef]](http://doi.org/10.1109/JIOT.2016.2579198) 4. Wang, X.; Han, Y.; Wang, C.; Zhao, Q.; Chen, X.; Chen, M. In-edge ai: Intelligentizing mobile edge computing, caching and [communication by federated learning. IEEE Netw. 2019, 33, 156–165. [CrossRef]](http://dx.doi.org/10.1109/MNET.2019.1800286) 5. Lu, Y.; Huang, X.; Dai, Y.; Maharjan, S.; Zhang, Y. Federated learning for data privacy preservation in vehicular cyber-physical [systems. IEEE Netw. 2020, 34, 50–56. [CrossRef]](http://dx.doi.org/10.1109/MNET.011.1900317) ----- _Electronics 2022, 11, 314_ 15 of 15 6. Yu, Z.; Hu, J.; Min, G.; Lu, H.; Zhao, Z.; Wang, H.; Georgalas, N. Federated learning based proactive content caching in edge computing. In Proceedings of the 2018 IEEE Global Communications Conference (GLOBECOM), Abu Dhabi, United Arab Emirates, 9–13 December 2018; pp. 1–6. 7. Guo, X.; Liu, Z.; Li, J.; Gao, J.; Hou, B.; Dong, C.; Baker, T. VeriFL: Communication-Efficient and Fast Verifiable Aggregation for [Federated Learning. IEEE Trans. Inf. Forensics Secur. 2021, 16, 1736–1751. 10.1109/TIFS.2020.3043139. [CrossRef]](http://dx.doi.org/10.1109/TIFS.2020.3043139) 8. Vonderwell, S. An examination of asynchronous communication experiences and perspectives of students in an online course: A [case study. Internet High. Educ. 2003, 6, 77–90. [CrossRef]](http://dx.doi.org/10.1016/S1096-7516(02)00164-1) 9. Chen, Y.; Ning, Y.; Slawski, M.; Rangwala, H. Asynchronous online federated learning for edge devices with non-iid data. In Proceedings of the 2020 IEEE International Conference on Big Data (Big Data), Atlanta, GA, USA, 10–13 December 2020; pp. 15–24. 10. Lu, Y.; Huang, X.; Dai, Y.; Maharjan, S.; Zhang, Y. Differentially Private Asynchronous Federated Learning for Mobile Edge [Computing in Urban Informatics. IEEE Trans. Ind. Inform. 2020, 16, 2134–2143. [CrossRef]](http://dx.doi.org/10.1109/TII.2019.2942179) 11. Liu, X.; Qin, X.; Chen, H.; Liu, Y.; Liu, B.; Zhang, P. Age-aware Communication Strategy in Federated Learning with Energy Harvesting Devices. In Proceedings of the 2021 IEEE/CIC International Conference on Communications in China (ICCC), Xiamen, China, 28–30 July 2021; pp. 358–363. 12. Damaskinos, G.; Guerraoui, R.; Kermarrec, A.M.; Nitu, V.; Patra, R.; Taïani, F. Fleet: Online federated learning via staleness awareness and performance prediction. In Proceedings of the 21st International Middleware Conference, Delft, The Netherlands, 7–11 December 2020; pp. 163–177. 13. Wu, W.; He, L.; Lin, W.; Mao, R.; Maple, C.; Jarvis, S. SAFA: A semi-asynchronous protocol for fast federated learning with low [overhead. IEEE Trans. Comput. 2020, 70, 655–668. [CrossRef]](http://dx.doi.org/10.1109/TC.2020.2994391) 14. Chai, Z.; Chen, Y.; Zhao, L.; Cheng, Y.; Rangwala, H. Fedat: A communication-efficient federated learning method with asynchronous tiers under non-iid data. arXiv 2020, arXiv:2010.05958. 15. Karimireddy, S.P.; Kale, S.; Mohri, M.; Reddi, S.; Stich, S.; Suresh, A.T. Scaffold: Stochastic controlled averaging for federated learning. In International Conference on Machine Learning; PMLR: McKees Rocks, PA, USA, 2020; pp. 5132–5143. 16. Yu, F.; Rawat, A.S.; Menon, A.; Kumar, S. Federated learning with only positive labels. In International Conference on Machine _Learning; PMLR: McKees Rocks, PA, USA, 2020; pp. 10946–10956._ 17. McMahan, B.; Moore, E.; Ramage, D.; Hampson, S.; y Arcas, B.A. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics; PMLR: McKees Rocks, PA, USA, 2017; pp. 1273–1282. 18. Khaled, A.; Mishchenko, K.; Richtárik, P. Tighter theory for local SGD on identical and heterogeneous data. In International _Conference on Artificial Intelligence and Statistics; PMLR: McKees Rocks, PA, USA, 2020; pp. 4519–4529._ 19. Zhao, Y.; Li, M.; Lai, L.; Suda, N.; Civin, D.; Chandra, V. Federated learning with non-iid data. arXiv 2018, arXiv:1806.00582. 20. Sahu, A.K.; Li, T.; Sanjabi, M.; Zaheer, M.; Talwalkar, A.; Smith, V. On the convergence of federated optimization in heterogeneous networks. arXiv 2018, arXiv:1812.06127. 21. Chen, M.; Yang, Z.; Saad, W.; Yin, C.; Poor, H.V.; Cui, S. A joint learning and communications framework for federated learning [over wireless networks. IEEE Trans. Wirel. Commun. 2020, 20, 269–283. [CrossRef]](http://dx.doi.org/10.1109/TWC.2020.3024629) 22. Chen, M.; Mao, B.; Ma, T. FedSA: A staleness-aware asynchronous Federated Learning algorithm with non-IID data. Future _[Gener. Comput. Syst. 2021, 120, 1–12. [CrossRef]](http://dx.doi.org/10.1016/j.future.2021.02.012)_ 23. Li, T.; Sahu, A.K.; Talwalkar, A.; Smith, V. Federated learning: Challenges, methods, and future directions. IEEE Signal Process. _[Mag. 2020, 37, 50–60. [CrossRef]](http://dx.doi.org/10.1109/MSP.2020.2975749)_ 24. Xie, C.; Koyejo, S.; Gupta, I. Asynchronous federated optimization. arXiv 2019, arXiv:1903.03934. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/electronics11030314?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/electronics11030314, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/2079-9292/11/3/314/pdf?version=1642669141" }
2,022
[]
true
2022-01-20T00:00:00
[ { "paperId": "fb1a85443b4fcddba3ba38cc78c02073f18aa0c8", "title": "Stochastic Controlled Averaging for Federated Learning with Communication Compression" }, { "paperId": "bafbb4115c71631746ffaca77a2bfe94a1d3cee1", "title": "Age-aware Communication Strategy in Federated Learning with Energy Harvesting Devices" }, { "paperId": "81c3c02f8d82c826833c7ac74746f65564930feb", "title": "Federated Learning on Non-IID Data: A Survey" }, { "paperId": "11e6675cad1c82f8066b07a392396151322b9063", "title": "FedAT: A Communication-Efficient Federated Learning Method with Asynchronous Tiers under Non-IID Data" }, { "paperId": "0f6ca44d390cf83a533a62f5925e0b9f20d081c3", "title": "FLeet: Online Federated Learning via Staleness Awareness and Performance Prediction" }, { "paperId": "8954778efe3bb122b7ce5924d9b25a33cf05d754", "title": "Federated Learning for Data Privacy Preservation in Vehicular Cyber-Physical Systems" }, { "paperId": "8452a1317237ddebffd80e610ecc773bfb678e9c", "title": "Federated Learning with Only Positive Labels" }, { "paperId": "38fbc724bafeb1b4908cd9608f866e2bd76871ca", "title": "Differentially Private Asynchronous Federated Learning for Mobile Edge Computing in Urban Informatics" }, { "paperId": "d1a7d35ac84502c14737567b72249e952e824d72", "title": "Asynchronous Online Federated Learning for Edge Devices with Non-IID Data" }, { "paperId": "fc7b1823bd8b59a590d0bc33bd7a145518fd71c5", "title": "SCAFFOLD: Stochastic Controlled Averaging for Federated Learning" }, { "paperId": "51b6c71899ad2416b8904a099a8bf5cca1e77139", "title": "SAFA: A Semi-Asynchronous Protocol for Fast Federated Learning With Low Overhead" }, { "paperId": "47883960418480d89c0d180c1c78171b149d9b3a", "title": "A Joint Learning and Communications Framework for Federated Learning Over Wireless Networks" }, { "paperId": "4f783752a59c28df08bad9b22dd9c7bafe4efb08", "title": "Tighter Theory for Local SGD on Identical and Heterogeneous Data" }, { "paperId": "49bdeb07b045dd77f0bfe2b44436608770235a23", "title": "Federated Learning: Challenges, Methods, and Future Directions" }, { "paperId": "8f6602b6ebe2962dacf0b563e73852183e628ddf", "title": "Asynchronous Federated Optimization" }, { "paperId": "3370d724bbba3b8793c7a743bd7ef8b57a88195b", "title": "On the Convergence of Federated Optimization in Heterogeneous Networks" }, { "paperId": "91f7e3856b9ac81bdaceb67e322084c811ed22b3", "title": "Federated Learning Based Proactive Content Caching in Edge Computing" }, { "paperId": "6f89a632ceb8fcb81eac3d7b52e937099659cc6a", "title": "In-Edge AI: Intelligentizing Mobile Edge Computing, Caching and Communication by Federated Learning" }, { "paperId": "5cfc112c932e38df95a0ba35009688735d1a386b", "title": "Federated Learning with Non-IID Data" }, { "paperId": "e3a442aa24e5df7e6b2a25e21e75c4c325f9eedf", "title": "Edge Computing: Vision and Challenges" }, { "paperId": "d1dbf643447405984eeef098b1b320dee0b3b8a7", "title": "Communication-Efficient Learning of Deep Networks from Decentralized Data" }, { "paperId": "03c49f3072f5f98f6272f78b3baa06fe555317dd", "title": "Fixing Issues and Achieving Maliciously Secure Verifiable Aggregation in \"VeriFL: Communication-Efficient and Fast Verifiable Aggregation for Federated Learning\"" }, { "paperId": "040039a3d84f0741d76a67727588f83c0bd53eb5", "title": "FedSA: A staleness-aware asynchronous Federated Learning algorithm with non-IID data" }, { "paperId": "22449b1e182f8aef13ee921ce0917a0c3fe02a49", "title": "Challenges and Solutions of Ubiquitous User Modeling" }, { "paperId": null, "title": "Fog computing and its role in the internet of things" }, { "paperId": "b2d721eb5fe3452a4c44e0b32b3081750a4c73d7", "title": "An examination of asynchronous communication experiences and perspectives of students in an online course: a case study" } ]
12,698
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0141e74a0695b544b3fd99677de1723dd1fe1548
[ "Computer Science" ]
0.857179
Formal Correctness of an Automotive Bus Controller Implementation at Gate-Level
0141e74a0695b544b3fd99677de1723dd1fe1548
IFIP Working Conference on Distributed and Parallel Embedded Systems
[ { "authorId": "2112861", "name": "Eyad Alkassar" }, { "authorId": "2070528983", "name": "P. Böhm" }, { "authorId": "39961125", "name": "Steffen Knapp" } ]
{ "alternate_issns": null, "alternate_names": [ "DIPES", "IFIP Work Conf Distrib Parallel Embed Syst" ], "alternate_urls": null, "id": "35a64850-665d-4cc1-8588-b8f3dbb93be8", "issn": null, "name": "IFIP Working Conference on Distributed and Parallel Embedded Systems", "type": "conference", "url": null }
null
# Formal Correctness of an Automotive Bus Controller Implementation at Gate-Level Eyad Alkassar, Peter B¨ohm, and Steffen Knapp **Abstract We formalize the correctness of a real-time scheduler in a time-triggered** architecture. Where previous research elaborated on real-time protocol correctness, we extend this work to gate-level hardware. This requires a sophisticated analysis of analog bit-level synchronization and message transmission. Our case-study is a concrete automotive bus controller (ABC). For a set of interconnected ABCs we formally prove at gate-level, that all ABCs are synchronized tight enough such that messages are broadcast correctly. Proofs have been carried out in the interactive theorem prover Isabelle/HOL using the NuSMV model checker. To the best of our knowledge, this is the first effort formally tackling scheduler correctness at gatelevel. ## 1 Introduction As more and more safety-critical functions in modern automobiles are controlled by embedded computer systems, formal verification emerges as the only technique to ensure the demanded degree of reliability. When analyzing correctness, as a bottom layer, often, only some synchronous model of distributed electronic control units (ECUs) sharing messages in lock-step is assumed. However, such models are im Eyad Alkassar[1] _· Steffen Knapp[1]_ Saarland University, Dept. of Computer Science, 66123 Saarbr¨ucken, Germany e-mail: {eyad,sknapp}@wjpserver.cs.uni-sb.de Peter B¨ohm[1] Oxford University Computing Laboratory, Wolfson Building, Oxford, OX1 3QD, England e-mail: peter.boehm@comlab.ox.ac.uk 1 Work partially funded by the German Research Foundation (DFG), by the German Federal Ministry of Education and Research (BMBF), and by the International Max Planck Research School (IMPRS). _Please use the following format when citing this chapter:_ ----- plemented at gate-level as highly asynchronous time-triggered systems. Hence it can not suffice to verify certain aspects of a system, as algorithms or protocols only. In this paper we examine a distributed system implementation consisting of ECUs connected by a bus. Our study has to combine arguments from three different areas: (i) asynchronous bit-level transmission, (ii) scheduling correctness, and (iii) classical digital hardware verification at gate-level. Our contribution is to show, by an extended case-study, how analog, real-time and digital proofs can be integrated into one pervasive correctness statement. The hardware model has been formalized in the Isabelle/HOL theorem prover [11] based on boolean gates. It can be translated to Verilog and run on a FPGA. All lemmata relating to scheduling correctness have been formally proven in Isabelle/HOL. We have made heavy use of the model checker NuSMV [5] and automatic tools, e.g. IHaVeIt [18], especially for the purely digital lemmata. Most lemmata dealing with analog communication (formalized using reals) have been shown interactively. **Overview. The correctness of our gate-level implementation splits in two main** parts: (i) the correctness of the transmission of single messages and (ii) the correctness of the scheduling mechanism initiating the message transmission and providing a common time base. Next we outline these two verification goals in detail. The verification of asynchronous communication systems must, at some point, deal with the low-level bit transmission between two ECUs connected to the same bus. The core idea is to ensure that the value broadcast on the bus is stable long enough such that it can be sampled correctly by the receiver. To stay within such a so-called sampling window, the local clocks on the ECUs should not drift apart more than a few clock ticks and therefore need to be synchronized regularly. This is achieved by a message encoding that enforces the broadcast of special bit sequences to be used for synchronization. The correctness of this low-level transmission mechanism cannot be carried out in a digital, synchronous model. It involves asynchronous and real-time-triggered register models taking setup and hold-times of registers as well as metastability into account. Our efforts in this respect are based on [3,8,16]. Ensuring correct message transmission between two ECUs is only a part of the overall correctness. Let us consider a set of interconnected ECUs. The scheduler has to avoid bus contention, i.e. to ensure that only one ECU is allowed to broadcast at a time and that all others are only listening. For that, time is divided into rounds, which are further subdivided into slots. A fixed schedule assigns a unique sender to a given slot number. The gate-level implementation of the scheduler has to ensure that all ECUs have roughly the same notion of the slot-start and end times, i.e. they must agree on the current sender and the transmission interval. Due to drifting clocks some synchronization algorithm becomes necessary. We use a simple idea: A cycle offset is added at the beginning and end of each slot. This offset is chosen large enough to compensate the maximal clock drift that can occur during a full round. The local timers are synchronized only once, at the beginning of each round. This is done by choosing a distinguished master ECU, being the first sender in a round. ----- The combination of the results into a lock-step and synchronous view of the system is now simple. The scheduler correctness ensures that always only one ECU is sending and all other ECUs do listen. Then we can conclude from the first part that the broadcast data is correctly received by all ECUs. Organization of the paper: In the remainder of this section we discuss the re lated work. In Section 2 we introduce our ABC implementation. Our verification approach is detailed in Section 3. Finally we conclude in Section 4. **Related Work. Serial interfaces were subject to formal verification in the work** of Berry et al. [1]. They specified a UART model in a synchronous language and proved a set of safety properties regarding FIFO queues. Based on that a hardware description can be generated and run on a FPGA. However, data transmission was not analyzed. A recent proof of the Biphase-Mark protocol has been proposed by Brown and Pike [4]. Their models include metastability but verification is only done at specification level, rather than at the concrete hardware. The models were extracted manually. Formal verification of clock synchronization in timed systems has a long his tory [9, 12, 17]. Almost all approaches focused on algorithmic correctness, rather than on concrete system or even hardware verification. As an exception Bevier and Young [2] describe the verification of a low-level hardware implementation of the Oral Message algorithm. The presented hardware model is quite simplified, as synchronous data transmission is assumed. Formal proofs of a clock-synchronization circuit were reported by Miner [10]. Based on abstract state machines, a correctness proof of a variant of the WelchLynch algorithm was carried out in PVS. However, the algorithm is only manually translated to a hardware specification, which is finally refined semi-automatically to a gate-level implementation. No formal link between both is reported. Besides, low-level bit transmission is not covered in the formal reasoning. The formal analysis of large bus architectures was tackled among others by Rushby [15] and Zhang [19]. Rushby worked on the time-triggered-architecture (TTA), and showed correctness of several key algorithms as group membership and clock synchronization. Assuming correct clock synchronization, Zhang verified properties of the Flexray bus guardian. Both approaches do not deal with any hardware implementation. The respective standard is translated to a formal specification by hand. In [14] Rushby proposes the separation of the verification of timing-related prop erties (as clock synchronization) and protocol specifications. A set of requirements is identified, which an implementation of a scheduler (e.g. in hardware) has to obey. In short (i) clock synchronization and (ii) a round offset large enough to compensate the maximum clock drift are assumed. The central result is a formal and generic PVS simulation proof between the real-time system and its lock-step and synchronous specification. Whereas the required assumptions are similar to ours, they have not been discharged for concrete hardware. ----- In [12] Rushby’s framework is instantiated with the time triggered protocol (TTP). Pike [13] corrects and extends Rushby’s work, and instantiates the new framework with SPIDER, a fly-by-wire communication bus used by NASA. The time-triggered model was extracted from the hardware design by hand. But neither approaches proved correctness of any gate-level hardware. ## 2 Automotive Bus Controller (ABC) Implementation We consider a time-triggered scenario. Time is divided into so-called rounds each consisting of ns slots. We uniquely identify slots by a tuple consisting of a roundnumber r ∈ N and a slot-number s ∈ [0 : ns _−_ 1]. Predecessors (r, _s)_ _−_ 1 and successors (r, _s)+_ 1 are computed modulo ns. The ABC is split in four main parts: (a) the host-interface provides the connec tion to the host, e.g. a microprocessor, and contains configuration registers (b) the send-environment performs the actual message broadcast and contains a send-buffer (c) the receive-environment takes care of the message reception and contains a receive-buffer (d) the schedule-environment is responsible for the clock synchronization and the obedience to the schedule. **Configuration Parameter. Unless synchronization is performed, slots are locally** _T hardware cycles long. A slot can be further subdivided into three parts; an initial_ as well as a final offset (each off hardware cycles) and a transmission window (tc hardware cycles). The length of the transmission window is implicitly given by the slot-length and the offset. Within each slot a fixed-length message of ℓ bytes is broadcast. The local schedule sendl, that is implemented as a bit-vector, indicates if the ABC is the sender in a given slot. Intuitively, in slot s, if sendl[s] = 1 then the ABC broadcasts the message stored in the send-buffer. Note that the ABC implementation is not aware of the round-number. It simply operates according to the slot-based fixed schedule, that is repeated time and again. The special parameter iwait indicates the number of hardware cycles to be awaited before the ABC starts executing the schedule after power-up. All parameters introduced so far are stored in configuration registers that need to be set by the host (we support memory mapped I/O) during an initialization phase. The host indicates that it has finished the initialization by invoking a setrd command. We do not go into details here, the interested reader may consult [7,8]. **Message Broadcast. The send-environment starts broadcasting the message con-** tained in the send-buffer sb if the schedule-environment raises the startsnd signal. The receive-environment permanently listens on the bus. At an incoming mes sage, indicated by a falling edge (the bus is high-active), it signals the start of a reception to the schedule-environment by raising the startedrcv signal for one cycle. In addition it decodes the broadcast frame and writes the message into the receive buffer rb. ----- **Fig. 1 Schedule Automaton** **Scheduling. The schedule-environment maintains two counters: The cycle counter cy** and the current slot counter csn. Both counters are periodically synchronized at the beginning of every round. All ECUs except the one broadcasting in slot 0 (we call the former slaves and the latter master) synchronize their counters to the incoming transmission in slot 0. Hence, the startedrcv signal from the receive environment is used to provide a synchronized time base (see below). Furthermore, the scheduleenvironment initiates the message broadcast by raising the startsnd signal for one cycle. The schedule environment implements the automaton from Fig. 1. The automa ton takes the following inputs: The startedrcv signal as described above. The signal _setrd denotes the end of the configuration phase. The signal sendl0 indicates if the_ ECU is the sender in the first slot and thus the master. Three signals are used to categorize the cycle counter; eqiwait indicates if the initial iwait cycles have been reached, similar to eqoff and eqT. The signal eqns indicates that the end of a round has been reached, i.e. that the slot counter equals ns 1. Finally sendlcur indicates _−_ if the ABC is the sender in the current slot, i.e. sendlcur = sendl[csn]. The automaton has six states and is clocked each cycle. Its functionality can be summarized as follows: If the reset signal is raised (which is assumed to happen only at power-up) the automaton is forced into the idle-state. If the host has finished the initialization and thus invoked setrd we split cases depending on the sendl0 signal. If the ABC is the master, i.e. if sendl0 holds, the ABC waits first iwait hardware cycles (in the iwait-state), then an additional off cycles (in the offwait-state) before it starts broadcasting the message (in the startbroad-state) and proceeds to the Twait-state. If the ABC is a slave (sendl0 = 0), it waits in the rcvwait-state for an active _startedrcv signal and then proceeds to the Twait-state. There all ABCs await the end_ of a slot indicated by eqT. Then we split cases if the round is finished or not. If the round is not finished yet (indicated by _eqns), all ABCs proceed to the offwait-_ _¬_ state. Furthermore, the sender in the current slot (indicated by sendlcur) proceeds to the startbroad-state, initiates the message broadcast and then proceeds to the _Twait-state; all other ABCs skip the startbroad-state and proceed directly to the_ _Twait-state. At the end of a round, the master simply repeats the ‘normal’ sender_ ----- cycle (from the Twait-state to the offwait-state and finally to the Twait-state again). All other ABCs proceed to the rcvwait-state to await an incoming transmission. Once initialized, the master ABC follows the schedule without any synchroniza tion. At the beginning of a round it waits off many cycles and initiates the broadcast. The clock synchronization on the slave ABCs is done in the rcvwait-state. In this state the cycle counter is not altered but simply stalls in its last value. At an incoming transmission (from the master) the slaves clear their slot-counter and set their cycle counter to off, i.e. the number of hardware cycles at which the master initiated the broadcast. After this all ABCs are (relatively) synchronized to the masters clock. **Hardware Construction. The number of ECUs connected to the bus is denoted ne.** Thus an ECU number is given by u [0 : ne 1]. We use subscript ECU numbers _∈_ _−_ to refer to single ECUs. We denote the hardware configurations of ECUu by hu. If the index u of the ECU does not matter, we drop it. The hardware configuration is split into a host configuration and an ABC configuration. Since we do not go into details regarding the host, we stick to h to denote the configuration of our ABC. Its essential components are: Two single bit-registers, one for sending and one for receiving. Both are directly _•_ connected to the bus. We denote them h.S and h.R. A second receiver register, denoted h.R[ˆ], to deal with metastability (see Sect. 3). _•_ Send buffer h.sb and receive buffer h.rb each capable of storing one message. _•_ The current slot counter h.csn and the cycle counter h.cy. _•_ The schedule automaton is implemented straight-forward as a transition sys _•_ tem on an unary coded bit-vector. We use h.state to code the current state (see Fig. 1). Configuration registers. _•_ The configuration registers are written immediately after reset / power-up. They contain in particular the locally relevant portions of the scheduling function. To simplify arguments regarding the schedule we define a global scheduling function send. Given a slot-number s it returns the number of the ECU sending in this slot. Let sendlu denote the local schedule of ECUu, then send(s) = u ⇔ _sendlu[s] = 1. Note that this definition implicitly requires a unique sender definition_ for each slot. Otherwise correct message broadcast becomes impossible due to bus contention. Thus if ECUu is (locally) in a slot with slot index s and send(s) = u then ECUu will transmit the content of the send buffer h.sb via the bus during some transmission interval. A serial interface that is not actively transmitting during slot (r, _s) puts by_ construction the idle value (the bit 1) on the bus. If we can guarantee that during the transmission interval all ECUs are locally in slot (r, _s), then transmission will be successful. The clock synchronization algorithm_ together with an appropriate choice of the transmission interval will ensure that. ----- _e (j)r_ _ts_ _th_ _clkr_ _cer_ _Rr_ _y_ _Ss,dinr_ _Ω_ _clks_ |ts|e ( r|(j)|)|th|Col6| |---|---|---|---|---|---| |ts||||th|| ||||||| ||||||| |y||||Ω|| |||||Ω|x| ||||||| _e (i)s_ _tpd_ **Fig. 2 Clock Edges** ## 3 Verification **Fig. 3 Schedule** To argue about asynchronous distributed communication systems we have to formalize the behavior of the digital circuits connected to the analog bus. Using the formalization of digital clocks we introduce a hardware model for continuous time. In the remainder of this section we sketch the message transmission correctness, detail the scheduling correctness and combine both into a single correctness statement. **Clocks. The hardware of each ECU is clocked by an oscillator having a nominal** clock period of τref . The individual clock period τu of an ECUu is allowed to deviate by at most δ = 0.15% from τref, i.e. ∀u. | τu _−τref |≤_ _τref ·δ_ . Note that this limitation can be easily achieved by current technology. Thus the relative deviation of two individual clock periods compared to a third clock period is bounded by | τu − _τv |≤_ _τw ·_ _∆_ where ∆ = 2δ _/(1_ _−_ _δ_ ). Given some clock-start offset ou < τu the date of the clock edge eu(i) that starts cycle i on ECUu is defined by eu(i) = ou + _i_ _·_ _τu._ In our scenario all ECUs are connected to a bus. The sending ECUs broadcasts data which is sampled by all other ECUs. Due to clock drift it is not guaranteed, that the timing parameter of the sampling registers are obeyed. This problem is solved by serial interfaces. To argue formally we first introduce a continuous time model for bits being broadcast. **Hardware Model with Continuous Time. The problems solved by serial inter-** faces can by their very nature not be treated in a standard digital hardware model with a single digital clock clk. Nevertheless, we can describe each ECUu in such a model having its own hardware configuration hu. To argue about the sender register h.S of a sending ECU transmitting data via the bus to a receiver register h.R of a receiving ECU, we have to extend the digital model. For the registers connected to the bus –and only for those– we extend the hard ware model such that we can deal with the concepts of propagation delay (tpd), setup time (ts), hold time (th), and metastability of registers. In the extended model used near the bus we therefore consider time to be a real valued variable t. Next we define in the continuous time model the output of the sender register hu.S during cycle i of ECUu, i.e. for t ∈ (eu(i) : eu(i + 1)]. The content of hu.S at time t is ----- denoted by Su(t). In the digital hardware model we denote the value of some register, e.g. R, during cycle i by h[i].R which equals the value at the clock edge eu(i + 1). If in cycle i _−_ 1 the digital clock enable Sce(h[i]u[−][1]) signal was off, we see during the whole cycle the old digital value h[i]u[−][1].S of the register. If the register was clocked (Sce(h[i]u[−][1]) = 1) and the propagation delay tpd has passed, we see the new digital value of the register, which equals the digital input Sdin(h[i]u[−][1]) during the previous cycle (see Fig. 2). Otherwise we cannot predict what we see, which we denote by Ω : _Su(t) =_  h[i]u[−][1].S : Sce(h[i]u[−][1]) = 0 _∧_ _t ∈_ (eu(i) : eu(i + 1)] _Sdin(h[i]u[−][1])_ : Sce(h[i]u[−][1]) = 1 _∧_ _t ∈_ [eu(i)+ _tpd : eu(i_ + 1)] Ω : otherwise The bus is an open collector bus modeled as the conjunction over all registers Su(t) for all t and u. Now consider the receiver register hv.R on any ECUv. It is continuously turned on; thus the register always samples from the bus. In order to define the new digital value hv[j][.][R][ of register][ R][ during cycle][ j][ on][ ECU]v [we have to consider the value] of the bus in the time interval (ev( _j) −_ _ts,_ _ev(_ _j) + th). If during that time the bus_ has a constant digital value x, the register samples that value, i.e. _x_ 0, 1 _._ _t_ _∃_ _∈{_ _}_ _∀_ _∈_ (ev( _j)_ _−_ _ts,_ _ev(_ _j)+_ _th). bus(t) = x ⇒_ _hv[j][.][R][ =][ x][. Otherwise we define][ h]v[j][.][R][ =][ Ω]_ [.] We have to argue how to deal with unknown values Ω as input to digital hard ware. We will use the output of register hu.R only as input to a second register hu.R[ˆ] whose clock enable is always turned on, too. If Ω is clocked into hu.R[ˆ] we assume that hu.R[ˆ] has an unknown but digital value, i.e. hu[j] _[.][R][ =][ Ω]_ _[⇒]_ _[h]u[j][+][1].R[ˆ] ∈{0,_ 1}. In real systems the counterpart of register R[ˆ] exists. The probability that R be comes metastable for an entire cycle and that this causes R[ˆ] to become metastable too is for practical purposes zero. **Continuous Time Lemmata for the Bus. Consider ECUs is the sender and ECUr** is a receiver in a given slot. Let i be a sender cycle such that Sce(h[i]s[−][1]) = 1, i.e. the output of S is not guaranteed to stay constant at time es(i). This change can only affect the value of register R of ECUr in cycle j if it occurs before the sampling edge er( _j) plus the hold time th, i.e. es(i) < er(_ _j)+th. The first cycle that is possibly_ being affected is denoted by cyr,s(i) = min{ j | es(i) < er( _j)+_ _th}._ In what follows we assume that all ECUs other than the sender unit ECUs put the value 1 on the bus and keep their Sce signal off (hence bus(t) = Ss(t) for all t under consideration). Furthermore, we consider only one receiving unit ECUr. Because the indices r and s are fixed we simply write cy(i) instead of cyr,s(i). **Theorem 1 (Message Broadcast Correctness). Let the broadcast start in sender-** _cycle i. The value of the send buffer of ECUsend(s) is copied to all receive buffers on_ _the network side within tc sender cycles, i.e. ∀u. h[cy]u_ [(][i][+][tc][)].rb = h[i]send(s)[.][sb.] This theorem is proven by an in-depth analysis of the send-environment and the receive-environment. For details see [8]. We do not go into details regarding the message transmission here. Instead we focus on the scheduling correctness. ----- **Scheduling. We assume w.l.o.g. that the ECU with number 0 is the master, i.e.** _send(0) = 0. Let pu be the point in time when ECUu is switched on. We assume_ that at most cpmax hardware cycles have passed on the master ECU from the point in time it was switched on until all other ECUs are switched on, too. Thus _u._ _∀_ _|_ _pu −_ _p0 | ≤_ _cpmax ·_ _τ0._ Once initialization is done, all hosts invoke a setrd command. The master ECU waits iwait hardware cycles before it starts executing the schedule. We assume that that there exists a point in time denoted Imax at which all slaves have invoked the _setrd command and await the first incoming message. This assumption can be easily_ discharged by deriving an upper bound for the duration of the initialization phase, say imax hardware cycles in terms of the master ECU, and choosing iwait to be _cpmax + imax. The upper bound can be obtained by industrial worst case execution_ time (WCET) analyzers [6] for the concrete processor and software. We introduce some notation to simplify the arguments regarding single slots. The start time of slot (r, _s) on an ECUu is denoted by αu(r,_ _s). Initially, for all u we_ define αu(0, 0) = Imax. To define the slot start times greater than slot (0, 0) we need a predicate schedexec that indicates if the schedule automaton is in one of three _executing states, i.e. schedexec(h[i]u[) =][ h][i]u[.][state][ ∈{][offwait][,]_ _[Twait][,]_ _[startbroad][}][. Let][ c]_ be the smallest local hardware cycle such that eu(c) is greater than αu((r, _s) −_ 1), _schedexec(h[c]u[)][ holds,][ h][c]u[,]_ _[cy][ =][ 0, and][ h][c]u[.][csn][ =][ s][. Moreover let][ c][′][ be the smallest]_ cycle sucht that eu(c[′]) is greater than αu((r, _s)_ _−_ 1) and h[c]u[′][.][state][ =][ rcvwait][.] _αu(r,_ _s) =_ � _eeuu((cc)[′]) : : otherwise u = 0_ _∨_ _s > 0_ Using the definition of a clock edge we obtain the hardware cycle corresponding to αu(r, _s), denoted by αtu(r,_ _s)._ The local timers are synchronized each round. Next we define the point in time when the synchronization is done in round r. The synchronization end time of round _r on ECUu, denoted by βu(r), is defined similar to the slot start time. Let c be_ the smallest hardware cycle such that that schedexec(h[c]u[)][ holds,][ cycle][c]u [=][ off] [, and] _slotu[c]_ [=][ 0. Then][ β][u][(][r][)][ is defined by][ e][u][(][c][)][.] **Lemma 1 (Synchronization Times Relation). For all u the synchronization of** _ECUu to the master is completed within the adjustment time ad = 10 cycles rel-_ _ative to an arbitrary clock period τw, i.e. β0(r) = α0(r,_ 0) + off · τ0 and βu(r) < _β0(r)+_ 10 _·_ _τw_ The proof of this lemma is split in two parts. First, an analysis of the sender bounds the delay between an active startsnd signal and the actual transmission start. Second, we need to bound the delay on the receiver side until the startedrcv signal is raised after an incoming transmission plus an additional cycle to update the counters and the schedule control automaton. Next we relate the start times of slots on the same ECU. **Lemma 2 (Slot Start Times Relation). The start of slot (r,** _s) on the master ECU_ _depends only on the progress of the local counter, i.e. α0(r,_ _s) = α0((r,_ _s)−1)+T ·τ0._ _The start of slot (r,_ _s) on all other ECUs is given by:_ ----- _αu(r,_ _s) =_ � _βu(r)+(T −_ _off_ ) _·_ _τu_ : s = 1 _αu((r,_ _s)_ _−_ 1)+ _T ·_ _τu_ : s ̸= 1 Proof by induction on r and s using arguments for the concrete hardware. The transmission is started in slot (r, _s) by ECUsend(s) if the local cycle count_ equals off . This point in time is denoted by ts(r, _s) = αsend(s)(r,_ _s)+_ _off ·_ _τsend(s). Ac-_ cording to Theorem 1 the transmission ends at time te(r, _s) = ts(r,_ _s)+_ _tc_ _·_ _τsend(s) =_ _αsend(s)(r,_ _s)+(off +_ _tc)_ _·_ _τsend(s)._ The schedule is correct if the transmission interval [ts(r, _s),te(r,_ _s)] is contained_ in the time interval, when all ECUs are in slot (r, _s), as depicted in Fig. 3._ **Theorem 2 (Schedule Correctness). All ECUs are in slot (r,** _s) before the transmis-_ _sion starts. Furthermore, the transmission must be finished before any ECU thinks_ _it is in the next slot, i.e. αu(r,_ _s) < ts(r,_ _s) and te(r,_ _s) < αu((r,_ _s)+_ 1) This theorem is proven by a case split on (r, _s) using Lemmata 1 and 2. Now we can_ state the overall transmission correctness in the digital hardware model: **Theorem 3 (Overall Transmission Correctness). Consider slot (r,** _s). The value of_ _the send buffer of ECUsend(s) at the start of slot (r,_ _s) is copied to all receive buffers_ _by the end of that slot, i.e. ∀u. h[α]u_ _[t][u][((][r][,][s][+][1][))][−][1].rb = hαsendtsend(s()s)(r,s).sb_ To prove this theorem we combined Theorem 1 and Theorem 2. According to Theorem 1 the actual broadcast is correct if the transmission window [ts(r, _s),te(r,_ _s)] is_ big enough. The latter is proven by Theorem 2. ## 4 Conclusion In this paper we present a formal correctness proof of a distributed automotive system at gate-level (Sect. 3) along with its hardware implementation (Sect. 2). The hardware model has been formalized in Isabelle/HOL on boolean gates. While a simple version of the message transmission correctness has already been published before [8,16], in this new work, we have formally analyzed the scheduler itself and have integrated both results into a single correctness statement. All lemmata relating to scheduling correctness have been formally proven in Isabelle/HOL which took about one person year. We used automatic tools as the symbolic, open source model checker NuSMV, to discharge properties related to bit-vector operations and the schedule automaton of the hardware. With our implementation heavily using bit-vectors, we ran into the infamous state explosion problem. By resorting to IHaVeIt (a domain-reducing preprocessor for model checkers) we were able to cope with this problem. However, missing support for real-linear arithmetic in the automatic tool landscape, made the verification of the analog and timed models tedious. Yet the integration of decision procedures of dense-order logic would be helpful. In short: automatic tools took a ----- heavy burden from us in the digital world but were almost useless for continoustimed analysis. Summing up, our work provides a strong argument for the feasibility of formal and pervasive verification of concrete hardware implementations at gate-level. ## References 1. Berry, G., Kishinevsky, M., Singh, S.: System level design and verification using a syn chronous language. In: ICCAD, pp. 433–440 (2003) 2. Bevier, W., Young, W.: The proof of correctness of a fault-tolerant circuit design. In: Second IFIP Conference on Dependable Computing For Critical Applications, pp. 107–114 (1991) 3. Beyer, S., B¨ohm, P., Gerke, M., Hillebrand, M., In der Rieden, T., Knapp, S., Leinenbach, D., Paul, W.J.: Towards the formal verification of lower system layers in automotive systems. In: ICCD ’05, pp. 317–324. IEEE Computer Society (2005) 4. Brown, G.M., Pike, L.: Easy parameterized verification of biphase mark and 8N1 protocols. In: TACAS’06, LNCS, vol. 3920, pp. 58–72. Springer (2006) 5. Cimatti, A., Clarke, E.M., Giunchiglia, E., Giunchiglia, F., Marco Pistore, M.R., Sebastiani, R., Tacchella, A.: NuSMV 2: An open source tool for symbolic model checking. In: CAV ’02, pp. 359–364. Springer-Verlag (2002) 6. Ferdinand, C., Martin, F., Wilhelm, R., Alt, M.: Cache Behavior Prediction by Abstract Inter pretation. Sci. Comput. Program. 35(2), 163–189 (1999) 7. Hillebrand, M., In der Rieden, T., Paul, W.: Dealing with I/O devices in the context of perva sive system verification. In: ICCD ’05, pp. 309–316. IEEE Computer Society (2005) 8. Knapp, S., Paul, W.: Realistic Worst Case Execution Time Analysis in the Context of Pervasive System Verification. In: Program Analysis and Compilation, LNCS, vol. 4444, pp. 53–81 (2007) 9. Lamport, L., Melliar-Smith, P.M.: Synchronizing clocks in the presence of faults. J. ACM **32(1), 52–78 (1985)** 10. Miner, P.S., Johnson, S.D.: Verification of an optimized fault-tolerant clock synchronization circuit. In: Designing Correct Circuits. Springer (1996) 11. Nipkow, T., Paulson, L.C., Wenzel, M.: Isabelle/HOL: A Proof Assistant for Higher-Order Logic, LNCS, vol. 2283. Springer (2002) 12. Pfeifer, H., Schwier, D., von Henke, F.W.: Formal verification for time-triggered clock syn chronization. In: DCCA-7, vol. 12, pp. 207–226. IEEE Computer Society, San Jose, CA (1999) 13. Pike, L.: Modeling Time-Triggered Protocols and Verifying Their Real-Time Schedules. In: FMCAD’07, pp. 231–238 (2007) 14. Rushby, J.: Systematic formal verification for fault-tolerant time-triggered algorithms. IEEE Transactions on Software Engineering 25(5), 651–660 (1999) 15. Rushby, J.: An overview of formal verification for the time-triggered architecture. In: FTRTFT’02, LNCS, vol. 2469, pp. 83–105. Springer-Verlag, Oldenburg, Germany (2002) 16. Schmaltz, J.: A Formal Model of Clock Domain Crossing and Automated Verification of Time-Triggered Hardware. In: FMCAD’07, pp. 223–230. IEEE/ACM, Austin, TX, USA (2007) 17. Shankar, N.: Mechanical verification of a generalized protocol for byzantine fault tolerant clock synchronization. In: FTRTFT’92, vol. 571, pp. 217–236. Springer, Netherlands (1992) 18. Tverdyshev, S., Alkassar, E.: Efficient bit-level model reductions for automated hardware ver ification. In: TIME 2008, to appear. IEEE Computer Society Press (2008) 19. Zhang, B.: On the Formal Verification of the FlexRay Communication Protocol. Automatic Verification of Critical Systems (AVoCS’06) pp. 184–189 (2006) -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-0-387-09661-2_6?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-0-387-09661-2_6, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://link.springer.com/content/pdf/10.1007/978-0-387-09661-2_6.pdf" }
2,008
[ "JournalArticle" ]
true
2008-09-07T00:00:00
[]
9,041
en
[ { "category": "Medicine", "source": "external" }, { "category": "Physics", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0141e776d7e5a2b7699f1b5cebf8af033b0c3e25
[ "Medicine" ]
0.872367
Bright and dark Talbot pulse trains on a chip
0141e776d7e5a2b7699f1b5cebf8af033b0c3e25
Communications Physics
[ { "authorId": "2116291604", "name": "Jiaye Wu" }, { "authorId": "2239218697", "name": "Marco Clementi" }, { "authorId": "15795306", "name": "E. Nitiss" }, { "authorId": "24355229", "name": "Jianqi Hu" }, { "authorId": "2237914606", "name": "C. Lafforgue" }, { "authorId": "1691785", "name": "C. Brès" } ]
{ "alternate_issns": [ "0868-3166" ], "alternate_names": [ "Commun Phys", "Communications in Physics" ], "alternate_urls": null, "id": "4c30a099-ac6f-4ed0-ac0e-0ca618941906", "issn": "2399-3650", "name": "Communications Physics", "type": "journal", "url": "http://www.nature.com/commsphys/" }
Temporal Talbot effect, the intriguing phenomenon of the self-imaging of optical pulse trains, is extensively investigated using macroscopic components. However, the ability to manipulate pulse trains, either bright or dark, through the Talbot effect on integrated photonic chips to replace bulky instruments has rarely been reported. Here, we design and experimentally demonstrate a proof-of-principle integrated silicon nitride device capable of imprinting the Talbot phase relation onto in-phase optical combs and generating the two-fold self-images at the output. We show that the GHz-repetition-rate bright and dark pulse trains can be doubled without affecting their spectra as a key feature of the temporal Talbot effect. The designed chip can be electrically tuned to switch between pass-through and repetition-rate-multiplication outputs and is compatible with other related frequencies. The results of this work lay the foundations for the large-scale system-on-chip photonic integration of Talbot-based pulse multipliers, enabling the on-chip flexible up-scaling of pulse trains’ repetition rate without altering their amplitude spectra. The generation of temporal Talbot effect, i.e., the formation of temporal self-imaging patterns, in integrated photonic devices, is limited by bulky setups that limit the repetition rate. The authors design a compact and tunable Talbot chip where the GHz-repetition-rate bright and dark pulse trains can be doubled without affecting their spectra.
### ARTICLE https://doi.org/10.1038/s42005-023-01375-x **OPEN** ## Bright and dark Talbot pulse trains on a chip #### Jiaye Wu 1, Marco Clementi 1, Edgars Nitiss1, Jianqi Hu 1, Christian Lafforgue 1 & Camille-Sophie Brès 1✉ Temporal Talbot effect, the intriguing phenomenon of the self-imaging of optical pulse trains, is extensively investigated using macroscopic components. However, the ability to manipulate pulse trains, either bright or dark, through the Talbot effect on integrated photonic chips to replace bulky instruments has rarely been reported. Here, we design and experimentally demonstrate a proof-of-principle integrated silicon nitride device capable of imprinting the Talbot phase relation onto in-phase optical combs and generating the two-fold self-images at the output. We show that the GHz-repetition-rate bright and dark pulse trains can be doubled without affecting their spectra as a key feature of the temporal Talbot effect. The designed chip can be electrically tuned to switch between pass-through and repetition-ratemultiplication outputs and is compatible with other related frequencies. The results of this work lay the foundations for the large-scale system-on-chip photonic integration of Talbotbased pulse multipliers, enabling the on-chip flexible up-scaling of pulse trains’ repetition rate without altering their amplitude spectra. 1 École Polytechnique Fédérale de Lausanne (EPFL), Photonic Systems Laboratory (PHOSL), STI-IEM, Station 11, Lausanne CH-1015, Switzerland. ✉ il ill b @ fl h ----- albot effect, initially described by Henry Fox Talbot in 1836 as the spatial-interference-induced formation of selfimaging patterns[1], has given rise to enormous interest for # T nearly two centuries due to its underlying physical mechanisms and potential applications. Since the invention of lasers and the implementation of modern ultrafast experimental techniques and instruments, there has been a rapid growth of the studies of the Talbot effect in the temporal[2][–][7], spectral[8][,][9], and azimuthal[10][,][11] domains in the last decades[12], which have consolidated the physical and mathematical understanding of this class of phenomena[13]. The temporal and spectral Talbot effects can be linked together through a space-time duality[14], and it was recently found and theorised that all these domains of observations can be unified by the duality by isomorphism[15], i.e., through space-time, position-momentum, and time-frequency dualities. The temporal Talbot effect, in particular, can be fruitfully employed in many ultrafast applications, such as integer, fractional, and arbitrary repetition rate multiplication (RRM)[4][,][5][,][16], broadband full-field invisibility[17], and noiseless intensity amplification[18]. Among these, RRM is highly appealing, especially in the case of extra-cavity scenarios like optical communications[19], passive amplification[20], and microwave photonics[21], where it can be implemented by spectral amplitude and/or phase filtering[2][,][3][,][5][,][7][,][12][,][22][–][24] and provides a train of highrepetition-rate pulses that is hard to achieve in intra-cavity geometries. Besides the commonly studied temporal Talbot effect of bright pulse trains, the mixing Talbot patterns of dark pulse trains at higher RRMs was also recently investigated[7]. Also not long ago, a novel method, based on the combination of Mach-Zehnder interferometers (MZIs) and delay-line interferometers (DLIs), to generate temporal Talbot effects has been experimentally demonstrated[5], shedding light on its suitability for scalable onchip integration. Realising photonic integration of optical functions, e.g., light sources (especially the well-established integrated microcomb sources)[25], amplifiers[26], or signal-processing components[27], is a necessary step towards future applications and has attracted a huge amount of interest. Currently, the Talbot effects are observed and utilised on-chip in microscopy[28], spectroscopy[29], and Talbot-cavity integrated lasers[30][–][32], serving as a coherent inphase coupling element[31]. However, to the best of our knowledge, the on-chip generation of the temporal Talbot effect has rarely been discussed. One of the latest state-of-the-art demonstrations, for example, is spectral phase filtering using a waveguide Bragg grating on silicon[33]. Yet, the limited tunability of the Bragg grating does not allow fine-tuning to produce a higher output quality and leverage the full potential of Talbot photonic chips. Also, the demonstration of an integrated Talbot processor at a lower repetition rate than the reported 10 GHz is of significant importance. The amount of dispersion required—and hence the size of the device—scales indeed quadratically with the inverse of the repetition rate[7], practically limiting the applicability of this approach to lower repetition rate scenarios. Moreover, the Talbot effect of the dark pulse trains has never been investigated in integrated optics to date. Manipulating bright and dark pulse trains with photonic integrated circuits (PICs), fully replacing complex and bulky combinations of instruments and equipment, is itself intriguing, enabling the possibility to seamlessly work with other PICs, and might pioneer potential all-optical systemon-chip (SoC) applications. In this work, we design and experimentally demonstrate, for the first time to the best of our knowledge, a silicon nitride (Si3N4) Talbot PIC based on a cascaded MZI and DLI geometry. Unlike more conventional rate multiplier methods, where multiplying the repetition rate always leads to an alteration of the ti l t k f t f th t l T lb t ff t i th t the comb spectra could stay unaffected by energy redistribution while achieving integer multiplications of the repetition rate. The proof-of-principle PIC can imprint the temporal Talbot phase relation onto the spectral components of an in-phase optical comb and therefore produce the two-fold (2×) Talbot self-image, doubling the repetition rates of bright or dark pulse trains. This technology can be eventually scaled up to N delay lines, resulting in an N-fold self-image. A key advantage over the conventional dispersion-based method in terms of scalability is that, by the proposed MZI-DLI scheme, the required length of the delay lines scales linearly with the inverse of the repetition rate rather than quadratically, making this approach particularly accessible for lower repetition rate combs. In addition, embedded, electrically controlled thermo-optic actuators enable the chip to arbitrarily switch between a pass-through or a 2× RRM output. In principle, it could also be designed for other repetition rates by changing the length of the delay line. We believe that the results of this work could serve as a reference design for large-scale photonic integration and broaden the understanding and the use of MZI onchip devices. Results Principle of temporal Talbot effect in delay line structures. Controlling the repetition rate of a pulse train inside a laser cavity, especially when integer multiplications are desired, can be very difficult. Conventional approaches to realise a multiplied repetition rate reduce the frequency components at the cost of significantly changing the amplitude spectra, e.g., spectral amplitude filtering[22][,][34][,][35]. The Talbot effect offers a solution to preserve the amplitude spectral profile (and, in particular, the number of comb lines) by linearly acting only on the phase spectrum of the input comb. The temporal Talbot phase relation takes the form ϕk = π(p/q)k[2], with k being the comb mode index (k = ± 1, 2, 3, . . . ) relative to the centre line (k = 0). p and q are mutually prime positive integers. By means of quadratic Gauss sums[36], the phase of each self-image also satisfies the Talbot phase relation[15]: �s � φn ¼ �π q [n][2][ þ][ c] : ð1Þ Here, the parameters s and c are related to p and q via Eq. (2), and n is the index of the pulse self-image (n = 0, 1, . . ., N − 1). We follow the notations from refs. [15][,][36] and denote [1/a]b as the a modular multiplicative inverse operation and � � as the Jacobi b symbol. The relation between the parameters s, c, p, q can now be expressed as[7]: 2 ð Þ 8 >>>>>>>>< >>>>>>>>: � � s 2 [1] ¼ 2p ; c [q][ �] [1] ¼ þ q 4 � p � 1 � q ; if q O 2 2 �� s [1] ¼ p ; c ¼ � [p] 2q 4 [�] � q � 1 � p ; if q E 2 2 where O and E are, respectively, the odd and even integer sets. The phase results of Eq. (2) rely on the parity of q. One can easily show that when this aforementioned phase relation is satisfied, the spectral shape is not affected. By implementing the temporal Talbot phase relation, the evolution diagram of temporal profiles of an optical pulse train can be acquired, which is known as the Talbot carpet shown in Fig. 1a. For clarity and without loss of generality, Fig. 1a is a halfperiod demonstration. The other half period (p/q ∈ [1, 2]) is the i f Fi 1 ith th ti (t l fil ) t ----- Fig. 1 The mechanisms of temporal Talbot effect. a The temporal Talbot carpet. Here, the Talbot carpet shown is half of its full period, with marks for 1×–5× self-images. A non-phase-shifted 1× self-image will appear at p/q = 2/1. p and q are phase parameters described in Eq. (1). The 2× self-image at p/q = 1/2 is expected for both bright and dark pulse trains on the designed Talbot photonic chip, as theoretically and experimentally shown in Fig. 3c, f. The physical mechanisms of b the conventional dispersion-based and c the proposed Mach-Zehnder-interferometer (MZI)-based Talbot repetition-ratemultiplication (RRM) realisation. In the transmission and phase spectra, the solid blue lines are the theoretical curves, and the red circles are the points where the comb lines are. p/q = 2 being exactly the same as at p/q = 0. The horizontal axis, p/q, can be understood as phase evolution or propagation distance, for it is also possible to launch a pre-assigned Talbotphased optical comb into an SMF and allow phase accumulation through propagation[7], which is illustrated in Fig. 1b. If a doubled repetition rate is desired, i.e., the 2× Talbot self-image, we seek the phase relation with p/q = 1/2 whose line-by-line phase relation of the comb should be [. . . π/2, 0, π/2, 0, π/2, 0, π/2, . . . ] according to Eq. (2). It is reported that by incorporating the N parallel optical tapped delay line structure, the Talbot phase relation can be losslessly imprinted by re-distributing energy into each delay line with different designed lengths, resulting in N× RRM[5]. The key idea of this mechanism is to combine[24] the effects of spectral amplitude filtering, which achieves N× RRM and phase filtering, which maintains the corresponding amplitude spectrum by altering the phase, to satisfy the temporal Talbot effect conditions. A schematic of this mechanism is shown in Fig. 1c. It is worth emphasising that the mechanism of this process is not as seemingly simple as just delaying the pulse long enough such that the delayed pulse train fits into the intervals of the non-delayed one when recombining and produces N× the repetition rate, as this operation would not, in general, satisfy the Talbot phase relation. On the contrary, the design for the phase is the key to preserving the amplitude spectra. After the energy splitting, the frequency components of the comb on each delay line accumulate diff t h d i ti D t th l li which is exactly [. . . π/2, 0, π/2, 0, π/2, 0, π/2, . . . ] for k = [. . . − 3, − 2, − 1, 0, + 1, + 2, + 3, . . . ]. Therefore, the N = 2 Talbot PIC is capable of imprinting the Talbot phase relation onto an input comb, doubling the repetition rate of the corresponding pulse train while keeping the amplitude spectrum unchanged. It is worth noting that, as one can observe from Eq. (3), the imparted Talbot phase scales linearly with the time delay τdelay. By recalling the expression of the free-spectral range (FSR) for a i DLI Δ /( L) b i L th b l d th nature of the process, each frequency component remains unchanged and maintains perfect coherence with those in other delay lines. When they recombine, the line-by-line interference gives a phase relation that satisfies Eq. (1). This can be regarded as a combination of spectral amplitude filtering and spectral phaseonly filtering[5], and for the N = 2 Talbot PIC designed in this work by using the generalised Landsberg-Schaar identity[37], the transfer function H of the combined filtering can be derived as: [1] N�1 � �[�] ¼ N n[∑]¼0 [exp][ �][2][π][nf][ k][τ][delay][ �] [i][φ][n] ���N¼2 ¼ p[1]ffiffiffi2 exp �� i [π]4 [þ][ i][ϕ][k]� ¼ p[1]ffiffiffi2 exp �� i [π]4 [þ][ i][ π]2[k][2]�; k H f k ¼ 2τdelay ! 3 ð Þ ----- Fig. 2 The Si3N4 Talbot chip. a Design schematic showing the components and the dimensions (cross-sectional view) of the photonic circuit; b Microscopic photograph of the chip; c Theoretical (blue dashed curves) and experimental (red solid curves) transmission spectrum of the device in the frequency multiplication configuration. group index, we observe that, here, the length required scales linearly with the inverse of repetition rate (L ∝ Δν[−][1]), rather than quadratically (L ∝ Δν[−][2]) as in the case of dispersion-based temporal Talbot effect (see the form factor comparison in “Discussion”), making our approach intrinsically advantageous in the perspective application to low repetition rate combs. Chip design and characterisation. To realise the N = 2 tapped delay line structure on a chip, we designed a two-stage cascaded MZI configuration, in which the optical length of the reference arms can be tuned by electrically-driven thermo-optic actuators. A schematic diagram of the device is shown in Fig. 2a. The firststage MZI ensures equal energy distribution into the second stage through local temperature control, while this control is also the key to the pass-through operation mode that allows a direct pass through the chip without manipulation. The second-stage MZI is a DLI with a delay line of 125 ps, corresponding to the half-period interval of a 4 GHz pulse train. The two branches are then recombined at a coupler, whereas the recombination phase is controlled by means of a thermo-optic actuator, which allows slight variations of the effective local index of the waveguide as a function of the bias current, ensuring the possibility of achieving the Talbot condition. A microscopic photograph of the device is shown in Fig. 2b. A sinusoidal interferometric pattern exists owing to the existence of the unbalanced DLI. The transmission spectrum of the device in this configuration is shown in Fig. 2c. The FSR of the interferometer is designed to match the output comb repetition rate of 8 GHz (blue dashed line), as confirmed by the experimental measurement (red solid line). Note that the visibility of the interferogram fringes can be varied between 0 and 1 by acting on the first phase shifter. Similarly, the interferogram transmission function can be shifted horizontally by acting on the DLI local heater, thus allowing the tuning of the alignment of the input comb lines with the transmission function of the device. Proof-of-principle experiments: repetition rate doubling of bright and dark pulse trains. We demonstrate the proof-ofprinciple experiments using the setup shown in Fig. 3a. We use a tunable 10 dBm continuous-wave (CW) laser at C-band as a light source. In light of the methods introduced in refs. [10][,][38], we utilise a harmonic combination of 4, 8, and 12 GHz sinusoidal waves to t th 4 GH d l ti di f (RF) i l f the 1 × 2 lithium niobate (LiNbO3) Mach-Zehnder modulator (MZM). The MZM generates bright and dark pulse trains, respectively, at its two outputs, whose patterns are complementary to each other. Both of such input pulse trains have a full width at half maximum intensity (FWHM) of 50 ps. In the experiment, we connect to one output at a time. The chip is mounted on a temperature-stabilised holder and set to work under a constant temperature of 25 °C. The local temperature of the MZI is also controlled by introducing a DC current through the embedded resistance. This local temperature change slightly affects the effective index and therefore the optical lengths of the reference arm of the first MZI, determining the light energy distribution into the second-stage DLI. By this control, the two modes of operation of this chip can be realised, namely, the pass-through mode, where the light goes through only one arm of the DLI (with negligible propagation loss), and the Talbot mode, where the light split evenly into the two arms of the DLI and recombined into a 2× Talbot self-image, which doubles the repetition rate. The output signal is collected and analysed by a highresolution optical spectrum analyser (OSA) and an oscilloscope (OSC) for the retrieval of its spectral and temporal profiles, respectively. Besides the conventional bright pulse experiments, the designed Talbot chip could also work with dark pulse trains by the same principles. The results of the chip operations are illustrated in Fig. 3b–g. To the best of our knowledge, this is the first demonstration of the temporal Talbot effect of dark pulse trains on a chip, following the discovery and discussions on the mixing Talbot patterns of the dark pulse trains[7]. In Fig. 3b, e, by tuning the local temperature control on the first MZI, the passthrough mode of chip operation presents the exact same temporal profile as the input. Under this circumstance, the optical comb is still an in-phase comb without any manipulation. The measured pulse train fits the theoretical curves well. By finely adjusting the DC voltage of the local temperature control, the DLI starts to provide a relative 125-ps (1/2 × 1/(4 GHz) = 125 ps) delay to one branch of the energy, and the chip can reach a state such that a relatively weaker secondary pulse train appears at the intervals of the original pulse train, finally stabilising at almost the same intensity. The created new pulse train has twice the repetition rate (8 GHz), as shown in Fig. 3c, f, which matches well with the theoretical 8-GHz pulse train in dashed lines The measured FWHM of the bright 8-GHz pulse ----- Fig. 3 Temporal Talbot effect of bright and dark pulse trains on a chip. a Schematic of the experimental setup. CW continuous wave, EDFA erbiumdoped fibre amplifier, PC polarisation controller, OSC oscilloscope, OSA optical spectrum analyser. Temporal profiles of b the original 4-GHz bright pulse trains at pass-through mode operation and c the corresponding 2× Talbot self-image of 8-GHz. Solid curves denote the experimental data, and the dashed curves represent the corresponding theoretical predictions. d Spectra comparison of these two pulse trains. e–g The corresponding temporal and spectral profiles of the dark pulse train. In (b, c, e, f), the theoretical curves are plotted in dashed blue lines, and the experimental ones are illustrated in solid red lines. In (d and g), the 4-GHz spectra are in yellow, and the 8-GHz (Talbot) spectra are in blue. h Temporal profile versus the dissipated power on the firststage interferometer, showing the two operations of the chip and the transition regions in between. train is also 50 ps, while the FWHM of its dark counterpart is slightly smaller with a more un-even DC component between each two dark pulses, as can be interpreted from Fig. 3f. This is due to the destructive intra-pulse interference happening between the overlapping part of the two recombined pulse trains. The spectral results are shown in Fig. 3d, g. The intervals between each comb line for the pass-through and the Talbot ti b th tl 0 032 di t th 4-GHz frequency difference at the C-band. In this figure, the spectra of the pass-through operation mode are plotted in light orange, and those of the Talbot operation mode are plotted in dark blue. The output spectra fit each other quite well without any frequency component loss or bandwidth change, denoting the presence of the temporal Talbot effect, as discussed in the previous section. The small observable spectral discrepancy b t th t ti d i d t l f d ----- power jittering, as well as possible amplifier noises, which is not a part of the Talbot effect and can be neglected, making the comparison between the theoretical analyses and experimental results accurate. The preservation of the spectral profile could be further improved by SoC-level packaging, taking advantage of a well-established technology trend in integrated photonics. We further experimentally investigate the temporal evolution of the 4 GHz pulse train into the 8 GHz one, as depicted in Fig. 3h for the bright pulse train and the chip temperature set to 25 °C. We record the dissipated electrical power on the first-stage local thermo-optic actuator while the DC voltage of the secondstage actuator is kept constant such that in the Talbot operation region, the chip produces an 8 GHz pulse train with equal intensity and unchanged spectrum. The dissipated power is monotonously linked to the effective phase shift, therefore, by measuring the dissipated power, we are equivalently following the evolution of the effective phase shift. It can be observed that the designed Talbot PIC works in Talbot operation mode within the small vicinity of 0 mW. When the DC power increases, the intensity of the secondary pulse train gradually vanishes. This is marked by the transition region in Fig. 3h. From 40 to 120 mW, the chip operates at the pass-through mode, producing the original 4 GHz pulse train. Near 150 mW, the 2× Talbot selfimage of the optical pulse train emerges again. For higher DC power, the chip is in pass-through mode but with a 125-ps delay with respect to the aforementioned pass-through mode region. This also proves that in the pass-through mode, the energy indeed flows mainly through one arm of the second-stage MZI. The power of 180 mW is the upper limit we set for the chip in order to protect it from potential damage. Therefore, Fig. 3h shows a nearly complete operation period of the Talbot PIC, giving a clear picture of its operating mechanism. To further investigate the link between our observations and the Talbot phase relation expressed by Eq. (3), we estimated the phase imparted to each comb line by reconstructing the phase transmission spectrum of our device when operated in the Talbot configuration. Such calculated phase spectrum, shown in Fig. 4, was inferred by exploiting the analytic relation between the real and imaginary parts of the device transfer function. In particular, the trace was obtained numerically from the amplitude transmission spectrum shown in Fig. 2c by a Hilbert transform[39]. Similar to the widely used Kramers-Kronig relations, this method is based on the principle of causality, leading to an analytical relation between the real and imaginary parts of the transfer function. This allows the extraction of the phase from the /2 measured phase theoretical phase theoretical FSR 3 /8 /4 /8 0 0 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 Comb line index (frequency: 4 GHz/div) Fig. 4 The phase of each comb line of the output at Talbot operation. The measured phase is extracted from the free-spectral range (FSR) data shown in Fig. 2c. The grey vertical grid shows the spectral locations of each comb line with a 4 GHz separation. The red dashed line is the theoretical FSR, the purple circles are the theoretical phase, and the blue solid line denotes the i t l h d t transmission spectrum. The grey vertical lines represent the spectral location of the comb lines with respect to the transmission spectrum, and the purple circles are the theoretical Talbot phase calculated by Eq. (1). As expected, the comb lines experience a nearly [. . . π/2, 0, π/2, 0, π/2, 0, π/2, . . . ] phase shift, confirming the matching of the Talbot phase relation in our experiment. The small deviations between the measured phase and the theoretical one are due to the limited wavelength control resolution of the laser used for measuring the transmission spectrum (Fig. 2c). In principle, the designed Talbot PIC can work as a broadband processing device, the main limiting factor being the 2 × 2 embedded multi-mode interference (MMI) couplers, whose nominal 3 dB bandwidth is around 40 nm. For input wavelengths outside this band, the MMI couplers splitting ratio is altered, resulting in an imbalance in the MZI splitting and recombination ratios. Besides, the DLI acts like a double-edged sword. On the one hand, it provides the essential delay to realise Talbot effect, but on the other hand, when operating on ultrashort pulses with a broad spectrum, additional dispersion compensation might be needed to counteract distortion. Discussion The proposed chip design can not only work with the standard 4-GHz pulse train but also with any other repetition frequencies that are the odd multiples of 4 GHz, i.e., 12 GHz, 20 GHz, 28 GHz, etc. Limited by the technical specifications of the RF components, here we discuss theoretically the possibility to operate at these frequencies. The illustrations in Fig. 5a show intuitively how and when the two trains of pulses align to form the doubled repetition rate. The repetition rates of the pulse trains in Fig. 5a are 12 GHz, 20 GHz, and 28 GHz, respectively, and the first pulses in each panel can be regarded as the very first input pulse. The repetition rate requirement, fr = 1/trep, for the designed chip to work is ξ/2 ⋅ trep = τdelay = 125 ps, with ξ being an odd number, as for even values, the delayed pulse train will overlap with the non-delayed one. This simple rule ensures that the delayed pulse train falls precisely in the middle of the original pulse intervals stably. Furthermore, the temporal Talbot phase relation, as shown in Eq. (3), is still satisfied as it is independent of the input repetition rate. Theoretically, at a given carrier wavelength (e.g., in the C-band), the only physical limitations for higher frequencies are pulse generation and temperature control accuracy since the DC power region of Talbot operation might vary. However, the latter can be better controlled in fully integrated platforms. From an engineering perspective, the length of the delayed arm of the second MZI could be extremely long if the design is altered for a much lower frequency (e.g., in the regime of MHz), which can cause a large amount of propagation loss. Fortunately, the state-of-the-art SiN photonic platform provides a <0.2 dB ⋅ cm[−][1] propagation loss[25], making it possible to realise even metre-long delay lines which can work with significantly lower frequencies than the demonstrated 4 GHz. As for the scaling to higher RRMs, the challenge lies in concatenating multiple levels of MZI-DLI interferometers, either in series (cascading) or in parallel[5]. Each level needs to be controlled separately, and due to the potential error and non-uniformity in fabrication, the DC power control can be different from level to level. For higher RRM, the size of the chip grows accordingly, and the system complexity also increases with the RRM factor N. A careful device engineering, together with new cross-disciplinary technologies—such as computer-controlled self-tuning enabled by machine learning[40]—may be useful in such high RRM li ti ----- Fig. 5 Comparisons with other operation frequencies and different solutions. a Illustration of the compatibility with other operating frequencies, with the examples being 12, 20, and 28 GHz inputs and forming the corresponding 24, 40, and 56 GHz outputs. The delay between the two vertical dashed lines is 125 ps, as in the chip design. The blue solid lines denote the original pulse trains, and the red dash-dotted lines represent the delayed pulse trains. b Form factor comparison among the other solutions to double the repetition rate of a pulse train. The schematic sketch of the waveforms (red) is in the temporal domain, and the comb spectra (yellow) are in the frequency domain. MZ DLI Mach-Zehnder delay line interferometer, SMF single-mode fibre. It is intuitive to think that performing Talbot operations in the designed device is not a lossless process due to interference. If we consider the chip to be a one-input one-output device, for an input K-line optical comb, the total power of the source should be Pin / E�� 0��2K �1∑k A�� k�� with k = 0, 1, . . ., K − 1 and Ak being the envelope of the signal. Due to the MZI interference, the total power at the output should be Pout / E�� 0��2ðNKÞ�1∑k A�� k�� which is half of the original input comb (N = 2). This is exactly what happened in the proof-of-principle setup shown in Fig. 3a since the focus lens could only pick up one of the two outputs at a time. In this single-output implementation, the loss scales with 1/N, which is non-negligible for higher N (e.g., 10 dB for N = 10 and 30 dB for N = 1000). However, the efficiency can reach nearly 100% (without taking into account the chip coupling loss) if both of the two outputs are collected. The output coupler can be regarded as a 2 × 2 discrete Fourier transform star network[41], which provides two 2× RRM Talbot pulse trains[5], which may be beneficial for applications such as optical clock distribution among many components. It should be noted that the two output channels are coherent with each other, carrying exactly the same optical signal up to a constant phase factor. Consequently, they can be recombined at a 50:50 directional coupler, recovering the ideally lossless operation that is a peculiarity of Talbot effect. In other words, the total output power of the system is preserved among all the output modes; these can be ideally recombined whenever a single-output mode and low loss are required. When our design is integrated with other optical functional components at an SoC level, the chip coupling loss can be greatly d d d th t t t b ffi i tl tili d Thi again emphasises the importance of miniaturising macroscopic optical setups and integrating them onto a PIC. Another main characteristic of this design is its form factor, made possible by photonic integration in silicon nitride technology. To assess the potential and practicality in future applications, we compare the sizes of approaches for repetition rate doubling, which are plotted in Fig. 5b. In addition to the conventional split and delay practice, which affects the spectrum, there are several methods to assign the Talbot phase relation to each frequency line of an optical comb. The tapped delay line structure can be realised by commercial MZIs and DLIs with conventional centimetre- to decimetre-scale components. The system is therefore 10–20 times larger than the 4.94 mm Talbot PIC and can be more expensive owing to a lower wafer yield. Additionally, they are not suitable for SoC-level integration or seamlessly working with other PICs. Another solution could be assigning the Talbot phase line-by-line using programmable wave-shapers (pulseshapers)[7]. However, these devices are currently bulky (>20× larger in their longest dimension) and less likely to be scaled down and integrated. More importantly, they can be limited by the resolution of their pixels when working with low repetition rate combs. The aforementioned SMF propagation method is simple and relatively convenient when the repetition rate of the pulse train is high. This technique allows the pulse train to naturally evolve to the p/q = 1/2 state by dispersion-induced accumulated phase, as shown in Fig. 1b. However, when it comes to low repetition rate, the length of the SMF needed can be impractically long[13] according to L = 2πp/(q∣β2∣Δω[2]). For a brief comparison, here we t k th d d di i ( l it di i ) ----- β2 = −21.6 ps[2]km[−][1] for silica SMF. The required length for the 4-GHz pulse train at p/q = 1/2 is 230.26 km, which is 4.6 × 10[7] times longer than a chip and would likely display impractically high propagation losses. The lower the frequency is, the greater the advantage of going on-chip. On the PIC level comparison, the current state-of-the-art Talbot chip using Bragg grating waveguide measures a length of ≈8 mm for 2× RRM at 10 GHz input, while for linearly chirped waveguide Bragg gratings, it would be ≈4 cm (c.f., ref. [33]). These values are respectively 1.6× and 8.1× larger than our proposed design, that also operates at a much lower repetition rate of 4 GHz. If the previously reported state-of-the-art designs were to be scaled for the same 4 GHz repetition rate we demonstrate, their devices should be made >4× larger. Conclusions In this work, we propose, design, and experimentally demonstrate a PIC with cascaded MZIs that converts in-phase optical combs into Talbot-phased ones with their repetition rates doubled while keeping their spectra unaltered. The chip is compatible with both bright and dark pulse trains and works with a variety of different repetition rates. The embedded temperature control also allows the chip to work in a pass-through mode, such that switching of the functions of this chip is readily realised. The evolution of the temporal profiles from the original input to the 2× Talbot selfimage is analysed. Additionally, we show, for the first time, to the best of our knowledge, the temporal Talbot effect of dark pulse trains on a chip. On-chip makes our device more compact than other repetition rate doubling solutions. We believe that the results of this work could be very useful in the applications of onchip Talbot effects and could provide a deeper insight into the physics of this phenomenon. The results discussed in this work also have the potential to be applied in optical communications, amplification, and imaging. Methods Chip design. The chip was fabricated through a multi-project wafer run at a commercial foundry (LIGENTEC SA). The Si3N4 waveguides have a dimension of 1.6 × 0.8 μm, encapsulated in the silicon dioxide (SiO2) layer on top of the silicon (Si) substrate, as shown in the cross-sectional view of Fig. 2. According with simulations, the fundamental transverse electric (TE) mode has an effective refractive index of 1.6769 at 1550 nm. The separation between the two parallel waveguides is 22 μm, while the coupler structures enable a 50:50 power splitting and therefore a continuously tunable splitting ratio at the output of the first-stage MZI. The second-stage MZI is a DLI with a delay line corresponding to the half-period interval of a 4 GHz pulse train (125 ps). The electrical resistance of the two thermo-optic phase shifters is ≈50 Ω. Pulse train generation and instruments. The RF signals are generated by an Anritsu MG3692C, an Agilent E8257D, and an Agilent MXG N5183A. The RF amplifier is a ZVA-0.5W303G+ working with a 10 MHz–20 GHz frequency range from MiniCircuits. The MZM LiNbO3 intensity modulator is an AX-1x20MSS-20-PFA-PFA-LV from EOSPACE. The pulse train is generated by a quasi-Nyquist method adapted from ref. [38]. The EDFA is an Optilab EDFA-16-LC-M. The tunable laser source is Yenista Tunics-T100S-HP. The chip is mounted on a custom stage with a Peltier array and sensors, monitored and controlled by a Thorlabs TED200C temperature controller, and set to stabilise at 25 °C. The spectra are collected by an APEX AP2043B OSA d th t l fil d b A il t Infiniium DCA 86100A wide-bandwidth oscilloscope with an Agilent 86105A 20 GHz optical module. Data availability [The data that support the plots within this paper are available at https://doi.org/10.5281/](https://doi.org/10.5281/zenodo.8272168) [zenodo.8272168.](https://doi.org/10.5281/zenodo.8272168) Code availability [The codes used to produce the results of this paper are available at https://doi.org/10.](https://doi.org/10.5281/zenodo.8272168) [5281/zenodo.8272168.](https://doi.org/10.5281/zenodo.8272168) Received: 15 March 2023; Accepted: 6 September 2023; References 1. Talbot, H. LXXVI. Facts relating to optical science. No. IV. Lond. Edinb. Dublin Philos. Mag. J. Sci. 9, 401–407 (1836). 2. Huang, C.-B. & lai, Y. Loss-less pulse intensity repetition-rate multiplication using optical all-pass filtering. IEEE Photonics Technol. Lett. 12, 167–169 (2000). 3. Caraquitena, J., Jiang, Z., Leaird, D. E. & Weiner, A. M. Tunable pulse repetition-rate multiplication using phase-only line-by-line pulse shaping. Opt. Lett. 32, 716 (2007). 4. Maram, R., Cortes, L. R., Van Howe, J. & Azana, J. Energy-preserving arbitrary repetition-rate control of periodic pulse trains using temporal Talbot effects. J. Light. Technol. 35, 658–668 (2017). 5. Hu, J., Fabbri, S. J., Huang, C.-B. & Brès, C.-S. Investigation of temporal Talbot operation in a conventional optical tapped delay line structure. Opt. Express 27, 7922 (2019). 6. Wu, J., Hu, J. & Brès, C.-S. Demonstration of temporal Talbot effect of dark pulse trains. In Conference on Lasers and Electro-Optics. Paper STh5E.6. [https://opg.optica.org/abstract.cfm?uri=cleo_si-2022-STh5E.6 (Optica](https://opg.optica.org/abstract.cfm?uri=cleo_si-2022-STh5E.6) Publishing, 2022). 7. Wu, J., Hu, J. & Brès, C.-S. Temporal Talbot effect of optical dark pulse trains. Opt. Lett. 47, 953–956 (2022). 8. Caraquitena, J., Beltrán, M., Llorente, R., Martí, J. & Muriel, M. A. Spectral self-imaging effect by time-domain multilevel phase modulation of a periodic pulse train. Opt. Lett. 36, 858 (2011). 9. Maram, R. & Azaña, J. Spectral self-imaging of time-periodic coherent frequency combs by parabolic cross-phase modulation. Opt. Express 21, 28824 (2013). 10. Hu, J., Brès, C.-S. & Huang, C.-B. Talbot effect on orbital angular momentum beams: azimuthal intensity repetition-rate multiplication. Opt. Lett. 43, 4033 (2018). 11. Lin, Z., Hu, J., Chen, Y., Yu, S. & Brès, C.-S. Spectral self-imaging of optical orbital angular momentum modes. APL Photonics 6, 111302 (2021). 12. Azana, J. & Muriel, M. Temporal self-imaging effects: theory and application for multiplying pulse repetition rates. IEEE J. Sel. Top. Quantum Electron. 7, 728–744 (2001). 13. Romero Cortés, L., Maram, R., Guillet de Chatellus, H. & Azaña, J. Arbitrary energy-preserving control of optical pulse trains and frequency combs through generalized Talbot effects. Laser Photonics Rev. 13, 1900176 (2019). 14. Kolner, B. Space-time duality and the theory of temporal imaging. IEEE J. Quantum Electron. 30, 1951–1963 (1994). 15. Cortés, L. R., Guillet de Chatellus, H. & Azaña, J. On the generality of the Talbot condition for inducing self-imaging effects on periodic objects. Opt. Lett. 41, 340 (2016). 16. Maram, R., Romero Cortes, L. & Azana, J. Programmable fiber-optics pulse repetition-rate multiplier. J. Light. Technol. 34, 448–455 (2016). 17. Romero Cortés, L., Seghilani, M., Maram, R. & Azaña, J. Full-field broadband invisibility through reversible wave frequency-spectrum control. Optica 5, 779 (2018). 18. Maram, R., Van Howe, J., Li, M. & Azaña, J. Noiseless intensity amplification of repetitive signals by coherent addition using the temporal Talbot effect. Nat. Commun. 5, 5163 (2014). 19. Zheng, B., Xie, Q. & Shu, C. Comb spacing multiplication enabled widely spaced flexible frequency comb generation. J. Light. Technol. 36, 2651–2659 (2018). 20. Pepino, V. M., da Mota, A. F., Borges, B.-H. V. & Teixeira, F. L. Terahertz passive amplification via temporal Talbot effect in metamaterial-based Bragg fibers J Opt Soc Am B 39 1763 (2022) ----- 21. Wun, J.-M. et al. Photonic high-power continuous wave THz-wave generation by using flip-chip packaged uni-traveling carrier photodiodes and a femtosecond optical pulse generator. J. Light. Technol. 34, 1387–1397 (2016). 22. Hillerkuss, D. et al. Simple all-optical FFT scheme enabling Tbit/s real-time signal processing. Opt. Express 18, 9324 (2010). 23. A. Preciado, M. & A. Muriel, M. All-pass optical structures for repetition rate multiplication. Opt. Express 16, 11162 (2008). 24. Chuang, H.-P. & Huang, C.-B. Generation and delivery of 1-ps optical pulses with ultrahigh repetition-rates over 25-km single mode fiber by a spectral lineby-line pulse shaper. Opt. Express 18, 24003 (2010). 25. Lihachev, G. et al. Platicon microcomb generation using laser self-injection locking. Nat. Commun. 13, 1771 (2022). 26. Liu, Y. et al. A photonic integrated circuit-based erbium-doped amplifier. Science 376, 1309–1313 (2022). 27. Feldmann, J. et al. Parallel convolutional processing using an integrated photonic tensor core. Nature 589, 52–58 (2021). 28. Han, C., Pang, S., Bower, D. V., Yiu, P. & Yang, C. Wide field-of-view on-chip Talbot fluorescence microscopy for longitudinal cell culture monitoring from within the incubator. Anal. Chem. 85, 2356–2360 (2013). 29. Katiyi, A. & Karabchevsky, A. Deflected Talbot-mediated overtone spectroscopy in near-infrared as a label-free sensor on a chip. ACS Sens. 5, 1683–1688 (2020). 30. Wang, L. et al. Phase-locked array of quantum cascade lasers with an integrated Talbot cavity. Opt. Express 24, 30275 (2016). 31. Meng, B. et al. Coherent emission from integrated Talbot-cavity quantum cascade lasers. Opt. Express 25, 3077 (2017). 32. Xu, Y. et al. Phase-locked terahertz quantum cascade laser array integrated with a Talbot cavity. Opt. Express 30, 36783 (2022). 33. Kaushal, S. & Azaña, J. On-chip dispersive phase filters for optical processing of periodic signals. Opt. Lett. 45, 4603 (2020). 34. Geng, Z. et al. Photonic integrated circuit implementation of a sub-GHzselectivity frequency comb filter for optical clock multiplication. Opt. Express 25, 27635 (2017). 35. Xie, Y., Zhuang, L. & Lowery, A. J. Picosecond optical pulse processing using a terahertz-bandwidth reconfigurable photonic integrated circuit. Nanophotonics 7, 837–852 (2018). 36. Fernández-Pousa, C. R. On the structure of quadratic Gauss sums in the Talbot effect. J. Opt. Soc. Am. A 34, 732 (2017). 37. Berndt, B. C. & Evans, R. J. The determination of Gauss sums. Bull. Am. Math. Soc. 5, 107–129 (1981). 38. Soto, M. A. et al. Optical sinc-shaped Nyquist pulses of exceptional quality. Nat. Commun. 4, 2898 (2013). 39. Dicaire, M.-C. N., Upham, J., De Leon, I., Schulz, S. A. & Boyd, R. W. Group delay measurement of fiber Bragg grating resonances in transmission: Fourier transform interferometry versus Hilbert transform. J. Opt. Soc. Am. B 31, 1006 (2014). 40. Baumeister, T., Brunton, S. L. & Nathan Kutz, J. Deep learning and model predictive control for self-tuning mode-locked lasers. J. Opt. Soc. Am. B 35, 617 (2018). 41. Marhic, M. E. Discrete Fourier transforms by single-mode star networks. Opt. Lett. 12, 63 (1987). Acknowledgements This work is supported by the Swiss National Science Foundation (Grant No. 200021188605). Author contributions J.H. conceived the original idea of this work. E.N. and J.H. designed the chip. J.W., M.C., and C.-S.B. designed the experiment. J.W. conducted the main experiments with the assistance of M.C. C.L. helped with the automation of the transmission spectrum measurement. C.-S.B. provided the experimental resources. All authors took part in analysing the data. J.W. wrote the manuscript with inputs from others. M.C. E.N., J.H., C.L., and C.-S.B. provided in-depth reviews and discussions in the revision of the manuscript. C.S.B. supervised the project. All authors have proofread the manuscript. Competing interests The authors declare no competing interests. Additional information Correspondence and requests for materials should be addressed to Camille-Sophie Brès. Peer review information Communications Physics thanks Jose Azaña and the other anonymous reviewer(s) for their contribution to the peer review of this work. [Reprints and permission information is available at http://www.nature.com/reprints](http://www.nature.com/reprints) Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from [the copyright holder. To view a copy of this license, visit http://creativecommons.org/](http://creativecommons.org/licenses/by/4.0/) [licenses/by/4.0/.](http://creativecommons.org/licenses/by/4.0/) © The Author(s) 2023 -----
{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC11041698, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.nature.com/articles/s42005-023-01375-x.pdf" }
2,023
[ "JournalArticle" ]
true
2023-09-13T00:00:00
[ { "paperId": "a36974c20235c066351ad0a42914e353680983dd", "title": "Phase-locked terahertz quantum cascade laser array integrated with a Talbot cavity." }, { "paperId": "13050b7bcf11bdfc41bb82785ab83bc5536f17f1", "title": "Demonstration of Temporal Talbot Effect of Dark Pulse Trains" }, { "paperId": "8d12687c2527ace9398f96d353d53e209d9f3fec", "title": "A photonic integrated circuit–based erbium-doped amplifier" }, { "paperId": "1cc4ade9580ff726e1069b40035afd4c5fbe41cb", "title": "Temporal Talbot effect of optical dark pulse trains." }, { "paperId": "6dee497fba90d418e8a1a52fc7cf09d2888fc23f", "title": "Spectral self-imaging of optical orbital angular momentum modes" }, { "paperId": "02f94b036da23b154df3921c2b7db40424fa7dc9", "title": "Platicon microcomb generation using laser self-injection locking" }, { "paperId": "f27c78f1ff6a8fcba34b6f85d92404cc85575655", "title": "Parallel convolutional processing using an integrated photonic tensor core" }, { "paperId": "77f6e83beeb54b6aa45355394f6b74e71453b98a", "title": "On-chip dispersive phase filters for optical processing of periodic signals." }, { "paperId": "17452a90d818777c54fe18176832e866e1dcc88c", "title": "Deflected Talbot mediated overtone spectroscopy in near-infrared as a label-free sensor on a chip." }, { "paperId": "95fc04c0e76f7c72de9112449f83854f1285b6d7", "title": "Arbitrary Energy‐Preserving Control of Optical Pulse Trains and Frequency Combs through Generalized Talbot Effects" }, { "paperId": "3df751b2d2bbca7e4f4779516cf27bc455d7eda2", "title": "Investigation of temporal Talbot operation in a conventional optical tapped delay line structure." }, { "paperId": "75e5578dd31ffc7d5a72e5907c7215b928216b1a", "title": "Talbot effect on orbital angular momentum beams: azimuthal intensity repetition-rate multiplication." }, { "paperId": "83be251797b93207ed75c207aa0e0059600e9ef2", "title": "Comb Spacing Multiplication Enabled Widely Spaced Flexible Frequency Comb Generation" }, { "paperId": "25fde7be42380d37ec9211c5babeba0375f63b7f", "title": "Full-field broadband invisibility through reversible wave frequency-spectrum control" }, { "paperId": "a82f7a94be34d366804d27a340ab902ef49cb353", "title": "Picosecond optical pulse processing using a terahertz-bandwidth reconfigurable photonic integrated circuit" }, { "paperId": "35f35e2fd41b73592b2863e0f755942e9dde74c7", "title": "Photonic integrated circuit implementation of a sub-GHz-selectivity frequency comb filter for optical clock multiplication." }, { "paperId": "7cb572dc89301820b58e46293ac5157912760339", "title": "Coherent emission from integrated Talbot-cavity quantum cascade lasers." }, { "paperId": "3907ece14c17a5e42e10da6403f731e1c0f1dc28", "title": "Energy-Preserving Arbitrary Repetition-Rate Control of Periodic Pulse Trains Using Temporal Talbot Effects" }, { "paperId": "de8013b5806c50c5e80b2638201fdaddc69e8b51", "title": "On the structure of quadratic Gauss sums in the Talbot effect." }, { "paperId": "7a82e573ef33f3091f3705d6de1166b735bca00f", "title": "Phase-locked array of quantum cascade lasers with an integrated Talbot cavity." }, { "paperId": "787c134abfc01703cf0bc5059f7a9e6dee9f897a", "title": "Photonic High-Power Continuous Wave THz-Wave Generation by Using Flip-Chip Packaged Uni-Traveling Carrier Photodiodes and a Femtosecond Optical Pulse Generator" }, { "paperId": "b6b354224672f65ab57c9f3a90d2b9875ee60658", "title": "On the generality of the Talbot condition for inducing self-imaging effects on periodic objects." }, { "paperId": "8a1942e4047256830dafecde51ad0f337b828a50", "title": "Programmable Fiber-Optics Pulse Repetition-Rate Multiplier" }, { "paperId": "86ae2258b12d6b8f43a4e6a513ce8636b686fe55", "title": "Noiseless intensity amplification of repetitive signals by coherent addition using the temporal Talbot effect" }, { "paperId": "bed8e491e8e9a50753c28d610af05ccee836ebe9", "title": "Group delay measurement of fiber Bragg grating resonances in transmission: Fourier transform interferometry versus Hilbert transform" }, { "paperId": "cf75c88e54a43ff9a8e1ae2cfe7197fe0f3ed383", "title": "Optical sinc-shaped Nyquist pulses of exceptional quality" }, { "paperId": "3972b0ae8b76b511f56fad4fa23041ad6d73a0a3", "title": "Spectral self-imaging of time-periodic coherent frequency combs by parabolic cross-phase modulation." }, { "paperId": "e2cf6402dcd422894c9cb83e9e80a7497c63ae5d", "title": "Wide field-of-view on-chip Talbot fluorescence microscopy for longitudinal cell culture monitoring from within the incubator." }, { "paperId": "87f14b79fe519d1b39da29c4b687dc3d04b640e3", "title": "Spectral self-imaging effect by time-domain multilevel phase modulation of a periodic pulse train." }, { "paperId": "335a85ddf5e33db4225cdfc21a534a35408150fc", "title": "Generation and delivery of 1-ps optical pulses with ultrahigh repetition-rates over 25-km single mode fiber by a spectral line-by-line pulse shaper." }, { "paperId": "ace13ab20c7c21f669c53d8e787554eb65f888c0", "title": "Simple all-optical FFT scheme enabling Tbit/s real-time signal processing." }, { "paperId": "58866fe859e560cd45a8c57c200f12823723a3c2", "title": "All-pass optical structures for repetition rate multiplication." }, { "paperId": "a6f813fd2b6ff29a571f4ea112cc237bee6cfedd", "title": "Tunable pulse repetition-rate multiplication using phase-only line-by-line pulse shaping." }, { "paperId": "a3e40b734cb9a776713e8fc0c74d0ff54113b810", "title": "Temporal self-imaging effects: theory and application for multiplying pulse repetition rates" }, { "paperId": "b18fa39dc6b9f74c24d81cdda141d7306a2fae05", "title": "Loss-less pulse intensity repetition-rate multiplication using optical all-pass filtering" }, { "paperId": "c9cd9780b47f1c9e89a2923c0d40cc28886b73b2", "title": "Space-time duality and the theory of temporal imaging" }, { "paperId": "001d692abfe32832cdaa5ad8a23a9a263d297c09", "title": "Communications in Physics" }, { "paperId": "4eb2835cff3675ad7d6441a8caea5a75d07a9579", "title": "The determination of Gauss sums" }, { "paperId": "7cf7bb922b9fe3adb7fc175d144d495338377856", "title": "LXXVI. Facts relating to optical science. No. IV" }, { "paperId": null, "title": "Deep learning and model predictive control for self-tuning mode-locked lasers. J. Opt. Soc. Am. B 35 , 617" }, { "paperId": "9b6cc9fccc61ea062997bb3e6e2257ca856bedbc", "title": "Loss-less pulse intensity repetition-rate multiplication using optical all-pass filtering" }, { "paperId": "8c9e89ca84c15315e13418ced45c6e6562d3e0f8", "title": "Discrete Fourier transforms by single-mode star networks." }, { "paperId": null, "title": "published maps and institutional af fi liations" }, { "paperId": null, "title": "Terahertz passive ampli fi cation via temporal Talbot effect in metamaterial-based Bragg fi bers" } ]
11,749
en
[ { "category": "Law", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01456d7ead48b17af151463d73aee8611ad0f450
[]
0.923111
NFTs AND COPYRIGHT LAW
01456d7ead48b17af151463d73aee8611ad0f450
SCIENCE International Journal
[ { "authorId": "2221475551", "name": "Belma Mujević" }, { "authorId": "52647070", "name": "M. Mujević" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
NFT stands for “non-fungible token” and refers to a cryptographically protected asset that represents a unique object, work of art, real estate, ticket, or certificate. Like cryptocurrencies, each NFT contains ownership data to facilitate identification and transfer between token holders. By purchasing a token, the owner gets ownership of an asset in the digital or real world. NFTs are often minted by artists and creatives, but collectors and investors who purchase and use them may have a different perspective on who owns the copyrights to the content associated with the NFTs. Attorneys specializing in art law, although they have not yet fully explored it, are developing a familiarity with the current crypto community and future metaverse in order to understand how a public ledger registration tool has created scarcity and value for digital assets, such as digital artwork.
UDK: 347.78:004.031.4(4 672EU) # NFTs AND COPYRIGHT LAW _Belma Mujević[1*], Mersad Mujević[2]_ [1The University of Szeged, Faculty of Law and Political Sciences, Hungary, e-mail: mujevic5@gmail.com](mailto:mujevic5%40gmail.com?subject=) [[2]International University in Novi Pazar, Republic of Serbia, e-mail: mersadm@t-com.me](mailto:mersadm%40t-com.me?subject=) **Abstract: NFT stands for “non-fungible token” and refers to a cryptographically protected asset that represents a** unique object, work of art, real estate, ticket, or certificate. Like cryptocurrencies, each NFT contains ownership data to facilitate identification and transfer between token holders. By purchasing a token, the owner gets ownership of an asset in the digital or real world. NFTs are often minted by artists and creatives, but collectors and investors who purchase and use them may have a different perspective on who owns the copyrights to the content associated with the NFTs. Attorneys specializing in art law, although they have not yet fully explored it, are developing a familiarity with the current crypto community and future metaverse in order to understand how a public ledger registration tool has created scarcity and value for digital assets, such as digital artwork. Keywords: non-fungible tokens, blockchain, smart contracts, copyright, digital art Field: Law sciences ## 1. INTRODUCTION Blockchain technology has advanced so much over time that it has mainstreamed some new trends in digital shopping and commerce on the Internet, such as cryptocurrencies that have “taken over” the world. However, hold on for a moment, the world is getting very familiar with a new trend - NFT tokens. Although cash and payment cards will continue to be the main payment processes for products and services, the fact is that cryptocurrencies and blockchain technology itself have offered the world a completely new payment value and the possibility of exchanging goods. Namely, we have been familiar with cryptocurrencies for a long time, we talked about their mining and trading, and for most people, they are not new. Cryptocurrencies have been around for a while now and their market value is followed daily, like the stock market. However, there is a new trend that has been appearing in recent years that has to do with digital money. As we mentioned in the introduction of the text, we are talking about NFT tokens or simply speaking non-fungible tokens which, in a way, represent the definition of digital assets. NFTs are a type of cryptocurrency that allows various works of art on various media and sites to be “tokenized” and sold through digital commerce mechanisms, such as Bitski.com. ## 2. BLOCKCHAIN TECHNOLOGY Blockchains are databases that store records on computers all over the world. This makes the blockchain a distributed database with a peer-to-peer architecture. The term ‘distributed’ means that the data is stored in multiple locations, and the term ‘peer-to-peer’ means that there is no central authority that holds the main copy of the data. The thing that makes blockchain so special is that once something is written into the blockchain, it can never be changed or deleted. Therefore, blockchain has become such a popular topic - because it provides a secure way to store information about assets. In the future, blockchain will be used to store data about who owns which house, apartment, car, insurance policy, etc. ‘There are four main characteristics of blockchain technology: 1. Transparency - All participants in the chain can see all records that have previously been entered in “blocks”. 2. Decentralization and data forwarding - the existence of a certain number of computers that coexist, having individually each of them the possibility of equal insight into data entered, and the possibility of introducing new data. [*Corresponding author: mujevic5@gmail.com](mailto:mujevic5%40gmail.com?subject=) © 2023 by the authors. This article is an open access article distributed under the terms and conditions of the [Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).](https://creativecommons.org/licenses/by/4.0/) https://scienceij.com ----- UDK: 347.78:004.031.4(4 672EU) 3. Non-refund - once the data is in the “block” chain it stays there forever. 4. Lack of intermediation - The central body, which would be managed, and closely regulated transactions that occur online, does not exist. Everything on the network that represents a blockchain system takes place and regulates between the participants in the “blocks” that are equal.’ (Strabac, M., 2021). **2.1. Tokens on a blockchain technology** ‘Tokens are assets that are encrypted into blocks on a blockchain. The process of creating a token is commonly known as ‘minting’ the token. A blockchain serves as the foundation onto which a token is encrypted, creating an immutable record of the existence and ownership of digital assets, such as artwork’ (Murray, DM 2022). **2.2. NFTs: Unique digital assets** ‘Non-fungible tokens (NFTs) are cryptographically unique tokens that are linked to digital (and sometimes physical) content, providing proof of ownership’ (Kramer, Graves, Philips, 2022). **2.3. Fungible vs non-fungible** ‘A fungible token can be exchanged one for one with any other token of its kind’ (Murray, DM 2022). A fungible token is a unit of cryptocurrency that can be exchanged for any other unit of cryptocurrency. For example, one bitcoin on the Bitcoin blockchain can be exchanged for any other bitcoin. Because they are fungible, each one of them has the same value. In other words, you cannot exchange one NFT with another NFT, nor can you sell parts of it. For example, one NFT that records the existence and ownership of a 30X40 oil painting will not be presumed to be exchangeable for another NFT that records the existence and ownership of a different 30X40 oil painting. ## 3. NFTS AND SMART CONTRACTS ‘Non-fungible tokens facilitate a registration process that makes tokenized digital artwork a unique asset. Although other similar works may exist and may even have NFT registrations on the blockchain, each NFT creation produces a unique blockchain record for a unique asset’ (Murray, DM 2022). ‘NFTs usually exist on a blockchain, which is, as noted above, a distributed ledger that records transactions. The main difference between NFTs and smart contracts is that NFTs are digital assets powered through smart contracts, meaning that smart contracts control the transferability and ownership of NFTs. In other words, smart contracts are not the same as NFTs but are vital to their use. Furthermore, both run on the blockchain, and so many of the disputes that arise with respect to NFTs go back to the smart contracts that control them’ (Schmitz, JA 2022). ‘This is important because some argue that smart contracts will create efficiencies and may largely eliminate the need for complicated and costly letters of credit, bonds, and security agreements by digitizing automatic enforcement or payment’ (Schmitz, JA 2022). As defined by Nick Szabo, ‘a smart contract is a computerized transaction protocol that executes the terms of a contract’ (Szabo, N. 1994). ‘The key characteristic of these contracts is that they can be represented in the program code and executed by computers: this differs from traditional contracts that are established through negotiations, written documents, and concessional actions. Smart contracts are self-promotional and self-adjusting computer programs based on a program algorithm’ (Cvetkovic, P. (2020). ‘Smart contracts are used to write and store non-fungible tokens (NFTs), which have basic sale terms defined within them. The Ethereum blockchain has differing levels of token fungibility depending on the technical standard applied. The ERC-20 standard is used for fungible tokens, which means that one token is always equal to all other tokens. In essence, an ERC-20 token is the same as Ether (ETH)’ (Aksoy, Üner 2021). **3.1. The rise of NFT** Why would someone pay millions for a JPEG image? ‘There is no limit to what NFTs can represent. They can represent digital images, films, audio, or something entirely intangible (such as an invisible sculpture)’ (Aksoy, Üner 2021). https://scienceij.com ----- UDK: 347.78:004.031.4(4 672EU) The concept of non-fungible tokens (NFTs) has gained widespread attention in the blockchain community due to the recent deployment of the NFT Standard on the Ethereum platform. NFT became popular in 2017 when the game “CryptoKittes” was launched, and it is the first game that used the ERC-21 protocol. In that game, virtual cats are unique and therefore tokenized and sold. Eventually, the market for CryptoKitties had the highest transaction volume on the Ethereum blockchain. ‘Arguably the most astonishing price tag came with a piece of digital art by artist Beeple named ‘Everydays’ : The first 5000 days, an NFT of which sold for $ 69,346,250 through an auction at Christie’s in early March 2021’ (Lantwin, T. 2021). The buyer who received this NFT, got a jpeg file and a unique piece on the Ethereum blockchain and this NFT did not include copyright ownership of its piece. Then, ‘Twitter founder Jack Dorsey’s firstever tweet has been sold for the equivalent of $2.9m’ (Harper, J. 2021). As the buyer, Mr. Dorsey will be digitally signing and verifying a certificate for Mr. Estavi, which will also include the metadata of the original tweet. The data will include information such as the time the tweet was posted and its text content. ‘Also, The New York Times has entered the NFT game, selling the chance to own ‘the first article in the almost 170-year history of The Times to be distributed as an NFT’ — and it’s sold for right around $560,000’ (Clark, M. 2021). **3.2. Are NFTs the future?** As I wrote previously, the blockchain contains a collection of data stored in electronic form that can be accessed quickly and easily by any number of users at the same time, which means that it is the same as the NFT protocol. In addition to the recording function, upon purchasing an NFT, it can be stored in a digital asset wallet and shared virtually. The owner can then display their NFT to showcase their ownership of the asset. It may even be considered a portable form of art (Aksoy, Üner 2021). ## 4. NFTs PLATFORMS ‘Once, when NFTs are minted, they become available for transactions, which are primarily facilitated by intermediary platforms. To exemplify the type of private ordering that shapes NFT transactions and is relevant for copyright purposes, the terms and conditions of particular platforms, as well as EU law instruments, play a significant role. (Bodó, Giannopoulou, Quintais, Mezei 2022). Based on the NFT function and the subject matter interest, the following is being presented: 1. platforms that function as open marketplaces for all minted NFTs; 2. platforms that function as collection-based marketplaces; (i) Platforms that function as open marketplaces ‘Open marketplaces enable the creation and exchange of NFTs by anyone. They function as the eBay of the NFT ecosystem. Dominated by a few major players, including OpenSea, Rarible, and Foundation. The growth of NFT marketplaces can be attributed to several factors. The streamlined mining process is particularly appealing to creators and companies, regardless of their technical experience. NFTs generated outside can be conveniently listed, and these factors combine to increase the variety and quantity of NFT supply. These variables could lead to a vicious cycle and eventually consolidate this sector into a few dominant players. Category (1) platforms impose the least amount of restrictions with respect to third-party minted NFTs and different types of NFTs. This openness enables them to operate on a larger scale. (Bodó, Giannopoulou, Quintais, Mezei 2022). (ii) Platforms that function as collection-based marketplaces ‘With the advent of blockchain technology which enabled the creation of digital collectibles in the form of NFT, many investors have diverted their attention to the NFT collectibles market in the last two years, creating a FOMO in this space. The NFT craze can be seen from its sales volumes in the last two years. The sales of non-fungible tokens (NFTs) were just $81.1 million in the first half of 2020 but surged to $2.5 billion in the first half of 2021. NFT collectibles share the same characteristics as https://scienceij.com ----- UDK: 347.78:004.031.4(4 672EU) traditional collectibles like scarcity, uniqueness and more. However, it has added advantages like the ability to authenticate the originality of an NFT collectible as well as in proving the ownership’ (Liew, VK 2021). ## 5. COPYRIGHT LAW AND NFTs The fundamental idea is that just because you own the NFT doesn’t mean you own the copyright as well. In other words, you can have possession of the object, but you might not have the copyright associated with that object. Copyright is not just a single right, but a collection of rights, and most of these rights are retained by the original creator of the work. ‘An explanation of the idea of ‘Digital Exhaustion’ and its connection to the first-sale doctrine, which maintains that the first authorized sale of a product with intellectual property attached exhausts the rightsholder’s capacity to allege infringement by further sales of that product’ (Bjarnason, C. 2021). So, what right are you getting when you buy an NFT? Oftentimes, people are unaware that they do not obtain a full transfer of copyright when they purchase an NFT. Although this is the case, the NFT still belongs to the buyer. They can then trade it, sell it, or give it away as they choose. Does owning an NFT grant you every right? Usually, no. When purchasing an NFT, it is important to ensure that the original copyright holder has expressly agreed, in writing, to convey the right to you. Without this agreement, you may only be granted certain rights to the NFT. There are, however, cases in which the original copyright holder grants full rights to the buyer of an NFT. This information can be checked and verified by reading the description of the NFT listing. It’s not surprising that there are so many legal implications that come with NFTs.The following chapter of this paper will center its focus on the matter of copyright law as it pertains to NFTs, including but not limited to the Information Society Directive, the Resale Rights Directive, and the Digital Single Market Directive. This is with specific reference to artwork and NFTs where the IP address is owned by the original creator. **5.1 EU Copyright Law Applied to NFTs** The Information Society (InfoSoc) Directive Under InfoSoc Directive, when you own a copyright, you typically own reproduction right, the right of communication to the public of works, and right of making available to the public other subject-matter and distribution right. ‘The first relevant clause of the InfoSoc Directive is Article 2 “Right to Reproduce”, which confers the copyright holder the exclusive right to reproduce and make copies of the artwork’ (Bjarnason, C. 2021). ‘Article 3 of the InfoSoc Directive provides the creator with the “right of communication to the public of works and the right of making available to the public”. This can be described as the ‘right to display’ (Bjarnason, C. 2021). The original creator always has the right to display the artwork, regardless of exhaustion. Furthermore, the owner of the NFT will also have the right to display the underlying artwork. This clause, which is both relevant and key, outlines the various rights given to NFT holders. One of these rights is the right to display the linked artwork. ‘Article 4 of the InfoSoc Directive grants the creator the exclusive right to authorize or prohibit the distribution of their work to the public, whether through a transaction or otherwise. In this case, the doctrine of first sale applies, which results in the creator’s rights being exhausted upon the sale of a specific version of a particular creation. As exhaustion is applicable, if an original creation is sold, the purchaser has the right to resell this creation. However, when it comes to NFTs, the NFT and its accompanying rights are the items that can be sold. The underlying artwork, on the other hand, is not necessarily sold unless this right is embedded in the terms of sale of the NFT. Nevertheless, it is common for the artwork to be transferred along with the NFT, as demonstrated by the Beeple auction and the ‘Disclosure Face’ sale mentioned above’ (Bjarnason, C. 2021). https://scienceij.com ----- UDK: 347.78:004.031.4(4 672EU) **5.1.1. The Resale Right Directive and the Digital Single Market Directive** ‘The Resale Right Directive in EU copyright law requires that any type of seller who subsequently sells an artist’s original artwork must provide a royalty-based commission to the artist. According to Article 2 of this directive, original works of art can include paintings, sculptures, ceramics, and photographs’ (Bjarnason, C. 2021). **5.1.2. Digital Single Market (DSM) Directive** ‘Article 17 of the directive specifies that online content-sharing service providers are responsible for any illicit content, including copyright infringement, on their platform. NFTs and their marketplaces have inherent copyright protection mechanisms, such as the marketplace’s terms of sale and transaction execution through smart contracts, that ensure compliance with the rules established in the DSM directive’ (Bjarnason, C. 2021). **5.2 NFTs and Copyright Ownership** ‘In an ideal world, the copyright owner of an artwork would also be the creator of its NFT.’ However, as one would expect, individuals who engage in infringement find their way around in the digital sphere and create more intellectual property-related problems. Additionally, a new problem has arisen in recent years. These individuals “mint” NFTs based on copied artwork without permission and put them up for sale. As the decentralization, encryption, and anonymity features that are inherent in blockchain ecosystems make it hard to find the copyright holder, this can be a big issue. Coming to the question, how can someone sell work that is not theirs? ‘NFTs and copyright law have two significant zones of interaction. The first is related to the ‘minting’ when NFTs are created, and the second is focused on the dissemination of the digitized work’ (Idelberger, Mezei 2022). ‘The concept of NFTs is such that the original content is not included in them. Rather, they are compiled with standard contracts, resulting in unique metadata that can be written to the blockchain. Essentially, an NFT functions as a digital receipt that links to the original content, much like a deed would for a house. It is worth noting that the NFT itself is not a copy of the content’ (Bodó, Giannopoulou, Quintais, Mezei 2022). ‘When it comes to NFTs, there is no real copyright ownership title over the tokenized work. This means that the original creator retains control over the work, even after it has been sold on an online marketplace. However, the metadata associated with the NFT may grant certain rights to the acquirer of the token. These rights are usually quite limited, and they often restrict the commercial use of the work’ (Bodó, Giannopoulou, Quintais, Mezei 2022). The validity and execution of these online agreements (aided by smart contracts) should ideally be without issue and within the parties’ freedom of contract if these online agreements meet the formal requirements of national copyright contract rules. ‘When it comes to NFTs, the sellers have the power to determine their own terms. These terms can include options like transferring traditional rights, using the NFT to unlock additional content, or implementing a digital royalty for resale. In any case, creators and owners of NFTs have significant control over the destiny of their creations’(Lapatoura, 2021, p. 171). While those who sell NFTs are free to establish their own licensing agreements for the tokenized work, these agreements will have little impact from a copyright standpoint regarding the exhaustion of the distribution right, as is frequently seen in collection-based or curated marketplaces. ‘For instance, Mike Shinoda from the band Linkin Park, who successfully sold the audio clip “Happy Endings” accompanied by his artwork, published the terms of his NFT sales as follows: “Only limited personal non-commercial use and resale rights in the NFT are granted and you have no right to license, commercially exploit, reproduce, distribute, prepare derivative works, publicly perform, or publicly display the NFT or the music or the artwork therein. All copyright and other rights are reserved and not granted.” ## 6. CONCLUSION NFTs give their holders the illusion of ownership; in other words, they are a ‘cryptographically signed receipt that you own a unique version of a work’ (Guadamuz, 2021c). The global market has been significantly impacted by NFTs due to their disruption of the traditional https://scienceij.com ----- UDK: 347.78:004.031.4(4 672EU) model of auctioning art. Cost-effectiveness is increased with NFTs because there is no need to worry about storage or insurance expenses. Non-fungible tokens should not be isolated to digital artwork or even the art industry. They will have a certain degree of influence on physical assets and result in their tokenization. Of course, like any other technological advancement, many questions and uncertainties are raised regarding NFTs. The minting of protectable artistic works and their sale have led to copyright issues that must be addressed. Many of these uncertainties can be clarified through license agreements. While NFTs are a new and exciting technological advancement, there are many questions and uncertainties raised in regard to them, especially in regard to copyright law. It is important to clarify some of these issues in order to ensure that the minting and sale of NFTs is done in a legal and protected manner. ## 7. BIBLIOGRAPHY Andrew (2022) ‘NFTs use ‘smart’ contracts—but what exactly are they?’ Available on: https:// www.theartnewspaper.com Aksoy, Üner (2021) ‘NFTs and copyright and opportunities. Available on: https://www. deepdyve.com/lp/oxford-university press/nfts Bjarnason, C. (2021) ‘NFT Explained: In the Eyes of EU Copyright Law’. Available on: https:// medium.com/@casperbjarnason/ nfts Bodó, Giannopoulou, Quintais, Mezei (2022) ‘The Rise of NFTs: These Aren’t the Droids You’re Looking For’. Available on: The Rise of NFTs: These Aren’t the Droids You’re Looking For by Balázs Bodó, Alexandra Giannopoulou, João Pedro Quintais, Péter Mezei: SSRN Clark, M. (2021) ‘The New York Times just sold an NFT for more than half a million dollars.’ Available on:https://www.theverge. com/ Cvetkovic, P. (2020) “ Legal aspects applications blockchain: An example of smart contracts”. Available on: Legal The word legal aspects of Blockchain.pdf Guadamuz, A. (2021c). The treachery of images: Non-fungible tokens and copyright. Journal of Intellectual Property Law & Practice, 16(12), 1367–1385. https://doi.org/10.1093/jiplp/jpab Harper, J. (2021) ‘Jack Dorsey’s first ever tweet sells for $2.9m’, BBC News, 23 March. Available on: https :// www.bbc.com/ news/business Idelberger, Mezei (2022) ‘Non-fungible tokens’. Available on: Non-fungible tokens by Florian Idelberger, Péter Mezei: SSRN Lantwin, T. (2021) ‘Beyond the hype: NFTs, digital and copyright’. Available on: https://www. dusip.de /en/2021/06/16/beyond the-hype-nfts-digital-art-and-copyright/ Lapatoura, I. (2021). Creative digital assets as NFTs: A new means for giving artists their power back? Entertainment Law Review, 32(6), 169–172. Liew, VK (2021) ‘ DeFi, NFT and GameFi Made Easy: A Beginner’s Guide to Understanding and Investing in DeFi, NFT and GameFi Project Kramer, Graves, Philips, (2022) ‘Beginner’s Guide to NFTs: What are Non-Fungible Tokens?’. Available on: https://decrypt.co/ resources/non-fungible-tokens-nfts-explained-guide-learn blockchain, Lapatoura (2021) ‘Copyright & NFTs of Digital Artworks’. Available on: https://ipkitten.blogspot .com /2021/03/guest-post copyright-nfts-of-digital.html Murray, DM (2022) ‘NFT Ownership and Copyrights’. Available on: NFT Ownership and Copyrights by Michael D. Murray: SSRN Purtill (2021) ‘Artists report discovering their work is being stolen and sold as NFTs’. Available on: https://www.abc.net.au/news/ science/ Schmitz, JA (2022) ‘Resolving NFT and Smart Contract Disputes’. Available on: Resolving NFT and Smart Contract Disputes by Amy J. Schmitz: SSRN Strabac, M. (2021) “Protection of personal data in the blockchain”. Available on: https://www.milic.rs/ zastita-podataka/zastita podataka-o-licnosti-u-blockchain-u/ Szabo, N. (1994) ‘Smart Contract’. Available on: https://www.fon.hum.uva.nl/rob/ Courses/ Information InSpeech/CDROM/ Literature/LOTwinterschool2006/szabo.best.vwh.net/idea.html Sephton (2021) ‘Copyright infringement and NFTs: How artists can protect themselves’ Available on: https:// cointelegraph. com/news/copyright-infringement-and-nfts-how-artists-can-protect-themselves Vrbanus (2021) ‘Emily Ratajkowski prodaje vlastitu sliku kao NFT da bi “preuzela kontrolu” nad vlasništvom’. Available on: https://www.bug.hr/blockchain/emily-ratajkowski-prodaje-vlastitu-sliku-kao-nft-da-bi-preuzela-kontrolu-nad-21067 https://scienceij.com -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.35120/sciencej020215m?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.35120/sciencej020215m, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "HYBRID", "url": "https://scienceij.com/index.php/sij/article/download/20/17" }
2,023
[ "JournalArticle" ]
true
2023-05-31T00:00:00
[]
6,592
en
[ { "category": "Economics", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0148207de2efe9b34947ed00528c7d09f7af7249
[]
0.91426
Is Bitcoin the 'Digit Gold'A Potential Safe-haven Asset?
0148207de2efe9b34947ed00528c7d09f7af7249
Advances in Economics, Management and Political Sciences
[ { "authorId": "2214470875", "name": "Shupeng Guan" }, { "authorId": "2214533470", "name": "Han Jiang" }, { "authorId": "2214509451", "name": "Muyang Zhou" }, { "authorId": "49721770", "name": "Jianuo Liu" } ]
{ "alternate_issns": null, "alternate_names": [ "Adv Econ Manag Political Sci" ], "alternate_urls": null, "id": "efee6e0c-f1db-4556-8c10-5a827a83e1b3", "issn": "2754-1169", "name": "Advances in Economics, Management and Political Sciences", "type": "journal", "url": null }
In this paper, we examine whether bitcoin has the potential to become safe-haven asset that can rival gold in the future. We observed, compared and analyzed and the performance of bitcoin and gold in face of a falling market and inflation pressure. We can see if investors can rely on bitcoin to reduce risk exposure significantly through empirical tests. At the end of our research, we found that bitcoin did not perform as well as gold did when faced with market crash and inflation. Therefore, we conclude that bitcoin does not yet show the potential to possess risk-proof merits as gold, the traditional high-quality hedge asset. Gold would probably remain the preferred hedge asset against cryptocurrency for now.
The 6th International Conference on Economic Management and Green Development (ICEMGD 2022) DOI: 10.54254/2754-1169/4/2022910 # ***Is Bitcoin the 'Digit Gold'——A Potential Safe-haven Asset? *** ## **Shupeng Guan [1,] *, Han Jiang [2], Muyang Zhou [3], and Jianuo Liu [4]** *1 School of Mathamatics, University of Birmingham,B15 2TT,United Kingdom * *2 Wuhan Britain-China School,430030, China * *3 Information School, University of Washington, 98195, United States * *4 Material Science&Engineering, Nanyang Technology University,639798, Singapore* * wayneg0530@163.com * **corresponding author * Abstract: In this paper, we examine whether bitcoin has the potential to become safe-haven asset that can rival gold in the future. We observed, compared and analyzed and the performance of bitcoin and gold in face of a falling market and inflation pressure. We can see if investors can rely on bitcoin to reduce risk exposure significantly through empirical tests. At the end of our research, we found that bitcoin did not perform as well as gold did when faced with market crash and inflation. Therefore, we conclude that bitcoin does not yet show the potential to possess risk-proof merits as gold, the traditional high-quality hedge asset. Gold would probably remain the preferred hedge asset against cryptocurrency for now. Keywords: Bitcoin, Sharpe ratio, portfolio, gold. **1. Introduction ** In the view of a segment of investors, holding physical assets may all be at risk in the future, even gold, which is traditionally considered the most safe-haven asset. For this group of investors, some of the attributes of virtual currencies such as bitcoin are deeply favored. For example, Bitcoins are not subject to national monetary policies, meaning they are not influenced or controlled by governments. Also, as virtual property, bitcoins are not at risk of being destroyed or lost in a war or natural disaster. So, is it possible that virtual currencies could become a more desirable asset for investors in a potentially volatile time in the future? In the last two years, our world has experienced an event unlike anything before in this century, an worldwide epidemic that continues today. In fact, it has changed our world in many ways during last two years, including our ways of thinking, and has allowed us to observe and interpret the performance of different assets in the financial markets differently. In response to our questions, we gathered several articles which were published at different times, and their proposals differ from each other. Wong's paper argues that virtual currencies bring higher portfolio risk because of their high intrinsic volatility but bring higher Sharpe ratios to gold and stock portfolios at the same time [1]. While Corbet's research suggested that virtual currencies could provide investors with many benefits and safety during an epidemic [2]. Conlon argued that virtual assets like bitcoin could not protect investors' assets during critical times[3]. And Hasan's paper suggested that safe-haven assets may vary over time [4]. So, how has the situation changed since then? © 2023 The Authors. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0 (https://creativecommons.org/licenses/by/4.0/). 392 ----- The 6th International Conference on Economic Management and Green Development (ICEMGD 2022) DOI: 10.54254/2754-1169/4/2022910 ## The first quarter of 2022 has been a very turbulent period for international financial markets, immediately following a wave of worldwide outbreaks of the new omicron epidemic at the beginning of the year, a military conflict broke out in Eastern Europe, which has continued to this day (late May). Markets worldwide suffered a harsh test under double pressure of epidemic and geopolitical events. The general result is a series of market shakes and inflation spikes. When the world enters a period of uncertainty, the prices of commodities such as food and oil rise, stock markets fall, and gold prices rise. But what is interesting is that most virtual currencies have seen their prices up and down to a greater or lesser extent during this period. In order to investigate whether bitcoin can be considered as ‘digit gold’ through empirical evidence, we gathered weekly data sets covering a five-year period on bitcoin price, gold price, S&P 500 index, US 10-year treasury yields, and US 1-month treasury yields. We aim to combine them into portfolios and observe, compare the performance of gold and bitcoin in different market situations. **2. Data ** To conduct the whole research we used several data sets, including weekly bitcoin prices in US dollars, weekly gold price in US dollars as two manipulated variables. Values of S&P 500 were attained as the representatives of the stock market, while the yield of US 10-year treasury act as a typical sample of the general bond market. In addition, the yield of US 1-month treasury was set as the risk-free rate. Bitcoin was chosen as the representative of the cryptocurrency market because the market capacity makes up nearly 40% of the total cryptocurrency market and hence should give a powerful indication of the overall characteristics of the whole cryptocurrency market. As gold is widely acknowledged as an ideal hedge or safe haven asset against economic recession or inflation, it is anchored to be the comparison object. Data for bitcoin are collected from Blockchain.com, a leading platform for cryptocurrency exchange.Data for gold price are collected from LBMA (London Bullion Market Association). Data for S&P 500 are collected from WSJ (The Wall Street Journal). Data for the US 10-year treasury(US10Y) and the US 1-month(US1M) treasury are sourced from FRED. |Col1|Table 1: Summary statistics of research data.| |---|---| |Variable|Obs Mean Std. Dev. Min Max Sharpe| |BTC price Gold S&P500 US10Y US1M|256 18,306 17820.2 1,448 66,954 256 1,542 261.52 1,182 2,048 256 3,262 711.64 2,357 4,793 256 1.93% 0.75% 0.55% 3.22% 256 1.00% 0.90% 0.00% 2.45%| |Ln(BTC) Ln(Gold) Ln(S&P500)|256 9.37 0.94 7.28 11.11 256 7.33 0.17 7.07 7.62 256 8.07 0.21 7.77 8.47| |Re(BTC) Re(Gold) Re(S&P500) Re(US10Y) Re(US1M)|255 1.28% 11.01% -38.65% 32.83% 0.83 255 0.16% 2.00% -9.88% 6.91% 0.11 255 0.22% 2.48% -13.38% 10.72% 0.14 255 0.02% 0.86% -3.43% 2.92% 0.01 255 0.02% 0.02% 0.00% 0.05% 0.01| The data are all weekly observed back from 03/05/2017 to 27/04/2022, yielding 256 observations after processing. Table 1 provides a summary report of the whole dataset, including price of cryptocurrency(bitcoin),gold,and stocks(S&P500).The top part shows the statistics about closing 393 ----- The 6th International Conference on Economic Management and Green Development (ICEMGD 2022) DOI: 10.54254/2754-1169/4/2022910 ## price in US dollars of bitcoin,gold and stocks(S&P 500), and yields of US 10-year treasury as well as 1-month treasury.The middle part shows the log price of bitcoin,gold and stocks(S&P500).The bottom part shows the log return of all five distinct assets. All the data transformations are intended for further calculation of optimal portfolio maximising the Sharpe ratio, which is shown in Table 2. Table 2: Optimal Sharpe portfolios, weight add to 100%. |Col1|Bitcoin Gold S&P500 US 10Y| |---|---| |S&P500/10Y S&P500/10Y/Gold S&P500/10Y/Gold/BTC|0% 0% 44% 56% 0% 82% 71% -53% 24% 81% 46% -51%| **3. Methodology ** In this section methodologies used to conduct our research will be described. To analyze whether bitcoin, the so-called ‘digit gold’, has the potential to be as good as gold when considered as a safe- haven investment, especially in when stock market is down and inflation worsen. We tried to separate the data into four sets--weeks both S&P500 and 10Y went up, weeks when both went down, weeks when S&P500 went up and 10Y went down, weeks when S&P500 went down and 10Y went up, that is, we considered all four possible market situations. Next we calculate the optimal Sharpe ratio portfolios for each of the four sets individually, initially the portfolio only consists of stocks(S&P500) and bonds(US 10-year treasury), then adding gold, then adding bitcoin, so that it would be convenient to compare the amount of bitcoin and gold we would like to hold if we knew what the stock and bond market would do. Especially if stock and bond market dropped, compare between gold and bitcoin, specifically, compare how much do gold and bitcoin improve the Sharpe ratios of the portfolio, to verify which is the ideal safe-haven asset. The specific data analysis is carried out in following steps.First step is the demeaning processing, where the mean equation is: 𝑅𝑑= (𝑅−𝑅) −(𝑟−𝑟) (1) where Rd is the demeaned excess returns, R is the return of the asset and 𝑅 is the mean return, r is the risk-free ratio while 𝑟 is the mean risk-free ratio. The second step is to calculate the optimal portfolio based on maximizing Sharpe ratio. Sharpe ratio is originally developed by Nobel laureate William F Sharpe [5], an indicator of calculating risk adjusted return. It is defined as ( 𝑅(𝑥)−𝑟 ) ## (2) 𝜎 ## where R(x) is the expected return of portfolio, r is the risk-free ratio, 𝜎 is the standard deviation of R(x). Normally, a higher Sharpe ratio indicates better investment performance, given the risk. If a Sharpe ratio is negative, it means the risk-free return is greater than expected return of portfolio, and Sharpe ratio conveys nothing meaningful. The mathematical model for the Sharpe Ratio based Portfolio optimization is given by 394 ----- The 6th International Conference on Economic Management and Green Development (ICEMGD 2022) DOI: 10.54254/2754-1169/4/2022910 𝑁 ∑ 𝑖 = 1 𝑊 𝑖 ’· 𝜇 𝑖 −𝑟 ## max (3) 𝑖 𝑗 𝑊 𝑖 - 𝑊 𝑗 - 𝜎 𝑖𝑗 ## ( √ [∑∑] ) subject to ∑ 𝑁𝑖=1 𝑊 𝑖 = 1 (4) where 0 ≤𝑊 𝑖 ≤1 (5) 𝑊 𝑖 is the weight of each asset.The numerator of the objective function denotes the excess returns of the investment over that of a risk-free asset 𝑟 and the denominator the risk of the investment. The objective is to maximize the Sharpe Ratio. The basic constraints indicate that this is a fully invested portfolio, in other words, weights adds to 100%. In the third step, since the optimal Sharpe portfolios of different combinations were attained, we computed the growth of $1 in all three portfolios, scaling each one to the same volatility(standard deviation mathematically), and visualize their growth to compare apples to apples. We will mainly focus on the situations where one of stock and bonds markets falls, so that we can compare bitcoin and gold to verify whether bitcoin also has the property to hedge against falling stock market or inflation as physical gold does. **4. Results ** We calculated the weight matrices and optimal Sharpe portfolio statistics of four sets. The numbers are shown in following tables—Table 3(weeks both S&P500 and 10Y went up),Table 4(weeks when both went down),Table 5(weeks when S&P500 went up and 10Y went down), Table 6(weeks when S&P500 went down and 10Y went up). And growth of portfolios(value of $1) are visualized. Table 3: Both S&P500 and 10Y went down. |Col1|Bitcoin Gold S&P500 US Mean Std. Sharpe 10Y Dev.|Col3| |---|---|---| |S&P500/10Y S&P500/10Y/Gold S&P500/10Y/Gold/BTC Standard deviation|0% 0% -19% -81% 0% 13% -21% -93% 2% 15% -19% -98% 12.63% 2.47% 2.54% 0.71%|1.05% 0.92% 8.04 1.13% 0.89% 8.97 1.10% 0.83% 9.49| ## (*Mean and standard deviation are weekly calculated while Sharpe ratio is annualized, same below.) Table 3 considers the situation when both S&P500 and 10Y went down. As tables above shown, the portfolio improved when gold is introduced, Sharpe ratio increased by 0.93, as gold is a well- known classic hedge against stock market and inflation. Bitcoin also adds subtle improvement to the portfolio, thought it doesn’t account for much in dollar size to the whole portfolio compared to gold, it makes a difference to the portfolio owing to its relatively high volatility, and the Sharpe ratio is raised by 0.52. Looking at the Figure 1, the growth trend of three different combinations basically follows the same mode, while combinations with more assets perform better in value over time. 395 ----- The 6th International Conference on Economic Management and Green Development (ICEMGD 2022) DOI: 10.54254/2754-1169/4/2022910 ## Figure 1: Growth of value $1 when both S&P500 and 10Y went down. Table 4: Both S&P500 and 10Y went up. |Col1|Bitcoin Gold S&P500 US Mean Std. Sharpe 10Y Dev.|Col3| |---|---|---| |S&P500/10Y S&P500/10Y/Gold S&P500/10Y/Gold/BTC Standard deviation|0% 0% 36% 64% 0% 0% 36% 63% 0% 0% 33% 67% 10.92% 1.86% 1.12% 0.52%|0.81% 0.50% 11.29 0.83% 0.52% 11.32 0.83% 0.52% 11.33| Table 4 considers the situation where both S&P500 and 10Y went up, bull market without inflation. It seems that when you are in a bull market and have no pressure from inflation, bitcoin and gold seldom be considered in your portfolio, as the portfolios basically remain the same structure, simply because you know you can make a lot of money out of stocks and bonds investment as they are going to be unstoppable, in reality, that is a big if. 396 ----- The 6th International Conference on Economic Management and Green Development (ICEMGD 2022) DOI: 10.54254/2754-1169/4/2022910 ## Figure 2: Growth of value $1 when both S&P500 and 10Y went up. Table 5: S&P500 went up and 10Y went down. |Col1|Bitcoin Gold S&P500 US Mean Std. Sharpe 10Y Dev.|Col3| |---|---|---| |S&P500/10Y S&P500/10Y/Gold S&P500/10Y/Gold/BTC Standard deviation|0% 0% 28% - 126% 0% -6% 34% - 127% 1% -6% 32% - 128% 9.87% 1.69% 1.48% 0.48%|1.21% 0.82% 10.48 1.32% 0.89% 10.57 1.30% 0.87% 10.59| Table 5 considers the situation when S&P500 went up and 10Y went down. Considering the second portfolio combination, gold shows subtle negative correlation with stock markets, which could be the result of the offset between a rising stock market and deflation. From the statistics, Sharpe ratio hardly changes, we can conclude bitcoin didn’t make much difference to the portfolio. Looking at the Figure 3, the claim is further supported. 397 ----- The 6th International Conference on Economic Management and Green Development (ICEMGD 2022) DOI: 10.54254/2754-1169/4/2022910 ## Figure 3: Growth of value $1when S&P500 went up and 10Y went down. Table 5: S&P 500 went down and 10Y went up. |Col1|Bitcoin Gold S&P500 US Mean Std. Sharpe 10Y Dev.|Col3| |---|---|---| |S&P500/10Y S&P500/10Y/Gold S&P500/10Y/Gold/BTC Standard deviation|0% 0% -14% 114% 0% 6% -19% 114% 0% 6% -18% 112% 11.38% 2.00% 2.46% 0.67%|1.17% 0.95% 8.68 1.30% 1.05% 8.75 1.27% 1.03% 8.75| Table 6 considers when S&P500 went down and 10Y went up. As is well known, it is acknowledged when stocks did not perform well, gold should be introduced into the portfolio to make life easier. However, bitcoin does not function as gold helps, as the Sharpe ratio remains the same, indicating it is of trivial role to this situation. 398 ----- The 6th International Conference on Economic Management and Green Development (ICEMGD 2022) DOI: 10.54254/2754-1169/4/2022910 ## Figure 4: Growth of value $1 S&P500 went down and 10Y went up. |Col1|Table 6: The whole period.|Col3| |---|---|---| ||Bitcoin Gold S&P500 US Mean Std. Sharpe 10Y Dev.|| |S&P500/10Y S&P500/10Y/Gold S&P500/10Y/Gold/BTC Standard deviation|0% 0% 44% 56% 0% 82% 71% -53% 24% 81% 46% - 130% 11.01% 2.00% 2.48% 0.86%|0.11% 1.13% 0.56 0.28% 2.57% 0.73 0.53% 3.73% 0.98| Table 7 gives the optimal portfolio statistics considering the whole dataset covering 256 weeks. As the numbers display, the Sharpe ratio rises from 0.56 to 0.73 when gold added, and increases to 0.98 after bitcoin added. Looking at the Figure 5, based on the blue line(S&P500/10Y), there are 2 obvious downturns across the period, the first happened around 03/05/2022, all three combinations fail to protect, the second happened near the end, compare the trend of green line(S&P500/10Y/Gold) and blue line(S&P500/10Y), one can notice that portfolio with gold reversed the falling trend of blue line and remains the general growing trend. However, bitcoin does not seem to protect in these downturns(compare red line and blue line), rather it helps by posting some spectacular returns during general good times. 399 ----- The 6th International Conference on Economic Management and Green Development (ICEMGD 2022) DOI: 10.54254/2754-1169/4/2022910 ## Figure 5: Growth of value $1. **5. Conclusion ** This paper investigates the potential of bitcoin as ‘digit gold’, a safe-haven asset. We try to verify the usefulness of bitcoin in terms of hedging compared to physical gold. From the Results, we can conclude it is not that persuasive to view bitcoin as 'digit gold'. Compared to physical gold, bitcoin is not a better holding in down markets. Specifically, in all four possible situations, the contribution bitcoin made are not more significant than gold did. Further, from the charts we can see portfolio containing bitcoin does not generally outperform its counterparts significantly without bitcoin in volatility and return growth during hard times, and it hardly shows the potential to protect portfolio against bad times. However, from Figure 5, it implies that cryptocurrencies can be useful as a supplement asset to raise returns in a portfolio during general good times. Considering its hedging role is so limited that much less than gold, and basically it made no difference to the growth trend, we can reasonably conclude bitcoin is uncorrelated with stock and bond markets. In summary, bitcoin's hedging role is deficient compared to gold in bear markets, while in bull markets, as the theory of diversification of portfolio indicates, it has more of a volatility-reducing effect than it does a significant increase in returns. Bitcoin can be used as a hedge against stocks in portfolio simply because it is uncorrelated to the stock market, but it is not plausible that bitcoin is as a good safe-haven asset as physical gold. Bitcoin is characterised by high volatility and high returns compared to gold though, risk-seeking investors can increase their risk reward by investing in cryptocurrencies. Our research has limitations due to the narrow selection of data. Later we hope to examine a wider range of results and investment opportunities by combining more asset classes in portfolio.To date, there are still many uncertainty around cryptocurrencies. Regulators may further suppress cryptocurrencies, leading to the often predicted bursting of the cryptocurrency bubble, on the other hand, many investors view bitcoin as a speculative asset, which helps its widespread acceptance. Bitcoin may be still in its infancy but derivatives like crypto options are growing. It is unclear if bitcoin will be the cryptocurrency of choice in the future. 400 ----- The 6th International Conference on Economic Management and Green Development (ICEMGD 2022) DOI: 10.54254/2754-1169/4/2022910 ## **References ** *[1]* *Wong, W. S., Saerbeck, D., & Delgado Silva, D. (2018, February 18). Cryptocurrency: A new investment* *opportunity? an investigation of the hedging capability of cryptocurrencies and their influence on stock, Bond and* *gold portfolios.* *[2]* *Corbet, S., Hou, Y. (G., Hu, Y., Larkin, C., & Oxley, L. (2020, July 7). Any port in a storm: Cryptocurrency safe-* *havens during the COVID-19 pandemic. Economics Letters.* *[3]* *Conlon, T., & McGee, R. (2020, May 24). Safe haven or risky hazard? bitcoin during the COVID-19 bear market.* *Finance Research Letters.* *[4]* *Hasan, M. B., Hassan, M. K., Rashid, M. M., & Alhenawi, Y. (2021, August 13). Are Safe Haven assets really safe* *during the 2008 global financial crisis and covid-19 pandemic? Global Finance Journal.* *[5]* *Sharpe, William F. Mutual Fund Performance, Journal of Business, January 1966, pp. 119–138.* *[6]* *Henriques, Irene, and Perry Sadorsky. 2018. "Can Bitcoin Replace Gold in an Investment Portfolio?" Journal of* *Risk and Financial Management 11, no. 3: 48.* *[7]* *Bessler, W., Taushanov, G., & Wolff, D. (2021, May 29). Factor investing and asset allocation strategies: A* *comparison of factor versus sector optimization - Journal of Asset Management.* *[8]* *Yoshino, N., Taghizadeh-Hesary, F., & Otsuka, M. (2020, July 12). Covid-19 and optimal portfolio selection for* *investment in Sustainable Development Goals. Finance Research Letters.* 401 -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.54254/2754-1169/4/2022910?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.54254/2754-1169/4/2022910, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "HYBRID", "url": "https://www.ewadirect.com/proceedings/aemps/article/view/744/pdf" }
2,023
[ "JournalArticle" ]
true
2023-03-21T00:00:00
[]
5,563
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/014981e1454105af6a6275bef4baf47afbeb7377
[ "Computer Science" ]
0.888705
AODV-Miner: Consensus-Based Routing Using Node Reputation
014981e1454105af6a6275bef4baf47afbeb7377
2022 18th International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob)
[ { "authorId": "2132854826", "name": "Edward Staddon" }, { "authorId": "1803713", "name": "V. Loscrí" }, { "authorId": "1801854", "name": "N. Mitton" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
With the increase of Internet of Things (IoT) applications, securing their communications is an important task. In multi-hop wireless networks, nodes must unconditionally trust their neighbours when performing routing activities. However, this is often their downfall as malicious nodes can infiltrate the network and cause disruptions during routing. We grant nodes the ability to evaluate the behaviour of their neighbours and, through consensus inspired from blockchain's miners, agree on the credibility of each node. The resulting metric is expressed as a node's reputation allowing, in the case of a malicious node, to isolate it from network operations. By illustrating this in an AODV-like multi-hop routing protocol, we can influence route selection no longer based solely upon the shortest number of hops, but also the highest overall reputation. Simulation results revealed that our approach can decrease packet drop rates by ≈ 48% in a static context when subjected to multiple black hole attacks compared to the original routing protocol.
## AODV-Miner: Consensus-Based Routing Using Node Reputation ### Edward Staddon, Valeria Loscrì, Nathalie Mitton To cite this version: #### Edward Staddon, Valeria Loscrì, Nathalie Mitton. AODV-Miner: Consensus-Based Routing Using Node Reputation. WiMob 2022 - The 18th International Conference on Wireless and Mobile Com- puting, Networking and Communications, Oct 2022, Thessaloniki, Greece. ￿hal-03787034￿ ### HAL Id: hal-03787034 https://inria.hal.science/hal-03787034 #### Submitted on 23 Sep 2022 #### HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. #### L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. ----- # AODV-Miner: Consensus-Based Routing Using Node Reputation #### Edward Staddon, Valeria Loscri and Nathalie Mitton Inria, France - firstname.lastname @inria.fr _{_ _}_ **_Abstract—With the increase of Internet of Things (IoT)_** **applications, securing their communications is an important task.** **In multi-hop wireless networks, nodes must unconditionally trust** **their neighbours when performing routing activities. However,** **this is often their downfall as malicious nodes can infiltrate the** **network and cause disruptions during routing. We grant nodes** **the ability to evaluate the behaviour of their neighbours and,** **through consensus inspired from blockchain’s miners, agree on** **the credibility of each node. The resulting metric is expressed** **as a node’s reputation allowing, in the case of a malicious node,** **to isolate it from network operations. By illustrating this in an** **AODV-like multi-hop routing protocol, we can influence route** **selection no longer based solely upon the shortest number of** **hops, but also the highest overall reputation. Simulation results** **revealed that our approach can decrease packet drop rates by** _≈_ 48% in a static context when subjected to multiple black hole **attacks compared to the original routing protocol.** **_Index Terms—IoT, Reputation, Cyber Security, AODV, Blockchain_** I. INTRODUCTION The Internet of Things (IoT) is becoming more prominent in every day life, processing increasing amounts of sensitive data. Indeed, some applications come with extreme risks which could result in severe consequences, from data breach to loss of life. However, in some cases, devices are deployed in sparse hostile environments, forcing them to employ other network paradigms to communicate. This is the case of multi-hop networks, where data is forwarded by intermediate devices to reach its destination. Therefore, securing this exchange is paramount as entrusting data to unknown nodes is a huge risk. Many solutions combating threats in IoT networks have been proposed in the literature. Such approaches as trust based [1], grant the capability to identify different nodes, based on their previous actions in the network. This allows routing protocols to adapt to the situation, granting routing capabilities only to the most trustworthy [2], be it based on cooperation [3] or through the use of signatures [4] to identify potential threats. Furthermore, by analysing node actions and assigning them a trust value, it is possible to influence how routing is performed [5]. By using the notion of reputation, inspired from the human psyche, we can evaluate the actions of surrounding nodes, allowing protocols to avoid attacks [6]. However, many approaches have been designed as part of existing protocols and, therefore, only work in conjunction with them. Another area is that of blockchain, which has long been a point of interest in security and when coupled with trust-based methods, can be highly fruitful [7]. Indeed, the blockchain, a decentralised immutable ledger [8] well known for its uses in cryptocurrencies such as Bitcoin [9], can be levied to distribute reputation values in a secure way to avoid tampering. All data is stored in structures called ”blocks”, each containing a reference to its predecessor in the form of a block hash. As a result, any modifications would ripple through the chain and be detected, rendering it immutable, its main advantage [10]. New blocks must also go through an extensive validation process prior to their insertion. This is performed by Miners, which calculate a Proof of Work (PoW), confirming a blocks credibility which is subsequently verified by other miners before it can be inserted. Although computationally intensive, this consensus-based mining process is the backbone which keeps the blockchain functioning and secure. Due to its immutability, it has been used in other areas such as IoT Security, serving as a decentralised and secure medium for IoT applications [11]. Furthermore, in recent years it has also become an influence in securing routing techniques [12] [13], and has even been used in aviation to secure Unmaned Aircraft Systems against potential threats [14]. However, to access the blockchain it must be stored, which can become a heavy process the more blocks are added. This is further emphasised when the devices using the blockchain possess limited resources, such as storage, computational or even energy capacity. In this paper, we propose a consensus-based reputation module, providing behavioural analysis to network activities, inspired by [15]. We employ a lightweight version of blockchain, reducing its functionalities to a dissemination tool only, repurposing its Miners with the extra responsibility of behavioural validation. Furthermore, we redefine the PoW method with our own consensus-based confirmation scheme, corresponding to the specifications and constraints of our network and validation models. As a consequence, these new _validation miners, significantly different from their blockchain_ origins, hold a key position in the network. By designing our module with adaptability in mind, it can be used by different routing protocols to influence the route selection process. We illustrate this with the Ad hoc On-Demand Distance Vector (AODV) reactive routing protocol [16], in a new implementation called AODV-Miner. By using such a well known protocol as AODV, we can illustrate the functionalities of our approach, and how it interacts with the chosen protocol. The rest of this paper is organised as follows: Section II defines our system model before presenting AODV-Miner in Section III. Section IV presents our results before finally discussing future endeavours and concluding this work in Section V. ----- II. SYSTEM MODEL _A. Network Model_ We consider a connected wireless network scenario with _N static nodes possessing omnidirectional antenna’s with a_ fixed transmission range. Each node is aware of all traffic on the wireless medium in proximity to them at all times. They also possess the ability to determine their own role for the lifetime of a route during discovery, making them either a router or a validation miner, with priority given to routing. As a result, receiving a route request identifies the node as a router, whereas overhearing the request, identifies them as a miner. Subsequently, nodes can participate in multiple routes and can, as a consequence, take on multiple roles. _B. Validation Model_ The role of validation miners is 1) to ”mine a route”, validating routing behaviour between neighbours; and 2) to ”mine a block”, confirming and distributing the results using the blockchain. For their first objective, each miner has the ability to validate the behaviour of its neighbours. By overhearing passing route requests, they can construct both forwards (src _dst) and reverse (dst_ _src) Route_ _→_ _→_ _Validation Tables (RVT) containing the expected hops in order._ Each ”good” or ”bad” action is categorised by the miner for each neighbouring routing node of a specific route. Since the miners parse and extract the expected next hop from their RVT to verify the activities, we can determine that the computation and spatial complexities are linked, resulting in O(n), with n nodes in the table. Once the route has expired, the miners begin their second objective and take on blockchain style responsibilities. Firstly, they aggregate their results into a temporary block which is then shared with neighbouring miners for confirmation. Once complete, the resulting confirmed blocks are again shared with neighbours, updating their network status. As stated previously, we use a lightweight blockchain approach to share blocks, which uses a custom PoW method, where miners simply analyse the block’s contents and check if the actions are inline with their own vision, responding if an error has been detected. This reduces the number of exchanges needed and makes the miners assume their work is valid if no response is received, disseminating thereafter. In this case, two data structures are explored, increasing the structural complexity to O(m _n), with m entries in the received block. However,_ _×_ with a worse case scenario of m = n, we can deduce the computational complexity to be O(n[2]). _C. Threat Model_ **Routing threat. A malicious node can either simply destroy** a packet, or send it elsewhere [17]. In the first case, be it either a complete destruction (black hole) or selective (grey _hole), the concerned data no longer traverses the network. In_ the second case, the malicious node can either transmit the data to another node using another medium, called Wormhole or redirect the packet by simply modifying its destination. In case, the malicious node deviates from the expected behaviour and their action’s are flagged as bad. **Packet threat. By modifying a packet, a malicious node can** change its contents. To resolve this, each miner keeps a CRC16 hash of passing packets during validation, thus detecting any modification mid route. Furthermore, if a node re-transmits a packet which has already been seen, known as replay, the miners can detect an unexpected hop for the corresponding hash and label the nodes behaviour as bad. III. OUR CONTRIBUTION: AODV-MINER This section introduces our consensus-based reputation module, illustrated with AODV, named AODV-Miner _A. Node Reputation_ A node’s reputation is calculated based upon their previous actions. If a node acts as expected, by routing a valid packet towards the correct next hop, it is considered to have performed a ”good” action. Any other action taken is considered to be malicious and flagged as ”bad”. By keeping a record of all actions taken by a single node, it is possible to determine their reliability. We define Sgoodn and Sbadn as the sum of good and bad actions respectively for node n with Wn the action window size, i.e. the number of previous actions taken into account. By varying this value, we can change its precision, limiting it to the most recent actions, or opening it up to a larger portion of node’s history. where β = 8 is the sensitivity factor influencing the sigmoid function as in [15] and α the weight of malicious actions. By adjusting the value of α, we modify the severity of bad actions upon the reputation as is shown in Fig.1a. We can see that the value of α influences the reputation, illustrating that the higher the value, the more unforgiving the network becomes. _1) Link Cost: To identify the best route, AODV uses a_ hop counter which is incremented on each hop. In a similar fashion to [15], we replace the hop count with a different metric called link cost, corresponding to the network ”cost” of using a specific node based upon their reputation. This method allows us to differentiate between good and bad nodes, where a low reputation will ensue a higher cost. When an RREQ or RREP packet is received, the node calculates the cost of the link between itself and the transmitter. By using the same base functionality of selecting the lowest hop count, we encourage the network to select the lowest link cost, thus containing as fewer malicious nodes as possible, increasing route integrity at the potential cost of longer routes. We define Cn as the cost of the link between n and its neighbours: _Cn = ⌊(1 −_ _Rnt_ ) × (Cmax − (Cmin − 1)) + Cmin⌋ (5) _Wn_ � _bad actionsni_ (2) _i=1_ _Sgoodn =_ _Wn_ � _good actionsni_ (1) _Sbadn =_ _i=1_ For a specific node, the reputation, Rn ∈ [0, 1], is defined as a sigmoid function where the exponent δn ∈ [−1, 1] corresponds to the weighted value of the relation between _Sgoodn and Sbadn_ : 1 _Rn =_ (3) _δn = β ×_ _[S][good][n][ −]_ _[α][ ×][ S][bad][n]_ (4) 1 + e[−][δ][n] _Sgoodn + α × Sbadn_ ----- 1 0.75 5 4 1 0.75 0.5 0.25 3 2 0.5 0.25 |Col1|α = 0.5 α = 1 α = 2 α = 5 α = 10| |---|---| ||| ||| ||| |Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13| |---|---|---|---|---|---|---|---|---|---|---|---|---| |||||||||||||| |||||||||||||| |||||||||||||| |||||||||||||| |Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12| |---|---|---|---|---|---|---|---|---|---|---|---| ||||||||||||| ||||Exponential Decay Linear Decay Static Decay = 0.1 Static Decay = 0.2 Static Decay = 0.5||||||||| ||||||||||||| 0 5 10 15 20 25 30 35 40 45 50 55 60 Time Index (minutes) 0 0% 25% 50% 75% 100% Malicious Activity 1 _α = 2_ _α = 5_ _α = 10_ 0 0% 25% 50% 75% 100% Malicious Activity (b) Link Cost evolution 0 (c) Reputation Decay (a) Reputation evolution Fig. 1: Impact of α on reputation and link cost, and reputation decay with a half life of 15min where Rnt is the reputation of node n at time t. Since Rnt is normalised between 0 and 1, we can scale the resulting cost by defining minimum and maximum values, Cmin and Cmax. We use Cmin = 1 meaning that even using a trustworthy node possesses a cost. Furthermore, the resulting cost is then reduced to the greatest natural number less than or equal to the calculated value. With a maximum value of 255, we can calculate the maximum possible cost based upon the number of potential nodes in the network: 255 _Cmax =_ (6) _Lmax_ _[−]_ [1 +][ C][min] By decreasing Lmax (i.e. the maximum possible route length), we can increase the precision of the link cost function. For example, Lmax = 32 would result with a maximum value of 8, whereas Lmax = 64 would only allow for 4 values. The resulting graph can be observed in Fig.1b, corresponding to the link cost of Fig.1a with Lmax = 64. The influence of _α is once again visible where we can see that, similar to_ Fig.1a, the higher its value, the steeper the climb in cost. By using a dynamic scaling function based upon the size of the network, the precision of the link-cost metric can be adapted to the situation. Furthermore, by associating it with the NET_DIAMETER configuration variable used by AODV, our system can be integrated in a seamless manner. _2) Reputation Decay: Once a node’s reputation has been_ calculated, it will only evolve if the node participates in another route. However, if a node possesses a reputation of 0, it may not be used again in the near future even if it is no longer malicious. In many cases, the malicious device is abandoned by the attacker once it is no longer useful, thus no longer posing a threat bur remaining excluded from routing operations due to its low reputation. To overcome this issue, we propose a new metric called _Reputation Decay, where a node’s reputation decays overtime_ when not used towards 0.5, a neutral reputation. By doing so, these abandoned or cleansed nodes can once again be used in routing, allowing them to prove their intentions. However, the decay value does not modify the list of good or bad actions, simply modifying the calculated reputation, making it easier to reincorporate nodes without changing their history: _Rdnt = (t −_ _tRn_ ) × ( _t 1[λ]_ ) (7) _Rnt = Rn −_ _Rdnt_ (8) 2 _[R]_ where Rdnt is the reputation decay of node n at time t, λ the decay factor, t 1 2 _[R][ the half life of the reputation and][ R][n][t]_ (as seen in (5)) the reputation of n at time t, after decay. Fig.1c presents different decay functions used by λ with t 1 2 _[R][ =]_ 15 min. We can see that each function impacts the decay rate in different fashions, from the classic exponential half-life to a more direct Linear or static approach. We decided to use a linear decay function with λ = 0.25, meaning the reputation will return to neutral after 2 _t 1_ _×_ 2 _[R][.]_ _B. RREP-2Hop_ To accurately identify good and bad behaviour, miners need to know the next expected hop for a route. By overhearing RREP’s transmitted between neighbours, it is possible to construct both forwards and reverse RVTs containing the exact sequence of hops. However, Fig.2a demonstrates the limitation of RREPs, where an RREP from n to n 1 only informs of _−_ the hop between them, but not the following towards n + 1. To remedy this, we propose an update to the RREP packet format called RREP-2Hop (Fig.3) to include the addresses of the next hop. Fig.2b illustrates the difference where, when compared to 2a, the hop n + 1 is known thanks to its layer 2 address. Furthermore, by also providing the layer 3 address, 2-Hop routes can be constructed if so desired. We also add a new flag which allows the receiving node to be informed if the 2Hop protocol is in use, allowing AODV to function with or without this new addition. _C. Behavioural Validation_ In our approach, any node can be a router or a miner for a specific route. Role selection is performed during the route discovery phase by listening for and analysing passing RREP_2Hop packets. If a received packet is destined for that node, it_ is processed as normal, marking the node as a router for that route. If not, the node checks it isn’t already a router for this route, as routing and mining for the same route would result in a conflict of interest. This is to reduce the risk of malicious routing nodes injecting false information during the validation phase, thus corrupting the reputation. If all is well, it then extracts the different addresses from the packet header and adds the corresponding forward and reverse hops to the RVTs as seen in Fig.2b. By ”mining a route”, the miners are responsible for overhearing passing data packets and verifying both the packet’s integrity and hop. For this, ----- Sniffed RREP-2Hop Packet Route Verification Tables Forward Reverse Data Flow **From** **To** **Next** **From** **To** **From** **To** Packet Sniffing _ni ni−1 ni+1_ _ni_ _ni+1_ _ni_ _ni−1_ Reverse Forward **Verified** **Verified** _ni_ _ni−1_ _ni+1_ (a) Validation with RREP Sniffed RREP Packet Route Verification Tables Forward Reverse Data Flow **From** **To** **From** **To** **From** **To** Packet Sniffing _ni ni−1_ _ni_ ? _ni_ _ni−1_ Reverse Forward **Verified** **Unverified** _ni_ _ni−1_ _ni+1_ (b) Validation with RREP-2Hop Fig. 2: Illustration of the need for RREP-2Hop 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 Type R AM Reserved Prefix Sz Hop Count � Miner Flag  Next Hop IP Address (if needed)  Next Hop MAC Address (if needed)  Fig. 3: RREP-2Hop packet structure 2Hop |0 1 2 3 4 5 6 7|8|9|10|11 12 13 14 15 16 17 18|Col6|19 20 21 22 23|24 25 26 27 28 29 30 31| |---|---|---|---|---|---|---|---| |Type|R|A|M|Reserved||Prefix Sz|Hop Count| |Destination IP Address|||||||| |Destination Sequence Number|||||||| |Originator IP Address|||||||| |Lifetime|||||||| |Next Hop IP Address (if needed)|||||||| |Next Hop MAC Address (if needed)|||||||| ||||||||| **Algorithm 1 Route validation run at node n upon reception** of pkt(llsrc,lldst,src,dst) 1: if New packet detected then 2: Create new bufpkt entry with hashpkt 3: set bufpkt as valid 4: else Previous malicious activity detected ; Exit ; 5: end if 6: RTE = Get route entry for [src _dst]_ _→_ 7: RV T = get validation tables from RTE for llsrc 8: if RTE & RV T both empty then 9: _▷_ No route validation table, Malicious behaviour 10: Increment badllsrc; Set bufpkt as invalid 11: else 12: _nextHoppkt = get the next hop from RV T_ 13: **if nextHoppkt ̸= lldst then** _▷lldst is not the next_ expected hop - Malicious behaviour 14: Increment badllsrc; Set bufpkt as invalid 15: **else** _▷_ Valid behaviour 16: Increment goodllsrc 17: **end if** 18: end if each miner levies the corresponding RVT to determine the next expected hop. However, we are only able to validate packets originating from either end of the route and not intermediate transmissions. As packets traverse the network, the different miners evaluate the behaviour of each routing node by performing _Route Validation (see Alg. 1) so long as the route remains_ active. Once expired, each miner checks their Packet Buffer for potentially dropped packets which haven’t completed all expected hops. If drops are detected, the corresponding node’s bad actions are updated, before the miner begins preparations for blockchain dissemination. Before dissemination can take place, the data must first be confirmed using a custom consensus-based method. This method replaces the standard _PoW used in blockchain, where instead of competing to solve_ a puzzle, the nodes simply request confirmation from their neighbours. By using such an approach, we not only reduce _PoW’s heavy computation, but also provide a simple method_ for sharing only valid data. The miner, therefore, creates a temporary block containing all calculated actions which is then broadcast up to two hops to reach only miners who could have mined the same portion of route (i.e. the same nodes). Upon receiving a block, each miner proceeds with two calculations: First, they compute the difference ratio in common nodes between the received block and their own. If this value reaches a certain threshold (i.e 80%), the received block is considered invalid and the miner transmits their own instead. If, however, it is valid then the miner determines the efficiency factor by calculating the percentage of nodes in common in the received block, PB with their own PM, with _M the nodes mined and B the nodes in the received block._ _PB =_ _[|][M][ ∪]_ _[B][|]_ (9) _PM =_ _[|][M][ ∪]_ _[B][|]_ (10) _|B|_ _|M_ _|_ If PM >= PB, we consider B to be more efficient as it contains more nodes overall. By using the efficiency factor, we can send as few blocks as possible, thus increasing efficiency and reducing overhead. However, this process relies on other miners to ”overrule” previously transmitted blocks, indicating that they are no longer considered valid and theirs should take its place. This serves two purposes: correcting miners and determining the most efficient block. If, however, no response is received, the transmitter miner considers their block valid. They then hash the contents, including the hash of the previous block, before adding it to the blockchain. It is then broadcast up to two hops so all neighbouring nodes can extract the list of actions. _D. Implementation_ Further to the two RVTs, each node contains a Packet _Buffer and a Node Reputation Table. The former stores the_ CRC16 hashes of passing packets during routing, with their next expected hop. The latter contains the list of node actions extracted from the blockchain used to calculate the reputation with Eq. (1) - (4). In our implementation, we emulate a lightweight blockchain, where the blocks are not stored but ----- 1 0.75 1 0.75 1.00 0.5 0.25 0.5 0.25 0 0 0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14 Time Index (minutes) (a) Reputation overtime with varying malicious activity Time Index (minutes) (b) Impact of _α_ with 25% malicious activity Fig. 4: Evolution of reputation TABLE I: Simulation Parameters |Parameter|Value|Parameter|Value| |---|---|---|---| |Area Max length (Lmax) Malicious Activity Reputation Decay Initial Reputation Simulation Duration|150m×150m 64 100% Linear 0.5 15 min.|Transmission Range Number of Nodes (N) Malicious Weight (α) Window Size (Wn) Number of Simulations|50m 30 2 5 100| broadcast up to two hops, reaching only the neighbours of the nodes contained in the block. Another change is the redefinition of the PoW consensus-based block validation method. This functionality is integrated directly into the validation miners, allowing them to automatically confirm their work, without interactions with the blockchain. As a result, only confirmed blocks are ”inserted” into the chain, keeping the contained information as valid as possible. With each passing RREQ and RREP-2Hop, the receiver calculates the reputation decay and link cost using Eq. (8) and (5) of the transmitter. By checking that the new cost is higher than the previous, we can protect against potential field overflow. By only forwarding RREQs with lower link costs, we can propagate more reliable routes towards the destination, which waits a certain amount of time for as many RREQs as possible, before responding only to the most reputable path. IV. RESULTS 0.00 Fig. 5: Visualisation of route reputation after 15 min with 25% malicious nodes. directly after the first route expires at around the 1 min. mark and stay in the same overall vicinity. Fig.4b extends this and illustrates the influence of α with 25% malicious activities. We can see that, conforming to our initial hypothesis, the higher the value of α, the quicker the reputation drops, and vice-versa. We can, therefore, actively influence the weight of bad behaviour, instantly punishing a node for misbehaving or forgiving them quickly. Fig.5 presents the status of one of the simulated networks of 30 nodes after 15 mins. with 25% of them acting as black holes (thick circled), superimposing node reputation and the most used route by AODV and AODV-Miner. We can see that where AODV uses the most direct route via a malicious node, _AODV-Miner is able to select a clear path. We can also see that_ we are able to assign a bad reputation to three malicious nodes, allowing them to be avoided, whereas all nodes in the route have received a good reputation. Furthermore, we can see that other nodes also possesses varying levels of good reputation, meaning that at some point in time they were used to route data, with all other nodes possessing a neutral reputation of 0.5. As a result, we can conclude that the reputation metric is crucial in our approach to limit malicious nodes from impacting routing activities. _AODV-Miner was implemented with Contiki-NG [18] and_ simulated using Cooja. The different parameters used in the simulations are presented in Table I. Each node possesses a wireless interface using the IPv6 netstack with 6LoWPAN and a non-beacon-enabled always on CSMA radio to reduce potential collisions. For our analysis, we consider Malicious Nodes to perform black hole attacks, which are distributed throughout the network at random. This preliminary study allows us to validate our implementation using a simple form of attack, paving the way for more advanced attack types in the future. We also consider that this attack drops only data packages, leaving AODV or block related traffic intact. Our analysis pitches AODV-Miner against its older brother, AODV. _A. Reputation Analysis_ Fig.4a shows the calculated reputation based on malicious activities. With α = 2, a 25% malicious node has a reputation close to neutral, whereas higher rates of activity rapidly decrease the value. We can also see that they are attributed _B. Route Analysis_ Fig.6 compares the efficiency of AODV-Miner to AODV. We use the number of packets dropped ( _Sent_ _Received_ ) _|_ _| −|_ _|_ to determine the network throughput, visible in Fig.6a and 6b. We can see that the number of packets dropped is reduced by 48% with 10% of nodes being malicious, resulting in a clear _≈_ increase in the corresponding throughput. It is also noticeable that whatever the percentage of malicious nodes, AODV-Miner possesses a higher throughput than AODV. It must be noted, however, that not all drops are prevented since reputations are computed based upon malicious activities, allowing time for nodes to wreak havoc. Furthermore, in many cases traversing a malicious node with a link cost of 4, still has a lower cost than five nodes with a link cost of 1. That being said, this better efficiency comes at a cost as confirmed by Fig.6c, where we can see that routes are on average longer in AODV-Miner. Another cost is related to the activity of the miners, where sharing blocks results in an uptake in packet transmissions. ----- 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 35 30 25 20 15 10 5 0 100% 75% 50% 25% 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14 Time Index (minutes) (a) Packets dropped with 10% malicious nodes (b) Throughput with varying presence of malicious nodes Malicious Nodes Time Index (minutes) (d) Normalised overhead with 10% malicious nodes Time Index (minutes) (c) Average route length with 10% malicious nodes Fig. 6: Routing efficiency between AODV-Miner and AODV in the presence of malicious nodes Fig.6d puts this into perspective by normalising the overhead of AODV-Miner in relation to AODV. We can see that our overhead is higher than AODV’s, confirming that there is a compromise where the increase in security comes at a cost. V. CONCLUSION & FUTURE WORKS In this paper, we introduce a consensus-based reputation metric to identify the most trustworthy route available. By exploiting blockchain technology to disseminate the reputation in the network, we assure that no intruders can falsely modify a nodes reputation. Furthermore, by introducing a validation technique based on miner consensus, we can quickly and accurately identify both highly trustworthy nodes as well as malicious entities. Finally, by incorporating a reputation decay functionality, we also reduce the risk associated with compromised trustworthy nodes as well as assure the reintegration of sanitised nodes back into the network. By analysing its functionalities in conjunction with a reactive routing protocol such as AODV, we demonstrate that our system can detect and avoid malicious devices. Through extensive simulations pitching this protocol AODV-Miner against routing based threats, we have not only proved the efficiency of our approach against AODV, but also the importance of reputation-based routing in multi-hop networks. However, our increased efficiency comes at a cost with a rise in packet overhead due to the lightweight implementation of blockchain. Indeed, although our module uses as few communications as possible for block validation, storage is still an issue, meaning each block must be broadcast to the other nodes. Due to the static nature of our scenarios, broadcasting blocks up to two hops is sufficient to inform nodes of the status of their neighbours. By extending this preliminary analysis with variable probability based attacks, such as grey holes, we provide more of a challenge compared to the situation provided by black holes. Furthermore, since our module was developed outside of a specific routing protocol, it can be adapted onto other protocols for further in-depth analysis. ACKNOWLEDGEMENTS REFERENCES [1] F. Bao, I.-R. Chen, M.J. Chang, and J.-H. Cho. Hierarchical trust management for wireless sensor networks and its applications to trustbased routing and intrusion detection. IEEE Transactions on Network _and Service Management, 9(2):169–183, 2012._ [2] D. K. Bangotra, Y. Singh, A. Selwal, N. Kumar, and P. K. Singh. A trust based secure intelligent opportunistic routing protocol for wireless sensor networks. Wireless Personal Communications, pages 1–22, 2021. [3] N. Djedjig, D. Tandjaoui, F. Medjek, and I. Romdhani. Trust-aware and cooperative routing protocol for iot security. Journal of Information _Security and Applications, 52:102467, 2020._ [4] J. Tang, A. Liu, M. Zhao, and T. Wang. An aggregate signature based trust routing for data gathering in sensor networks. Security and _Communication Networks, 2018, 2018._ [5] Weidong Fang, Wuxiong Zhang, Wei Yang, Zhannan Li, Weiwei Gao, and Yinxuan Yang. Trust management-based and energy efficient hierarchical routing protocol in wireless sensor networks. _Digital_ _Communications and Networks, 7(4):470–478, 2021._ [6] L. Guillaume, J. van de Sype, L. Schumacher, G. Di Stasi, and R. Canonico. Adding reputation extensions to aodv-uu. In IEEE Symp. _on Comm. and Vehicular Technology in the Benelux (SCVT), 2010._ [7] A. Moinet, B. Darties, and J.-L. Baril. Blockchain based trust & authentication for decentralized sensor networks. _ArXiv,_ abs/1706.01730, 2017. [8] NARA. Blockchain white paper. White paper, National Archives and Records Administration, February 2019. [9] A. M Antonopoulos. _Mastering Bitcoin: Programming the open_ _blockchain. ” O’Reilly Media, Inc.”, 2017._ [10] X. Li, P. Jiang, T. Chen, X. Luo, and Q. Wen. A survey on the security of blockchain systems. Future Generation Computer Systems, 107, 2020. [11] M. S. Ali, M. Vecchio, M. Pincheira, K. Dolui, F. Antonelli, and M. H. Rehmani. Applications of blockchains in the internet of things: A comprehensive survey. IEEE Com. Surveys Tutorials, 21(2), 2019. [12] C. Machado and C. M. Westphall. Blockchain incentivized data forwarding in manets: Strategies and challenges. _Ad Hoc Networks,_ 110:102321, 2021. [13] H. Lazrag, A. Chehri, R. Saadane, and M. D. Rahmani. A blockchainbased approach for optimal and secure routing in wireless sensor networks and iot. In Int. Conf. on Signal-Image Technology Internet_Based Systems (SITIS), 2019._ [14] J. Wang, Y. Liu, S. Niu, and H. Song. Lightweight blockchain assisted secure routing of swarm uas networking. Computer Communications, 165:131–140, 2021. [15] M. A. A. Careem and A. Dutta. Reputation based routing in MANET using Blockchain. In Int. Conference on COMmunication Systems _NETworkS (COMSNETS), 2020._ [16] S. R. Das, C. E. Perkins, and E. M. Belding-Royer. Ad hoc On-Demand Distance Vector (AODV) Routing. RFC 3561, July 2003. [17] E. Staddon, V. Loscri, and N. Mitton. Attack categorisation for iot applications in critical infrastructures, a survey. _Applied Sciences,_ 11(16), 2021. [18] G. Oikonomou, S. Duquennoy, A. Elsts, J. Eriksson, Y. Tanaka, and N. Tsiftes. The contiki-ng open source operating system for next generation IoT devices. SoftwareX, 18:101089, 2022. This work was partially supported by a grant from CPER DATA and by the European Union’s H2020 Project “CyberSANE” -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/WiMob55322.2022.9941558?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/WiMob55322.2022.9941558, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://hal.inria.fr/hal-03787034/file/AODV-Miner_WiMob_2022.pdf" }
2,022
[ "JournalArticle", "Conference" ]
true
2022-10-10T00:00:00
[]
9,375
en
[ { "category": "Business", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Economics", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0149a1792d4ed2dfb4972b5a49c089d2f0bead9e
[]
0.903394
The Evaluation of Block Chain Technology within the Scope of Ripple and Banking Activities
0149a1792d4ed2dfb4972b5a49c089d2f0bead9e
Journal of Central Banking Theory and Practice
[ { "authorId": "2082925186", "name": "Erdogan Kaygin" }, { "authorId": "2059898696", "name": "Yunus Zengin" }, { "authorId": "47422914", "name": "Ethem Topçuoğlu" }, { "authorId": "2125741273", "name": "Serdal Ozkes" } ]
{ "alternate_issns": null, "alternate_names": [ "J Central Bank Theory Pract" ], "alternate_urls": null, "id": "b665ae3f-05b5-443a-98de-ecbe4779954d", "issn": "1800-9581", "name": "Journal of Central Banking Theory and Practice", "type": "journal", "url": "http://www.degruyter.com/view/j/jcbtp" }
Abstract Technological developments have always led to changes in all aspects of our lives. Crypto currency is one of those changes. As a result of those changes, thousands of currencies such as bitcoin, ripple, litecoin and ethereum have evolved and have found a use in business. The present study focuses upon Ripple and tries to explain its effects on banks and business theoretically. It has been stated that the money transfer performed through Ripple is faster and more economical when compared to present systems. Additionally, it has been realised that the present SWIFT system has been influenced by that speed and economy, and therefore taken considerable technologic steps with an effort to improve its system.
The Evaluation of Block Chain Technology within the Scope of Ripple and Banking Activities **153** *UDK: 336.71:004* *DOI: 10.2478/jcbtp-2021-0029* *Journal of Central Banking Theory and Practice, 2021, 3, pp. 153-167* *Received: 26 February 2020; accepted: 06 October 2020* ## ***Erdogan Kaygin [*], Yunus Zengin [**],*** ***Ethem Topcuoglu [***], Serdal Ozkes [****]*** # **The Evaluation of Block Chain ** **Technology within the Scope of ** **Ripple and Banking Activities** **Abstract** : Technological developments have always led to changes in all aspects of our lives. Crypto currency is one of those changes. As a result of those changes, thousands of currencies such as bitcoin, ripple, litecoin and ethereum have evolved and have found a use in business. The present study focuses upon Ripple and tries to explain its effects on banks and business theoretically. It has been stated that the money transfer performed through Ripple is faster and more economical when compared to present systems. Additionally, it has been realised that the present SWIFT system has been influenced by that speed and economy, and therefore taken considerable technologic steps with an effort to improve its system. **Keywords:** Ripple, Bitcoin, block chain, crypto currencies, banking and Ripple, SWIFT. **JEL Classification:** M15, M21, M48. ## **1. INTRODUCTION ** The internet, computer, mobile phones and other technologic activities have great impact upon daily lives of people and the technology increasingly penetrates into our lives. Technology alters existing habits over time. While reading a printed newspaper was a tool and an indicator of cultural level in our society, currently reading a printed newspaper instead of using mobile phones provides learn ** Kafkas University,* *Kars,Turkey* *E-mail:* *kaygin@kafkas.edu.tr* *** Kafkas University,* *Kars, Turkey* *E-mail:* *yunuszengin@kafkas.edu.tr* **** Kafkas University,* *Kars,Turkey* *E-mail:* *ethemtopcuoglu@kafkas.edu.tr* ***** Kafkas University,* *Kars, Turkey* *E-mail:* *serdalozkes@jandarma.gov.tr* ----- **154** Journal of Central Banking Theory and Practice ing later rather than being up to date. This change cannot only be observed on this example but also by looking at the number of people who go to a pay office to pay their gas and electricity bills as a result of mobile and electronic banking activities, which constitutes the changes easily observed in daily life. Business life has undergone changes as did our daily lives. Entrepreneurs have a tendency to open virtual stores on websites instead of physical shops or stores. Thanks to this method, rents for shops and stores are not paid, a store is not restricted to a single area, the restriction of space and place goes away, transactions can be controlled by computers or even by mobile phones, strong companies are addressed in terms of refund and the refund is definitely taken (Wong, Lau & Yip, 2020). In many societies, it has been common to purchase extra-territorial things through the use of foreign websites such as Alibaba, Ali Express, Gearbest, Geekbuying, Geek, Amazon, and eBay. The increasing spread of electronic shopping systems has resulted in an increase in commissions paid to the banks and duplicated spending; secure electronic systems of spending and payment have been needed due to slow processing speed of the banks, being the third party and the incidents of stealing credit card information (Luburić, 2020). The Bitcoin and block chain technology had been formed in response in 2008 by Satoshi Nakamoto and the first Bitcoin transaction was effected in 2009 (Gulec & Aktas, 2019). The technology of block chain has developed in time and paved the way for companies of crypto money and block chain solution such as Ripple, Ethereum, Litecoin, Corda, Nexledger, and Hyperledger using the same technology as Bitcoin and being privatized in accordance with its area of use. ## **2. CONCEPTUAL ENVIRONMENT** Throughout the present study, the money transfer systems used in our country and in the world, Bitcoin and block chain technology and the negative effects of Bitcoin upon business have been explained within the scope of conceptual environment. Furthermore, the opportunities to be provided by Ripple showing performance in the field of banking by privatizing the same technological elements as Bitcoin have been considered. ----- The Evaluation of Block Chain Technology within the Scope of Ripple and Banking Activities **155** ### **2.1. The Present Banking System ** There are three different practices in terms of money transfer. The first one involves remittance process in which money is transferred from bank A of the deposit account to another account in bank A. The remittance is allowed every day for twenty four hours and the time for money transfer between accounts is stated with seconds. The second practice involves EFT (Electronic Fund Transfer) in which money is transferred from deposit account of bank A to a deposit account in bank B in Turkey. Through this method, transactions can be performed every weekday between 8:30 a.m and 05:30 p.m except during bank holidays (TCMB, 2019). When compared to the remittance system, the process of transaction is longer and more costly. The third practice involves SWIFT (Society of Worldwide Interbank Financial Telecommunication) in which money is transferred from a bank account in Turkey to another bank account abroad. The transaction through this system can be performed every weekday until 05:00 p.m except during bank holidays (Kuwait Turk, 2019). It can be performed by authorized banks. The transaction actualizes within 3-4 days while it changes from one bank to another (Sanlısoy & Ciloglu, 2019). The SWIFT process is more expensive in comparison to EFT and remittance practice. In the world, approximately two millions daily and over 7.8 billions yearly SWIFT transactions are conducted in 200 regions. The value of SWIFT transactions performed in a day is above 300 billion dollars (SWIFT, 2018a). ### **2.2. Bitcoin and Block Chain Technology ** Many innovations and inventions exist and develop as a result of building new and better ones upon present ones rather than finding out nonexistent ones. Indeed, the block chain technology was built upon peer-to-peer (P2P) technology. The P2P technology was used in programs such as Napster, LimeWear, Bittorrent in 1990s to enable videos, music and other data to be shared without a central authority. Together with block chain technology located upon P2P, Bitcoin is shared instead of films and data. The security of Bitcoin system is provided through crypto technology and interpersonal money transfer is performed without need for third parties such as banks and financial institutions. This system provides opportunity for faster and cheaper money transfer (Kaygin, Topcuoglu & Ozkes, 2018). In Bitcoin system, the currency is named Bitcoin and abbreviated as BTC. One BTC is separated into smaller currencies corresponding to 100 millions Satoshi (Bonneau et al., 2015). For purchase and sale of Bitcoin, stock markets such as ----- **156** Journal of Central Banking Theory and Practice Binance, KuCoin, BitFinex, Coinbase and Kraken exist and transactions are conducted on those stock markets in exchange for dollar, euro, yen, and renminbi whereas in Turkey there are stock markets like BtcTurk in which transactions are performed through Turkish Lira (TRY). As no central authority exists, reliable nodes in more than one point (i.e. computer systems) are needed to provide maintenance of the system and to perform transactions. Called miners, those nodes take the responsibility of mathematical calculations so as to operate the system, complete the blocks and form new bitcoins. The miners are given incentive payments so that they can cover CPU power (electricity power) and other costs that they have spent during transactions. This incentive payment involves giving 50 BTC award (12.5 BTC since July 2016) to the first miner forming successful block (Khalilov, Gündebahar, & Kurtulmuşlar, 2017). The system continues by interpenetrating in the form of a chain created by blocks coming together. In nearly every ten minutes, a block having 1 MB processing limit is formed and 7 transactions are performed in a second in each block (Zheng et al, 2017). The one who wants to transfer Bitcoin signs digitally the hash (proofing keys) of the previous transaction and public key (anonymous name) of the one who will get the money and form transaction by adding those to the end of records. The credit side can confirm signatures, tenure and the chain via system (Nakamoto, 2008). To operate the system mentioned above, some critical elements are required as follows; *Security* ; the security of the system is provided by Secure Hash Algorithm (SHA), a system which enables storing demanded information by separating into insignificant pieces with a definite algorithm and resolving them by combining those insignificant pieces when demanded. It is known as a cryptographic system SHA2 (SHA-256) used by Bitcoin. At present, many applications utilise from SHA-1, which could be broken formally in 2011 by the USA National Institute of Standards and Technology (NIST). Furthermore, over 9 quintillion SHA-1 accounting (9.223.372.036.854.775.808) in total was made by the cryptology group in Google and Centrum Wiskunde & Informatica (CWI) and it was proved that SHA-1 was able to be broken with 6500 years CPU (processing unit) accounting to complete the first phase of the attack and 110 years GPU (graphic unit) accounting to complete the second phase (Karakose, 2017). It is impossible to break SHA-256 cryptographic technology for now. ----- The Evaluation of Block Chain Technology within the Scope of Ripple and Banking Activities **157** *Distributed Accounts Recording Book (public ledger)* ; It is compulsory for businesses and banks to keep accounts or records, which are generally kept and recorded at a single central point. Moreover, the records kept are secret and the businesses do not want them to be known by anyone except for stakeholders. While the present system is like that, those records are used, kept and seen by all nodes included in the system with block chain technology rather than being at a single point. It is possible to see all records belonging to transactions performed as of January 3, 2009 when the system started to be used. Those records are accessible to public and the vendors and purchasers can be viewed through anonymous names (i.e. nickname). *Cyber Security* ; while collecting records at a single centre or in a hand pose a risk to cyber attacks, keeping the records in a distributed way at various nodes prevents the risk of a central attack. While attacking the banks and business systems practicing single, common and certain security protocols is quite easy for cyber pirates, it seems impossible to succeed in attacking a system consisting of ten thousands or perhaps hundred of thousand users, renewing itself every ten minutes (every few seconds for Ripple) and working instantaneously over the same record book. After a block is closed, both transactions of the previous block and those belonging to the new block should be blocked until the other block is reached (i.e. ten minutes). As stated by Nakamoto, such interference is impossible without gaining 51 % of the system (Nakamoto, 2008). In Bitcoin system, gaining 51% of the system by one segment is called Byzantine Generals, which refers to the fact that some generals betray during a war and work for the benefits of the enemies by participating in their folds. In this respect, Byzantine Generals problem appears as the matter in which some miners in nodes cooperate and gain 51% of the management (Schwartz, Youngs & Britto, 2014). The shares of miners over mining pools system transactions are found to be F2Pool (18.2 %), Poolin (15 %), BTC.com (11.1 %), Antpool (9 %) (btc.com). In the light of this data, it is thought that experiencing the Byzantine Generals problem will not be as difficult as expected. *Double Posting*, the mistakes of double posting disappear as a result of making instantaneous entries and approving within approximately ten minutes by majority of the users. It is not possible for the one not having money to spend and for a spending to be taken in twice (Aggarwal et al., 2019). *Time* , confirming the accuracy of transactions and controls performed by the banks sometimes takes hours and even days while complete accuracy of hashes ----- **158** Journal of Central Banking Theory and Practice providing confirmation in this system can be obtained within 30 minutes (Monrat, Schelén & Andersson, 2019). The use of Bitcoin in processes such as laundering, illegal sale of drugs and weapons, child pornography is quite common. 46% of transactions performed by 26% of total Bitcoin users consist of illegal transactions (Foley, Karlsen & Putniņš, 2018). It has been realized that the purchase and sale processes practiced with Tether, a crypto currency, lead to speculative movements upon and increase its prices (Griffin & Shams, 2018). The value of Bitcoin can increase or decrease instantly as it has not been produced in exchange for a value (i.e. gold, silver etc.). The fact that one Bitcoin was traded for 12$ in October 2012, 266$ in April 2013, 1240$ in December 2013, and 339$ in April 2014 in free stock market draws an inconsistent graph (ECB, 2015). In free stock markets, a Bitcoin costing 13,854$ on average during January 2018 was traded for 10,125$ as of February 9, 2020. The fact that the wishes of the states to collect taxes from earnings gained through Bitcoin and to keep those markets in order increases day by day and it will make it difficult to provide maintenance of the system. Especially observing the presence of terrorist organisations, illegal groups and laundering enforces the demand for controlling those markets. Carrying taxation into practice in France in 2014 and the legal arrangements performed in Sweden, Germany, the USA, and Japan prove these cases (Uzer, 2017,). When the matters mentioned above are considered, the currency of Bitcoin is found to pose great risks for business and banks and is thought to bring about damages rather than benefits. In this respect, the issue will be evaluated through Ripple which will provide facilities for business and banks using block chain technology and will not create problems in terms of legal procedures (Koc, 2019). Ripple was founded by Jed McCaleb, Arthur Britto, David Schwartz, and Ryan Fugger in 2012, which gives service under the name of Ripple (XRP), a crypto money dealt in stock markets, and RippleNet as a supplier of infrastructure for financial service institutions. The systems of XRP and RippleNet utilise from the same infrastructure properties. Through RippleNet, the contracted banks and its offices in different countries all around the world provide the services of fast and safe money transfer from one point to another point with the help of block chain technology. The company has offices in San Francisco, New York, London, Sidney, India, Singapore and Luxemburg. Contrary to other crypto currencies, the service points, company managers and those making investment in the ----- The Evaluation of Block Chain Technology within the Scope of Ripple and Banking Activities **159** company are displayed transparently on the authorised website. While security and distributed ledger systems are the same as those of Bitcoin, it differs from Bitcoin in terms of the ones keeping the distributed ledger, the company’s own nodes and the nodes determined by the commission beforehand. In Ripple system, because of the fact that all the nodes have been solved and are known, the problem of Byzantine Generals has been completely solved. This system is called as UNL (Unique Node List). In order for transactions to be performed upon Ripple system, transaction instructions of at least 40% of more than one hundred UNLs should match up with each other. Those matching data are carried into draft blocks and more than one voting is performed so as to be approved by UNL nodes. When matching at the rate of 80% is enabled as a result of voting, a new block exists in distributed ledger (Ali et al., 2019). As in the Bitcoin system, transaction records are open to public and are anonymous (Jani, 2018). XRP is the third greatest crypto currency following Bitcoin and Ethereum in terms of market value (Gupta & Sadoghi, 2018). Ripple is dealt with the abbreviation of XRP in various crypto money stock markets. Ciaian, Rajcaniova and Kancs (2018) measured the change observed in values of Bitcoin and sixteen subcoins between the years of 2013-2016 with Autoregressive Distributed Lag (ARDL) model analysis. As a result, it was found that the macro economic and financial developments did not form a significant difference upon the value of XRP and the changes experienced in Bitcoin prices did not influence XRP. Fry (2018) practised rational bubble model for crypto currencies and detected bubble in Bitcoin and Ethereum while no bubble was detected in XRP. He explains that the reason why no bubble exists in Ripple stems from technological superiority of Ripple over Bitcoin. XRP can perform 50,000 transactions while Bitcoin can make 7 transactions and Ethereum 14 ones in a second (Koens & Poll, 2018). When an increasing number of users is added into the inefficacy of Bitcoin in that it performs 7 transactions in a second, waits and losses of time become indispensable (Monrat, Schelén & Andersson, 2019). As well as waiting for 10 minutes to perform a transaction in Bitcoin, three blocks are required to be formed so as to understand that the transaction has been executed and become definite and six blocks are required in order to see that it is impossible to turn back. Briefly, the transaction becomes definite and irrevocable. Through XRP system, this transaction is performed only in four seconds (Armknecht et al., 2015). There is no central unit to be applied when an incorrect operation is performed in Bitcoin. It is nearly impossible to get the money back when you have sent money to an unwanted person. While there are no systems to be addressed in other ----- **160** Journal of Central Banking Theory and Practice crypto currencies, a firm named Ripple Lab. exists in XRP, in which the banks of Santander and Standard Chartered make investment. The headquarters of this firm is in the USA, with offices in various countries. In the event that you perform an incorrect operation, it is possible to apply to the banks operating with Ripple headquarters and offices. The miners are not needed to operate XRP as in Bitcoin. A 100 million in crypto money was prepared during foundation phase (Jani, 2017). The miners constitute one of the most criticized issues of Bitcoin. The electricity spent for a transaction performed by miners is found to be equal to monthly electricity consumption of a house in the UK (Truby, 2018). In RippleNet system, although it is not obligatory to buy XRP for money transfer, the money transfer operations are charged. Along with the increasing number of users in Bitcoin system, the money transfer operations which were free at the beginning have become 0.10 € (Boucher, Nascimento, & Kritikos, 2017), which can increase and change in accordance with the amount of Bitcoin to be transferred. In the event that the demand for Bitcoin continues, it is estimated that the price will rise much more. In the present study, the banking actions focus on Ripple rather than XRP actions. In this respect, RippleNet provides fast and safe money transfer through contracted banks by working based upon block chain technology. RippleNet is a solution partner that will provide benefits and opportunities for banks and business. The present customers of Ripple are mostly comprised of companies and financial institutions (Xiao, Zhang, Lou, & Hou, 2020). Ripple provides service with more than 300 institutions in forty countries, and offers fast and economical money transfer to companies and institutions through contracted banks. The money transfer system has been divided into two different categories; the first of which involves members (i.e. banks and financial institutions) and the second one involves users (companies and customers) (Wang et al., 2019). Giving information about the extent of the service provided and the institutions worked together through some examples will be useful for understanding the issue. An agreement was signed between Ripple and Standard Chartered (the UK), National Australia Bank (Australia), Mizuho Financial Group (Japan), BMO Financial Group (Canada), Siam Commercial Bank (Thailand) and Shanghai Huarui Bank (People’s Republic of China) for pilot scheme in September 15, 2016 (Patterson, 2016). Furthermore, an agreement was made between ten financial ----- The Evaluation of Block Chain Technology within the Scope of Ripple and Banking Activities **161** institutions and Ripple in April 26, 2017. The agreed institutions involve financial institutions such as MUFG (Japan), BBVA (Spain), SEB (Sweden), Akbank (Turkey), Axis Bank (India), YES BANK (India), SBI Remit (Japan), Cambridge Global Payments (Canada), Star One Credit Union (the USA) and eZforex.com (the USA). A partnership agreement was signed between American Express and Ripple in terms of money transfer except for card actions on November 16, 2017 (Ripple Team, 2017). Moreover, a partnership agreement was signed between Ripple and Moneygram on January 11, 2018 (Truby, 2018). An agreement was made between Ripple and Saudi Arabia Money Authority (SAMA) and Kingdom of Saudi Arabia Bank on February 2, 2018 over shifting to pilot scheme. SAMA and KSA are formal central banks of the Kingdom of Saudi Arabia and the institutions managing monetary policies (Sanlısoy & Ciloglu, 2019). ## **3. CONCLUSION** The present study is believed to have made a contribution to understanding the block chain technology and Ripple, improvement of long waiting periods experienced during money transfer, payment and banking actions of the businesses, evaluation of views regarding protection from risks of exchange rate and to the related studies (Al-Rjoub, 2021). The study utilised the method of literature search, which was found to contribute to determination of the scope of research problems, development of new research topics, elimination of useless methods, finding possible future studies and forming an idea about the methods to be used (Gultekin & Bulut, 2017). As a result of literature review conducted throughout the present study, only one study examining the relationship between SWIFT and Ripple was confronted. The study carried out by Qiu, Zhang and Gao (2019) suggests that the new systems like Ripple will change greatly the market of offshore transfer within 5 or 10 years. The globalisation of the world has paved the way for removal of the borders and made it possible to have access to all geographies from China to the USA via the internet in the living room. While business has removed national borders and works through internet network for 7 days and 24 hours, the fact that the actions of EFT and SWIFT are performed only during weekdays between 08:00 a.m and 17:00 p.m gives damage not only to domestic trade but also to foreign trade. The mobile application developed by Ripple and Santander Bank provides opportunity to send money at an amount between 10 or 10.000 Sterlin to twenty one countries through Euro exchange and to the USA through dollars exchange (Santander Bank, 2019). The increase in those apllication will bring about increase in ----- **162** Journal of Central Banking Theory and Practice satisfaction of customers for banks and the opportunity of instant transaction for businesses. The business transfer or receive money as a result of export and import actions to other countries, and hence restricting those actions to office hours affects the business negatively. When you buy something abroad for your business, the product arrives and you want to pay for it, you will have to wait for 2-4 working days. Armknecht et al. (2015) realized that Ripple created a new distributed accounting records book within a few seconds at a rate of 99%. In the remaining 1%, this duration changes between 30 or 40 seconds, while this percentile declined below twenty seconds in the first two months of 2015. In transactions performed through the RippleNet, it will be avoided that the institutions are blamed for laundering and tax evasion due to transactions performed through Bitcoin. Thus, the business and institutions will not be discredited. Moreover, the incidences of tax loss and post taxation will not exist since the states collect taxes through banks. The losses which are possible to be experienced between the first price of a commercial item purchased from international markets and its price when the payment is performed are called exchange rate risk. The fluctuation in foreign currency in our country is a factor affecting the business negatively. In market conditions where 1 $ was worth 4.57 ₺ on July 6, 2018; 4.87 ₺ on July 11, 2018; 4.73 ₺ on July 23, 2018; 5.06 ₺ on August 2, 2018, the return of trade in the amount of 1 million ₺ or 1 million $ displays daily change. Whereas 1 million ₺ in the pocket of a tradesman costs 218,819 $ on July 6, 2018, it cost 205,339 $ on July 11, 2018 (doviz.com). Because the business will not have to wait and get money instantly thanks to fast money transfer through Ripple, they will not be influenced by exchange risk. In the same vein, the banks will minimise customer objections and loss of customers and remove the costs of supplementary staff for transaction follow-up. The action of double accounting of one transfer or spending stemming from the banking system is called as double posting. The distributed accounts book developed by Ripple will prevent the incidences of double posting. Furthermore, the use of distributed account book will prevent the need for the business and the banks to search for how the money has been spent. Since this system also performs the reconciliation actions with record book, the time, expenditure and workforce spent for interbank reconciliation actions will be reduced, which in turn will contribute to the banks and business positively. ----- The Evaluation of Block Chain Technology within the Scope of Ripple and Banking Activities **163** The use of Ripple transfer system in international payments will enable an opportunity for more affordable (approximately 60%) and faster money transfer with the help of direct transactions (Ripple Team, 2017). Furthermore, the idea of using block chain for interbanks money transfer by Ripple has influenced other business and institutions. For instance, SWIFT tries to perform money transfer through use of Hyperledger block chain tecnology with the participation of 34 banks (SWIFT, 2018b). The system formed by Ripple becomes a turning point in terms of banking actions. However, transfer from one country to another is limited due to low number of members of RippleNet system. Even if the country has this system, the number of banks and financial institutions is limited as well. For example, in Turkey, this system is practiced only by Akbank. For the customers who do not have an account at Akbank, the use of this system may not seem practical and applicable. For the business, not only Ripple but also block chain technolgies enabling different solutions are available. Handling containers used by Maersk with block chain technology achieve 300 $ saving per container (Diordiiev, 2018). With the agreement made by Maersk with IBM upon block chain issue, the product named TradeLens existed. According to the explanation made by IBM (IBM, 2018), TradeLens has reduced packaging costs of the products carried by the ships of Maersk performing in USA line in the ratio of 40% and made thousands of dollars profit. The block chain technology is an innovation presenting many opportunities and capabilities together and providing diversity in terms of practice for the business and banks. The fact that the business focus upon block chain technology instead of crypto money which is not based upon any authority and has unsteady market price and banks find appropriate solutions for themselves will increase the profitability as in the example of Maersk and enhance competitiveness by achieving saving in terms of time and workforce. ----- **164** Journal of Central Banking Theory and Practice ## **References ** 1. Aggarwal, S., Chaudhary, R., Aujla, G.S., Kumar, N., Choo, K.-K.R. & Zomaya, A.Y. (2019). Blockchain for smart communities: Applications, challenges and opportunities. *Journal of Network and Computer* *Applications*, 144, 13-48. 2. Ali, M.S., Vecchio, M., Pincheira, M., Dolui, K., Antonelli, F. & Rehmani, M.H. (2019). Applications of blockchains in the internet of things: A comprehensive survey. *IEEE Communications Surveys & Tutorials*, 21 (2), 1676-1717. 3. Al-Rjoub, A.M.S. (2021). A financial Stability Index for Jordan. *Journal* *of Central Banking Theory and Practice*, 10 (2), 157-178. http://dx.doi. org/10.2478/jcbtp-2021-0018 4. Armknecht, F., Karame, G.O., Mandal A., Youssef F. & Zenner E. (2015). Ripple: Overview and Outlook. *In Trust and Trustworthy Computing*, ed. Conti M., Schunter M. & Askoxylakis I., 163-180. Basel: Springer International Publishing. 5. Bonneau, J., Miller, A., Clark, J., Narayanan, A., Kroll, J.A. & Felten, E.W. (2015). SoK: Research perspectives and challenges for Bitcoin and cryptocurrencies. In Proc. IEEE Symp. Secur. Privacy, May 2015, 104–121. *How Blockchain* 6. Boucher, P., Nascimento, S. & Kritikos, M. (2017). *Technology Could Change Our Lives* . Brussels: European Parliamentary Research Service (EPRS). 7. BTC.com (2020). Pool Distribution, https://btc.com/stats/pool (Accessed 11.10.2019). 8. Ciaian, P., Rajcaniova, M. & Kancs, A. (2018). Virtual relationships: Short- and long-run evidence from Bitcoin and altcoin markets. *Journal of* *International Financial Markets, Institutions and Money*, 52, 173-195. 9. Diordiiev, V. (2018). Blockchain Technology and Its Impact on Financial and Shipping Services. *Institute for Market Problems and Economic and* *Ecological Research of National Academy of Sciences of Ukraine*, 2 (1), 51-63. 10. Doviz.com (Noktacom Medya İnternet Hiz. San. ve Tic. A.Ş.) (2019). https:// kur.doviz.com/serbest-piyasa/amerikan-dolari#, (Accessed 11.10.2019). 11. European Central Bank (ECB) (2015). *Virtual Currency Schemes–A Further* *Analysis*, Frankfurt.: European Central Bank (ECB). 12. Foley, S., Karlsen, J.R. & Putniņš, T.J. (2018). Sex, Drugs, and Bitcoin: How much illegal activity is financed through cryptocurrencies?. *Review of* *Financial Studies*, Advance online publication. http://dx.doi.org/10.2139/ ssrn.3102645 ----- The Evaluation of Block Chain Technology within the Scope of Ripple and Banking Activities **165** 13. Fry, J. (2018). Booms, busts and heavy-tails: The story of Bitcoin and cryptocurrency markets?. *Economics Letters*, 171, 225-229. 14. Griffin, J.M. & Shams, A. (2018). Is Bitcoin Really Un-Tethered? Advance online publication. http://dx.doi.org/10.2139/ssrn.3195066 (Accessed 11.10.2019) 15. Gupta S. & Sadoghi M. (2018). Blockchain Transaction Processing. In Encyclopedia of Big Data Technologies. ed. Sakr S. & Zomaya A. Basel: Springer International Publishing AG. https://doi.org/10.1007/978-3-31963962-8_333-1 16. Gulec, T. & Aktaş, H. (2019). Kripto para birimi piyasalarında etkinliğin uzun hafıza ve değişen varyans özelliklerinin testi yoluyla analizi. *Eskişehir* *Osmangazi Üniversitesi İktisadi ve İdari Bilimler Dergisi*, 14 (2), 491-510. 17. Gultekin Y. & Bulut, Y. (2017). Bitcoin Ekonomisi: Bitcoin Eko-Sisteminden doğan yeni sektörler ve analizi. *Adnan Menderes Üniversitesi SBE Dergisi*, 3 (3), 82-92. 18. IBM (2018). Trade Lens, http://newsroom.ibm.com/2018-08-09-Maerskand-IBM-Introduce-TradeLens-Blockchain-Shipping-Solution, (Accessed 11.10.2019) 19. Jani S. (2018). An Overview of Ripple Technology & Its Comparison with Bitcoin Technology. https://www.researchgate.net/publication/322436263, (Accessed 11.10.2019) 20. Karakose, İ.S. (2017), Elektronik Ödemelerde Blok Zinciri Sistematiği ve Uygulamaları, Erciyes Üniversitesi Sosyal Bilimler Enstitüsü, Master Thesis, Kayseri. *Bitcoin sistem ve özelliklerinin* 21. Kaygın, E., Topçuoğlu, E. & Ozkes, S. (2018). *iş ahlakı kapsamında incelenmesi. İş Ahlakı Dergisi*, 11 (2), 165-192. *Bitcoin ile* 22. Khalilov, K.M.C., Gündebahar, M. & Kurtulmuşlar, İ. (2017). *Dünya ve Türkiye’deki Dijital Para Çalışmaları Üzerine Bir İnceleme* . https:// ab.org.tr/ab17/bildiri/100.pdf, (Accessed 11.10.2019). 23. Koc, C. (2019). Türk Ceza Kanununda Kişisel Verilerin Kaydedilmesi Sucu TCK m 135. *Legal Hukuk Dergisi*, 17 (199), 2839-2867. 24. Koens, T. & Poll, E. (2018). What Blockchain Alternative Do You Need?, In Data Privacy Management, Cryptocurrencies and Blockchain Technology, ed. Alfaro, J.G., Joancomartí, J.H., Livraga, G. and Rios, R., 113-129. Basel:Springer International Publishing. 25. Kuwait Turk (Kuveyt Türk Katılım Bankası A.Ş.) (2019). https://www. kuveytturk.com.tr/kobi/nakit-yonetimi/odeme-yonetimi/doviz-transferiswift, (Accessed 11.10.2019). 26. Luburić, R. (2020). Crisis Prevention and the Coronavirus Pandemic as a Global and Total Risk of Our Time. *Journal of Central Banking Theory and* *Practice*, 10 (1), 55-74. http://dx.doi.org/10.2478/jcbtp-2021-0003 ----- **166** Journal of Central Banking Theory and Practice 27. Monrat, A.A., Schelén, O. & Andersson, K. (2019). A survey of blockchain from the perspectives of applications, challenges, and opportunities. *IEEE* *Access*, 7, 117134-117151. 28. Nakamoto, S. (2008), Bitcoin: A Peer-to-Peer Electronic Cash System, www. bitcoin.org, (Accessed 11.10.2019). 29. Patterson, D. (2016), Ripple Adds Several New Banks to Global Network, https://ripple.com/ripple_press/ripple-adds-several-new-banks-globalnetwork/, (Accessed 11.10.2019) 30. Qiu, T., Zhang, R. & Gao, Y. (2019). Ripple vs. SWIFT: Transforming cross border remittance using blockchain technology. *Procedia Computer Science*, 147, 428-434. 31. Ripple Team (2017). American Express Introduces Blockchain-enabled, Cross-border Payments, https://ripple.com/ripple_press/american-expressintroduces-blockchain-enabled-cross-border-payments/, (Accessed 11.10.2019). 32. Sanlisoy, S. & Ciloglu, T. (2019). An investigation on the crypto currencies and its future. *International Journal of eBusiness and eGovernment Studies*, 11 (1), 69-88. 33. Santander Bank (2019). https://www.santander.com/csgs/ Satellite?appID=santander.wc.CFWCSancomQP01&canal=CSCORP&cid= 1278712674240&empr=CFWCSancomQP01&leng=pt_PT&pagename=CF WCSancomQP01%2FGSNoticia%2FCFQP01_GSNoticiaDetalleImpresion_ PT48, (Accessed 11.10.2019) 34. Schwartz, D., Youngs, N. & Britto, A. (2014). The ripple protocol consensus algorithm. https://ripple.com/files/ripple_consensus_whitepaper.pdf. (Accessed 11.10.2019). 35. SWIFT, (2018a). Annual Review, https://www.swift.com/file/62596/ download?token=5cf760oV (Accessed 11.10.2019). 36. SWIFT, (2018b) SWIFT completes landmark DLT proof of concept, https:// www.swift.com/news-events/news/swift-completes-landmark-dlt-proof-ofconcept), (Accessed 11.10.2019). 37. TCMB (Türkiye Cumhuriyeti Merkez Bankası) (2019). Ödeme Sistemleri http://eftemkt.tcmb.gov.tr/odemeSistemleri_TR.htm, (Accessed 11.10.2019). Reference in the paper? 38. Truby, J. (2018), Decarbonizing Bitcoin: Law and policy choices for reducing the energy consumption of Blockchain technologies and digital currencies. *Energy Research & Social Science*, 44, 399-410. 39. Uzer, B. (2017), *Sanal para birimleri*, Ankara: Türkiye Cumhuriyet Merkez Bankası Ödeme Sistemleri Genel Müdürlüğü. ----- The Evaluation of Block Chain Technology within the Scope of Ripple and Banking Activities **167** 40. Xiao, Y., Zhang, N., Lou, W. & Hou, T.Y. (2020). A survey of distributed consensus protocols for blockchain networks. I *EEE Communications* *Surveys & Tutorials (Early Access)*, https://doi.org/10.1109/ COMST.2020.2969706 41. Wang, Q., Zhu, X., Ni, Y., Gu, L. & Zhu, H. (2019). Blockchain for the IoT and industrial IoT: A review. *Internet of Things*, Available Online: https://doi. org/10.1016/j.iot.2019.10 0 081 42. Wong, T.L., Lau, W.Y. & Yip, T.M. (2020). Cashless Payments and Economic Growth: Evidence from Selected OECD. Countries. *Journal of Central* *Banking Theory and Practice*, 9 (SI), 189-213. http://dx.doi.org/10.2478/ jcbtp-2020-0028 43. Zheng Z., Xie S., Dai, H., Chen, X. & Wang, H. (2017). An Overview of Blockchain Technology: Architecture, Consensus, and Future Trends, 2017 *IEEE 6* *[th]* *International Congress*, 557-564. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.2478/jcbtp-2021-0029?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.2478/jcbtp-2021-0029, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GOLD", "url": "https://sciendo.com/pdf/10.2478/jcbtp-2021-0029" }
2,021
[]
true
2021-09-01T00:00:00
[ { "paperId": "c33ac1bb6f0b0542a3c43064e30a91245118bde8", "title": "Blockchain Transaction Processing" }, { "paperId": "5d62512ca4512b672cd3fe77422b40a0d8317ca7", "title": "A financial Stability Index for Jordan" }, { "paperId": "489716461be34b325dc42c35720f248cbba31997", "title": "Crisis Prevention and the Coronavirus Pandemic as a Global and Total Risk of Our Time" }, { "paperId": "3ed0baf901d55d3533a7e358673821530f700c61", "title": "Cashless Payments and Economic Growth: Evidence from Selected OECD Countries" }, { "paperId": "8c7bf247d1bfea85b68c646aa2eb786b352ee093", "title": "Blockchain for the IoT and industrial IoT: A review" }, { "paperId": "91b307dab6b5da230190ae6f7e3b126daac100c0", "title": "Is Bitcoin Really Un-Tethered?" }, { "paperId": "d2b5eec036d9b5094a1e242e0d9a7c3e26c00d4b", "title": "Blockchain for smart communities: Applications, challenges and opportunities" }, { "paperId": "4b2b5b8252dd2bdcb93e619065501eae9efae98e", "title": "Kripto Para Birimi Piyasalarında Etkinliğin Uzun Hafıza Ve Değişen Varyans Özelliklerinin Testi Yoluyla Analizi" }, { "paperId": "627341e1a5872676798c3ab56355157c1fe78bcc", "title": "A Survey of Blockchain From the Perspectives of Applications, Challenges, and Opportunities" }, { "paperId": "bca90c73d32bb7cd268d983390f4846cd7aa28e1", "title": "AN INVESTIGATION ON THE CRYPTO CURRENCIES AND ITS FUTURE" }, { "paperId": "20d82e2cbf460df9fd7d1b461511e729d0e54f90", "title": "A Survey of Distributed Consensus Protocols for Blockchain Networks" }, { "paperId": "b58ba438747f77530ecadf3667feef34d66f29e9", "title": "How Blockchain Technology Could Change Our Lives" }, { "paperId": "1fbf77970e8bb129b9597e65a58d85fb1d76e230", "title": "Sex, Drugs, and Bitcoin: How Much Illegal Activity Is Financed Through Cryptocurrencies?" }, { "paperId": "15c69c4ece24b82105fc41d270d642593ce4318f", "title": "Decarbonizing Bitcoin: Law and policy choices for reducing the energy consumption of Blockchain technologies and digital currencies" }, { "paperId": "e50b9eebdff480013d47eed1a84a7a580b45d465", "title": "What Blockchain Alternative Do You Need?" }, { "paperId": "e538f221a3ff45df561dcef25ebed552d8a8e39a", "title": "Booms, busts and heavy-tails: The story of Bitcoin and cryptocurrency markets?" }, { "paperId": "6302676e3c48181341495520efb0f859bc338fc5", "title": "Blockchain technology and its impact on financial and shipping services" }, { "paperId": "8fffb3e09abeb2dc6afd5b4cf43f24b5c8e2e489", "title": "Bitcoin Ekonomisi: Bitcoin Eko-Sisteminden Doğan Yeni Sektörler Ve Analizi" }, { "paperId": "ee177faa39b981d6dd21994ac33269f3298e3f68", "title": "An Overview of Blockchain Technology: Architecture, Consensus, and Future Trends" }, { "paperId": "e68a72995419964d2fa42f2de34c5b1eef7cb51d", "title": "Virtual Relationships: Short- and Long-run Evidence from BitCoin and Altcoin Markets" }, { "paperId": "cc9b027bd1b8c0ee854992bb56e64d72ebf3355e", "title": "Ripple: Overview and Outlook" }, { "paperId": "9584863217908a530f7bc1bff2e9e19f423d8204", "title": "SoK: Research Perspectives and Challenges for Bitcoin and Cryptocurrencies" }, { "paperId": "a5f7e2d583dd64dae9952ef5533c252e269a70c5", "title": "Annual Review" }, { "paperId": null, "title": "Pool Distribution" }, { "paperId": "3d267bbcce5a599ac9cc42964fefb40e7b49cbb1", "title": "Applications of Blockchains in the Internet of Things: A Comprehensive Survey" }, { "paperId": null, "title": "Kuwait Turk (Kuveyt Türk Katılım Bankası A.Ş.)" }, { "paperId": null, "title": "Doviz.com (Noktacom Medya İnternet Hiz. San. ve Tic. A.Ş.)" }, { "paperId": null, "title": "TCMB (Türkiye Cumhuriyeti Merkez Bankası)" }, { "paperId": null, "title": "Türk Ceza Kanununda Kişisel Verilerin Kaydedilmesi Sucu TCK m 135" }, { "paperId": null, "title": "Ödeme Sistemleri" }, { "paperId": "4e48e909a20869f9cde931f5b8f364c8cb11ac2c", "title": "Ripple vs. SWIFT: Transforming Cross Border Remittance Using Blockchain Technology" }, { "paperId": null, "title": "SWIFT completes landmark DLT proof of concept" }, { "paperId": null, "title": "An Overview of Ripple Technology & Its Comparison with Bitcoin" }, { "paperId": null, "title": "Trade Lens" }, { "paperId": null, "title": "Bitcoin sistem ve özelliklerinin iş ahlakı kapsamında incelenmesi" }, { "paperId": null, "title": "Elektronik Ödemelerde Blok Zinciri Sistematiği ve Uygulamaları, Erciyes Üniversitesi Sosyal Bilimler Enstitüsü, Master Thesis, Kayseri" }, { "paperId": null, "title": "Ripple Team" }, { "paperId": null, "title": "Sanal para birimleri , Ankara: Türkiye Cumhuriyet Merkez Bankası Ödeme Sistemleri Genel Müdürlüğü" }, { "paperId": null, "title": "American Express Introduces Blockchain-enabled, Cross-border Payments" }, { "paperId": "ad26e43d2a03ca9f7d817907c83ba6b803b945e3", "title": "Bitcoin ile Dünya ve Türkiye ’ deki Dijital Para Çalışmaları Üzerine Bir İnceleme Merve Can" }, { "paperId": null, "title": "Ripple Adds Several New Banks to Global Network" }, { "paperId": null, "title": "Virtual Currency Schemes–A Further Analysis , Frankfurt" }, { "paperId": "bff4ecdd2c40bb67abab8d49e99c81287a7b2810", "title": "The Ripple Protocol Consensus Algorithm" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" } ]
9,426
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0149bc70d355eaf564fd6bf2a25587b73c634961
[ "Computer Science" ]
0.869611
A Comparative and Comprehensive Analysis of Smart Contract Enabled Blockchain Applications
0149bc70d355eaf564fd6bf2a25587b73c634961
International Journal on Recent and Innovation Trends in Computing and Communication
[ { "authorId": "2175657136", "name": "Vishalkumar Langaliya" }, { "authorId": "1500396861", "name": "Jaypalsinh A. Gohil" } ]
{ "alternate_issns": null, "alternate_names": [ "Int J Recent Innov Trends Comput Commun" ], "alternate_urls": null, "id": "30f717c3-b4cb-461f-9152-25a2fc4208e8", "issn": "2321-8169", "name": "International Journal on Recent and Innovation Trends in Computing and Communication", "type": "journal", "url": "http://www.ijritcc.org/" }
Blockchain is a disruptive innovation that is already reshaping corporate, social, and political connections, as well as any other form of value exchange. Again, this isn't simply a shift; it's a fast-moving phenomenon that has already begun. Top financial institutions and a large number of businesses have begun to investigate blockchain in order to cut transaction costs, speed up transaction times, reduce fraud risk, and eliminate the need for middlemen or intermediate services. Blockchain is believed to be the component that completes the Internet puzzle and makes it more open, more accessible, and more reliable. In this article, we first introduced the blockchain technology and smart contracts and their merits and demerits. Second, we present a comparative and comprehensive analysis of smart contract-enabled blockchain applications. Toward the end, we discussed the future development trends of smart contract enabled blockchain applications. This document is intended to serve as a guide and resource for future research initiatives.
**_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ # A Comparative and Comprehensive Analysis of Smart Contract Enabled Blockchain Applications ### 1. INTRODUCTION Blockchain is model for delivering the information because it provides immediate, shareable, and completely transparent data stored on an immutable ledger that can only be read by permissioned network members. A blockchain network can track orders, payments, accounts, production, and much more. Because affiliates share a single view of the fact, you can see all the facts of a transaction from beginning to end, giving you more confidence as well as new efficiencies and opportunities. **_Blockchain[1]_** is a decentralised, immutable database that simplifies the recording of transactions and asset tracking in a corporate network. A tangible item (such as a house, car, cash, or land) can also be an intangible asset (intellectual property, patents, copyrights, branding). Almost everything of value may be recorded and traded on the blockchain network, which reduces risk and lowers costs for all parties involved. **_Bitcoin[2]_** is a peer-to-peer payment system that eliminates the need for trusted third parties. Bitcoin is a decentralized cryptocurrency that is not restricted to any nation and is a global currency. It is decentralized in every aspect— technical, logical, as well as political [1]. **_Ethereum[3]_** is a piece of software that runs on a network of computers and ensures that data and small computer programmes known as smart contracts are duplicated and processed across the entire network without the need for a central controller. ## Vishalkumar Langaliya[1] Research Scholar, Department of Computer Application, Marwadi University, Rajkot, Gujarat 360003. India. https://orcid.org/0000-0001-9581-547X, vishal.langaliya@gmail.com ## Jaypalsinh A. Gohil[2] Assistant Professor, Department of Computer Application, Marwadi University, Rajkot, Gujarat 360003. India. jaypalsinh.gohil@marwadieducation.edu.in https://orcid.org/0000-0003-0925-6646 **Abstract- Blockchain is a disruptive innovation that is already reshaping corporate, social, and political connections, as well as any other** form of value exchange. Again, this isn't simply a shift; it's a fast-moving phenomenon that has already begun. Top financial institutions and a large number of businesses have begun to investigate blockchain in order to cut transaction costs, speed up transaction times, reduce fraud risk, and eliminate the need for middlemen or intermediate services. Blockchain is believed to be the component that completes the Internet puzzle and makes it more open, more accessible, and more reliable. In this article, we first introduced the blockchain technology and smart contracts and their merits and demerits. Second, we present a comparative and comprehensive analysis of smart contract-enabled blockchain applications. Toward the end, we discussed the future development trends of smart contract enabled blockchain applications. This document is intended to serve as a guide and resource for future research initiatives. **Keywords- Blockchain, Smart contracts, Blockchain Applications, Comparative Analysis.** ### 1. INTRODUCTION of value may be recorded and traded on the blockchain Blockchain is model for delivering the information because network, which reduces risk and lowers costs for all parties it provides immediate, shareable, and completely transparent involved. data stored on an immutable ledger that can only be read by **_Bitcoin[2]_** is a peer-to-peer payment system that eliminates permissioned network members. A blockchain network can the need for trusted third parties. Bitcoin is a decentralized track orders, payments, accounts, production, and much cryptocurrency that is not restricted to any nation and is a more. Because affiliates share a single view of the fact, you global currency. It is decentralized in every aspect— can see all the facts of a transaction from beginning to end, technical, logical, as well as political [1]. giving you more confidence as well as new efficiencies and opportunities. **_Ethereum[3]_** is a piece of software that runs on a network of **_Blockchain[1]_** is a decentralised, immutable database that computers and ensures that data and small computer simplifies the recording of transactions and asset tracking in programmes known as smart contracts are duplicated and a corporate network. A tangible item (such as a house, car, processed across the entire network without the need for a cash, or land) can also be an intangible asset (intellectual central controller. 1. Blockchain. https://www.ibm.com/in-en/topics/what-is-blockchain 2. Bitcoin. https://bitcoin.org/en/ 3. Ethereum. https://ethereum.org/ **_16_** ----- **_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ It builds on the Bitcoin blockchain principle of validating, storing, and replicating transaction data across multiple computers all over the world (thus the term "distributed ledger"). Ethereum goes a step farther by running computer code on numerous computers around the globe in the same way. [2]. The word was invented in the 1990s by cryptographer Nick Szabo, who defined it as "a set of promises, expressed in digital form, including mechanisms within which the parties fulfil on the other promises." **_Smart contracts[4]_** have grown since then, particularly since the introduction of decentralised blockchain platforms with the birth of Bitcoin in 2009[3]. The majority of _smart contracts are written in a_ high-level language like Solidity. However, in order to run, ### 2. LITERATURE SURVEY Fahim Ullah et al [2021], The authors employed the systematic review method to examine and analyse material published between 2000 and 2020. The literature focuses on the application of blockchain smart contracts in smart real estate and presents a conceptual framework for their implementation in smart cities that govern real estate negotiations [4]. Ten major characteristics of blockchain smart contracts are addressed in the article, which are organised into six tiers for smart real estate adoption. To 4 Solidity. https://docs.soliditylang.org/en/latest/ 5 Hyperledger. https://www.hyperledger.org/ they must be assembled to the EVM's low-level bytecode. They are installed on the Ethereum platform using a specific contract creation transaction, which is identifiable as such by being submitted to the special contract creation address, after they have been compiled. **_Hyperledger[5]_** is an open source project that aims to develop blockchain technology across industries. It's a worldwide cooperation that includes leaders in banking, finance, IoT, manufacturing, supply chains, and technology. Hyperledger is hosted by the Linux Foundation, a non-profit organisation dedicated to facilitating mass creativity through open source. The Linux Foundation also facilitates collaboration and sharing of ideas, infrastructure, and code across a global developer community. demonstrate the development of a smart contract that may be utilised for blockchain smart contracts in real estate, the decentralised application and its interactions with the Ethereum Virtual Machine (EVM) are described. Real estate owners and users as smart contract parties benefit from a sophisticated design and engagement mechanism. A stepby-step approach for establishing and ending smart contracts is described, as well as a list of functions for initiating, generating, changing, or terminating smart contracts. The suggested framework is a contractual process that is more immersive, user-friendly, and visualised. **_17_** **_Hyperledger[5]_** is an open source project that aims to develop The word was invented in the 1990s by cryptographer Nick blockchain technology across industries. It's a worldwide Szabo, who defined it as "a set of promises, expressed in cooperation that includes leaders in banking, finance, IoT, digital form, including mechanisms within which the parties manufacturing, supply chains, and technology. Hyperledger fulfil on the other promises." **_Smart contracts[4]_** have grown is hosted by the Linux Foundation, a non-profit organisation since then, particularly since the introduction of dedicated to facilitating mass creativity through open decentralised blockchain platforms with the birth of Bitcoin source. The Linux Foundation also facilitates collaboration in 2009[3]. The majority of _smart contracts are written in a_ and sharing of ideas, infrastructure, and code across a global high-level language like Solidity. However, in order to run, developer community. _Fig. 1. Structure of Blockchain_ ### 2. LITERATURE SURVEY demonstrate the development of a smart contract that may be utilised for blockchain smart contracts in real estate, the Fahim Ullah et al [2021], The authors employed the decentralised application and its interactions with the systematic review method to examine and analyse material Ethereum Virtual Machine (EVM) are described. Real estate published between 2000 and 2020. The literature focuses on owners and users as smart contract parties benefit from a the application of blockchain smart contracts in smart real sophisticated design and engagement mechanism. A step- estate and presents a conceptual framework for their by-step approach for establishing and ending smart contracts implementation in smart cities that govern real estate is described, as well as a list of functions for initiating, negotiations [4]. Ten major characteristics of blockchain generating, changing, or terminating smart contracts. The smart contracts are addressed in the article, which are suggested framework is a contractual process that is more organised into six tiers for smart real estate adoption. To immersive, user-friendly, and visualised. ----- **_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ Adarsh kumar et al [2020], The authors propose a smart healthcare system with a blockchain data network and healthcare 4.0 processes, which include industry 4.0 processes such as the internet of things (IoT), industrial IoT (IIoT), cognitive computing, artificial intelligence, cloud computing, fog computing, edge computing to provide transparency, easy and faster accessibility, security, efficiency[5]. The Ethereum network, as well as associated programming languages, tools, and techniques such as solidity, web3.js, Athena, and others, are used to construct the smart healthcare system. The learning created a comprehensive and comparative survey of cutting- edge blockchain-based smart healthcare systems. A simulationoptimization approach using JaamSim simulator is proposed to improve the performance of the overall system and subsystems. The proposed approach is tested, verified and validated through simulation and implementation. Mayank Raikwar et al [2018], An experimental prototype was built on Hyperledger fabric, an open source permissioned blockchain design platform, by the authors [6]. They discussed the most important design needs and design propositions, as well as how to encode various insurance procedures into smart contracts. Extensive experiments were conducted to analyse performance of framework and security of the proposed design and transactions based on a blockchain-enabled platform. Hoai Luan Pham et al [2018], Patients, healthcare providers (such as hospitals), and healthcare professionals (doctors) formed a remote healthcare system. Sensors were used to control the health of patients, and this information was automatically put into the blockchain [7]. In addition, they presented a processing technique for efficiently and sparingly storing medical device information based on the patient's health status. In specific, they filtered sensor data before deciding whether or not to send it to blockchain. As a result, they will be able to minimise the size of the blockchain and save a significant quantity of coins for transaction efficiency. However, abnormal data from sensors would be promptly written to blockchain, triggering an emergency call to a doctor and hospital for prompt treatment. They tested the proposed smart contract on Ethereum's TESTRPC test environment and built the system in a real-world setting with real devices. At a small scale, this system functioned successfully. Toqeer Ali et al [2020], Authors proposed a Transparent and Reliable Property Registration System on Permissioned Blockchain which provide the solution of the problem within this mechanism is that there is no transparency regarding the integrity of data about a property. i.e. the person-in-charge can manipulate the information within the database and provide misinformation to the involved stakeholders as being a manual process [8]. They took Saudi Arabia as a use-case and designed the system accordingly to transform the property registration on blockchain for the country. This study delivers a solution for governing transparency and satiates in the provision of a trusted property registration system over the Blockchain for the kingdom of Saudi Arabia. The infrastructure offers many features to the stakeholders related to the purchasing and retailing of property. The transparency, integrity of the record, and trust factor is ensured via a tamper-proof ledger Olawande Daramola et al [2020], demonstrates how the Architecture Trade-off Analysis Method (ATAM) may help stakeholders in national elections assess the risks, opportunities, and challenges that a blockchain e-voting system for national elections could bring. Using a study context of South Africa as a case study, a proposed blockchain e-voting architecture was used to assist election stakeholders in reasoning on the concept of blockchain evoting in order to educate them on the possible hazards, security concerns, important requirements qualities, and flaws related with deploying blockchain e-voting for national elections [9]. According to the report, blockchain evoting can prevent security breaches, internal vote manipulation, and boost transparency. Valentina Gatteschi et al [2018], The author of the paper uses blockchain to illustrate the process of decision- making by actors in the insurance system, analysing its benefits and drawbacks, and discussing many use examples from the insurance industry that may easily be expanded to other areas [10]. Sujit Biswas et al [2020], they first analyse and explain how business blockchain can be effectively used in healthcare, followed by the unique requirements of a healthcare system. In the latter parts of this article, they discuss the migration challenges and possible solution, the trade-off between unified and multi-chain environments, consensus algorithms for healthcare, users & access privileges, smart contracts, and e-healthcare specific industry regulations [11]. The goal of this paper was to show how difficult it is to establish a blockchain solution for e-healthcare systems and to explore for potential alternatives. Friorik P. Hjalmarsson et al [2018], They propose a unique electronic voting system based on blockchain in this research paper, which tackles some of the shortcomings of existing systems and examines some of the most wellknown blockchain frameworks for the aim of building a blockchain-based e-voting system [12]. They unveiled a blockchain-based electronic voting system that makes use of **_18_** transparency, easy and faster accessibility, security, property registration system over the Blockchain for the efficiency[5]. The Ethereum network, as well as associated kingdom of Saudi Arabia. The infrastructure offers many programming languages, tools, and techniques such as features to the stakeholders related to the purchasing and solidity, web3.js, Athena, and others, are used to construct retailing of property. The transparency, integrity of the the smart healthcare system. The learning created a record, and trust factor is ensured via a tamper-proof ledger comprehensive and comparative survey of cutting- edge Olawande Daramola et al [2020], demonstrates how the blockchain-based smart healthcare systems. A simulation- Architecture Trade-off Analysis Method (ATAM) may help optimization approach using JaamSim simulator is proposed stakeholders in national elections assess the risks, to improve the performance of the overall system and sub- opportunities, and challenges that a blockchain e-voting systems. The proposed approach is tested, verified and system for national elections could bring. Using a study validated through simulation and implementation. context of South Africa as a case study, a proposed Mayank Raikwar et al [2018], An experimental prototype blockchain e-voting architecture was used to assist election was built on Hyperledger fabric, an open source stakeholders in reasoning on the concept of blockchain e- permissioned blockchain design platform, by the authors [6]. voting in order to educate them on the possible hazards, They discussed the most important design needs and design security concerns, important requirements qualities, and propositions, as well as how to encode various insurance flaws related with deploying blockchain e-voting for procedures into smart contracts. Extensive experiments were national elections [9]. According to the report, blockchain e- conducted to analyse performance of framework and voting can prevent security breaches, internal vote security of the proposed design and transactions based on a manipulation, and boost transparency. blockchain-enabled platform. Valentina Gatteschi et al [2018], The author of the paper Hoai Luan Pham et al [2018], Patients, healthcare providers uses blockchain to illustrate the process of decision- making (such as hospitals), and healthcare professionals (doctors) by actors in the insurance system, analysing its benefits and formed a remote healthcare system. Sensors were used to drawbacks, and discussing many use examples from the control the health of patients, and this information was insurance industry that may easily be expanded to other automatically put into the blockchain [7]. In addition, they areas [10]. presented a processing technique for efficiently and Sujit Biswas et al [2020], they first analyse and explain how sparingly storing medical device information based on the business blockchain can be effectively used in healthcare, patient's health status. In specific, they filtered sensor data followed by the unique requirements of a healthcare system. before deciding whether or not to send it to blockchain. As a In the latter parts of this article, they discuss the migration result, they will be able to minimise the size of the challenges and possible solution, the trade-off between blockchain and save a significant quantity of coins for unified and multi-chain environments, consensus algorithms transaction efficiency. However, abnormal data from for healthcare, users & access privileges, smart contracts, sensors would be promptly written to blockchain, triggering and e-healthcare specific industry regulations [11]. The goal an emergency call to a doctor and hospital for prompt of this paper was to show how difficult it is to establish a treatment. They tested the proposed smart contract on blockchain solution for e-healthcare systems and to explore Ethereum's TESTRPC test environment and built the system for potential alternatives. in a real-world setting with real devices. At a small scale, this system functioned successfully. Friorik P. Hjalmarsson et al [2018], They propose a unique electronic voting system based on blockchain in this ----- **_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ smart contracts to ensure a secure and cost-effective election while also protecting voters' privacy. Tanesh Kumar et al [2018], Exploring the possible applications of blockchain technology in present healthcare systems, as well as the most critical needs for such systems, such as trustless and transparent healthcare systems [13]. In addition, this report also outlines the hurdles and roadblocks that must be overwhelmed before blockchain technology can be successfully implemented in healthcare systems. they also introduce the smart contract for blockchain-based healthcare systems, which is critical for setting pre-defined agreements among various players. Ioannis Karamitsos et al [2018], The goal of this article is to present Blockchain and smart contracts in the context of real estate. A full smart contract design is described, followed by an examination of a use case for renting residential and commercial premises [14]. They present a smart contract design technique that allows for the development of various use cases using Blockchain technology. A full description of state finite functions and processes is provided for a specific use case that makes significant contributions to the real estate domain. Rohan Bennett et al [2021], The authors show how comparative analysis may be done utilising a variety of frameworks, such as business requirements adherence, technology eagerness and maturity assessment, and strategic grid analysis. The findings suggest that the hybrid strategy allows for compliance with land dealing business criteria, and that proofs-of-concept are an important phase in the development process [15]. Finally, a maturity model for the use of blockchain and smart contracts in land transactions is offered. Vinay Thakura et al [2019], It highlights concerns such as nominal transparency, accountability, incoherent data sets with several government departments pertaining to the same piece of land, and delays in the current Land Records management procedure, as well as how to fix these issues utilising Blockchain Technology [16]. The authors also demonstrate a system design for the deployment of a Land Titling system utilising Blockchain Technology, so that land titles are tamper-proof and give legitimate and conclusive rights on ownership. The research report recommends utilising Blockchain's inherent benefits, with a focus on smart contracts. Each transaction, whether it is a property sale, an inheritance, a court order, or a land acquisition, will be captured and permanently recorded by the system. This means you get near-real-time updated records with accurate traceability and visibility into the state of your property records. **_19_** such as trustless and transparent healthcare systems [13]. In offered. addition, this report also outlines the hurdles and roadblocks that must be overwhelmed before blockchain technology Vinay Thakura et al [2019], It highlights concerns such as can be successfully implemented in healthcare systems. they nominal transparency, accountability, incoherent data sets also introduce the smart contract for blockchain-based with several government departments pertaining to the same healthcare systems, which is critical for setting pre-defined piece of land, and delays in the current Land Records agreements among various players. management procedure, as well as how to fix these issues utilising Blockchain Technology [16]. The authors also Ioannis Karamitsos et al [2018], The goal of this article is to demonstrate a system design for the deployment of a Land present Blockchain and smart contracts in the context of real Titling system utilising Blockchain Technology, so that land estate. A full smart contract design is described, followed by titles are tamper-proof and give legitimate and conclusive an examination of a use case for renting residential and rights on ownership. The research report recommends commercial premises [14]. They present a smart contract utilising Blockchain's inherent benefits, with a focus on design technique that allows for the development of various smart contracts. use cases using Blockchain technology. A full description of state finite functions and processes is provided for a specific Each transaction, whether it is a property sale, an use case that makes significant contributions to the real inheritance, a court order, or a land acquisition, will be estate domain. captured and permanently recorded by the system. This means you get near-real-time updated records with accurate Rohan Bennett et al [2021], The authors show how traceability and visibility into the state of your property comparative analysis may be done utilising a variety of records. frameworks, such as business requirements adherence, ----- **_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ ## 3. COMPARATIVE AND COMPREHENSIVE REVIEW ANALYSIS BASED ON SELECTIVE CRITERIA. **Whet** **Blockchai** **Researc** **her** **Deployme** **Year of** **n** **Problem** **h** **Soluti** **nt or** **Future** **Authors** **Publica** **Applicati** **Solution** **Statement** **Method** **on is** **Implement** **scope** **tion** **on** **tested** **ation** **Domain** **?** Fahim 2021 Real blockchain conceptu propose a Yes Yes To Ullah, Estate smart al new smart illustrate Fadi Al- Deals in contract framewo contract the smart Turjman smart adoption rk architectur contract cities to manage e for real process in real estate estate smart real deals in transaction estate, a smart s practical cities framewor k in the form of a sophisticat ed website or app can be establishe d. Adarsh 2020 Novel Create a conceptu Integration Yes Implement The Kumar Smart smart al and ed on proposed Rajalaksh Healthcar healthcare framewo interopera Ethereum work will mi e 4.0 system rk bility of using tools be Krishnam Processes using Blockchai solidity, expanded urthi, Blockchai n 3.0 and web3.js, to include Anand n 3.0 and Healthcare Athena, implement Nayyar, Healthcare 4.0 to etc. ation Kriti 4.0 create a across Sharma, connectivit smart many Vinay y and healthcare blockchai Grover interopera system n and Eklas bility. networks Hossain using various tools and methodolo gies. Mayank 2018 Insurance Design an Design Prepare a Yes No Each Raikwar Processes experimen Experim model for smart Subhra tal ental a cost- contract in Mazumdar prototype Prototyp effective the model , Sushmita on e way to has its Ruj, Hyperledg processing own Sourav er fabric, insurance- collection Sen an open related of Gupta, source transaction endorsing **_20_** |Authors|Year of Publica tion|Blockchai n Applicati on Domain|Problem Statement|Researc h Method|Solution|Whet her Soluti on is tested ?|Deployme nt or Implement ation|Future scope| |---|---|---|---|---|---|---|---|---| |Fahim Ullah, Fadi Al- Turjman|2021|Real Estate Deals in smart cities|blockchain smart contract adoption to manage real estate deals in smart cities|conceptu al framewo rk|propose a new smart contract architectur e for real estate transaction s|Yes|Yes|To illustrate the smart contract process in smart real estate, a practical framewor k in the form of a sophisticat ed website or app can be establishe d.| |Adarsh Kumar Rajalaksh mi Krishnam urthi, Anand Nayyar, Kriti Sharma, Vinay Grover and Eklas Hossain|2020|Novel Smart Healthcar e 4.0 Processes|Create a smart healthcare system using Blockchai n 3.0 and Healthcare 4.0 connectivit y and interopera bility.|conceptu al framewo rk|Integration and interopera bility of Blockchai n 3.0 and Healthcare 4.0 to create a smart healthcare system|Yes|Implement ed on Ethereum using tools solidity, web3.js, Athena, etc.|The proposed work will be expanded to include implement ation across many blockchai n networks using various tools and methodolo gies.| |Mayank Raikwar Subhra Mazumdar , Sushmita Ruj, Sourav Sen Gupta,|2018|Insurance Processes|Design an experimen tal prototype on Hyperledg er fabric, an open source|Design Experim ental Prototyp e|Prepare a model for a cost- effective way to processing insurance- related transaction|Yes|No|Each smart contract in the model has its own collection of endorsing| **Authors** **Publica** **Applicati** **Solution** **Statement** **Method** **on is** **Implement** **scope** **tion** **on** **tested** **ation** **Domain** **?** Fahim 2021 Real blockchain conceptu propose a Yes Yes To Ullah, Estate smart al new smart illustrate Fadi Al- Deals in contract framewo contract the smart Turjman smart adoption rk architectur contract cities to manage e for real process in real estate estate smart real deals in transaction estate, a smart s practical cities framewor k in the form of a sophisticat ed website or app can be establishe d. Adarsh 2020 Novel Create a conceptu Integration Yes Implement The Kumar Smart smart al and ed on proposed Rajalaksh Healthcar healthcare framewo interopera Ethereum work will mi e 4.0 system rk bility of using tools be Krishnam Processes using Blockchai solidity, expanded urthi, Blockchai n 3.0 and web3.js, to include Anand n 3.0 and Healthcare Athena, implement Nayyar, Healthcare 4.0 to etc. ation Kriti 4.0 create a across Sharma, connectivit smart many Vinay y and healthcare blockchai Grover interopera system n and Eklas bility. networks Hossain using various tools and methodolo gies. Mayank 2018 Insurance Design an Design Prepare a Yes No Each Raikwar Processes experimen Experim model for smart Subhra tal ental a cost- contract in Mazumdar prototype Prototyp effective the model , Sushmita on e way to has its Ruj, Hyperledg processing own Sourav er fabric, insurance- collection ----- **_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ Anupam permission s on a peers, Chattopad ed blockchain which hyay, and blockchain network. may be Kwok- design extended Yan Lam framework to the . transactio n level to allow for a separate group of endorsing peers for each transactio n. Hoai Luan 2018 Secure Proposed Design Prepared Yes verified the For the Pham, Thi Remote remote of and tested proposed suggested Hong Healthcar healthcare Experim a Remote smart remote Tran, e System system ental Healthcare contract on healthcare Yasuhiko for using processi System Ethereum system, Nakashim Hospital blockchain ng based on test decentralis a based on mechani Smart environme ed storage the sm Contract nt called can be Ethereum on TESTRPC developed. protocol blockchain and technolog implemente y d the system on an experiment al environme nt with real devices. Toqeer 2020 Property Discover a Use For the No No After Ali, Registrati Permission case- Kingdom testing, Adnan on System ed based of Saudi the Nadeem, Blockchai study Arabia, recommen Ali n-based and this ded Alzahrani, Transpare design solution framewor Salman nt and framewo controls k and Jan Trusted rk. transparen system Property cy and design Registratio satisfies in will be n System. the implement provision ed and of a improved trusted utilising property the registratio appropriat n system e on the platform. Blockchai n. The infrastruct ure provides **_21_** |Anupam Chattopad hyay, and Kwok- Yan Lam|Col2|Col3|permission ed blockchain design framework .|Col5|s on a blockchain network.|Col7|Col8|peers, which may be extended to the transactio n level to allow for a separate group of endorsing peers for each transactio n.| |---|---|---|---|---|---|---|---|---| |Hoai Luan Pham, Thi Hong Tran, Yasuhiko Nakashim a|2018|Secure Remote Healthcar e System for Hospital|Proposed remote healthcare system using blockchain based on the Ethereum protocol|Design of Experim ental processi ng mechani sm|Prepared and tested a Remote Healthcare System based on Smart Contract on blockchain technolog y|Yes|verified the proposed smart contract on Ethereum test environme nt called TESTRPC and implemente d the system on an experiment al environme nt with real devices.|For the suggested remote healthcare system, decentralis ed storage can be developed.| |Toqeer Ali, Adnan Nadeem, Ali Alzahrani, Salman Jan|2020|Property Registrati on System|Discover a Permission ed Blockchai n-based Transpare nt and Trusted Property Registratio n System.|Use case- based study and design framewo rk.|For the Kingdom of Saudi Arabia, this solution controls transparen cy and satisfies in the provision of a trusted property registratio n system on the Blockchai n. The infrastruct ure provides|No|No|After testing, the recommen ded framewor k and system design will be implement ed and improved utilising the appropriat e platform.| allow for a separate group of endorsing peers for each transactio n. Hoai Luan 2018 Secure Proposed Design Prepared Yes verified the For the Pham, Thi Remote remote of and tested proposed suggested Hong Healthcar healthcare Experim a Remote smart remote Tran, e System system ental Healthcare contract on healthcare Yasuhiko for using processi System Ethereum system, Nakashim Hospital blockchain ng based on test decentralis a based on mechani Smart environme ed storage the sm Contract nt called can be Ethereum on TESTRPC developed. protocol blockchain and technolog implemente y d the system on an experiment al environme nt with real devices. Toqeer 2020 Property Discover a Use For the No No After Ali, Registrati Permission case- Kingdom testing, Adnan on System ed based of Saudi the Nadeem, Blockchai study Arabia, recommen Ali n-based and this ded Alzahrani, Transpare design solution framewor Salman nt and framewo controls k and Jan Trusted rk. transparen system Property cy and design Registratio satisfies in will be n System. the implement provision ed and of a improved trusted utilising property the registratio appropriat n system e on the platform. ----- **_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ various benefits to those involved in the purchasing and selling of real estate. A tamperproof ledger ensures openness, record integrity, and trustworthi ness. Olawande 2020 E-Voting To gain a Architect Through a No No This Daramola, System better ure collaborati research Darren for understand Trade-of ve was Thebus National ing of the Analysis architectur limited to Elections risks, Method al South opportuniti (ATAM) assessmen Africa. It es, and t and could be challenges documenta improved involved tion in the with a process, future for blockchain demonstra various e-voting ted how countries system for the throughou national architectur t the elections. e trade-of world. analysis technique (ATAM) might be used to enable election stakeholde rs to understand the possible risks, problems, and prospects of blockchain e-voting. **_22_** |Col1|Col2|Col3|Col4|Col5|various benefits to those involved in the purchasing and selling of real estate. A tamper- proof ledger ensures openness, record integrity, and trustworthi ness.|Col7|Col8|Col9| |---|---|---|---|---|---|---|---|---| |Olawande Daramola, Darren Thebus|2020|E-Voting System for National Elections|To gain a better understand ing of the risks, opportuniti es, and challenges involved with a blockchain e-voting system for national elections.|Architect ure Trade-of Analysis Method (ATAM)|Through a collaborati ve architectur al assessmen t and documenta tion process, demonstra ted how the architectur e trade-of analysis technique (ATAM) might be used to enable election stakeholde rs to understand the possible risks, problems, and prospects of blockchain e-voting.|No|No|This research was limited to South Africa. It could be improved in the future for various countries throughou t the world.| of real estate. A tamper- proof ledger ensures openness, record integrity, and trustworthi ness. Olawande 2020 E-Voting To gain a Architect Through a No No This Daramola, System better ure collaborati research Darren for understand Trade-of ve was Thebus National ing of the Analysis architectur limited to Elections risks, Method al South opportuniti (ATAM) assessmen Africa. It es, and t and could be challenges documenta improved involved tion in the with a process, future for blockchain demonstra various e-voting ted how countries system for the throughou national architectur t the elections. e trade-of world. analysis technique (ATAM) might be used to enable election stakeholde rs to understand the possible risks, problems, and prospects of blockchain e-voting. ----- **_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ Valentina 2018 Insurance To provide SWOT Outline No No Based on Gatteschi, Processes assistance Analysis the the results Fabrizio to those benefits of the Lamberti, involved and SWOT Claudio in the drawbacks analysis, Demartini decision-, as well as blockchai Chiara making explore n Pranteda process, as specific technolog and Víctor well as to applicatio y might be Santamarí discuss n instances easily a many use from the extended examples insurance to from the industr different insurance industries. industry that might easily be applied to other domains. Sujit 2020 E- To gain a Literatur examine No No The Biswas, Healthcar better e Survey and findings of Kashif e understand explain the study Sharif, Systems. ing of how how can be Fan Li, difficult it business applied to Saraju P. is to blockchain the Mohanty establish a can be developm blockchain used ent of a solution effectively blockchai for e- in n healthcare healthcare, applicatio systems as well as n. and to the special seek for needs of potential the solutions. healthcare sector. Frorik Þ. 2018 E-Voting developing Design Proposed a Yes Yes Additional Hjalmarss System an e- of blockchain measures on, voting experime -based would be Gunnlaug system ntal electronic required ur K. based on framewo voting for Hreioarsso the rk system countries n, blockchain that uses of greater Mohamma smart size to d contracts accommo Hamdaqa, to ensure date Gisli that higher Hjalmtyss elections transactio on are secure n volume and cost- per effective second. while maintainin g voter anonymity . **_23_** |Valentina Gatteschi, Fabrizio Lamberti, Claudio Demartini Chiara Pranteda and Víctor Santamarí a|2018|Insurance Processes|To provide assistance to those involved in the decision- making process, as well as to discuss many use examples from the insurance industry that might easily be applied to other domains.|SWOT Analysis|Outline the benefits and drawbacks , as well as explore specific applicatio n instances from the insurance industr|No|No|Based on the results of the SWOT analysis, blockchai n technolog y might be easily extended to different industries.| |---|---|---|---|---|---|---|---|---| |Sujit Biswas, Kashif Sharif, Fan Li, Saraju P. Mohanty|2020|E- Healthcar e Systems.|To gain a better understand ing of how difficult it is to establish a blockchain solution for e- healthcare systems and to seek for potential solutions.|Literatur e Survey|examine and explain how business blockchain can be used effectively in healthcare, as well as the special needs of the healthcare sector.|No|No|The findings of the study can be applied to the developm ent of a blockchai n applicatio n.| |Frorik Þ. Hjalmarss on, Gunnlaug ur K. Hreioarsso n, Mohamma d Hamdaqa, Gisli Hjalmtyss on|2018|E-Voting System|developing an e- voting system based on the blockchain|Design of experime ntal framewo rk|Proposed a blockchain -based electronic voting system that uses smart contracts to ensure that elections are secure and cost- effective while maintainin g voter anonymity .|Yes|Yes|Additional measures would be required for countries of greater size to accommo date higher transactio n volume per second.| Pranteda process, as specific technolog and Víctor well as to applicatio y might be Santamarí discuss n instances easily a many use from the extended examples insurance to from the industr different insurance industries. industry that might easily be applied to other domains. Sujit 2020 E- To gain a Literatur examine No No The Biswas, Healthcar better e Survey and findings of Kashif e understand explain the study Sharif, Systems. ing of how how can be Fan Li, difficult it business applied to Saraju P. is to blockchain the Mohanty establish a can be developm blockchain used ent of a solution effectively blockchai for e- in n healthcare healthcare, applicatio systems as well as n. and to the special seek for needs of potential the solutions. healthcare sector. Frorik Þ. 2018 E-Voting developing Design Proposed a Yes Yes Additional Hjalmarss System an e- of blockchain measures on, voting experime -based would be Gunnlaug system ntal electronic required ur K. based on framewo voting for Hreioarsso the rk system countries n, blockchain that uses of greater Mohamma smart size to d contracts accommo Hamdaqa, to ensure date Gisli that higher Hjalmtyss elections transactio on are secure n volume and cost- per effective second. ----- **_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ Tanesh 2018 healthcare To outline Literatur present the No No Facts may Kumar, systems the issues e survey smart be put into Vidhya and contract practise Ramani, roadblocks for with the Ijaz that must blockchain right tools. Ahmad, be -based An overcome healthcare Braeken, before systems, Erkki blockchain which is Harjula, technology critical for Mika can be setting Ylianttila successfull pre y defined implement agreement ed in s among healthcare the systems. numerous stakeholde rs involved. Ioannis 2018 Real offer a Concept A full Yes Yes must Karamitso Estate smart ual descriptio evaluate s, Maria System contract framewo n of state the impact Papadaki, design rk finite of various Nedaa process functions platforms Baker Al that allows and such as Barghuthi for the processes Hyperledg developme is er Fabric nt of provided various for a use cases specific using use case Blockchai that makes n significant technology contributio . ns to the real estate domain. Rohan 2021 Land Proposed a Compara A maturity Yes Yes Institution Bennett, Administr Hybrid tive model for al trust, Todd ation Approache analysis the use of legal, and Miller, s for Smart blockchain policy Mark Contracts and smart challenges Pickering, in Land contracts are some Al-Karim Administr in land of the Kara ation transaction major s. issues that can be addressed. **_24_** |Tanesh Kumar, Vidhya Ramani, Ijaz Ahmad, An Braeken, Erkki Harjula, Mika Ylianttila|2018|healthcare systems|To outline the issues and roadblocks that must be overcome before blockchain technology can be successfull y implement ed in healthcare systems.|Literatur e survey|present the smart contract for blockchain -based healthcare systems, which is critical for setting pre- defined agreement s among the numerous stakeholde rs involved.|No|No|Facts may be put into practise with the right tools.| |---|---|---|---|---|---|---|---|---| |Ioannis Karamitso s, Maria Papadaki, Nedaa Baker Al Barghuthi|2018|Real Estate System|offer a smart contract design process that allows for the developme nt of various use cases using Blockchai n technology .|Concept ual framewo rk|A full descriptio n of state finite functions and processes is provided for a specific use case that makes significant contributio ns to the real estate domain.|Yes|Yes|must evaluate the impact of various platforms such as Hyperledg er Fabric| |Rohan Bennett, Todd Miller, Mark Pickering, Al-Karim Kara|2021|Land Administr ation|Proposed a Hybrid Approache s for Smart Contracts in Land Administr ation|Compara tive analysis|A maturity model for the use of blockchain and smart contracts in land transaction s.|Yes|Yes|Institution al trust, legal, and policy challenges are some of the major issues that can be addressed.| Braeken, before systems, Erkki blockchain which is Harjula, technology critical for Mika can be setting Ylianttila successfull pre- y defined implement agreement ed in s among healthcare the systems. numerous stakeholde rs involved. Ioannis 2018 Real offer a Concept A full Yes Yes must Karamitso Estate smart ual descriptio evaluate s, Maria System contract framewo n of state the impact Papadaki, design rk finite of various Nedaa process functions platforms Baker Al that allows and such as Barghuthi for the processes Hyperledg developme is er Fabric nt of provided various for a use cases specific using use case Blockchai that makes n significant technology contributio . ns to the real estate domain. Rohan 2021 Land Proposed a Compara A maturity Yes Yes Institution Bennett, Administr Hybrid tive model for al trust, Todd ation Approache analysis the use of legal, and Miller, s for Smart blockchain policy Mark Contracts and smart challenges Pickering, in Land contracts are some Al-Karim Administr in land of the Kara ation transaction major s. issues that can be addressed. ----- **_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ Vinay 2015 Land Land Concept Illustrates Yes Yes to Thakura,, Titling records on ual a system combine M.N. System Blockchai framewo design for Blockchai Dojab, n for rk and implement n Yogesh K. implement design of ing a Land technolog Dwivedic, ation of system Titling y with Tanvir Land system in artificial Ahmadd, Titling in the intelligenc Ganesh India country e (AI) in Khadanga utilising order to e Blockchai make the n entire land Technolog manageme y, so that nt land titles ecosystem are safer, tamper- faster, proof and more give transparen authentic t, and and more conclusive responsive ownership . rights. _Table 1. Comparative and Comprehensive Review Analysis Based on Selective Criteria_ _._ 3. Andreas M. Antonopoulos, Gavin Wood, "What is a Smart Contract?" 2018. ## 4. CONCLUSION AND DISCUSSION 4. Fahim Ullah, Fadi Al-Turjman "A conceptual framework This article began by stating that the advent of for blockchain smart contract adoption to manage real estate "blockchain" technology prompted conceptual and deals in smart cities," 2021. design work in a variety of fields aimed at realising 5. Adarsh Kumar Rajalakshmi Krishnamurthi, Anand Nayyar, Kriti Sharma, Vinay Grover and Eklas Hossain, "A Novel the previous "smart contract" notion. The research Smart Healthcare Design, Simulation, and Implementation focused on a Systemic Literature Review of Using Healthcare 4.0 Processes," 2020. contemporary research work conducted between 2015 6. Mayank Raikwar Subhra Mazumdar, Sushmita Ruj, Sourav and 2021. Its current analysis is a comparative and Sen Gupta, Anupam Chattopadhyay, and Kwok-Yan Lam. "A Blockchain Framework for Insurance Processes," 2018. comprehensive review of block chain applications in 7. Hoai Luan Pham, Thi Hong Tran, Yasuhiko Nakashima, "A numerous domains. According to current research, Secure Remote Healthcare System for Hospital Using only a few application sectors have been covered for Blockchain Smart Contract," 2018. 8. Toqeer Ali, Adnan Nadeem, Ali Alzahrani, Salman Jan “A blockchain technology deployment, such as the health Transparent and Trusted Property Registration System on sector, insurance sector, e-voting sector, or land sector. Permissioned Blockchain," 2020. In the future, Blockchain technology combined with 9. Olawande Daramola, Darren Thebus "Architecture-Centric smart contacts could be used in a variety of sectors Evaluation of Blockchain-Based Smart Contract E- Voting for National Elections," 2020. that are currently untapped. 10. Valentina Gatteschi, Fabrizio Lamberti, Claudio Demartini Chiara Pranteda and Víctor Santamaría, “Blockchain and ## FUNDING: This study was not funded by any Smart Contracts for Insurance: Is the Technology Mature organization. Enough?” 2018. 11. Sujit Biswas, Kashif Sharif, Fan Li, Saraju P. Mohanty, “Blockchain for E-Healthcare Systems:Easier Said Than ## CONFLICT OF INTEREST: The authors declare Done,” 2020. that they have no conflicts of interest. 12. Frorik Þ. Hjalmarsson, Gunnlaugur K. Hreioarsson, Mohammad Hamdaqa, Gisli Hjalmtysson, “BlockchainBased E-Voting System,” 2018. ## REFERENCES 13. Tanesh Kumar, Vidhya Ramani, Ijaz Ahmad, An Braeken, 1. Satoshi Nakamoto "Bitcoin: A Peer-to-Peer Electronic Cash Erkki Harjula, Mika Ylianttila “Blockchain Utilization in System," 2008. Healthcare: Key Requirements and Challenges” 2018. 2. Antony Lewis "A Gentle Introduction to Ethereum," 2016. 14. Ioannis Karamitsos, Maria Papadaki, Nedaa Baker Al **_25_** |Vinay Thakura,, M.N. Dojab, Yogesh K. Dwivedic, Tanvir Ahmadd, Ganesh Khadanga e|2015|Land Titling System|Land records on Blockchai n for implement ation of Land Titling in India|Concept ual framewo rk and design of system|Illustrates a system design for implement ing a Land Titling system in the country utilising Blockchai n Technolog y, so that land titles are tamper- proof and give authentic and conclusive ownership rights.|Yes|Yes|to combine Blockchai n technolog y with artificial intelligenc e (AI) in order to make the entire land manageme nt ecosystem safer, faster, more transparen t, and more responsive .| |---|---|---|---|---|---|---|---|---| Ahmadd, Titling in the intelligenc Ganesh India country e (AI) in Khadanga utilising order to e Blockchai make the n entire land Technolog manageme y, so that nt land titles ecosystem are safer, tamper- faster, proof and more give transparen authentic t, and and more conclusive responsive ownership . rights. _Table 1. Comparative and Comprehensive Review Analysis Based on Selective Criteria_ _._ 3. Andreas M. Antonopoulos, Gavin Wood, "What is a Smart Contract?" 2018. ## 4. CONCLUSION AND DISCUSSION 4. Fahim Ullah, Fadi Al-Turjman "A conceptual framework This article began by stating that the advent of for blockchain smart contract adoption to manage real estate "blockchain" technology prompted conceptual and deals in smart cities," 2021. design work in a variety of fields aimed at realising 5. Adarsh Kumar Rajalakshmi Krishnamurthi, Anand Nayyar, Kriti Sharma, Vinay Grover and Eklas Hossain, "A Novel the previous "smart contract" notion. The research Smart Healthcare Design, Simulation, and Implementation focused on a Systemic Literature Review of Using Healthcare 4.0 Processes," 2020. contemporary research work conducted between 2015 6. Mayank Raikwar Subhra Mazumdar, Sushmita Ruj, Sourav and 2021. Its current analysis is a comparative and Sen Gupta, Anupam Chattopadhyay, and Kwok-Yan Lam. "A Blockchain Framework for Insurance Processes," 2018. comprehensive review of block chain applications in 7. Hoai Luan Pham, Thi Hong Tran, Yasuhiko Nakashima, "A numerous domains. According to current research, Secure Remote Healthcare System for Hospital Using only a few application sectors have been covered for Blockchain Smart Contract," 2018. 8. Toqeer Ali, Adnan Nadeem, Ali Alzahrani, Salman Jan “A blockchain technology deployment, such as the health Transparent and Trusted Property Registration System on sector, insurance sector, e-voting sector, or land sector. Permissioned Blockchain," 2020. In the future, Blockchain technology combined with 9. Olawande Daramola, Darren Thebus "Architecture-Centric smart contacts could be used in a variety of sectors Evaluation of Blockchain-Based Smart Contract E- Voting for National Elections," 2020. that are currently untapped. 10. Valentina Gatteschi, Fabrizio Lamberti, Claudio Demartini Chiara Pranteda and Víctor Santamaría, “Blockchain and ## FUNDING: This study was not funded by any Smart Contracts for Insurance: Is the Technology Mature organization. Enough?” 2018. 11. Sujit Biswas, Kashif Sharif, Fan Li, Saraju P. Mohanty, “Blockchain for E-Healthcare Systems:Easier Said Than ## CONFLICT OF INTEREST: The authors declare Done,” 2020. ----- **_ISSN: 2321-8169 Volume: 9 Issue: 9_** **_DOI: https://doi.org/10.17762/ijritcc.v9i9.5489_** **_Article Received: 20 June 2021 Revised: 23 July 2021 Accepted: 30 August 2021 Publication: 30 September 2021_** ______________________________________________________________________________________________________________________ Barghuthi “Design of the Blockchain Smart Contract: A Use Case for Real Estate.” 2018. 15. Rohan Bennett, Todd Miller, Mark Pickering, Al-Karim Kara “Hybrid Approaches for Smart Contracts in Land Administration: Lessons from Three Blockchain Proofs-ofConcept” 2021. 16. Vinay Thakura,, M.N. Dojab, Yogesh K. Dwivedic, Tanvir Ahmadd, Ganesh Khadangae “Land records on Blockchain for implementation of Land Titling in India, ” 2015. **_26_** Ahmadd, Ganesh Khadangae “Land records on Blockchain for implementation of Land Titling in India, ” 2015. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.17762/ijritcc.v9i9.5489?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.17762/ijritcc.v9i9.5489, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GOLD", "url": "https://www.ijritcc.org/index.php/ijritcc/article/download/5489/5376" }
2,021
[]
true
2021-09-30T00:00:00
[ { "paperId": "b045bed810823b2667c8325e015fb1bb6b1a459f", "title": "A conceptual framework for blockchain smart contract adoption to manage real estate deals in smart cities" }, { "paperId": "29bdd042b7f34b7d412a2961a0b54febf8e20d2d", "title": "Hybrid Approaches for Smart Contracts in Land Administration: Lessons from Three Blockchain Proofs-of-Concept" }, { "paperId": "99df24d3f9fa3b524b2cf04a0a05f611dad42c5d", "title": "A Novel Smart Healthcare Design, Simulation, and Implementation Using Healthcare 4.0 Processes" }, { "paperId": "04ea37512c74a290148693d2749ad2080feb8ecb", "title": "Land records on Blockchain for implementation of Land Titling in India" }, { "paperId": "e5982a60d60d28dc6ded3c034bef7d66862714e8", "title": "Architecture-Centric Evaluation of Blockchain-Based Smart Contract E-Voting for National Elections" }, { "paperId": "b7f13d6c6f8dfcc1d070230731672ca8d37e6c11", "title": "A Transparent and Trusted Property Registration System on Permissioned Blockchain" }, { "paperId": "3903ace178b79f35ed08987ae0276cce6b24ccd9", "title": "A Secure Remote Healthcare System for Hospital Using Blockchain Smart Contract" }, { "paperId": "dc8384ac62e420e358e3bfaf2f93dfa075ada1f4", "title": "Blockchain Utilization in Healthcare: Key Requirements and Challenges" }, { "paperId": "54d50269928dafc6a0744e46044c17d973fdb01c", "title": "Blockchain-Based E-Voting System" }, { "paperId": "a857db244890540325950efe1f15e3772c76c50b", "title": "Design of the Blockchain Smart Contract: A Use Case for Real Estate" }, { "paperId": "a338e7181374da20ae974bbe7025bf759d3e7547", "title": "Blockchain and Smart Contracts for Insurance: Is the Technology Mature Enough?" }, { "paperId": "40cbc5913249fce1d56c9558649e53074f4ccfe8", "title": "A Blockchain Framework for Insurance Processes" }, { "paperId": "ccc0ba2add9ed38b735224c519707f15429c407a", "title": "Smart contract" }, { "paperId": null, "title": "Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169 Volume: 9 Issue: 9 DOI: https://doi.org/10.17762/ijritcc.v9i9.5489" }, { "paperId": null, "title": "A Gentle Introduction to Ethereum" }, { "paperId": "ecdd0f2d494ea181792ed0eb40900a5d2786f9c4", "title": "Bitcoin : A Peer-to-Peer Electronic Cash System" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": null, "title": "Mohanty , “ Blockchain for E - Healthcare Systems : Easier Said Than Done" } ]
13,749
en
[ { "category": "Medicine", "source": "external" }, { "category": "Biology", "source": "external" }, { "category": "Biology", "source": "s2-fos-model" }, { "category": "Medicine", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/014a35c538928a59935d1940bad737afa4159dfa
[ "Medicine", "Biology" ]
0.89658
Morphological and Biochemical Correlations of Abnormal Tau Filaments in Progressive Supranuclear Palsy
014a35c538928a59935d1940bad737afa4159dfa
Journal of Neuropathology and Experimental Neurology
[ { "authorId": "2111133033", "name": "Makio Takahashi" }, { "authorId": "4730818", "name": "K. Weidenheim" }, { "authorId": "2278202", "name": "D. Dickson" }, { "authorId": "1399784349", "name": "H. Ksiȩżak-Reding" } ]
{ "alternate_issns": null, "alternate_names": [ "J Neuropathol Exp Neurol" ], "alternate_urls": [ "http://journals.lww.com/jneuropath/pages/default.aspx" ], "id": "a09d6660-d07d-4847-a4e7-160d8c06b0ab", "issn": "0022-3069", "name": "Journal of Neuropathology and Experimental Neurology", "type": "journal", "url": "http://www.aanp-jnen.com/jnenframes.html" }
null
**Journal of Neuropathology and Experimental Neurology** Vol. 61, No. 1 Copyright � 2002 by the American Association of Neuropathologists January, 2002 pp. 33 45 # Morphological and Biochemical Correlations of Abnormal Tau Filaments in Progressive Supranuclear Palsy MAKIO TAKAHASHI, MD, KAREN M. WEIDENHEIM, MD, DENNIS W. DICKSON, MD, AND HANNA KSIEZAK-REDING, PHD **Abstract.** Progressive supranuclear palsy (PSP) is characterized by specific filamentous tau inclusions present in 3 types of cells including oligodendrocytes (coiled bodies), astrocytes (tufted astrocytes), and neurons (neurofibrillary tangles; NFTs). To correlate the morphological features and biochemical composition of tau in the inclusions, we examined tau filamentenriched fractions isolated from selected brain regions. Frontal and cerebellar white matter manifested a predominance of coiled bodies. The isolated fractions contained straight, 14-nm-wide filaments of relatively smooth appearance. Caudate nucleus and motor cortex with numerous tufted astrocytes contained mostly straight, but irregular, 22-nm-wide filaments with jagged contours. Perirhinal cortex and hippocampus, rich in NFTs, contained 22-nm-wide filaments that were twisted at 80nm intervals. Among the regions, those with tufted astrocytes showed the most heterogeneity in the ultrastructure of filaments. In all regions, isolated filaments were immunolabeled with PHF-1, Tau 46, and AT8. Fractions from all regions showed 2 PHF-1 immunoreactive bands of 64 and 68 kDa, while an additional band of 60 kDa was detected in NFT-enriched regions. All fractions, in varying extents, showed Tau-1-immunoreactive bands between 45–64 kDa. The results indicate that the 3 types of PSP tau inclusions vary in the ultrastructure although with some overlapping features. Neuronal and glial inclusions also vary in the biochemical profile of tau protein. These differences may depend on the metabolism of tau in the diseased oligodendrocytes, astrocytes, and neurons. **Key Words:** Biochemical tau mapping; Glial lesions; Paired helical filaments; Progressive supranuclear palsy; Tau inclusions; Tau phosphorylation; Ultrastructure. INTRODUCTION Progressive supranuclear palsy (PSP) is one of the rare neurodegenerative disorders that is clinically characterized by supranuclear ocular palsy, pseudobulbar palsy, parkinsonism with axial dystonia and postural instability, and progressive subcortical dementia (1–3). The presence of glial as well as neuronal pathology has recently been highlighted in PSP (4–8). The pathological features have been described as abnormal intracellular tau inclusions in specific anatomical areas involving astrocytes, oligodendrocytes, and neurons. Astrocytic pathology is seen in 3 recognizable forms: tufted astrocytes, astrocytic plaques, and thorn-shaped astrocytes. Specific tau-positive astrocytic inclusions in PSP are tufted astrocytes. For example, tufted astrocytes found in frontal lobe (areas 4 and 6) and putamen (6) are highly suggestive of PSP. These are extremely rare in the basal region (temporal lobe and insular cortex) and limbic system (amygdaloid nucleus, cingulate gyrus, and hippocampus). On the other hand, astrocytic plaques are rare in PSP and more common in corticobasal degeneration (CBD) (9, 10). Thorn-shaped From the Department of Pathology (Neuropathology) (MT, KMW, HK-R), Montefiore Medical Center and Albert Einstein College of Medicine, Bronx, New York; Department of Pathology (DWD), Neuropathology Laboratory, Mayo Clinic, Jacksonville, Florida; Visiting scientist from the Department of Brain Pathophysiology (MT), University of Kyoto, Graduate School of Medicine, Kyoto, Japan Correspondence to: Dr. Makio Takahashi, D1–801, Shinsenriminami 3-7, Toyonaka, Osaka 565-0084, Japan. Grant support: Awarded to HKR by the Society for Progressive Supranuclear Palsy and the Alzheimer’s Disease Association. astrocytes are argyrophilic tau-positive astrocytes detected in subpial and subependymal regions of some PSP cases (11). They resemble reactive astrocytes containing argyrophilic-rich cytoplasm and are mostly located in the vicinity of subpial and perivascular areas corresponding to astrocytic end feet. An unusual type of PSP tau-positive astrocyte has also been reported, resembling the Alzheimer-type 1 glial cell commonly seen in hepatic encephalopathy and Wilson’s disease (12). Tau inclusions seen in oligodendrocytes of PSP are described as coiled bodies and similar structures are detected also in other degenerations, e.g. CBD (13). Neurofibrillary tangles (NFTs) in PSP resemble either cortical flame-shaped neuronal inclusions as seen in Alzheimer disease (AD) or subcortical globose-shaped inclusions often referred to as globose tangles (4, 14). The abnormal tau inclusions are composed of aggregated and highly phosphorylated tau protein, which is a microtubule-associated protein playing a role in microtubule assembly and stability. Normal brain tau contains 6 isoforms, with the 1:1 ratio between isoforms expressing 3 repeats (3Rtau) and 4 repeats (4Rtau) in the microtubule binding domain (15). The isoforms are generated by the alternative splicing of a single tau gene on chromosome 17. Some of the neurodegenerative disorders designated as tauopathies are linked to specific abnormalities in the tau gene affecting the splicing ratio of 3Rtau vs 4Rtau and binding of tau to microtubules (16). Tauopathies are characterized by accumulation of intracellular tau inclusions in the absence of amyloid pathology and also include PSP. In selected regions of PSP ----- 34 TAKAHASHI ET AL MATERIALS AND METHODS Brain Tissue Case # Diagnosis TABLE 1 Patient Data Clinical course Age Postmortem (yr) (yr) Sex interval (h) 1 2 3 4 PSP PSP PSP AD 7 8 6 — 62 70 69 72 F M M F 23 12 3 — Abbreviations: yr, years; h, hours. brain, 4Rtau isoforms appear to predominate as reported at mRNA (17) and protein (18, 19) levels. Tau gene mutations favoring the 4Rtau splicing, however, have been only rarely associated with familial PSP (20). Instead, the _tau A0 allele may represent a genetic risk factor for PSP_ (21). Analysis of tau levels, phosphorylated epitopes, and tau isoform content clarified some of the morphological and biochemical differences among tauopathies (13, 22, 23). Tau may therefore serve as a potential biomarker identifying PSP and other tauopathies. Abnormal intracellular inclusions contain tau aggregated into filaments. Two kinds of tau filaments have been ultrastructurally demonstrated in PSP: straight filaments and paired helical filaments (PHFs), also seen as loosely intertwined paired filaments (24–26). It is unclear which kinds of filaments are present in the 3 morphologically distinct tau inclusions. It is also unclear whether tau content in these filaments is biochemically diverse. A minimal diversity in tau content is suggested by the biochemical studies of PSP brain homogenates showing a uniform composition of tau in several brain regions (27). In the present study, we examined both tau protein content and the ultrastructure of abnormal filaments isolated from pathologically distinct PSP brain regions. Our goals were to find correlations between the morphological appearance of abnormal tau inclusions in various types of cells and (i) the ultrastructure of tau filaments in these inclusions, and (ii) the composition of tau protein. The results of our studies suggest that the 3 kinds of PSP tau inclusions are distinct from each other in ultrastructure. Glial and neuronal inclusions also differ in the biochemical composition of tau protein. We conclude that these differences depend on the type of affected cells. Data on PSP and AD patients used in the present studies are listed in Table 1. An AD patient (case 4) had moderately advanced pathological stage of AD. Frontal lobe used in our studies had 2 to 3 NFTs per 40� field. Brains were obtained from brain banks at Albert Einstein College of Medicine/Montefiore Medical Center (Bronx, NY) and Mayo Clinic (Jacksonville, FL). In all cases, left hemispheres were fixed in 4% formalin and used for neuropathological examination, and right hemispheres were kept frozen at �85�C until used for biochemical and ultrastructural studies. Our PSP cases were typical and were not expected to show asymmetrical changes. It is, however, difficult to completely exclude the possibility that the burden of a particular lesion was different in the frozen and fixed tissue. Primary Antibodies Tau antibodies, their characteristics, and the location of epitopes on the tau molecule are shown in Figure 1 and Table 2. Tau antibodies obtained commercially were Tau 14 and Tau 46 (Zymed Laboratories, South San Francisco, CA), Tau-1 (Boehringer Mannheim, Indianapolis, IN), and AT100 and AT8 (Polymedco, Inc, Cortlandt Manor, NY). Other tau antibodies included PHF-1, MC-1, and CP13 donated by Peter Davies (Albert Einstein College of Medicine, Bronx, NY), 12E8 donated by Peter Seubert (Elan Pharmaceuticals, Inc. formerly Athena Neurosciences, Inc., South San Francisco, CA), and E10 donated by Andre´ Delacourte and Luc Boue´e (Inserm 422, Lille, France). Non-tau antibodies included 2 polyclonal antibodies against glial fibrillary acidic protein (GFAP) (Sigma, St. Louis, MO and BioGenex, San Ramon, CA). Immunohistochemistry Paraffin-embedded 5-�m-thick sections were obtained from motor cortex and cerebellar white matter from all PSP brains. In addition, sections of frontal white matter, perirhinal cortex with amygdala, and caudate nucleus were obtained from case 1. All sections were stained with H&E, Bielschowsky, and Gallyas methods. Immunocytochemistry for tau was performed with selected antibodies (MC-1, PHF-1, CP13, Tau-1, and AT8) using standard methods and the Avidin-Biotin Vectastain kit (Vector Laboratories, Inc., Burlingame, CA). Double immunostaining was performed as described (35) using pairs of antibodies GFAP/MC-1, GFAP/PHF-1 and GFAP/AT8 and the Avidin-Biotin system. Two chromogen substrates for horseradish peroxidase included 3,3�-diaminobenzidine tetrahydrochloride (DAB; Sigma) and SG substrate kit (Vector� SG, Vector Laboratories, Inc.). These substrates resulted in either brown (DAB) or blue (SG) precipitates. **Fig. 1.** Diagram of the longest tau isoform and location of epitopes. ----- ULTRASTRUCTURE AND BIOCHEMISTRY OF PSP FILAMENTS 35 TABLE 2 Tau Antibodies and Location of Epitopes Antibody Dilution* Tau epitope Phosphate dependence Reference MC-1 1:50 7–9 Conformation-dependent (28) Tau 14 1:1000 141–178 None (29) Tau-1 1:2000 192–199 Ser199/Ser202** (29) AT8 1:200 202–205 Ser202/Thr205 (30) CP13 1:100 �202 Ser202 (a) AT100 1:250 212–214 Thr212/Ser214 (23) 12E8 1:200 262–356 Ser262/Ser356 (31) E-10 1:10,000 274–283 None (32) PHF-1 1:200 396–404 Ser396/Ser404 (33) Tau 46 1:2000 428–441 None (34) - Immunostaining and/or Western blotting. ** Phosphorylation inhibits the antibody binding. (a) Peter Davies, personnal communication. Isolation of Filament-Enriched Fractions Tau filament-enriched fractions were isolated from the same regions of PSP brains as described for immunohistochemistry. The amount of frozen brain tissue available was as little as 0.3 g (caudate nucleus and perirhinal cortex) up to 2 g or more (motor cortex and white matter). Tau filaments (PHFs) were also isolated from an AD brain (frontal lobe). The isolation procedure has been described previously (36) and used with minor modifications. In brief, brain tissue was homogenized in isolation buffer A (1:10 w/v; 20 mM MES/NaOH, pH 6.8, 80 mM NaCl, 1 mM MgCl2, 2 mM EGTA, 0.1 mM EDTA, 0.2 mM phenylmethylsulfonyl fluoride [PMSF]) and the homogenate was centrifuged for 20 min at 27,000� _g. The obtained_ supernatant, which contained most of the soluble tau protein (nonfilamentous tau), was separated and designated as S1 fraction. The remaining pellet was suspended in buffer B (1:10, w/ v; 10 mM MES/NaOH, pH 7.4, 0.8 M NaCl, 10% sucrose, 1 mM EGTA, and 0.2 mM PMSF) and centrifuged as above. The obtained supernatant was supplemented with sarcosyl (1%, w/ v), incubated at 4�C overnight and centrifuged for 2 hours (h) at 100,000� _g. The resulting pellet was suspended in 50 mM_ MES/NaOH, pH 7.2 (0.5 ml/g tissue) and designated as sarcosyl-insoluble fraction. This fraction was enriched in sarcosylinsoluble tau filaments as examined by electron microscopy. In fractions from AD, abundant twisted filaments typical of ADtype PHFs were observed. protein signals were detected with an enhanced chemiluminescence system (ECL) from Amersham Life Science (Arlington Heights, IL). Immunogold Labeling and Electron Microscopy Filament-enriched fractions (10–25 �l) were deposited for 5 min onto 200 mesh copper grids (Fullam, Latham, NJ) precoated with Formvar and carbon. Unfixed samples on grids were incubated for 30 min at 25�C in the blocking medium (phosphate buffered saline, pH 7.4 with 0.1% bovine serum albumin, 0.1% gelatin and 5% horse serum), immunolabeled as described below and then stained with 2% uranyl acetate (37). For immunogold labeling, grids were incubated for 1 h with each of the primary and secondary antibodies. Primary antibodies were 2- to 10-fold more concentrated than that for Western blotting (Table 2). Secondary antibodies were conjugated to 10-nm colloidal gold particles and used at 1:25 dilution (Amersham Life Science). Samples were examined using a JEOL 100CX electron microscope. RESULTS Western Blotting Samples of filament-enriched fractions were mixed with Laemmli buffer to obtain final concentrations of SDS (2%), �mercaptoethanol (2%), and Tris/HCl, pH 6.8 (62.5 mM). Samples were boiled for 5 min, then spun for 1 min at 12,000� _g_ before separation on SDS-PAGE using 10% polyacrylamide gels. Separated proteins were electrotransferred onto nitrocellulose membranes and incubated with 5% nonfat milk in 10 mM Tris/HCl, pH 7.4, and 150 mM NaCl (TBS) to block nonspecific protein binding sites. Membranes were incubated with primary antibodies overnight at 4�C and then with secondary antibodies conjugated to horseradish peroxidase (Vector Laboratories, Burlingame, CA) for 1 h at 25�C. Both primary and secondary antibodies were diluted in 5% milk in TBS. Specific Characterization of PSP Cases All 3 PSP patients clinically presented with parkinsonism, postural instability, supranuclear gaze palsy, and pseudobulbar palsy. These features were consistent with PSP, but inconsistent with mixed dementia because of fairly well-preserved cognitive functions. The diagnosis of PSP based on the clinical history was confirmed by a neuropathologic examination of autopsy material. All 3 cases followed the neuropathologic criteria for typical PSP (7) in that they had an abnormal semiquantitative distribution of NFTs and neuropil threads, especially in the basal ganglia and brainstem, and tau-positive astrocytes in the involved areas. Rare amyloid plaques typically found in PSP (38, 39) were present in cases 1 and 3 according to the expected frequency for age. In each brain, we identified anatomical regions containing a single predominant type of pathological tau inclusion: tufted shape inclusions in astrocytes, coiled bodies in oligodendrocytes, and NFTs in neurons. In a given ----- 36 TAKAHASHI ET AL section, the type of inclusion was determined by using biochemical and morphological criteria. Initially, to select for tufted astrocytes, double immunocytochemistry was performed using tau and the astrocytic cell marker, GFAP. Co-localization of both markers confirmed that inclusions were of astrocytic origin. The shape of the tufted astrocytes was very characteristic, resembling that described earlier (6, 9). Tufted astrocytes exhibited no apparent cytoplasm but had tufts of argyrophilic, long, fine radiating fibers. Tau-positive but GFAP-negative inclusions indicated cells of either neuronal or oligodendroglial origin. To differentiate these cells, the morphological appearance of inclusions was evaluated with Gallyas stain or tau immunohistochemistry. Globose or flame-like-shaped NFTs were indicative of neuronal pathology. Coiled body shape was indicative of oligodendroglia (40, 41). A small number of tufted astrocytes were also found to be immunonegative for GFAP. In these instances, the morphology of the inclusion material was a deciding factor. Immunohistochemistry of Tau Inclusions _Coiled Bodies: Coiled bodies were particularly prom-_ inent in both frontal and cerebellar white matter with Gallyas stain (Fig. 2A, insertion). They were immunoreactive with tau antibodies MC-1 (Fig. 2A), CP13, and AT8. Neither tau-immunoreactive astrocytes nor tau-positive neurons could be distinguished in the white matter (Fig. 2A), suggesting that oligodendroglial pathology was an exclusive tau pathology in these regions. _Tufted Astrocytes: Tufted astrocytes were numerous in_ motor and premotor cortex, caudate nucleus, nucleus accumbens, and putamen, particularly as shown on sections with Gallyas stain (Fig. 2B). In the motor cortex, the number of tufted astrocytes varied between cases in a decreasing order (case 1 � case 3 � case 2). Tufted astrocytes were immunoreactive with tau antibodies, including MC-1 (Fig. 2C), CP13 (Fig. 2D), and AT8. In double-stained tufted astrocytes (Fig. 2C), a strong tau immunoreactivity was detected at the periphery of the astrocytic processes, whereas the GFAP immunoreactivity predominated in the perinuclear region. In some cells, the presence of nuclei was difficult to discern either due to staining artifact or to intrinsic nuclear loss. In motor cortex of cases 2 and 3, tufted astrocytes were interspersed with a few NFTs and coiled bodies. In all 3 PSP cases, thorn-shaped astrocytes were absent. _Neurofibrillary Tangles: NFTs were predominantly_ found in perirhinal cortex and anterior hippocampus. Although these NFTs were easily detected with Gallyas stain, they were less abundant than those typically seen in AD. Two kinds of NFTs could be distinguished: classic flame-like NFTs seen particularly in perirhinal cortex (Fig. 2E, F) and globose type of NFTs found mostly in deep subcortical nuclei (Fig. 2G). Regardless of the shape, both kinds of NFTs were immunoreactive with PHF-1, MC-1 (Fig. 2E, G), CP13 (Fig. 2F), and AT8, but were not immunoreactive with the Tau-1 antibody (not shown). The absence of Tau-1 immunoreactivity suggests that this epitope is blocked by phosphorylation as seen in other neurodegenerative disorders. For example, Tau1 binding is inhibited in AD and this inhibition is considered a hallmark of neurodegeneration (42). In perirhinal cortex, in addition to NFTs, a few coiled bodies were also detected. In the vicinity of NFTs, classical senile plaques surrounded by astrocytes were interspersed (0–1 plaque/40� field) as identified using tau and thioflavin-S staining. Ultrastructure and Labeling of Tau Filaments The results of ultrastructural analysis are summarized in Table 3. In the present studies, twisted filaments were defined as those displaying periodicity in width with decisive crossovers of hemi-filaments. Straight filaments were those lacking periodicity in width. In addition, a third type of filament was identified. These filaments were neither straight nor twisted and were designated as jagged filaments (see ‘‘Tufted Astrocytes’’ below). The results of immunogold labeling were compatible with immunohistochemical data (Table 3). _Coiled Bodies: Coiled bodies were the only inclusions_ identified microscopically in frontal and cerebellar white matter regions (Fig. 2A). Filament-enriched fractions isolated from these regions contained predominantly, but not exclusively, straight filaments labeled at various density with PHF-1 (Fig. 3A, C, D, F, G) and only rarely with Tau-1 antibodies (Fig. 3B, E). These filaments were approximately 14-nm wide (n � 19), but their width varied extensively from 7 to 20 nm, often within a single filament. The variations in width appeared random rather than periodic, and decisive crossovers of hemi-filaments were absent. These features were distinct from those described in AD (36). Such morphology could result from PSP-specific or cell type-specific aggregation of tau. Most of the filaments had a smooth surface, with an occasional small bud extending from or attached to the filament proper (Fig. 3A, B, arrows). Rarely, approximately 7- to 8-nm-wide filaments were seen in tight, parallel bundles (Fig. 3F) or loose aggregates. _Tufted Astrocytes: Tufted astrocytes were seen predom-_ inantly in the caudate nucleus of case 1 and motor cortices from all 3 cases (Fig. 2B–D). Filament-enriched fractions isolated from these regions contained filaments that were highly heterogeneous and widely varied in width and in an overall appearance (Fig. 4). The majority was similar to straight filaments although their surface appeared rough and irregular and their contours were jagged. The ends of filaments were blunt or splayed. In splayed filaments, fan-like ends consisted of distinct fibrils (Fig. 4A, B, E, asterisks) as seen at a higher magnification (Fig. 5A, C). Most of the filaments had irregular, nonperiodic changes in width, which ranged from 7 ----- ULTRASTRUCTURE AND BIOCHEMISTRY OF PSP FILAMENTS 37 **Fig. 2.** Immunohistochemistry/Gallyas stain of tau inclusions in oligodendrocytes (A), astrocytes (B–D), and neurofibrillary tangles (E–G) in paraffin sections from PSP brains (A–C, E, G; case 1 and D, F; case 3). A: Frontal white matter, GFAP (blue)/ MC-1 (brown) and Gallyas stain (insert). B: Motor cortex, Gallyas stain. C: Caudate, GFAP (blue)/MC-1 (brown). D: Motor cortex, CP13 (brown). E: Perirhinal cortex, MC-1 (blue)/GFAP (brown). F: Perirhinal cortex, CP13 (brown). G: Perirhinal cortex, MC-1 (brown) and Gallyas stain (insert). Tufted astrocytes are clearly demonstrated with Gallyas stain (B) and GFAP (blue)/ MC-1 (brown) double immunoreactive as seen in (C), but not in (A), which shows typical coiled bodies of oligodendrocytes (brown). Note a number of GFAP-positive (blue) and tau-negative astrocytes in (A) and (C). Flame-like neurofibrillary tangles are seen in (E) and (F), whereas typical globose neurofibrillary tangles are seen in (G). Insert in (A) has a 2-fold magnification compared to background. ----- 38 TAKAHASHI ET AL TABLE 3 Summary of the Results Obtained Using 3 PSP Brains Ultrastructure of tau filaments (width � SD) Tau protein (Western blots) PHF-1 immunoreactivity Tau 46 immunoreactivity Brain regions Predominant tau pathology (inclusions) Frontal and cerebellar Coiled bodies Straight and smooth, 2 bands (64 and 68 2 bands (64 and 68 white matter 14 nm � 3.9 (n � kDa) kDa) 19)* Caudate and motor cor- Tufted astrocytes Straight and jagged, 2 bands (64 and 68 2–6 bands (45–68 tex 22 nm � 5.5 (n � kDa) kDa), vary between 22) cases Perirhinal cortex and NFT PHF-like, 22 nm � 3 bands (60, 64, and 68 Not determined hippocampus 4.8 (n � 24) kDa)—as in AD brain - p � 0.001 vs other regions (Student t-test). **Fig. 3.** Tau filaments from coiled body-predominant regions. Filament-enriched fractions were isolated from frontal white matter (A, C, D, G; case 1) and cerebellar white matter (B, E, F; cases 3, 2 and 3, respectively). Samples were immunolabeled for tau with PHF-1 (A, C, D, F, G) or Tau-1 (B, E) and 10-nm immunogold particles. Arrows in (A) and (B) indicate small buds projecting from the filament proper. Scale bar in (A) applies to all panels. JEOL 100CX. **Fig. 4.** Tau filaments from tufted astrocyte-predominant regions. Filament-enriched fractions were isolated from motor cortex (A, B, H; case 1), (F; case 2), (C, E, G, I; case 3), and caudate (D; case 1). Samples were immunolabeled with tau antibodies and 10-nm immunogold particles as follows: PHF-1 (A, D, F, H), Tau-1 (B, E, I), AT100 (C) and Tau 46 (G). Asterisks (A, B, E) denote splayed ends of filaments. Scale bar in (A) applies to all panels. JEOL 100CX. ----- ULTRASTRUCTURE AND BIOCHEMISTRY OF PSP FILAMENTS 39 **Fig. 5.** Details of splayed tau filaments. The filaments are from motor cortex, cases 3 and 1 (A, C) and perirhinal cortex, case 1 (B, D). The filaments in (A) and (C) are also shown in Figure 4 (A, E). Scale bar in (C) applies to all panels. JEOL 100CX. to 33 nm. An average maximal width of these filaments was approximately 22 nm (n � 22). Although changes in width were more regular in some filaments (Fig. 4A, F, G, arrowheads), these filaments lacked distinct crossovers and thus were not considered twisted. Since the filaments in tufted astrocyte-enriched fractions were neither straight nor twisted, they were regarded as a third type of filament and designated as jagged filaments. As an exception, a single 27-nm-wide filament was detected, which was regularly twisted at 62-nm intervals (Fig. 4I) and thus closely resembled PHFs of AD type. This filament might have originated from the rare NFTs in the deeper cortical neurons. The filaments in tufted astrocyte-enriched fractions were consistently labeled with PHF-1, AT8, AT100, and Tau 46 but not with Tau-1 (Fig. 4B, E, I). _Neurofibrillary Tangles: NFTs were predominantly_ seen in perirhinal cortex and hippocampus of case 1. Filament-enriched fractions from both regions contained the majority of regularly twisted filaments with clearly defined crossovers (Fig. 6A–E). The twisted filaments were 17- to 33-nm wide (average maximal width of 22 nm; n � 24). The twisting interval ranged from 50 to 110 nm, with an average value of 80 nm. Twisting in some exceptionally wide 27- to 33-nm filaments, however, was uncertain (Fig. 6A, insert). The appearance of twisted filaments from perirhinal cortex (Fig. 6A, C, E) closely resembled that of AD-PHFs (Fig. 6F). Interestingly, some of the twisted filaments were splayed into 2 or 3 fibrils (Fig. 5B, D). A small number of straight, 13-nm-wide filaments was also identified (Fig. 6D). The appearance of straight filaments was similar to that seen in frontal and cerebellar white matter. Both twisted and straight filaments were immunolabeled with PHF-1 (Fig. 6A–D) and AT8 (not shown), similar to those in AD (Fig. 6F). In Western blotting, PHF-1 was found to be the most sensitive of the tau antibodies used. In all regions, there were 2 major PHF-1 immunoreactive polypeptides of 64 and 68 kDa, except for perirhinal cortex and hippocampus, where an additional 60 kDa polypeptide was seen (Fig. 7A–C). The 3-band, but not 2-band, pattern was similar to that found in samples from AD (Fig. 7C). The Tau 46 and E-10 immunoreactivity was examined only in selected regions (Fig. 7A, B) due to a limited material. With E-10, the pattern of 2 major bands of 64 and 68 kDa was detected in all selected fractions, similar to that obtained with PHF-1. With Tau 46, besides the 64 and 68 kDa polypeptides, there were 4 additional bands detected between 45–60 kDa. These additional bands were particularly prominent in motor cortex of cases 2 and 3 (Fig. 7B-MTR2 and MTR3). The 45–60 kDa bands are likely to represent the Tau 46-positive tau protein lacking phosphorylation at the PHF-1 site. They are unlikely to Tau Protein: Band Pattern ----- 40 TAKAHASHI ET AL **Fig. 6.** Tau filaments from neurofibrillary tangle-predominant regions. Filament-enriched fractions were isolated from PSP perirhinal cortex (A, D, E; Case 1), PSP hippocampus (B, C; Case 1), and AD brain (F). Samples were immunolabeled with tau antibodies and 10-nm gold particles as follows: PHF-1 (A–D), None (E), and AT8 (F). Scale bar in (A) applies to all panels. JEOL 100CX. represent tau degradation products, e.g. C-terminal fragments retaining only the Tau 46, but not the PHF-1 epitope, since both epitopes are located close to each other (Fig. 1; Table 2). Tau Protein: Variable Phosphorylation The presence of Tau 46-positive, but PHF-1-negative, bands in only some samples suggested that phosphorylation of PHF-tau polypeptides varied among cases and regions. Such a possibility was confirmed by using phosphate-dependent tau antibodies, including Tau-1, which binds only to a nonphosphorylated epitope. Consistent with previous reports, Tau-1 showed no binding to PHFtau polypeptides in our AD samples (Fig. 7C). In contrast, Tau-1 detected 4–6 polypeptides migrating between 45–64 kDa in most of the PSP samples (Fig. 7A–C). These polypeptides co-migrated with Tau 46-positive bands between 45–64 kDa but not with the 68 kDa band, which was Tau 46-positive but Tau-1-negative. The extent of immunoreactivity with Tau-1 differed among the regions. For example, motor cortex and hippocampus of case 1 (Fig. 7B-MTR1 and Fig. 7C-Hipp1) showed the least, and cerebellar white matter of case 3 (Fig. 7ACbl3) showed the most intense binding. Although it was difficult to precisely estimate the extent of phosphorylation, the results suggest that the motor cortex of case 1 (Fig. 7B-MTR1) contains almost all tau phosphorylated at the Tau-1 site, whereas the motor cortex (Fig. 7BMTR3) and cerebellar white matter of case 3 (Fig. 7ACbl3) contain only a small fraction of tau phosphorylated at this site. Further comparisons made between motor cortex of the 2 most diverse cases (case 1 and case 3) clearly demonstrate that the total tau content (Tau 46 binding) is lower, but the phosphorylation is higher in ----- ULTRASTRUCTURE AND BIOCHEMISTRY OF PSP FILAMENTS 41 **Fig. 7.** Western blotting: all cases. Filament-enriched fractions were isolated from PSP regions (cases 1–3) enriched in coiled bodies (A), tufted astrocytes (B), and NFTs (C) and from AD brain, as indicated. Samples were immunoblotted with PHF-1, Tau1, Tau 46, and E-10 as marked. Samples on blots were equivalent in the amount of tissue. Major PHF-tau polypeptides (arrows) immunoreactive with PHF-1 and E-10 migrated at 64 and 68 kDa (coiled bodies and tufted astrocytes) and 60, 64, and 68 kDa (NFTs). Four major Tau 46 and Tau-1-positive bands migrated at 52–64 kDa. The 68-kDa band was also immunoreactive with Tau 46, but not with Tau-1, indicating that phosphorylation at the Tau-1 binding epitope blocked the antibody reactivity. Abbreviations: Front, frontal white matter; Cbl, cerebellar white matter; Caud, caudate nucleus; MTR, motor cortex; Perirh, perirhinal cortex; Hipp, hippocampus. ----- 42 TAKAHASHI ET AL **Fig. 8.** Western blotting: motor cortex of selected PSP cases. Filament-enriched fractions were isolated from the motor cortex of case 1 (lane a) and case 3 (lane b) and immunoblotted with PHF-1, AT8, 12E8, AT100, Tau-1, and Tau 46, as indicated. Although the total PHF-tau content as estimated by Tau 46 immunoreactivity is lower in case 1 than in case 3, more protein is phosphorylated in case 1, especially at the PHF-1, AT8 and Tau-1 sites. The number of PHF-tau polypeptides detected varies depending on the antibody used from 2 (PHF-1, AT8, 12E8, AT100) to 6 (Tau 46). case 1 than case 3 (Fig. 8). Higher phosphorylation was particularly evident with the PHF-1 and AT8 binding. A significant inhibition of Tau-1 binding also indicated higher phosphorylation at this site. With 2 other antibodies (12E8 and AT100), the difference in phosphorylation between the cases was less pronounced. In case 1, two bands of 56 and 60 kDa were Tau-1-positive but Tau 46negative, suggesting that they may represent tau degradation products devoid of the most C-terminal region. Alternatively, the differential binding may be due to a higher binding affinity of Tau-1 than Tau 46. DISCUSSION PSP appears to be unique among tauopathies by involving 3 types of cells in specific anatomical areas. Also, each cell pathology presents with morphologically distinct tau inclusions. These unusual features of PSP were employed in the present studies to correlate the morphology of inclusions with the tau protein content. We conclude that the morphological heterogeneity of inclusions is accompanied by the heterogeneity of tau protein at the ultrastructural and biochemical levels. Filament-enriched fractions from regions with the predominant glial pathology displayed structural heterogeneity depending on whether they derived from regions enriched in coiled bodies or tufted astrocytes. Filaments from coiled bodies were 14-nm wide, smooth, and on average 40% thinner than those from tufted astrocytes. They resembled abnormal thin tubules (13–15-nm diameter) of coiled bodies described in immunoelectron microscopy studies by Arima et al (41). In comparison, filaments isolated from regions enriched in tufted astrocytes had uneven jagged surface and were as wide (22 nm) as PHFs in AD. Previous studies described either 15-nm-wide, straight filaments occasionally narrowing to 9 nm without evidence of periodicity (43), 15- to 20-nmwide, straight tubular structures immunoreactive for tau as seen in GFAP-positive cells (44), or 18–22-nm-wide, straight filaments of tubular appearance (45). Such filaments appear to share similarity with our jagged filaments. The fine ultrastructure of jagged filaments was uncertain. If jagged filaments were indeed made of 2 helically twisted filaments, as are those in AD, the twisting was haphazard and lacked regular periodicity or distinct crossovers. Furthermore, jagged filaments appeared to consist of multiple thin fibrils arranged parallel to the long axis, particularly prominent at splayed ends. A similar dissociation into fibrils has rarely been noted in PHFs from AD, but is seen more frequently in filaments from other disorders, e.g. CBD (46, 47). In studies of PSP filaments, cross-sectional views revealed the presence of 6 or more protofilaments (45). On the other hand, it is unlikely that jagged filaments resulted from simple pairing of straight filaments from coiled bodies. Jagged filaments were thinner than 2 straight filaments (e.g. 22 nm vs 28 nm) and they were not smooth. Since jagged filaments resembled neither PHFs of AD nor smooth straight filaments from coiled bodies, they were considered a unique third type of filament in PSP. They also differed from filaments found in other disorders involving glia, e.g. CBD, by displaying only minimal twisting and 22nm width rather than 29-nm width (46). The estimation of mass per unit length of filaments using scanning transmission electron microscopy will be necessary to determine the regional and disease-specific characteristics of PSP filaments. Both straight and jagged filaments resolved into 2 PHF-1 immunoreactive polypeptides migrating at 64 kDa ----- ULTRASTRUCTURE AND BIOCHEMISTRY OF PSP FILAMENTS 43 and 68 kDa. The double-band pattern was similar to that reported by Vermersch et al (27) and others (48) for most of PSP anatomical areas except for hippocampus and entorhinal cortex, where an additional 60 kDa band was found. The present results indicate that besides PHF-1positive tau species, PHF-1-negative polypeptides are also components of straight and jagged filaments. Tau antibodies against nonphosphorylated epitopes, e.g. Tau1 and Tau 46, revealed up to 4 additional bands, migrating between 45 kDa and 60 kDa. The total number of tau bands differed between cases and regions, but we failed to demonstrate any consistent difference in the tau content between smooth and jagged filaments. Moreover, our preliminary analysis showed that both filaments contained similar isoforms of tau, which expressed predominantly, but not exclusively, exon 10, with or without exon 2 and lacking exon 3 (case 1, not shown). We conclude that the diversity of tau at the PHF-1 epitope has no apparent effect on the ultrastructure of filaments. Studies of tau assembly in vitro using nonphosphorylated recombinant tau (49) and phosphorylated species of bovine brain tau (50) reached a similar conclusion based on the ultrastructural similarity between assembled and authentic PHFs. Further studies will determine whether phosphorylation plays only a secondary role in the formation of tau filaments in neurodegenerative disorders. It is still unclear what are the factors responsible for ultrastructural heterogeneity between smooth and jagged filaments. Since these filaments originated in different cell types, it is possible that the type of cell is one of the determining factors. Both oligodendrocytes and astrocytes contain tau protein (51, 52) and may differ in the metabolism of tau and/or in factors stimulating aggregation of tau in pathological conditions. In studies of tau assembly in vitro, stimulatory compounds (49, 53) as well as unknown factor (55) were able to determine ultrastructure of filaments. The variability in phosphorylation of tau among regions and cases was unexpected, especially considering the similarity of the clinical history and postmortem delay. The motor cortex from 2 cases (case 1 and case 3) was most diverse. It was rich in tufted astrocytes and displayed the absence of morphological or immunohistochemical differences. In spite of differential phosphorylation of tau at PHF-1, AT8, 12E8, AT100, and Tau-1 epitopes in these cases (Fig. 7), both contained similar jagged filaments (compare Fig. 4A with 4G). These results emphasize an apparent lack of correlation between the ultrastructure of jagged filaments and the phosphorylation state of tau. However, the underlying reason for differential phosphorylation of tau among cases and regions remains unclear. Similar observations of tau lacking certain phosphoepitopes were made in other disorders (32, 55), including Pick’s disease (50, 56, 57). For example, only 2 of 4 PHF-tau bands were found to be PHF-1-immunoreactive in Pick’s disease (50). In PSP, a weak immunoreactivity with phospho-dependent but not phospho-independent antibodies was also noted (19). This was attributed to normal tau in homogenates. In our studies, we examined isolated filament-enriched fractions devoid of a soluble pool of tau. Therefore, tau lacking certain phospho-epitopes cannot be attributed to contamination with normal tau protein. Moreover, such a contamination should be detected consistently in all PSP samples rather than in selected cases and regions only. On the other hand, Iwatsubo et al probed several tauopathies, including PSP, and concluded that the widely occurring tau-positive inclusions share common phosphorylation characteristics irrespective of underlying disease or cell type (58). Since the latter observations were based on immunostaining of tissue sections, examination of tau by immunoblotting is needed to more precisely define disease- and/or cell-dependent phosphorylation patterns. It is clear that, in addition to neurons, both astrocytes and oligodendrocytes express tau protein (51, 52). It is uncertain whether similar neuronal kinases and phosphatase are found in these glial cells. Further study of the differential phosphorylation of tau in the tauopathies is necessary. PHF-1 bound to a subset of tau polypeptides suggests the possibility that 2 discrete subpopulations of filaments were present. By immunogold labeling, however, all the filaments were PHF-1-positive, suggesting that only a single population of filaments was present and that both PHF-1-positive and PHF-1-negative polypeptides were inter-mixed in the same filament population. Tau-1 also showed differential binding to tau polypeptides by blotting. With Tau-1, however, we failed to detect any significant immunogold labeling, suggesting that this epitope, readily accessible by blotting, was inaccessible to labeling in filaments. We speculate that the inaccessibility may be due to intertwining of Tau-1-positive and Tau-1negative polypeptides in the same filament effectively blocking this epitope from labeling. Alternatively, the density of Tau-1-positive polypeptides is below the level of detection by immunogold labeling. These results underline the importance of Western blotting in determining the phosphorylation state of PHF-tau. In contrast to glial pathology, neuronal pathology resulted in twisted filaments. For example, filament-enriched fractions from regions involving neurons, e.g. perirhinal cortex or hippocampus, contained an abundance of twisted filaments. As in glia, some filaments splayed into distinct fibrillary components. It will be of interest to determine the mass of the individual fibrils by scanning transmission electron microscopy. The maximal and minimal widths of these filaments, as well as their regular 80-nm interval of twisting, indicate that ultrastructurally ----- 44 TAKAHASHI ET AL these filaments closely resemble PHFs of AD. They resemble PHFs of AD also by immunoblotting as demonstrated by the presence of 3 major tau bands of 60, 64, and 68 kDa, similar to those described for AD. We speculate that PHFs from PSP and AD have a common ultrastructural and biochemical character. Such similarity between PSP and AD has been previously reported (27, 47) and attributed to the specific tau pathology of neurons. As in AD, twisted PHF-like filaments of PSP were found in anatomical areas containing NFTs. Unlike AD, however, 2 types of NFTs with globose or flame-shape could be distinguished morphologically. We were unable to differentiate the tau content of these NFTs because of their co-localization. It is likely that the 2 types of NFTs contain similar tau, since filaments isolated from NFTenriched regions appear ultrastructurally homogenous. In conclusion, the results of our studies suggest that 3 kinds of tau inclusions present in PSP are distinct from each other in ultrastructure although some overlap exists. Neuronal and glial inclusions also differ in biochemical composition. These differences may depend on cell-typespecific metabolism of tau. ACKNOWLEDGMENTS The authors express special thanks to Dr. TV Le (Mayo Clinic, Jacksonville, FL) for his expertise in application of double immunohistochemistry and Dr. A. Hirano (Montefiore Medical Center, Bronx, NY) for his valuable suggestions in the course of our studies. Supported by the Society for Progressive Supranuclear Palsy and the Alzheimer’s Disease Association (to HKR). REFERENCES 1. Steele JC, Richardson JC, Olszewski J. Progressive supranuclear palsy. Arch Neurol 1964;10:333–59 2. Golbe LI, Davis PH, Schoenberg BS, Duvoisin RC. Prevalence and natural history of progressive supranuclear palsy. Neurology 1988; 38:1031–34 3. Goedert M, Spillantini MG, Davies SW. Filamentous nerve cell inclusions in neurodegenerative diseases. Curr Opin Neurobiol 1998;8:619–32 4. Dickson DW. Neuropathologic differentiation of progressive supranuclear palsy and corticobasal degeneration. J Neurol (Suppl 2) 1999;246:6–15 5. Dickson DW. Astrocytes in other neurodegenerative diseases. In: Schipper HM, ed. Astrocytes in brain aging and neurodegeneration. Austin: Landes RG Co, 1998:165–88 6. Matsusaka H, Ikeda K, Akiyama H, Arai T, Inoue M, Yagishita S. Astrocytic pathology in progressive supranuclear palsy: Significance for neuropathological diagnosis. Acta Neuropathol 1998;96: 248–52 7. Hauw J-J, Daniel SE, Dickson DW, et al. Preliminary NINDS neuropathologic criteria for Steele-Richardson-Olszewski syndrome (progressive supranuclear palsy). Neurology 1994;44:2015–19 8. Komori T, Arai N, Oda M, et al. Morphologic difference of neuropil threads in Alzheimer’s disease, corticobasal degeneration and progressive supranuclear palsy: A morphometric study. Neuroscience Lett 1997;233:89–92 9. Komori T, Arai N, Oda M, et al. Astrocytic plaques and tufts of abnormal fibers do not coexist in corticobasal degeneration and progressive supranuclear palsy. Acta Neuropathol 1998;96:401–8 10. Bergeron C, Pollanen MS, Weyer L, Lang AE. Cortical degeneration in progressive supranuclear palsy. A comparison with corticalbasal ganglionic degeneration. J Neuropathol Exp Neurol 1997;56: 726–34 11. Ikeda K, Akiyama H, Haga C, Tanno E, Tokuda T, Ikeda S. Thornshaped astrocytes: Possibly secondary induced tau-positive glial fibrillary tangles. Acta Neuropathol 1995;90:620–25 12. Yamada T, Calne DB, Akiyama H, McGeer EG, McGeer PL. Further observations on Tau-positive glia in the brains with progressive supranuclear palsy. Acta Neuropathol 1993;85:308–15 13. Feany MB, Dickson DW. Neurodegenerative disorders with extensive tau pathology: A comparative study and review. Ann Neurol 1996;40:139–48 14. Cervos-Navarro J, Schumacher K. Neurofibrillary pathology in progressive supranuclear palsy (PSP). J Neural Transm Suppl 1994;42: 153–64 15. Ksiezak-Reding H, Shafit-Zagardo B, Yen SH. Differential expression of exon10 and 11 in normal tau and tau associated with paired helical filaments. J Neurosci Res 1995;41:583–93 16. Murrell JR, Spillantini MG, Zolo P, et al. Tau gene mutation G389R causes a tauopathy with abundant Pick body-like inclusions and axonal deposits. J Neuropathol Exp Neurol 1999;58:1207–26 17. Chambers LB, Lee JM, Troncoso JC, Reich S, Muma NA. Overexpression of four-repeat tau mRNA isoforms in progressive supranuclear palsy but not in Alzheimer’s disease. Ann Neurol 1999;46: 325–32 18. Litvan I, Dickson DW, Buttner-Ennever JA, et al. Research goals in progressive supranuclear palsy. Mov Disord 2000;15:446–58 19. Sergeant N, Wattez A, Delacourte A. Neurofibrillary degeneration in progressive supranuclear palsy and corticobasal degeneration: Tau pathologies with exclusive ‘‘exon 10’’ isoforms. J Neurochem 1999;72:1243–49 20. Stanford PM, Halliday GM, Brooks WS, et al. Progressive supranuclear palsy pathology caused by a novel silent mutation in exon 10 of the tau gene: Expansion of the disease phenotype caused by tau gene mutations. Brain 2000;123:880–93 21. Conrad C, Andreadis A, Trojanowski JQ. Genetic evidence for the involvement of tau in progressive supranuclear palsy. Ann Neurol 1997;41:277–81 22. Feany MB, Ksiezak-Reding H, Liu WK, Vincent I, Yen SH, Dickson DW. Epitope expression and hyperphosphorylation of tau protein in corticobasal degeneration: Differentiation from progressive supranuclear palsy. Acta Neuropathol 1995;90:37–43 23. Hoffmann R, Lee VMY, Leight S, Varga I, Otvos JL. Unique Alzheimer’s disease paired helical filament specific epitopes involve double phosphorylation at specific sites. Biochemistry 1997;36: 8114–24 24. Tabaton M, Whitehouse PJ, Perry G, Davies P, Autilio-Gambetti L, Gambetti P. Alz 50 recognizes abnormal filaments in Alzheimer’s disease and progressive supranuclear palsy. Ann Neurol 1988;24: 407–13 25. Tellez-Nagel I, Wisniewski HM. Ultrastructure of neurofibrillary tangles in Steele-Richardson-Olszewski syndrome. Arch Neurol 1973;29:324–27 26. Takauchi S, Mizuhara T, Miyoshi K. Unusual paired helical filaments in progressive supranuclear palsy. Acta Neuropathol (Berl) 1983;59:225–28 27. Vermersch P, Robitaille Y, Bernier L, Wattez A, Gauvreau D, Delacourte A. Biochemical mapping of neurofibrillary degeneration in a case of progressive supranuclear palsy: Evidence for general cortical involvement. Acta Neuropathol 1994;87:572–77 ----- ULTRASTRUCTURE AND BIOCHEMISTRY OF PSP FILAMENTS 45 28. Jicha GA, Berenfeld B, Davies P. Sequence requirements for formation of conformational variants of tau similar to those found in Alzheimer’s disease. J Neurosci Res 1999;55:713–23 29. Kosik KS, Orecchio LD, Binder LI, Trojanowski JQ, Lee VM-Y, Lee G. Epitopes that span the tau molecule are shared with paired helical filaments. Neuron 1998;1:817–25 30. Goedert M, Jakes R, Vanmechelen E. Monoclonal antibody AT8 recognizes tau protein phosphorylated at both serine202 and threonine205. Neurosci Lett 1995;189:167–69 31. Seubert M, Mawal-Dewan R, Barbour R, et al. Detection of phosphorylated Ser262 in fetal tau, adult tau, and paired helical filament tau. J Biol Chem 1995;270:18917–22 32. Sergeant N, David JP, Lefranc D, Vermersch P, Wattez A, Delacourte A. Different distribution of phosphorylated tau protein in Alzheimer’s and Pick’s diseases. FEBS Lett 1997;412:578–82 33. Otvos L, Feiner L, Lang E, Szendrei GI, Goedert M, Lee VMY. Monoclonal antibody PHF-1 recognizes tau protein phosphorylated at serine residues 396–404. J Neurosci 1994;39:669–73 34. Carmel G, Mager EM, Binder LI, Kuret J. The structural basis of monoclonal antibody Alz50‘s selectivity for Alzheimer’s disease pathology. J Biol Chem 1996;271:32789–95 35. Ishizawa K, Ksiezak-Reding H, Davis P, Delacourte A, Tiseo P, Yen SH, Dickson DW. A double labeling immunohistochemical study of tau exon 10 in Alzheimer’s disease, progressive supranuclear palsy and Pick’s disease. Acta Neuropathol 2000;100:235–44 36. Ksiezak-Reding H, Liu W-K, Wall JS, Yen S-H. Biochemical isolation and characterization of paired helical filaments. In: Avila J, Brandt R, Kosik KS, eds. Brain microtubule associated proteins; modifications in disease. Amsterdam: Harwood Academic Publishers, 1997:331–50 37. Yang LS, Ksiezak-Reding H. Ubiquitin immunoreactivity of paired helical filaments differs in Alzheimer’s disease and corticobasal degeneration. Acta Neuropathol 1998;96:520–26 38. Hauw JJ, Verny M, Delaere P, Cervera P, He Y, Duyckaerts C. Constant neurofibrillary changes in the neocortex in progressive supranuclear palsy. Basic differences with Alzheimer’s disease and aging. Neurosci Lett 1990;119:182–86 39. Hof PR, Delacourte A, Bouras C. Distribution of cortical neurofibrillary tangles in progressive supranuclear palsy. Acta Neuropathol 1992;84:45–51 40. Feany MB, Mattiace LA, Dickson DW. Neuropathologic overlap of progressive supranuclear palsy, Pick’s disease and corticobasal degeneration. J Neuropathol Exp Neurol 1996;55:53–67 41. Arima K, Nakamura M, Sunohara N, et al. Ultrastructural characterization of the tau-immunoreactive tubule in the oligodendroglial perikarya and their inner loop processes in progressive supranuclear palsy. Acta Neuropathol 1997;93:558–66 42. Grundke-Iqbal I, Iqbal K, Tung YC, Quinlan M, Wisniewski HM, Binder LI. Abnormal phosphorylation of the microtubule associated protein tau in Alzheimer cytoskeletal pathology. Proc Natl Acad Sci USA 1986;83:4913–17 43. Bugiani O, Mancardi GL, Brusa A, Ederli A. The fine structure of subcortical neurofibrillary tangles in progressive supranuclear palsy. Acta Neuropathol 1979;45:147–52 44. Abe H, Yagishita S, Amano N, Bise K. Ultrastructural and immunochemical study of ‘‘astrocytic tangles’’ (ACT) in patients with progressive supranuclear palsy [abstract]. Clin Neuropathol 1992; 11:278 45. Montpetit V, Clapin DF, Guberman A. Substructure of 20 nm filaments of progressive supranuclear palsy. Acta Neuropathol 1985; 68:311–18 46. Ksiezak-Reding H, Tracz E, Yang LS, Dickson DW, Martha S, Wall JS. Ultrastructural instability of paired helical filaments from corticobasal degeneration as examined by scanning transmission electron microscopy. Am J Pathol 1996;149:639–51 47. Tracz E, Dickson DW, Hainfeld JF, Ksiezak-Reding H. Paired helical filaments in corticobasal degeneration: The fine fibrillary structure with Nano Van. Brain Res 1997;773:33–44 48. Schmidt ML, Huang R, Martin JA, et al. Neurofibrillary tangles in progressive supranuclear palsy contain the same tau epitopes identified in Alzheimer’s disease PHFtau. J Neuropathol Exp Neurol 1996;55:534–39 49. Goedert M, Jakes R, Spillantini MG, Hasegawa M, Smith MJ, Crowther RA. Assembly of microtubule-associated protein tau into Alzheimer-like filaments induced by sulphated glycosaminoglycans. Nature 1996;383:550–53 50. King ME, Ghoshal N, Wall JS, Binder LI, Ksiezak-Reding H. Structural analysis of Pick’s disease derived and in vitro-assembled tau filaments. Am J Pathol 2001;158:1481–90 51. Ksiezak-Reding H, He D, Gordon-Krajcer W, Kress Y, Lee S, Dickson DW. Induction of Alzheimer-specific tau epitope AT100 in apoptotic human fetal astrocytes. Cell Motil Cytoskeleton 2000;47: 236–52 52. LoPresti P, Szuchet S, Papasozomenos SC, Zinkowski RP, Binder LI. Functional implications for the microtubule-associated protein tau: Localization in oligodendrocytes. Proc Natl Acad Sci USA 1995;92:10369–73 53. Arrasate M, Perez M, Valpuesta JM, Avila J. Role of glycosaminoglycans in determining the helicity of paired helical filaments. Am J Pathol 1997;151:1115–22 54. Ksiezak-Reding H, Yang G, Simon M, Wall JS. Assembled tau filaments differ from native paired helical filaments as determined by scanning transmission electron microscopy. Brain Res 1998;814: 86–98 55. Mailliot C, Sergeant N, Bussiere T, Caillet-Boudin ML, Delacourte A, Bouee L. Phosphorylation of specific sets of tau isoforms reflects different neurofibrillary degeneration processes. FEBS Lett 1998; 433:201–4 56. Rizzini C, Goedert M, Hodges JR, et al. Tau gene mutation K257T causes a tauopathy similar to Pick’s disease. J Neuropathol Exp Neurol 2000;59:990–1001 57. Probst A, Tolnay M, Langui D, Goedert M, Spillantini MG. Pick’s disease: Hyperphosphorylated tau protein segregates to the somatoaxonal compartment. Acta Neuropathol 1996;92:588–96 58. Iwatsubo T, Hasegawa M, Ihara Y. Neuronal and glial tau-positive inclusions in diverse neurologic diseases share common phosphorylation characteristics. Acta Neuropathol 1994;88:129–36 Received February 28, 2001 Revision received June 18, 2001 and September 6, 2001 Accepted September 11, 2001 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1093/JNEN/61.1.33?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1093/JNEN/61.1.33, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://academic.oup.com/jnen/article-pdf/61/1/33/9552913/61-1-33.pdf" }
2,002
[ "JournalArticle" ]
true
2002-01-01T00:00:00
[]
13,980
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/014ad6b5c003ecf7566570266d43154ac8683758
[ "Computer Science" ]
0.820874
Convolutional Transformer based Dual Discriminator Generative Adversarial Networks for Video Anomaly Detection
014ad6b5c003ecf7566570266d43154ac8683758
ACM Multimedia
[ { "authorId": "4738891", "name": "Xinyang Feng" }, { "authorId": "2451800", "name": "Dongjin Song" }, { "authorId": "2786131", "name": "Yuncong Chen" }, { "authorId": "1766853", "name": "Zhengzhang Chen" }, { "authorId": "2090567", "name": "Jingchao Ni" }, { "authorId": "2145225543", "name": "Haifeng Chen" } ]
{ "alternate_issns": null, "alternate_names": [ "MM" ], "alternate_urls": null, "id": "f2c85de5-7cfa-4b92-8714-a0fbdcf0274e", "issn": null, "name": "ACM Multimedia", "type": "conference", "url": null }
Detecting abnormal activities in real-world surveillance videos is an important yet challenging task as the prior knowledge about video anomalies is usually limited or unavailable. Despite that many approaches have been developed to resolve this problem, few of them can capture the normal spatio-temporal patterns effectively and efficiently. Moreover, existing works seldom explicitly consider the local consistency at frame level and global coherence of temporal dynamics in video sequences. To this end, we propose Convolutional Transformer based Dual Discriminator Generative Adversarial Networks (CT-D2GAN) to perform unsupervised video anomaly detection. Specifically, we first present a convolutional transformer to perform future frame prediction. It contains three key components, i.e., a convolutional encoder to capture the spatial information of the input video clips, a temporal self-attention module to encode the temporal dynamics, and a convolutional decoder to integrate spatio-temporal features and predict the future frame. Next, a dual discriminator based adversarial training procedure, which jointly considers an image discriminator that can maintain the local consistency at frame-level and a video discriminator that can enforce the global coherence of temporal dynamics, is employed to enhance the future frame prediction. Finally, the prediction error is used to identify abnormal video frames. Thoroughly empirical studies on three public video anomaly detection datasets, i.e., UCSD Ped2, CUHK Avenue, and Shanghai Tech Campus, demonstrate the effectiveness of the proposed adversarial spatio-temporal modeling framework.
## Convolutional Transformer based Dual Discriminator Generative Adversarial Networks for Video Anomaly Detection ### Xinyang Feng ##### Columbia University New York, New York, USA xf2143@columbia.edu ### Dongjin Song[∗] ##### University of Connecticut Storrs, Connecticut, USA dongjin.song@uconn.edu ### Zhengzhang Chen ##### NEC Laboratories America, Inc. Princeton, New Jersey, USA zchen@nec-labs.com #### ABSTRACT ### Jingchao Ni ##### NEC Laboratories America, Inc. Princeton, New Jersey, USA jni@nec-labs.com #### KEYWORDS ### Yuncong Chen ##### NEC Laboratories America, Inc. Princeton, New Jersey, USA yuncong@nec-labs.com ### Haifeng Chen ##### NEC Laboratories America, Inc. Princeton, New Jersey, USA haifeng@nec-labs.com Detecting abnormal activities in real-world surveillance videos is an important yet challenging task as the prior knowledge about video anomalies is usually limited or unavailable. Despite that many approaches have been developed to resolve this problem, few of them can capture the normal spatio-temporal patterns effectively and efficiently. Moreover, existing works seldom explicitly consider the local consistency at frame level and global coherence of temporal dynamics in video sequences. To this end, we propose Convolutional Transformer based Dual Discriminator Generative Adversarial Networks (CT-D2GAN) to perform unsupervised video anomaly detection. Specifically, we first present a convolutional transformer to perform future frame prediction. It contains three key components, i.e., a convolutional encoder to capture the spatial information of the input video clips, a temporal self-attention module to encode the temporal dynamics, and a convolutional decoder to integrate spatio-temporal features and predict the future frame. Next, a dual discriminator based adversarial training procedure, which jointly considers an image discriminator that can maintain the local consistency at frame-level and a video discriminator that can enforce the global coherence of temporal dynamics, is employed to enhance the future frame prediction. Finally, the prediction error is used to identify abnormal video frames. Thoroughly empirical studies on three public video anomaly detection datasets, i.e., UCSD Ped2, CUHK Avenue, and Shanghai Tech Campus, demonstrate the effectiveness of the proposed adversarial spatio-temporal modeling framework. #### CCS CONCEPTS Video anomaly detection; Generative adversarial networks; Transformer model; Convolutional neural network; Spatio-temporal modeling **ACM Reference Format:** Xinyang Feng, Dongjin Song, Yuncong Chen, Zhengzhang Chen, Jingchao Ni, and Haifeng Chen. 2021. Convolutional Transformer based Dual Discriminator Generative Adversarial Networks for Video Anomaly Detection. In Proceedings of the 29th ACM International Conference on Multimedia (MM _’21), October 20–24, 2021, Virtual Event, China. ACM, New York, NY, USA,_ [9 pages. https://doi.org/10.1145/3474085.3475693](https://doi.org/10.1145/3474085.3475693) #### 1 INTRODUCTION - Computing methodologies → **Scene anomaly detection;** **Adversarial learning; Anomaly detection; Neural networks.** ∗Corresponding author Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. _MM ’21, October 20–24, 2021, Virtual Event, China_ © 2021 Association for Computing Machinery. ACM ISBN 978-1-4503-8651-7/21/10...$15.00 [https://doi.org/10.1145/3474085.3475693](https://doi.org/10.1145/3474085.3475693) With the rapid growth of video surveillance data, there is an increasing demand to automatically detect abnormal video sequences in the context of large-scale normal (regular) video data. Despite a substantial amount of research effort has been devoted to this problem [3, 8, 13, 14, 16, 19, 22, 31, 34], video anomaly detection, which aims to identify the activities that do not conform to regular patterns in a video sequence, is still a challenging task. This is because real-world abnormal video activities can be extremely diverse while the prior knowledge about these anomalies is usually limited or even unavailable. With the assumption that a model can only generalize to data from the same distribution as the training set, abnormal activities in the test set will manifest as deviance from regular patterns. A common approach to resolve this problem is to learn a model that can capture regular patterns in the normal video clips during the training stage, and check whether there exists any irregular pattern that diverges from regular patterns in the test video clips. Within this framework, it is crucial to not only represent the regular appearances but also capture the normal spatio-temporal dynamics to differentiate abnormal activities from normal activities in a video sequence. This serves as an important motivation for our proposed methods. Early studies have used handcrafted features to represent video patterns [13, 16, 19, 29]. For instance, Li et al. [13] introduced mixtures of dynamic textures and defined outliers under this model as anomalies. These approaches, however, are usually not optimal for video anomaly detection since the features are extracted based upon a different objective. ----- Recently, deep neural networks are becoming prevalent in video anomaly detection, showing superior performance over handcrafted feature based methods. For instance, Hasan et al. [8] developed a convolutional autoencoder (Conv-AE) to model the spatio-temporal patterns in a video sequence simultaneously with a 2D CNN. The temporal dynamics, however, are not explicitly considered. To better cope with the spatio-temporal information in a video sequence, convolutional long short-term memory (LSTM) autoencoder (ConvLSTM-AE) [17, 27] was proposed to model the spatial patterns with fully convolutional networks and encode the temporal dynamics using convolutional LSTM (ConvLSTM). ConvLSTM, however, suffers from computational and interpretation issues. A powerful alternative for sequence modeling is the self-attention mechanism [33]. It has demonstrated superior performance and efficiency in many different tasks, e.g., sequence-to-sequence machine translation [33], time series prediction [24], autoregressive model based image generation [23], and GAN-based image synthesis [39]. However, it has seldom been employed to capture regular spatio-temporal patterns in the surveillance videos. More recently, adversarial learning has shown impressive progress on video anomaly detection. For instance, Ravanbakhsh et al. [25] developed a GAN based anomaly detection approach following conditional GAN framework [10]. Liu et al. [14] proposed an anomaly detection approach based on future frame prediction. Tang et al. [31] extended this framework by adding a reconstruction task. The generative models in these two works were based on U-Net [26]. Similar to Conv-AE, the temporal dynamics in the video clip were not explicitly encoded and the temporal coherence was enforced by a loss term on the optical flow. Moreover, the potential discriminative information in the form of consistency at frame-level and global coherence of temporal dynamics in video sequences were not fully considered in previous works. In this paper, to better capture the regular spatio-temporal patterns and cope with the potential discriminative information at frame-level and in video sequences, we propose Convolutional Transformer based Dual Discriminator Generative Adversarial Networks (CT-D2GAN) to perform unsupervised video anomaly detection. We first present a convolutional transformer to perform future frame prediction. The convolutional transformer is essentially a encoder-decoder framework consisting of three key components, _i.e., a convolutional encoder to capture the spatial patterns of the_ input video clip, a novel temporal self-attention module adapted for video temporal modeling that can explicitly encode the temporal dynamics, and a convolutional decoder to integrate spatio-temporal features and predict the future frame. Because of the temporal self-attention module, convolutional transformer can capture the underlying temporal dynamics efficiently and effectively. Next, in order to maintain the local consistency of the predicted frame and the global coherence conditioned on the previous frames, we adapt dual discriminator GAN to deal with video frames and employ an adversarial training procedure to further enhance the prediction performance. Finally, the prediction error is adopted to identify abnormal video frames. Thoroughly empirical studies on three public video anomaly detection datasets, i.e., UCSD Ped2, CUHK Avenue, and Shanghai Tech Campus, demonstrate the effectiveness of the proposed framework and techniques. #### 2 RELATED WORK The proposed Convolutional Transformer based Dual Discriminator Generative Adversarial Networks (CT-D2GAN) is closely related to deep learning based video anomaly detection and self-attention mechanism [33]. Note that we focus our discussions on methods based on unsupervised settings, which are efficient in generalization without the time-consuming and error-prone process of manual labeling. We are aware that there are numerous works on weakly supervised or supervised video anomaly detection, e.g., Sultani et al. (2018) proposed a deep multiple instance ranking framework using videolevel labels and achieves better performance than convolutional auto-encoder (Conv-AE) based method [8], but it employs both normal and abnormal video clips for training which is different from our setting. Deep neural networks based video anomaly detection methods demonstrate superior performance over traditional methods based on handcrafted features. Hasan et al. (2016) developed Conv-AE method to simultaneously learn the spatio-temporal patterns in a video with 2D convolutional neural networks by concatenating the video frames in the channel dimension. The temporal information is mixed with the spatial information in the first convolutional layer, thus not explicitly encoded. Xu et al. (2017) proposed appearance and motion DeepNet (AMDN) to learn video feature representations, which however still requires a decoupled one-class SVM classifier applied on learned representation to generate anomaly score. Dong et al. (2019) proposed a memory-augmented autoencoder (MemAE) that uses a memory module to constrain the reconstruction. More recently, adversarial learning has demonstrated flexibility and impressive performance in multiple video anomaly detection studies. A generative adversarial networks (GANs) based anomaly detection approach [25] was developed following cGAN framework of image-to-image translation [10]. Specifically, it employs image and optical flow as source domain and target domain, and vice versa, and trains cross-channel generation through adversarial learning. The reconstruction error is used to compute anomaly score. The only temporal constraint is imposed by the optical flow calculation. Liu et al. (2018) proposed an anomaly detection approach based on future frame prediction in GAN framework and U-Net [26]. Similar to Conv-AE, the temporal information is not explicitly encoded and the temporal coherence between neighboring frames is enforced by a loss term on the optical flow. Tang et al. (2020) extended the future frame prediction framework by adding a reconstruction task. One way to alleviate the temporal encoding issue in video spatio-temporal modeling is to use convolutional LSTM autoencoder (ConvLSTM-AE) based methods [4, 17, 27, 38], where the spatial and temporal patterns are encoded with fully convolutional networks and convolutional LSTM, respectively. Despite its popularity, ConvLSTM suffers from issues such as large memory consumption. The complex gating operations add to the computational cost and complicate the information flow, making interpretation difficult. A more effective and efficient alternative for sequence modeling is the self-attention mechanism [33], which is essentially an attention mechanism relating different positions of a single sequence to compute a representation of the sequence, in which the keys, values, ----- and queries are from the same set of features. Some related applications include autoregressive model based image generation [23], GAN-based image synthesis [39]. In this work, based on related works, we introduce the convolutional transformer by extending the self-attention mechanism to video sequence modeling and develop a novel self-attention module specialized for spatio-temporal modeling in video sequences. Compared to existing approaches for video anomaly detection, the proposed convolutional transformer model has the advantage of being able to explicitly and efficiently encode the temporal information in a sequence of feature maps, where the computation of attentions can be fully parallelized via matrix multiplications. Based on the convolutional transformer, a dual discriminator generative adversarial networks (D2GAN) approach is developed to further enhance the future frame prediction through enforcing local consistency of the predicted frame and the global coherence conditioned on the previous frames. Note that the proposed D2GAN differs from existing works on dual discriminator based GAN which have been applied to different scenarios [5, 21, 35, 37]. #### 3 CT-D2GAN In this section, we first introduce the problem formulation and input to our framework. Then, we present the motivation and technical details of the proposed CT-D2GAN framework including convolutional transformer, dual discriminator GAN, the overall loss function, and lastly the regularity score calculation. An overview of the framework is illustrated in Figure 1. In CT-D2GAN, a convolutional transformer is employed to generate future frame prediction based on past frames, an image discriminator and a video discriminator are used to maintain the local consistency and global coherence. #### 3.1 Problem Statement Given an input representation of video clip of length 𝑇, i.e., 𝐼 = (𝐼𝑡 −𝑇 +1, ..., 𝐼𝑡 ) ∈ R[ℎ][×][𝑤][×][𝑐][×][𝑇], where ℎ, 𝑤, 𝑐 are the height, width and number of channels, we aim to predict the (𝑡 + 1)-th frame as _𝐼[^]𝑡_ +1 ∈ R[ℎ][×][𝑤][×][𝑐] and identify abnormal activities based upon the prediction error, i.e., 𝑒MSE,𝑡 = _ℎ_ - 1 - �𝑐𝑖=1 [∥][𝐼][^][:][,][:][,𝑖,𝑡] [+][1][ −] _[𝐼][:][,][:][,𝑖,𝑡]_ [+][1] [∥]𝐹[2] [,] where 𝐼:,:,𝑖,𝑡 +1 ∈ R[ℎ][×][𝑤]. #### 3.2 Input As appearance and motion are two characteristics of video data, it is common to explicitly incorporate optical flow together with the still images to describe a video sequence [28], e.g. optical flow has been employed to represent video sequences in the cGAN framework [25] and used as a motion constraint [14]. In this work, we stack image with pre-computed optical flow maps [2, 9] in channel dimension as inputs similar to Simonyan et al. [28] for video action recognition and Ravanbakhsh et al. [25] for video anomaly detection. The optical flow maps consist of a horizontal component, a vertical component and a magnitude component. To be noted that, the optical flow map is computed from the previous image and current image, thus does not contain future frame information. Therefore, the input can be given as 𝐼 ∈ R[ℎ][×][𝑤][×][4][×][𝑇], and we used 5 consecutive frames as inputs, i.e., 𝑇 = 5, similar to Liu et al. [14]. #### 3.3 Convolutional Transformer Convolutional transformer is developed to obtain a future frame prediction based on past frames. It consists of three key components: a convolutional encoder to encode spatial information, a temporal self-attention module to capture the temporal dynamics, and a convolutional decoder to integrate spatio-temporal features and predict future frame. _3.3.1_ _Convolutional Encoder. The convolutional encoder [15] is_ employed to extract spatial features from each frame of the video. Each frame of the video is first resized to 256 × 256 and then fed into the convolutional encoder. The convolutional encoder consists of 5 convolutional blocks. And the convolutional block follows common structure in CNN. All the convolutional kernels are set as 3 × 3 pixels. For brevity, we denote a convolutional layer with stride 𝑠 and number of filters 𝑛 as Conv𝑠,𝑛, a batch normalization layer as BN, a scaled exponential linear unit [12] as SELU, and a dropout operation with dropout ratio 𝑟 as dropout𝑟 . The structure of the convolutional encoder is: [Conv1,64-SELUBN]-[Conv2,64-SELU-BN-Conv1,64-SELU]-[Conv2,128-SELU-BNConv1,128-SELU]-[Conv2,256-SELU-BN-dropout0.25-Conv1,256SELU]-[Conv2,256-SELU-BN-dropout0.25-Conv1,256-SELU], where each [ ] represents a convolutional block. At the 𝑙-th convolutional block conv[𝑙], 𝐹𝑡[𝑙]−𝑖 [∈] [R][ℎ][𝑙] [×][𝑤][𝑙] [×][𝑐][𝑙] _[,𝑖]_ [∈] [0, ...,𝑇 − 1] denotes the input feature maps to the self-attention module with ℎ𝑙, 𝑤𝑙, 𝑐𝑙 as the height, width, and number of channels, respectively. The temporal dynamics among the spatial feature maps of different time steps will be encoded with temporal selfattention module. _3.3.2_ _Temporal Self-attention Module. To explicitly encode the tem-_ poral information in the video sequence, we extend self-attention mechanism in the transformer model [33] and develop a novel temporal self-attention module to capture the temporal dynamics of the multi-scale spatial feature maps generated from the convolutional encoder. This section applies to all layers, thus we omit the layer for clarity. An illustration of the multi-head temporal selfattention module is shown in the upper panel of Figure 1. Spatial **Feature Vector. We first use global average pooling (GAP) to ex-** tract a feature vector f𝑡 from the feature map 𝐹𝑡 extracted in the convolutional encoder. The feature vector in current time step f𝑡 will be used as part of the query and each historical feature vector f𝑡 −𝑖, 𝑖 ∈[1,𝑇 − 1] will be used as part of the key to index spatial feature maps. **Positional Encoding. Different from sequence models such as** LSTM, self-attention does not model sequential information inherently, therefore it is necessary to incorporate temporal positional information into the model. We generate a positional encoding vector PE ∈ R[𝑑][𝑝] following [33]: PE𝑝,2𝑖 = sin(𝑝/10000[2][𝑖][/][𝑑][𝑝] ) (1) PE𝑝,2𝑖+1 = cos(𝑝/10000[2][𝑖][/][𝑑][𝑝] ) where 𝑑𝑝 denotes the dimension of PE, 𝑝 denotes the temporal position and 𝑖 ∈[0, ..., (𝑑𝑝 /2−1)] denotes the index of the dimension. Empirically, we fix 𝑑𝑝 = 8 in our study. **Temporal Self-Attention. We concatenate the positional encod-** ing vector with the spatial feature vector for each time step and ----- |Col1|,)(*) …,)(-)| |---|---| |Head-h|| |Discriminator 2D Conv ~ t+1 t+1 Real or Fake|~ Discriminator 3D Conv 1, 2, … t, t+1 Real or Fake 1, 2, … t, t+1| |---|---| () **_1, 2,_** **_…_** **_…_** **_t,_** **_…_** **_t,_** #### ~ **_t+1_** **_t,_** **_t+1_** #### ~ **_t,_** **_t+1_** **Figure 1: The architecture of the proposed CT-D2GAN framework. (Upper panel) The convolutional transformer generator is** **consisted of a convolutional encoder, a temporal self-attention module, and a convolutional decoder. Multi-head self-attention** **is applied on the feature maps 𝐹𝑡** **extracted from convolutional encoder: 𝐹𝑡** **is transformed to multi-head feature maps 𝐹𝑡[(][k][)]** **via** **a convolutional operation; within each head, we apply a global average pooling (GAP) operation on 𝐹𝑡[(][𝑘][)]** **to generate a spatial** **feature vector by aggregating over spatial dimension, and concatenate the positional encoding (PE) vector; we then compare** **the similarity 𝐷cos between query q𝑡[(][𝑘][)]** **and memory m𝑡[(][𝑘][)]** **feature vectors and generate the attention weights by normalizing** **across time steps using softmax 𝜎; the attended feature map 𝐻𝑡[(][ℎ][)]** **is a weighted average of the feature maps at different time** **steps; the final attended map 𝐻𝑡[MH]** **is the concatenation over all the heads; the final integrated map 𝑆𝑡** **is a weighted average of** **the query 𝐹𝑡[MH]** **and the attended feature maps according to a spatial selective gate (SSG). 𝑆𝑡** **is decoded to the predicted future** **frame with the convolutional decoder. (Lower panels) The image discriminator (left) and video discriminator (right) used in** **our dual discriminator GAN framework.** **_…_** **_t,_** **_t+1_** **_1, 2,_** **_…_** **_1, 2,_** **_…_** use the concatenated vectors as the queries and keys, and the feature maps as the values in the setting of self-attention mechanism. For each query frame at time 𝑡, the current concatenated feature vector q𝑡 = [f𝑡 ; PE] ∈ R[𝑐][𝑙] [+][𝑑][𝑝] is used as query, and compared to the feature vector of each frame from the input video clip i.e. memory m𝑡 −𝑖 = [f𝑡 −𝑖 ; PE] ∈ R[𝑐][𝑙] [+][𝑑][𝑝] _,𝑖_ ∈[1, ...,𝑇 − 1] using cosine similarity: _𝐷_ (q𝑡 _, m𝑡_ −𝑖 ) = ∥qq𝑡𝑡∥∥· mm𝑡𝑡−−𝑖𝑖 ∥ _[.]_ (2) Based on the similarity between q𝑡 and m𝑡 −𝑖, we can generate the normalized attention weights 𝑎𝑡,𝑖 ∈ R across the temporal dimension using a softmax function: _𝑎𝑡,𝑡_ −𝑖 = � exp(𝛽𝐷 (q𝑡 _, m𝑡_ −𝑖 )) (3) _𝑗_ ∈[1...𝑇 −1] [exp][(][𝛽𝐷] [(][q]𝑡 _[,][ m]𝑡_ −𝑗 [))][,] where a positive temperature variable 𝛽 is introduced to sharpen the level of focus in the softmax function and is automatically learned **_1, 2,_** **_…_** **_t+1_** in the model through a single hidden densely-connected layer with the query as the input. The final attended feature maps 𝐻𝑡 are a weighted sum of all feature maps 𝐹 using the attention weights in Eq. (3): ∑︁ _𝐻𝑡_ = _𝑎𝑡,𝑡_ −𝑖 - 𝐹𝑡 −𝑖 _._ (4) _𝑖_ ∈[1,...,𝑇 −1] **Multi-head** **Temporal** **Self-Attention.** Multi-head selfattention [33] enables the model to jointly attend to information from different representation subspaces at different positions. We adapt it to spatio-temporal modeling by first mapping the spatial feature maps to 𝑛ℎ = 8 groups, each using 32 1 × 1 convolutional kernels. For each group of feature maps with dimension 𝑐ℎ = 32, we then perform the single head self-attention as described in the previous subsection and generate attended feature maps for head 𝑘 ----- as 𝐻𝑡[(][𝑘][)] : ∑︁ _𝐻𝑡[(][𝑘][)]_ = _𝑎𝑡,𝑡[(][𝑘][)]−𝑖_ [·][ 𝐹]𝑡[(]−[𝑘]𝑖[)][,] (5) _𝑖_ ∈[1,...,𝑇 −1] where 𝐹𝑡[(]−[𝑘]𝑖[)] [∈] [R][ℎ][𝑙] [×][𝑤][𝑙] [×][𝑐][ℎ] [is the transformed feature map at frame] _𝑡_ − _𝑖_ for head 𝑘, 𝑎𝑡,𝑡[(][𝑘][)]−𝑖 [is the corresponding attention weight. The] final multi-head attended feature map 𝐻𝑡[MH] ∈ R[ℎ][𝑙] [×][𝑤][𝑙] [×(][𝑐][ℎ] [·][𝑛][ℎ][)] is the concatenation of the attended feature maps from all the heads along the channel dimension: _𝐻𝑡[MH]_ = Concat(𝐻𝑡[(][1][)], ..., 𝐻𝑡[(][𝑛][ℎ][)] ). (6) In this way, the final attended feature maps not only integrate spatial information from the convolutional encoder, but also capture temporal information from multi-head temporal self-attention mechanism. **Spatial Selective Gate. The aforementioned module extends the** self-attention mechanism to the temporal modeling of 2D image feature maps, however, it comes with the loss of fine-grained spatial resolution due to the GAP operation. To compensate this, we introduce spatial selective gate (SSG), which is a spatial attention mechanism to integrate the current and historical information. The attended feature maps from the temporal self-attention module and the feature maps of the current query are concatenated, on which we learn a spatial selective gate using a sub-network NSSG with structure: Conv1,256-BN-SELU-Conv1,256-BN-SELUConv1,256-BN-SELU-Conv1,256-Conv1,256-Sigmoid. The final output is a pixel-wise weighted average of the attended maps 𝐻𝑡[MH] and the current query’s multi-head transformed feature maps _𝐹𝑡[MH]_ ∈ R[ℎ][𝑙] [×][𝑤][𝑙] [×(][𝑐][ℎ] [·][𝑛][ℎ][)], according to 𝑆𝑆𝐺: _𝑆𝑡_ = 𝑆𝑆𝐺 ◦ _𝐹𝑡[MH]_ + (1 − _𝑆𝑆𝐺) ◦_ _𝐻𝑡[MH]_ (7) where denotes element-wise multiplication. ◦ We add SSG at each level of temporal self-attention module. As the spatial dimensions are larger at shallow layers and we want to include contextual information while preserving the spatial resolution, we use dilated convolution [36] with different dilatation factors at the 4 convolutional blocks in the sub-network NSSG, specifically from conv[2] to conv[5], the dilation factors are (1,2,4,1), (1,2,2,1), (1,1,2,1), (1,1,1,1). Note that SSG is computationally more efficient than directly forwarding the concatenated feature maps to the convolutional decoder. _3.3.3_ _Convolutional Decoder. The outputs of the temporal self-_ attention module 𝑆𝑡 are fed into the convolutional decoder. The convolutional decoder predicts the video frame using 4 transposed convolutional layers with stride 2 on the feature maps in a reverse order of the convolutional encoder. The fully-scaled feature maps then go through one convolutional layer with 32 filters and one convolutional layer with 𝑐 filters of size 1 × 1 that maps to the same size of channels 𝑐 in the input. In order to predict finer details, we utilize the skip connection [26] to connect the spatio-temporally integrated maps at each level of the convolutional encoder to the corresponding level of the convolutional decoder, which allows the model to further fine-tune the predicted frames. #### 3.4 Dual Discriminator GAN We propose a dual discriminator GAN using both an image discriminator and a video discriminator to further enhance the future frame prediction of convolutional transformer via adversarial training. The image discriminator 𝐷𝐼 critiques on whether the current frame is generated or real just on the basis of one single frame to assess the local consistency. The video discriminator 𝐷𝑉 performs critique on the prediction conditioned on the past frames to assess the global coherence. Specifically, we stack the past frames with current generated or real frame in the temporal dimension and the video discriminator is essentially a video classifier. This idea of combining local and global (contextual) discriminator is similar to adversarial image inpainting [37] but is used in a totally different context. The network structures of the two discriminators are kept the same except that we use 2D operations in image discriminator and the corresponding 3D operations in the video discriminator. We use PatchGAN architecture as described in [10] and use spectral normalization [20] in each convolutional layer. In the 3D version, the stride and kernel size in the temporal dimension were set at 1 and 2 respectively. The method in Liu et al. [14] is similar to using the image discriminator only. Different from the video discriminator in Tulyakov et al. [32], which applies on the whole synthetic video clip, our proposed video discriminator conditions on the real frames. #### 3.5 Loss For the adversarial training, we use the Wasserstein GAN with gradient penalty (WGAN-GP) setting [1, 7]. The generator 𝐺 is the mapping : 𝐼 → [�]𝐼𝑡 +1. For discriminators, 𝐷𝑉 : (𝐼, _𝐼[^]𝑡_ +1) → _𝑝_ [(𝐼, _𝐼[^]𝑡_ +1) is real] and 𝐷𝐼 : 𝐼[^]𝑡 +1 → _𝑝_ [𝐼[^]𝑡 +1 is real] are video and image discriminators respectively. The GAN loss is: _𝐿𝑎𝑑𝑣_ (𝐺, 𝐷𝐼 _, 𝐷𝑉_ ) = E𝐼,�𝐼𝑡 +1 [𝐷𝑉 (𝐼, [�]𝐼𝑡 +1)] − E𝐼,𝐼𝑡 +1 [𝐷𝑉 (𝐼, 𝐼𝑡 +1)] + 𝜆E𝐼,𝐼^𝑡 +1 [(∥∇𝐷𝑉 (𝐼, _𝐼[^]𝑡_ +1)∥2 − 1)[2]] (8) + E�𝐼𝑡 +1 [𝐷𝐼 ([�]𝐼𝑡 +1)] − E𝐼𝑡 +1 [𝐷𝐼 (𝐼𝑡 +1)] + 𝜆E𝐼^𝑡 +1 [(∥∇𝐷𝐼 (𝐼[^]𝑡 +1)∥2 − 1)[2]] where _𝐼[^]𝑡_ +1 = 𝜖𝐼𝑡 +1 +(1−𝜖)[�]𝐼𝑡 +1, 𝜖 ∼ _𝑈_ [0, 1]. The penalty coefficient _𝜆_ is fixed as 10 in all our experiments. In addition, we consider the pixel-wise 𝐿1 loss of the prediction. Therefore the total loss 𝐿 is: _𝐿_ = 𝐿𝑎𝑑𝑣 + ∥𝐼𝑡 +1 − [�]𝐼𝑡 +1 ∥1 (9) We trained our models on each dataset separately by minimizing the loss above using ADAM [11] algorithm with learning rate 0.0002 and a batch size of 5. #### 3.6 Regularity Score A regularity score based on the prediction error 𝑒𝑡 is calculated for each video frame: _𝑟𝑒𝑡_ = 1 − _𝑒𝑡_ − min𝜏𝑒𝜏 (10) max𝜏𝑒𝜏 − min𝜏𝑒𝜏 In Hasan et al. [8], 𝑒𝑡 is the frame-wise reconstruction 𝑒MSE,𝑡 . In Liu et al. [14], 𝑒𝑡 is equivalently negative frame-wise prediction ----- **Table 1: Video anomaly detection datasets details** |Dataset|Total # frames/clips|Training # frames/clips|Testing # frames/clips|Anomaly Types| |---|---|---|---|---| |UCSD Ped2|4,560/28|2,550/16|2,010/12|biker, skater, vehicle| |CUHK Avenue|30,652/37|15,328/16|15,324/21|running, loitering, object throwing| |ShanghaiTech|315,307/437|274,516/330|40,791/107|biker, skater, vehicle, sudden motion| PSNR (Peak Signal to Noise Ratio): PSNR = 10log10 max𝑒MSE([�],𝑡𝐼𝑡 ) [. In] this study, we use similar setting to the two methods above with: _𝑒𝑡_ = log10𝑒MSE,𝑡 . #### 4 EXPERIMENTS In this section, we first introduce the three public datasets used in our experiments, which follow the same setup as other similar unsupervised video anomaly detection studies. Then, we report the video anomaly detection performance and comparison with other methods. Finally, we perform ablation studies to demonstrate the contribution of each component and interpret the results based on the proposed CT-D2GAN. #### 4.1 Datasets We evaluate our framework on three widely used public video anomaly detection datasets, i.e., UCSD Ped2 dataset [13] [1], CUHK Avenue dataset [16] [2], and ShanghaiTech Campus (SH-Tech) dataset [18] [3]. We describe the dataset-specific characteristics and the effects on video anomaly detection performance, some details can be found in Table 1: _4.1.1_ _UCSD Ped2. UCSD Ped2 includes pedestrians, vehicles_ largely moving in parallel to the camera plane. _4.1.2_ _CUHK Avenue. CUHK Avenue includes pedestrians and ob-_ jects both moving parallel to or toward/away from the camera. Slight camera motion is present in the dataset. Some of the anomalies are staged actions. _4.1.3_ _ShanghaiTech. Different from the other datasets, the Shang-_ haiTech dataset is a multi-scene dataset (13 scenes), and includes pedestrians, vehicles, and sudden motions, and the ratios of each scene in the training set and test set can be different. #### 4.2 Evaluation The model was trained and evaluated on a system with an NVIDIA GeForce 1080 Ti GPU and implemented with PyTorch. To measure the effectiveness of our proposed CT-D2GAN framework for video anomaly detection, we report the area under the receiver operating characteristics (ROC) curve i.e., AUC. Specifically, AUC is calculated by comparing the frame-level regularity scores with frame-level ground truth labels. [1http://www.svcl.ucsd.edu/projects/anomaly/dataset.html](http://www.svcl.ucsd.edu/projects/anomaly/dataset.html) [2http://www.cse.cuhk.edu.hk/leojia/projects/detectabnormal/dataset.html](http://www.cse.cuhk.edu.hk/leojia/projects/detectabnormal/dataset.html) [3https://github.com/StevenLiuWen/sRNN_TSC_Anomaly_Detection#](https://github.com/StevenLiuWen/sRNN_TSC_Anomaly_Detection##shanghaitechcampus-anomaly-detection-dataset) [shanghaitechcampus-anomaly-detection-dataset](https://github.com/StevenLiuWen/sRNN_TSC_Anomaly_Detection##shanghaitechcampus-anomaly-detection-dataset) **Table 2: Frame-level video anomaly detection performance** **(AUC).** Method UCSD Ped2 CUHK SH-Tech MPPCA+SF [19] 61.3 - MDT [13, 19] 82.9 - Conv-AE [8] 85.0 [†] 80.0 [†] 60.9 [†] 3D Conv [40] 91.2 80.9 Stacked RNN [18] 92.2 81.7 68.0 ConvLSTM-AE [17] 88.1 77.0 memAE [6] 94.1 83.3 71.2 memNormality [22] 97.0 **88.5** 70.5 ClusterAE [3] 96.5 86.0 73.3 AbnormalGAN [25] 93.5 - Frame prediction [14] 95.4 85.1 72.8 Pred+Recon [31] 96.3 85.1 73.0 CT-D2GAN **97.2** 85.9 **77.7** † Evaluated in [14]; -: Not evaluated in the study. Ordered in publication year. The best performance in each dataset is highlighted in boldface. #### 4.3 Video Anomaly Detection To demonstrate the effectiveness of our proposed CT-D2GAN framework for video anomaly detection, we compare it against 12 different baseline methods. Among those, MPPCA (mixture of probabilistic principal component analyzers) + SF (social force) [19], MDT (mixture of dynamic textures) [13, 19] are handcrafted feature based methods; Conv-AE [8], 3D Conv [40], Stacked RNN [18], and ConvLSTM-AE [17] are encoder-decoder based approaches; MemAE [6], MemNormality [22] and ClusterAE [3] are recent encoder-decoder based methods enhanced with memory module or clustering; AbnormalGAN [25], Frame prediction [14], and Pred+Recon [31] are methods based on adversarial training. Table 2 shows the frame-level video anomaly detection performance (AUC) of various approaches. We observed that encoderdecoder based approaches in general outperform handcrafted feature based methods. This is because the handcrafted features are usually extracted based upon a different objective and thus can be sub-optimal. Within encoder-decoder based approaches, ConvLSTM-AE outperforms Conv-AE since it can better capture temporal information. We also notice that adversarial training based methods perform better than most baseline methods. Finally, our |Method|UCSD Ped2|CUHK|SH-Tech| |---|---|---|---| |MPPCA+SF [19]|61.3|-|-| |MDT [13, 19]|82.9|-|-| |Conv-AE [8]|85.0†|80.0†|60.9†| |3D Conv [40]|91.2|80.9|-| |Stacked RNN [18]|92.2|81.7|68.0| |ConvLSTM-AE [17]|88.1|77.0|-| |memAE [6]|94.1|83.3|71.2| |memNormality [22]|97.0|88.5|70.5| |ClusterAE [3]|96.5|86.0|73.3| |AbnormalGAN [25]|93.5|-|-| |Frame prediction [14]|95.4|85.1|72.8| |Pred+Recon [31]|96.3|85.1|73.0| |CT-D2GAN|97.2|85.9|77.7| ----- **Figure 2: Examples of video anomaly detection. The blue lines in the line graphs delineate frame-level regularity scores. The** **green and red shaded segments in the line graphs indicate the ground truth normal and abnormal video segments respectively.** **The frames in the green boxes are regular frames from the regular video segments, the frames in the red boxes are abnormal** **frames from abnormal video segments. The abnormal objects are annotated.** proposed CT-D2GAN framework achieves the best performance on UCSD Ped2 and SH-Tech, and close to the best performance in CUHK [22]. This is because our proposed model can not only capture the spatio-temporal patterns explicitly and effectively through convolutional transformer but also leverage the dual discriminator GAN based adversarial training to maintain local consistency at frame-level and global coherence in video sequences. Recent memory or clustering enhanced methods [3, 6, 22] show good performance and is orthogonal to our proposed framework and can integrate with our proposed framework in future work to further improve performance. Examples of video anomaly detection results overlaid on the abnormal activity ground truth of all three datasets are shown in Figure 2, along with example video frames from the regular and abnormal video segments. Due to the multi-scene nature of SH-Tech dataset, we also analyzed the most ample single scene that constitutes 25% (83/330 clips) of training set and 32% (34/107 clips) of test set, the AUC is 87.5 which is much better than the overall dataset and reach similar level with other single-scene datasets. This could imply that generalizing to less ample scenes is still a challenging task given unbalanced training set. Thanks to the convolutional transformer architecture and optimizations including spatial selective gate, our model is computationally efficient. At inference time, our model runs at 45 FPS on one NVIDIA GeForce 1080 Ti GPU. **Table 3: Video anomaly detection performance under differ-** **ent ablation settings on UCSD Ped2 dataset.** #### 4.4 Ablation Studies To understand how each component contributes to the anomaly detection task, we conducted ablation studies with different settings: (1) convolutional transformer only without the adversarial training (Conv Transformer), (2) Conv Transformer with image discriminator only, (3) Conv Transformer with video discriminator only, (4) U-Net based generator (as has been utilized in image-toimage translation [10] and video anomaly detection [14]) with dual discriminator, and compare with our full CT-D2GAN model. The performance comparison can be found in Table 3. We observed that adversarial training can enhance the performance for anomaly detection, either with the image discriminator or the video discriminator. Video discriminator alone achieves nearly similar performance as using dual discriminator, but we observed the loss decreased faster when combined with image discriminator. Using image discriminator alone was not as effective, and the loss was less stable. Finally, we observed that CT-D2GAN achieved superior performance than U-Net with dual discriminator, suggesting that convolutional transformer can better capture the spatio-temporal dynamics and thus can make a more accurate detection. #### 4.5 Interpretation We illustrate an example of predicted future frame _𝑡[�]+ 1 and com-_ pare it with the previous frame 𝑡 and the ground truth future frame _𝑡_ +1 in Figure 3. The prediction performance is poor for the anomaly (red box). And also we noted that the model is able to capture the temporal dynamics by predicting the future behavior in normal part of the image (green box). **Self-attention weights under perturbation. It is not straightfor-** ward to directly interpret the temporal self-attention weight vector, as temporal self-attention is applied to an abstract representation of the video. Therefore, to further investigate the effectiveness of temporal self-attention, we perturb two frames of the video and run the inference on this perturbed video segment. For one frame (Figure 4, red), we added a random Gaussian noise with zero mean |Ablation setting|AUC| |---|---| |Conv Transformer|94.2| |Conv Transformer + image discriminator|95.7| |Conv Transformer + video discriminator|96.9| |U-Net + dual discriminator|95.5| |CT-D2GAN|97.2| ----- **Figure 3: An example showing the future frame prediction** **in the normal part of the image (green box, pedestrian in this** **case) where we observe the model capturing the dynamics of** **the behavior, and abnormal part of the image (red box, bicy-** **cle in this case) where there is large prediction error. From** **left to right, we show the last frame in the input video clip** **(𝑡), the predicted future frame** _𝑡[�]+ 1, and the ground truth fu-_ **ture frame 𝑡** + 1. **Figure 4: Temporal self-attention weights in perturbed video** **clip.** and 0.1 standard deviation to the image to simulate the deterioration in video quality; for another frame (Figure 4, purple), we scaled the optical flow maps by 0.9 to simulate the frame rate distortion. We plot the temporal attention weights for the frame right after the two perturbed frames in Figure 4. The weights assigned to the perturbed frames are clearly lower than the others, implying less contribution to the attended map. This suggests that self-attention module can adaptively select relevant feature maps and is robust to input noise. #### 5 CONCLUSIONS In this paper, we developed Convolutional Transformer based Dual Discriminator Generative Adversarial Networks (CT-D2GAN) to perform unsupervised video anomaly detection. The convolutional transformer which consists of three components, i.e., a convolutional encoder to capture the spatial patterns of the input video clip, a temporal self-attention module to encode the temporal dynamics, and a convolutional decoder to integrate spatio-temporal features, was employed to perform future frame prediction. A dual discriminator based adversarial training approach was used to maintain the local consistency of the predicted frame and the global coherence conditioned on the previous frames. Thorough experiments on three widely used video anomaly detection datasets demonstrate that our proposed CT-D2GAN is able to detect anomaly frames with superior performance. #### REFERENCES [1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. 2017. Wasserstein generative adversarial networks. In International Conference on Machine Learning _(ICML). PMLR, 214–223._ [2] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. 2004. High accuracy optical flow estimation based on a theory for warping. In European _Conference on Computer Vision (ECCV). Springer, 25–36._ [3] Yunpeng Chang, Zhigang Tu, Wei Xie, and Junsong Yuan. 2020. Clustering Driven Deep Autoencoder for Video Anomaly Detection. In European Conference _on Computer Vision (ECCV). Springer, 329–345._ [4] Yong Shean Chong and Yong Haur Tay. 2017. Abnormal event detection in videos using spatiotemporal autoencoder. In International Symposium on Neural _Networks (ISNN). Springer, 189–196._ [5] Fei Dong, Yu Zhang, and Xiushan Nie. 2020. Dual Discriminator Generative Adversarial Network for Video Anomaly Detection. IEEE Access 8 (2020), 88170– 88176. [6] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. 2019. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In IEEE International Conference on Computer Vision (ICCV). IEEE, 1705–1714. [7] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. 2017. Improved training of Wasserstein GANs. In Advances _in Neural Information Processing Systems (NIPS). 5767–5777._ [8] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. 2016. Learning temporal regularity in video sequences. In IEEE _Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 733–742._ [9] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. 2017. Flownet 2.0: Evolution of optical flow estimation with deep networks. In IEEE Conference on Computer Vision and Pattern Recognition _(CVPR). IEEE, 2462–2470._ [10] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. 2017. Image-toImage Translation with Conditional Adversarial Networks. In IEEE Conference _on Computer Vision and Pattern Recognition (CVPR). IEEE, 5967–5976._ [11] Diederik P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. International Conference on Learning Representations (ICLR) (2015). [12] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. 2017. Self-normalizing neural networks. In Advances in Neural Information _Processing Systems (NIPS). 971–980._ [13] Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. 2014. Anomaly detection and localization in crowded scenes. IEEE Transactions on Pattern Analysis and _Machine Intelligence 36, 1 (2014), 18–32._ [14] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. 2018. Future Frame Prediction for Anomaly Detection – A New Baseline. In IEEE Conference on _Computer Vision and Pattern Recognition (CVPR). IEEE, 6536–6545._ [15] Jonathan Long, Evan Shelhamer, and Trevor Darrell. 2015. Fully convolutional networks for semantic segmentation. In IEEE Conference on Computer Vision and _Pattern Recognition (CVPR). IEEE, 3431–3440._ [16] Cewu Lu, Jianping Shi, and Jiaya Jia. 2013. Abnormal event detection at 150 FPS in MATLAB. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2720–2727. [17] Weixin Luo, Wen Liu, and Shenghua Gao. 2017. Remembering history with convolutional LSTM for anomaly detection. In IEEE International Conference on _Multimedia and Expo (ICME). IEEE, 439–444._ ----- [18] Weixin Luo, Wen Liu, and Shenghua Gao. 2017. A revisit of sparse coding based anomaly detection in stacked RNN framework. IEEE International Conference on _Computer Vision (ICCV) 1, 2 (2017), 3._ [19] Vijay Mahadevan, Weixin Li, Viral Bhalodia, and Nuno Vasconcelos. 2010. Anomaly detection in crowded scenes. In IEEE Conference on Computer Vision and _Pattern Recognition (CVPR). IEEE, 1975–1981._ [20] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. 2018. Spectral Normalization for Generative Adversarial Networks. In International _Conference on Learning Representations (ICLR)._ [21] Tu Nguyen, Trung Le, Hung Vu, and Dinh Phung. 2017. Dual discriminator generative adversarial nets. In Advances in neural information processing systems _(NIPS). 2670–2680._ [22] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. 2020. Learning Memoryguided Normality for Anomaly Detection. In IEEE Conference on Computer Vision _and Pattern Recognition (CVPR). 14372–14381._ [23] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. 2018. Image Transformer. In International _Conference on Machine Learning (ICML). PMLR, 4052–4061._ [24] Yao Qin, Dongjin Song, Haifeng Chen, Wei Cheng, Guofei Jiang, and Garrison W. Cottrell. 2017. A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction. In International Joint Conference on Artificial Intelligence _(IJCAI). 2627–26332._ [25] Mahdyar Ravanbakhsh, Moin Nabi, Enver Sangineto, Lucio Marcenaro, Carlo Regazzoni, and Nicu Sebe. 2017. Abnormal Event Detection in Videos using Generative Adversarial Nets. IEEE International Conference on Image Processing _(ICIP) (2017), 1577–1581._ [26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolutional networks for biomedical image segmentation. In International Conference on _Medical Image Computing and Computer-Assisted Intervention (MICCAI). Springer,_ 234–241. [27] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. 2015. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in Neural Information Processing _Systems (NIPS). 802–810._ [28] Karen Simonyan and Andrew Zisserman. 2014. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Pro_cessing Systems (NIPS). 568–576._ [29] Dongjin Song and Dacheng Tao. 2010. Biologically Inspired Feature Manifold for Scene Classification. IEEE Transactions on Image Processing 19, 1 (2010), 174–184. [30] Waqas Sultani, Chen Chen, and Mubarak Shah. 2018. Real-World Anomaly Detection in Surveillance Videos. In IEEE Conference on Computer Vision and _Pattern Recognition (CVPR). IEEE, 6479–6488._ [31] Yao Tang, Lin Zhao, Shanshan Zhang, Chen Gong, Guangyu Li, and Jian Yang. 2020. Integrating prediction and reconstruction for anomaly detection. Pattern _Recognition Letters 129 (2020), 123–130._ [32] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. 2018. MoCoGAN: Decomposing Motion and Content for Video Generation. In IEEE Conference on _Computer Vision and Pattern Recognition (CVPR). IEEE, 1526–1535._ [33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems (NIPS). 6000–6010. [34] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. 2017. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer _Vision and Image Understanding 156 (2017), 117–127._ [35] Han Xu, Pengwei Liang, Wei Yu, Junjun Jiang, and Jiayi Ma. 2019. Learning a Generative Model for Fusing Infrared and Visible Images via Conditional Generative Adversarial Network with Dual Discriminators.. In International _Joint Conference on Artificial Intelligence (IJCAI). 3954–3960._ [36] Fisher Yu and Vladlen Koltun. 2016. Multi-scale context aggregation by dilated convolutions. International Conference on Learning Representations (ICLR) (2016). [37] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S. Huang. 2018. Generative Image Inpainting With Contextual Attention. In IEEE Conference on _Computer Vision and Pattern Recognition (CVPR). IEEE, 5505–5514._ [38] Chuxu Zhang, Dongjin Song, Yuncong Chen, Xinyang Feng, Cristian Lumezanu, Wei Cheng, Jingchao Ni, Bo Zong, Haifeng Chen, and V. Nitesh Chawla. 2019. A Deep Neural Network for Unsupervised Anomaly Detection and Diagnosis in Multivariate Time Series Data. In Association for the Advancement of Artificial _Intelligence (AAAI). AAAI, 1409–1416._ [39] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. 2019. Selfattention generative adversarial networks. In International Conference on Machine _Learning (ICML). PMLR, 7354–7363._ [40] Yiru Zhao, Bing Deng, Chen Shen, Yao Liu, Hongtao Lu, and Xian-Sheng Hua. 2017. Spatio-Temporal AutoEncoder for Video Anomaly Detection. In ACM _International Conference on Multimedia (ACM MM). ACM, 1933–1941._ -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2107.13720, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://arxiv.org/pdf/2107.13720" }
2,021
[ "JournalArticle", "Book" ]
true
2021-07-29T00:00:00
[ { "paperId": "18f207d8dab7357f4f674211ec4f150de1c93a0e", "title": "Learning Memory-Guided Normality for Anomaly Detection" }, { "paperId": "fe09f7a379944444201552e952b910188c0aeaca", "title": "Integrating prediction and reconstruction for anomaly detection" }, { "paperId": "74fd298aa27b390194596a91f3db949a8ce8e72e", "title": "Learning a Generative Model for Fusing Infrared and Visible Images via Conditional Generative Adversarial Network with Dual Discriminators" }, { "paperId": "d65eb30e5f0d2013fd5e4f45d1413bc2969ee803", "title": "Memorizing Normality to Detect Anomaly: Memory-Augmented Deep Autoencoder for Unsupervised Anomaly Detection" }, { "paperId": "d9cb449af41b6da76c98475ff268b3ef06d6dec1", "title": "A Deep Neural Network for Unsupervised Anomaly Detection and Diagnosis in Multivariate Time Series Data" }, { "paperId": "1db9bd18681b96473f3c82b21edc9240b44dc329", "title": "Image Transformer" }, { "paperId": "84de7d27e2f6160f634a483e8548c499a2cda7fa", "title": "Spectral Normalization for Generative Adversarial Networks" }, { "paperId": "6b0bbf3e7df725cc3b781d2648e41782cb3d8539", "title": "Generative Image Inpainting with Contextual Attention" }, { "paperId": "96ed8ce9ef9fc475db9e02c79f984dc110409b62", "title": "Real-World Anomaly Detection in Surveillance Videos" }, { "paperId": "8a6acba7fb2aad1299fcf35701417e063d410ed4", "title": "Future Frame Prediction for Anomaly Detection - A New Baseline" }, { "paperId": "fef6f1e04fa64f2f26ac9f01cd143dd19e549790", "title": "Spatio-Temporal AutoEncoder for Video Anomaly Detection" }, { "paperId": "99dff291f260b3cc3ff190106b0c2e3e685223a4", "title": "A Revisit of Sparse Coding Based Anomaly Detection in Stacked RNN Framework" }, { "paperId": "46bc2a73ff36d00357eda4e8051f0b015b63e246", "title": "Dual Discriminator Generative Adversarial Nets" }, { "paperId": "9d5290fadb7625862a966e0330bd0f9e111fc99d", "title": "Abnormal event detection in videos using generative adversarial nets" }, { "paperId": "e76edb86f270c3a77ed9f5a1e1b305461f36f96f", "title": "MoCoGAN: Decomposing Motion and Content for Video Generation" }, { "paperId": "acd87843a451d18b4dc6474ddce1ae946429eaf1", "title": "Wasserstein Generative Adversarial Networks" }, { "paperId": "792250ae660b7c25f85eeea7dcae623e4301d97c", "title": "Remembering history with convolutional LSTM for anomaly detection" }, { "paperId": "204e3073870fae3d05bcbc2f6a8e263d9b72e776", "title": "Attention is All you Need" }, { "paperId": "424a6e62084d919bfc2e39a507c263e5991ebdad", "title": "Self-Normalizing Neural Networks" }, { "paperId": "76624f8ff1391e942c3313b79ed08a335aa5077a", "title": "A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction" }, { "paperId": "edf73ab12595c6709f646f542a0d2b33eb20a3f4", "title": "Improved Training of Wasserstein GANs" }, { "paperId": "e5366a704ffa3b41aacd385f3c087ec3fd566934", "title": "Detecting anomalous events in videos by learning deep representations of appearance and motion" }, { "paperId": "527cc8cd2af06a9ac2e5cded806bab5c3faad9cf", "title": "Abnormal Event Detection in Videos using Spatiotemporal Autoencoder" }, { "paperId": "edd846e76cacfba5be37da99c006e3ccc9b861b0", "title": "FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks" }, { "paperId": "8acbe90d5b852dadea7810345451a99608ee54c7", "title": "Image-to-Image Translation with Conditional Adversarial Networks" }, { "paperId": "97e7c94a78ae17cfb90848c1cfca8c431082a7b2", "title": "Learning Temporal Regularity in Video Sequences" }, { "paperId": "7f5fc84819c0cf94b771fe15141f65b123f7b8ec", "title": "Multi-Scale Context Aggregation by Dilated Convolutions" }, { "paperId": "f9c990b1b5724e50e5632b94fdb7484ece8a6ce7", "title": "Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting" }, { "paperId": "6364fdaa0a0eccd823a779fcdd489173f938e91a", "title": "U-Net: Convolutional Networks for Biomedical Image Segmentation" }, { "paperId": "a6cb366736791bcccc5c8639de5a8f9636bf87e8", "title": "Adam: A Method for Stochastic Optimization" }, { "paperId": "6fc6803df5f9ae505cae5b2f178ade4062c768d0", "title": "Fully convolutional networks for semantic segmentation" }, { "paperId": "67dccc9a856b60bdc4d058d83657a089b8ad4486", "title": "Two-Stream Convolutional Networks for Action Recognition in Videos" }, { "paperId": "869b17632ed4f19f93b3b58dcaa9f0b8e92108f3", "title": "Abnormal Event Detection at 150 FPS in MATLAB" }, { "paperId": "9d3f0d47449c7db37d1bae3b70db2928610a8db7", "title": "Anomaly detection in crowded scenes" }, { "paperId": "91228e00fe33ed6072cfe849ab9e98160461549d", "title": "High Accuracy Optical Flow Estimation Based on a Theory for Warping" }, { "paperId": "7c75203739f5f89e109b11144d170d4d3f2a6abc", "title": "Clustering Driven Deep Autoencoder for Video Anomaly Detection" }, { "paperId": "e37daeaa20bc8cd68db07641201faf1db3b8c31d", "title": "Dual Discriminator Generative Adversarial Network for Video Anomaly Detection" }, { "paperId": "aae932cf9c2434f52b03991fcab050a61a960d48", "title": "Anomaly Detection and Localization in Crowded Scenes" }, { "paperId": "5da739abaa85ba155c6aa76a29d5cbdefb6f018e", "title": "Biologically Inspired Feature Manifold for Scene Classification" }, { "paperId": null, "title": "Viral Bhalodia, and Nuno Vasconcelos" } ]
14,460
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/014be151dfab3a65e5da397810b0bd834b2aa06d
[ "Computer Science" ]
0.846703
A Survey of Local Differential Privacy and Its Variants
014be151dfab3a65e5da397810b0bd834b2aa06d
arXiv.org
[ { "authorId": "2215284412", "name": "Likun Qin" }, { "authorId": "2238048286", "name": "Nan Wang" }, { "authorId": "2234372112", "name": "Tianshuo Qiu" } ]
{ "alternate_issns": null, "alternate_names": [ "ArXiv" ], "alternate_urls": null, "id": "1901e811-ee72-4b20-8f7e-de08cd395a10", "issn": "2331-8422", "name": "arXiv.org", "type": null, "url": "https://arxiv.org" }
The introduction and advancements in Local Differential Privacy (LDP) variants have become a cornerstone in addressing the privacy concerns associated with the vast data produced by smart devices, which forms the foundation for data-driven decision-making in crowdsensing. While harnessing the power of these immense data sets can offer valuable insights, it simultaneously poses significant privacy risks for the users involved. LDP, a distinguished privacy model with a decentralized architecture, stands out for its capability to offer robust privacy assurances for individual users during data collection and analysis. The essence of LDP is its method of locally perturbing each user's data on the client-side before transmission to the server-side, safeguarding against potential privacy breaches at both ends. This article offers an in-depth exploration of LDP, emphasizing its models, its myriad variants, and the foundational structure of LDP algorithms.
# A Survey of Local Differential Privacy and Its Variants ### Likun Qin, Nan Wang, Tianshuo Qiu Department of Electrical and Computer Engineering Shandong University, Jinan, China Abstract—The introduction and advancements in Local Differential Privacy (LDP) variants have become a cornerstone in addressing the privacy concerns associated with the vast data produced by smart devices, which forms the foundation for datadriven decision-making in crowdsensing. While harnessing the power of these immense data sets can offer valuable insights, it simultaneously poses significant privacy risks for the users involved. LDP, a distinguished privacy model with a decentralized architecture, stands out for its capability to offer robust privacy assurances for individual users during data collection and analysis. The essence of LDP is its method of locally perturbing each user’s data on the client-side before transmission to the server-side, safeguarding against potential privacy breaches at both ends. This article offers an in-depth exploration of LDP, emphasizing its models, its myriad variants, and the foundational structure of LDP algorithms. I. INTRODUCTION Collecting and analyzing data introduces significant privacy concerns because it often includes sensitive user information. With the advent of sophisticated data fusion and analysis methods, user data becomes even more susceptible to breaches and exposure in this era of big data. For instance, by studying appliance usage, adversaries can deduce daily routines or behaviors of individuals, like when they are home or their specific activities such as watching TV or cooking. It’s crucial to prioritize the protection of personal data when gathering information from diverse devices. Currently, the European Union (EU) has released the GDPR [1], which oversees EU data protection laws for its citizens and outlines the specifics related to the handling of personal data. Similarly, the U.S. National Institute of Standards and Technology (NIST) is in the process of crafting privacy frameworks. These frameworks aim to more effectively recognize, evaluate, and address privacy risks, enabling individuals to embrace innovative technologies with increased trust and confidence [2], [3]. From a privacy-protection standpoint, differential privacy (DP) has been introduced over a decade ago [4], [5]. Recognized as a robust framework for safeguarding privacy, it’s often termed as global DP or centralized DP. DP’s strength lies in its mathematical rigor; it operates independent of an adversary’s background knowledge and assures potent privacy protection for users. It has found applications across various domains [6]. However, DP assumes the presence of a trustworthy server, which can be a challenge since many online platforms or crowdsourcing systems might have untrustworthy servers keen on user data statistics [7] [8] Emerging from the concept of DP, local differential privacy (LDP) was introduced [9]. LDP stands as a decentralized version of DP, offering individualized privacy assurances and making no assumptions about third-party server trustworthiness. LDP has become a focal point in privacy research due to its theoretical significance and practical implications [10]. Numerous corporations, including Apple’s iOS [11], Google Chrome, and the Windows operating system, have integrated LDP-driven algorithms into their systems. Owing to its robust capabilities, LDP has become a preferred choice to address individual privacy concerns during various statistical and analytical operations. This includes tasks like frequency and mean value estimation [12], the identification of heavy hitters [13], k-way marginal release, empirical risk minimization (ERM), federated learning, and deep learning. While LDP is powerful, it’s not without its challenges, notably in striking an optimal balance between utility and privacy [14]. To address this, there are two primary approaches. Firstly, by devising improved mechanisms - leading to the introduction of numerous LDP-based protocols and sophisticated mechanisms in academic circles. Secondly, by revisiting the definition of LDP itself, with researchers suggesting more flexible privacy concepts to better cater to the utility-privacy balance required for real-world applications. Given the growing significance of LDP, a thorough survey of the topic is both timely and essential. While there exists some literature reviewing LDP, the focus has often been narrow. They either focus on specific applications or certain types of mechanisms. In this paper, we delve deep into the world of LDP and its various offshoots, meticulously studying their recent advancements and associated mechanisms. We embark on a thorough exploration of the foundational principles that drive LDP and the evolutionary trajectories of its multiple variants. We aim to identify the cutting-edge developments, shedding light on the innovations that have shaped these privacy tools and the challenges they aim to address in our contemporary digital landscape. Furthermore, we analyze the specific mechanisms that support and enhance the capabilities of LDP, understanding their technical intricacies and the realworld applications they cater to. Through this comprehensive study, we aspire to provide readers with a panoramic view of the current state of LDP research, setting the stage for future inquiries and innovations in this critical domain ----- II. LOCAL DIFFERENTIAL PRIVACY, PROPERTIES AND MECHANISMS In this section, we study LDP and its properties, and LDP based mechanisms. We start from the definition of LDP. Definition 1 (ε-Local Differential Privacy (ε-LDP) . A randomized mechanism M satisfies ε-LDP if and only if for any pairs of input values v, v[′] in the domain of M, and for any possible output y ∈ Y, it holds P [M (v) = y] ≤ e[ε] - P [M (v[′]) = y], (1) where P [·] denotes probability and ε is the privacy budget. A smaller ε means stronger privacy protection, and vice versa. The basic properties of LDP include the followings: Composition: [15] Given two mechanisms M1 and M2 that provide ε1-LDP and ε2-LDP respectively, their sequential composition provides (ε1 + ε2)-LDP. M (v) = (M1(v), M2(v)) =⇒ M is (ε1 + ε2)-LDP (2) Post-processing: Any function applied to the output of an ε-LDP mechanism retains the ε-LDP guarantee. If M (v) is ε-LDP, then f (M (v)) is also ε-LDP. (3) Robustness to Side Information: LDP guarantees hold even if an adversary has access to auxiliary or side information. Utility-Privacy Tradeoff: Generally, a lower value of ε implies stronger privacy but might result in reduced utility of the perturbed data. Independence of Background Knowledge: The privacy guarantees of LDP mechanisms are designed to hold regardless of any background knowledge an adversary might have. Next, we study mechanisms that satisfy LDP: Randomize Response [16] The Randomized Response Mechanism is a simple yet effective approach to achieving LDP. It’s particularly used for binary data, i.e., when a user’s data item is either 0 or 1. The mechanism operates as follows: 1) With probability [1]2 [, the user truthfully answers a ques-] tion. 2) With probability 12 [, the user randomly answers the] question. Mathematically, given a user’s true data item v ∈{0, 1}, the mechanism outputs v with probability [1]2 [and outputs][ 1] [−] [v] (i.e., the opposite of v) with probability 2[1] [.] The probability mass function (pmf) is given by: P [M (v) = 1] = [1] (4) 2 [v][ + 1]2 [(1][ −] [v][) = 1]2 P [M (v) = 0] = [1] (5) 2 [(1][ −] [v][) + 1]2 [v][ = 1]2 where ∆f is the sensitivity of the function f and Lap(·) represents the Laplace distribution. Gaussian Mechanism Similar to the Laplace Mechanism, the Gaussian Mechanism adds noise but from the Gaussian distribution: Given a data item v, the mechanism outputs: M (v) = v + N (0, σ[2]) where σ[2] determines the amount of noise based on the desired ε and function sensitivity ∆f . Exponential Mechanism The Exponential Mechanism selects an output based on a scoring function and weights outputs with the exponential of their score. Given a set of possible outputs R, a data item v, and a scoring function q(v, r), the probability of selecting output r is proportional to: � εq(v, r) exp 2∆q � This mechanism ensures ε-LDP with ε = ln(2). Laplace Mechanism [17] The Laplace Mechanism adds noise drawn from the Laplace distribution to the true value of the data. For LDP, this mechanism can be adjusted as: Given a data item v, the mechanism outputs: M (v) = v + Lap( [∆][f] ) where ∆q is the sensitivity of q. Perturbed Histogram Mechanism For a set of items, instead of perturbing each item, this mechanism perturbs the histogram of the data items. Given a data item set V, the mechanism constructs a histogram H and then outputs: M (H) = H + Lap( [∆][H] ε [)] where ∆H is the sensitivity of the histogram construction. Observe that each mechanism’s efficacy is closely tied to the sensitivity of the query, denoted as ∆f . In the realm of Local Differential Privacy (LDP), this sensitivity can often grow significantly, especially when the input domain is vast. The larger the sensitivity, the more noise needs to be introduced by the mechanism to ensure the desired privacy level. This can lead to significant distortion in the data, compromising its utility. Furthermore, as the input support size increases, maintaining the desired privacy guarantee becomes a challenge. Noise calibrated to a high sensitivity can sometimes overshadow the actual data, rendering the results almost meaningless or leading to misinterpretations. The consequence of this is a pronounced tradeoff between utility and privacy. Achieving stronger privacy often means accepting reduced accuracy and utility in the results, and vice versa. For applications that require high precision, this can be problematic. It implies that while these mechanisms provide a robust privacy guarantee in theory, their practical applicability can be constrained, especially in scenarios where fine-grained insights from data are crucial. Hence, while the promise of LDP is enticing, its real-world implementation requires careful consideration of the utilityprivacy balance, pushing researchers to seek more efficient mechanisms or modified privacy models to better cater to practical needs ----- III. ADVANCED LDP MECHANISMS As we mentioned in the introduction, to improve the utilityprivacy tradeoff provided by LDP, there are typically two manners. One is to design dedicated mechanism or advanced protocols. The other is to relax the definition of LDP to enhance the data utility. In this section, we summarized several advanced LDP algorithms, aiming to improve the general utility-privacy tradeoff. RAPPOR [10] (Randomized Aggregatable PrivacyPreserving Ordinal Response): Introduced by Google. RAPPOR enhances the randomized response mechanism through the incorporation of Bloom filters. Each user’s value is hashed multiple times into a Bloom filter, which is then perturbed using the RR technique. This allows multiple string values to be encoded before randomization. Advantage: Its main strength lies in collecting statistics about low-frequency items in the user population. It can provide meaningful insights even when items are not commonly observed. Local Hashing [12]: Addressing the problem of efficiency in the RR technique when dealing with a large domain of inputs, local hashing maps the original vast domain into a smaller domain using hash functions. This condensed domain can then be analyzed using traditional RR techniques. Advantage: It substantially reduces the noise introduced in the randomization process, enabling accurate estimation of frequencies for individual items in the domain. This mechanism improves the utility, especially when the original domain is considerably large. Piecewise RR: Instead of applying the same randomization mechanism across the entire input domain, the Piecewise RR technique divides the domain into multiple segments or pieces. Each segment then gets its own randomization mechanism tailored to its characteristics. Advantage: This method achieves a more granular utility-privacy tradeoff. It can offer enhanced privacy in sensitive segments while improving utility in less-sensitive ones. Optimized RR: The protocol doesn’t just use a fixed randomization parameter; instead, it optimizes the parameters of the RR. This optimization is often based on real data distribution or some auxiliary information, ensuring that the randomization provides the best possible utility. Advantage: By adjusting the randomization according to data distribution, it achieves better accuracy in aggregate statistics. Fourier Perturbation Algorithm (FPA) [18]: Instead of perturbing the raw data directly, FPA operates in the frequency domain. The data undergoes a Fourier transformation, after which the perturbation is applied. This allows for randomization in a different space that might be more conducive to certain types of analyses. Advantage: Provides enhanced utility for specific query types, especially those that are frequency-based or need insights from periodic patterns in data. IV. LDP VARIANTS AND MECHANISMS In this section, we introduce LDP variants that aim to provide better utility privacy tradeoff in different applications A. Variants and Mechanisms of LDP 1) (ε, δ)-LDP: Drawing parallels with how (ε, δ)-DP [19] extends ε-DP, (ε, δ)-LDP (sometimes termed as approximate LDP) serves as a more flexible counterpart to ε-LDP (or pure LDP). Definition 1 (Approximate Local Differential Privacy). A randomized process M complies with (ε, δ)-LDP if, for all input pairs v and v[′] within M ’s domain and any probable output y ∈ Y, the following holds: P [M (v) = y] ≤ e[ε] - P [M (v[′]) = y] + δ. Here, δ is customarily a small value. In essence, (ε, δ)-LDP implies that M achieves ε-LDP with a likelihood not less than 1−δ. If δ = 0, (ε, δ)-LDP converges to ε-LDP. 2) BLENDER Model: BLENDER [20], a fusion of global DP and LDP, optimizes data utility while retaining privacy. It classifies users based on their trust in the aggregator into two categories: the opt-in group and clients. BLENDER enhances utility by balancing data from both. Its privacy measure mirrors that of (ε, δ)-DP [21]. 3) Geo-indistinguishability: Originally tailored for location privacy with global DP, Geo-indistinguishability [22] uses the data’s geographical distance. Alvim et al. [23] argued for metric-based LDP’s advantages in specific contexts. Definition 2 (Geo-indistinguishability). A randomized function M adheres to Geo-indistinguishability if, for any input pairs v and v[′] and any output y ∈ Y, the subsequent relation is met: P [M (v) = y] ≤ e[ε][·][d][(][v,v][′][)] - P [M (v[′]) = y], where d(., .) designates a distance metric. This model adjusts privacy depending on data distance, augmenting utility for datasets like location or smart meter consumption that are sensitive to distance. 4) Local Information Privacy: Local Information Privacy (LIP) was originally proposed in [24] as a prior-aware version of LDP, and then, in [25], Jiang et al relax the prior-aware assumption to partial prior-aware (Bounded Prior in their version). The definition of LIP is shown as follows: Definition 3. (ǫ, δ)-Local Information Privacy [26] A mechanism M satisfies (ǫ, δ)-LIP, if ∀x ∈X, y ∈ Range(M): P (Y = y) ≥ e[−][ǫ]P (Y = y|X = x) − δ, (6) P (Y = y) ≤ e[ǫ]P (Y = y|X = x) + δ. 5) Sequential Information Privacy: Sequential Information Privacy (SIP), built upon LIP, measures the privacy leakage for a data sequence, or time series data. SIP naturally decomposes using chain rule-similar techniques and is comparable to that of LDP. Definition 4. [(ǫ)-Sequence Information Privacy] [27] A mechanism M satisfies (ǫ)-SIP for some ǫ ∈ `R[+], if ∀X1[T]` [∈X] [,] Y1[T] [∈] [Range][(][M][)][:] 1 [) =][ y]1[T] []] e[−][ǫ] ≤ [P] [[][M][ [(][x]T[T] T ] ≤ e[ǫ] (7) ----- The operational meaning of LIP is, the output Y provides limited additional information about any possible input X, and the amount of the additional information is measured by the privacy budget ǫ and failure probability δ. In [28], multiple LIP mechanisms were proposed and testified, showing that even though ǫ-LIP is stronger than 2ǫ-LDP in terms of privacy protection. The mechanisms achieve more than 2 times of utility gain. 6) CLDP: Recognizing LDP’s diminished utility with fewer users, Gursoy et al. [29] introduced the metric-based model of condensed local differential privacy (CLDP). Definition 5 (α-CLDP). For all input pairs v and v[′] in M ’s domain and any potential output y ∈ Y, a randomized function M satisfies α-CLDP if: P [M (v) = y] ≤ e[α][·][d][(][v,v][′][)] - P [M (v[′]) = y], where α > 0. In CLDP, a decline in α compensates for a growth in distance d. Gursoy et al. employed an Exponential Mechanism variant to devise protocols, particularly benefitting scenarios with limited users. 7) PLDP: PLDP [30] offers user-specific privacy levels. Here, users can modify their privacy settings, denoted by ε. Definition 6 (ε-PLDP). For a user U, and all input pairs v and v[′] in M ’s domain and any potential output y ∈ Y, a randomized function M meets εU -PLDP if: P [M (v) = y] ≤ e[ε][U] - P [M (v[′]) = y]. Approaches like the personalized count estimation protocol and advanced combination strategy cater to users with varying privacy inclinations. 8) Utility-optimized LDP (ULDP): Traditional LDP assumes all data points have uniform sensitivity, often causing excessive noise addition. Recognizing that not all personal data have equivalent sensitivity, the Utility-optimized LDP (ULDP) model was proposed. In this model, let KS ⊆ K be the sensitive dataset and KN = K \ KS be the nonsensitive dataset. Let YP ⊆ Y be the protected output set and YI = Y \YP be the invertible output set. The formal definition of ULDP is: Definition 7. Given KS ⊆ K, YP ⊆ Y, a mechanism M adheres to (KS, YP, ǫ)-ULDP if: - For every y ∈ YI, there is a v ∈ KN with P [M (v) = y] > 0 and P [M (v[′]) = y] = 0 for any v[′] ̸= v. - For all v, v[′] ∈ K and y ∈ YP, P [M (v) = y] ≤ e[ǫ] P [M (v[′]) = y]. In simpler terms, (KS, YP, ǫ)-ULDP ensures that sensitive inputs are mapped only to the protected output set. 9) Input-Discriminative LDP (ID-LDP): While ULDP classifies data as either sensitive or non-sensitive, the ID-LDP model offers a more nuanced approach by acknowledging varying sensitivity levels among data It is defined as: Definition 8. Given a set of privacy budgets E = {ǫv}v∈K, a mechanism M adheres to E-ID-LDP if for all input pairs v and v[′], and any output y ∈ Y : P [M (v) = y] ≤ e[r][(][ǫ][v][,ǫ][v][′][ )] - P [M (v[′]) = y] where r(·, ·) is a function of two privacy budgets. The study in [31] primarily utilizes the minimum function between ǫv and ǫv′ and introduces the MinID-LDP as a specialized case. 10) Parameter Blending Privacy (PBP): PBP was proposed as a more flexible LDP variant [32]. In PBP, let Θ represent the domain of privacy parameters. Given a privacy budget θ ∈ Θ, let P (θ) denote the frequency with which θ is selected. PBP is defined as: Definition 9. A mechanism M adheres to r-PBP if, for all θ ∈ Θ, v, v[′] ∈ K, y ∈ Y, there exists a θ[′] ∈ Θ such that: P (θ)P [M (v; θ) = y] ≤ e[r][(][θ][)] - P (θ[′])P [M (v[′]; θ[′]) = y] B. A Summary of LDP variants Local Differential Privacy (LDP) is a foundational approach tailored for all data types and operates using the randomized response (RR) technique. Its primary advantage is its broad applicability, but it may add more noise than necessary, especially when not all data attributes have the same sensitivity levels. To address this, approximate LDP, which allows for minor violations in privacy guarantees, introduces flexibility. However, this relaxation can be a double-edged sword, potentially compromising privacy in highly sensitive scenarios. BLENDER, on the other hand, is crafted explicitly for categorical data. By synergizing aspects of both global Differential Privacy and LDP, it aims to improve data utility. Yet, its reliance on grouping user data might introduce challenges in dynamic or constantly changing environments. Local d-privacy is another variant, designed with metric spaces in mind. It’s particularly beneficial for data like location points, but may not be the first choice for other data structures due to its specific metric-based method. CLDP stands out for its unique approach to address challenges that arise with smaller user counts, an often overlooked but crucial aspect in privacy. However, while it addresses issues in smaller datasets, it might introduce complexities when the user base grows, making scalability a potential concern. PLDP, meanwhile, strives to provide a more granular level of privacy. While this granularity is its strength, the tradeoff might be a more significant computational overhead and intricate implementation details. ULDP takes a novel stance by focusing on optimizing utility through an emphasis on sensitive data. The premise here is that not all data pieces hold equal sensitivity. However, the challenge and responsibility of correctly categorizing which data is sensitive can be daunting. ID-LDP further refines this concept by providing protection based on the actual sensitivity of the input, using unary encoding to achieve this. Its main challenge is the intricate parameter setting required to ensure optimal performance. Lastly, PBP is distinct in its pursuit of robust privacy By maintaining the secrecy of provider ----- parameters, it bolsters privacy assurances. Yet, this added layer [16] S. L. Warner, “Randomized response: A survey technique for of secrecy might introduce complexities in implementation and eliminating evasive answer bias,” Journal of the American Statistical Association, vol. 60, no. 309, pp. 63–69, 1965. [Online]. Available: understanding. [http://www.jstor.org/stable/2283137](http://www.jstor.org/stable/2283137) [17] C. Dwork, M. Naor, T. Pitassi, and G. N. Rothblum, “Differential privacy under continual observation,” in Proceedings of the Forty-Second ACM V. CONCLUSION Symposium on Theory of Computing, ser. STOC ’10, New York, NY, In the realm of data privacy, Local Differential Privacy USA, 2010, p. 715–724. (LDP) stands out as a vital tool for preserving user data. This [18] E. Bozkir, O. G¨unl¨u, W. Fuhl, R. Schaefer, and E. Kasneci, “Differential privacy for eye tracking with temporal correlations,” PLOS ONE, vol. 16, research delves into various LDP mechanisms, protocols and p. e0255979, 08 2021. variants in definition, each addressing unique challenges. From [19] R. Bassily, “Linear queries estimation with local differential the foundational LDP to specialized versions like BLENDER privacy,” CoRR, vol. abs/1810.02810, 2018. [Online]. Available: [http://arxiv.org/abs/1810.02810](http://arxiv.org/abs/1810.02810) for categorical data and Local d-privacy for metrics, the [20] B. Avent, A. Korolova, D. Zeber, T. Hovden, and spectrum of solutions is vast. Techniques like CLDP tackle B. Livshits, “BLENDER: Enabling local search with a smaller datasets, while PLDP, ULDP, and ID-LDP optimize hybrid differential privacy model,” in 26th USENIX Security Symposium (USENIX Security 17). Vancouver, BC: USENIX data utility and privacy levels. The introduction of PBP em- Association, Aug. 2017, pp. 747–764. [Online]. Available: phasizes secrecy in privacy parameters. Ultimately, this paper [https://www.usenix.org/conference/usenixsecurity17/technical-sessions/presentation/aven](https://www.usenix.org/conference/usenixsecurity17/technical-sessions/presentation/avent) underscores the importance of selecting the right LDP variant, [21] C. Dwork, K. Kenthapadi, F. McSherry, I. Mironov, and M. Naor, “Our data, ourselves: Privacy via distributed noise generation,” in Advances given the specific nature of data and privacy needs. in Cryptology - EUROCRYPT 2006, 25th Annual International Conference on the Theory and Applications of Cryptographic Techniques, St. Petersburg, Russia, May 28 - June 1, 2006, REFERENCES Proceedings, ser. Lecture Notes in Computer Science, S. Vaudenay, [1] European Parliament and Council of the European Union. Regulation Ed., vol. 4004. Springer, 2006, pp. 486–503. [Online]. Available: (EU) 2016/679 of the European Parliament and of the Council. [https://doi.org/10.1007/11761679 29](https://doi.org/10.1007/11761679_29) [[Online]. Available: https://data.europa.eu/eli/reg/2016/679/oj](https://data.europa.eu/eli/reg/2016/679/oj) [22] M. E. Andr´es, N. E. Bordenabe, and K. Chatzikokolakis, [2] R. MEYER. (2014) Facebook’s mood manipulation “Geo-indistinguishability: Differential privacy for location-based experiment might have been illegal. [Online]. Available: systems,” CoRR, vol. abs/1212.1984, 2012. [Online]. Available: [https://www.theatlantic.com/technology/archive/2014/09/facebooks-mood-manipulation-experiment-might-be-illegal/380717/http://arxiv.org/abs/1212.1984](https://www.theatlantic.com/technology/archive/2014/09/facebooks-mood-manipulation-experiment-might-be-illegal/380717/) [3] “No free lunch in data privacy,” in Proceedings of SIGMOD 2011 [23] M. S. Alvim, K. Chatzikokolakis, C. Palamidessi, and and PODS 2011, ser. Proceedings of the ACM SIGMOD International A. Pazii, “Metric-based local differential privacy for statistical Conference on Management of Data. Association for Computing applications,” CoRR, vol. abs/1805.01456, 2018. [Online]. Available: Machinery, 2011, pp. 193–204. [http://arxiv.org/abs/1805.01456](http://arxiv.org/abs/1805.01456) [4] C. Dwork, F. McSherry, and K. Nissim, “Calibrating noise to sensitivity [24] B. Jiang, M. Li, and R. Tandon, “Context-Aware data aggregation with in private data analysis,” in Theory of Cryptography: Third Theory localized information privacy,” in 2018 IEEE Conference on Communiof Cryptography Conference, 2006, pp. 265–284. [Online]. Available: cations and Network Security (CNS), May 2018. [https://doi.org/10.1007/11681878 14](https://doi.org/10.1007/11681878_14) [25] ——, “Local information privacy with bounded prior,” in ICC 2019 [5] C. Dwork, “Differential privacy,” in Automata, Languages and 2019 IEEE International Conference on Communications (ICC), May Programming: 33rd International Colloquium, ICALP 2006, Part II, 2019, pp. 1–7. M. Bugliesi, B. Preneel, V. Sassone, and I. Wegener, Eds., 2006, pp. [26] ——, “Local information privacy and its application to privacy[1–12. [Online]. Available: https://doi.org/10.1007/11787006 1](https://doi.org/10.1007/11787006_1) preserving data aggregation,” IEEE Transactions on Dependable and [6] Y. Xiao and L. Xiong, “Protecting locations with differential privacy Secure Computing, vol. 19, no. 3, pp. 1918–1935, 2022. under temporal correlations,” in Proceedings of the 22nd ACM SIGSAC [27] ——, “Online context-aware data release with sequence information Conference on Computer and Communications Security, ser. CCS ’15, privacy,” 2023. New York, NY, USA, 2015, p. 1298–1309. [28] B. Jiang, M. Seif, R. Tandon, and M. Li, “Context-aware local informa [7] W. Primoff and S. Kess, “The equifax data breach: What cpas and firms tion privacy,” IEEE Transactions on Information Forensics and Security, need to know now,” The CPA Journal, vol. 87, no. 12, pp. 14–17, 2017. vol. 16, pp. 3694–3708, 2021. [8] J. Lu, “Assessing the cost, legal fallout of capital one data breach,” Legal [29] M. E. Gursoy, A. Tamersoy, S. Truex, W. Wei, and L. Liu, “Secure Fallout Of Capital One Data Breach (August 15, 2019), 2019. and utility-aware data collection with condensed local differential [9] C. Dwork, “Differential privacy: A survey of results,” in Theory and privacy,” CoRR, vol. abs/1905.06361, 2019. [Online]. Available: Applications of Models of Computation: 5th International Conference, [http://arxiv.org/abs/1905.06361](http://arxiv.org/abs/1905.06361) TAMC, M. Agrawal, D. Du, and Z. Duan, Eds., 2008, pp. 1–19. [30] Y. NIE, W. Yang, L. Huang, X. Xie, Z. Zhao, and S. Wang, “A utility [[Online]. Available: https://doi.org/10.1007/978-3-540-79228-4 1](https://doi.org/10.1007/978-3-540-79228-4_1) optimized framework for personalized private histogram estimation,” [10] ´Ulfar Erlingsson, V. Pihur, and A. Korolova, “Rappor: IEEE Transactions on Knowledge and Data Engineering, vol. 31, no. 4, Randomized aggregatable privacy-preserving ordinal response,” in pp. 655–669, 2019. Proceedings of the 21st ACM CCS, 2014. [Online]. Available: [31] T. Murakami and Y. Kawamoto, “Utility-optimized local differential [https://arxiv.org/abs/1407.6981](https://arxiv.org/abs/1407.6981) privacy mechanisms for distribution estimation,” in Proceedings of the [11] A. Greenberg, “Apple’s ‘differential privacy’ is about collecting 28th USENIX Conference on Security Symposium, ser. SEC’19. USA: your data—but not your data,” 2016. [Online]. Available: USENIX Association, 2019, p. 1877–1894. [https://www.wired.com/2016/06/apples-differential-privacy-collecting-data/](https://www.wired.com/2016/06/apples-differential-privacy-collecting-data/) [32] S. Takagi, Y. Cao, and M. Yoshikawa, “Poster: Data collection via [12] T. Wang, J. Blocki, N. Li, and S. Jha, “Locally differentially local differential privacy with secret parameters,” in Proceedings of private protocols for frequency estimation,” in 26th USENIX Security the 15th ACM Asia Conference on Computer and Communications 17. USENIX Association, 2017, pp. 729–745. [Online]. Available: Security, ser. ASIA CCS ’20. New York, NY, USA: Association [https://www.usenix.org/conference/usenixsecurity17/technical-sessions/presentation/wang-tianhaofor Computing Machinery, 2020, p. 910–912. [Online]. Available:](https://www.usenix.org/conference/usenixsecurity17/technical-sessions/presentation/wang-tianhao) [13] Z. Qin, Y. Yang, and T. Yu, “Heavy hitter estimation over set-valued [https://doi.org/10.1145/3320269.3405441](https://doi.org/10.1145/3320269.3405441) data with local differential privacy,” in Proceedings of the 2016 ACM SIGSAC, ser. CCS ’16, 2016, pp. 192–203. [Online]. Available: [http://doi.acm.org/10.1145/2976749.2978409](http://doi.acm.org/10.1145/2976749.2978409) [14] C. Huang, P. Kairouz, X. Chen, L. Sankar, and R. Rajagopal, “Contextaware generative adversarial privacy,” CoRR, vol. abs/1710.09549, [2017. [Online]. Available: http://arxiv.org/abs/1710.09549](http://arxiv.org/abs/1710.09549) [15] P. Kairouz, S. Oh, and P. Viswanath, “The composition theorem for differential privacy,” IEEE Transactions on Information Theory, vol. 63, 6 4037 4049 2017 -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2309.00861, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "CLOSED", "url": "https://arxiv.org/pdf/2309.00861" }
2,023
[ "JournalArticle", "Review" ]
true
2023-09-02T00:00:00
[]
8,538
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/014d53aa1795a4e229037afe328ae014e5e2c18f
[ "Computer Science" ]
0.841913
AmbientDB: P2P data management middleware for ambient intelligence
014d53aa1795a4e229037afe328ae014e5e2c18f
IEEE Annual Conference on Pervasive Computing and Communications Workshops, 2004. Proceedings of the Second
[ { "authorId": "1684284", "name": "Willem Fontijn" }, { "authorId": "1687211", "name": "P. Boncz" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
# AmbientDB: P2P Data Management Middleware for Ambient Intelligence ## Willem Fontijn Peter Boncz Philips Research CWI willem.fontijn@philips.com boncz@cwi.nl Abstract system has to present to the user a unified and consistent _The future generation of consumer electronics devices is_ view irrespective of the context and location of the user or _envisioned to provide automatic cooperation between_ the type of interaction. In an ad-hoc mobile environment, _devices and run applications that are sensitive to people's_ this requires novel synchronization procedures, _likings, personalized to their requirements, anticipatory of_ transparent to the user [4]. _their behavior and responsive to their presence. We see_ The _Connected Home [5] may be seen as a first step_ _this ‘Ambient Intelligence’ as a key feature of future_ towards AmI. Devices contain their own embedded _pervasive computing. We focus here on one of the_ DBMS, and operate in isolation below the network layer _challenges_ _in_ _realizing_ _this_ _vision:_ _information_ (see Fig.1a). If one device requires information located on _management._ _This_ _entails_ _integrating,_ _querying,_ another device it will first have to find the latter device _synchronizing and evolving structured data, on a_ and query it for the information. When the complexity of _heterogeneous and ad-hoc collection of (mobile) devices._ the network increases, i.e. more and more diverse devices _Rather than hard-coding data management functionality_ are introduced, the increasing number of possible _in each individual application, we argue for adding high-_ combinations will make it hard to create robust _level data management functionalities to the distributed_ applications. Also, the performance of such a system will _middleware layer. Our AmbientDB P2P database_ degrade rapidly. This could be countered by pre-emptive _management system addresses this by providing a global_ data aggregation on resource-rich central servers, but this _database abstraction over an ad-hoc network of_ has multiple drawbacks. It creates mounting overhead as _heterogeneous peers._ the complexity increases, implies that certain functionality is available only when in range of a server and makes the system vulnerable to point failures. Finally, it is questionable from a marketing viewpoint whether ## 1. Introduction customers will buy an expensive home-server that just improves the performance of other devices. Future generations of consumer electronic devices will An alternative to the central server approach is to keep the make computing power and connectivity omnipresent, yet participating devices fully autonomous but cooperative. A hidden from the users view. This pervasive computing collection of such devices we call a Device Society. Each infrastructure will be used to create an environment that device has its own responsibility to deliver certain anticipates users wishes instead of just responding to functionality and the collection as a whole delivers the direct commands. The aim is to improve quality of life by environment to run high-level applications. In such a creating the desired atmosphere and functionality via system, most data is stored distributed and only integrated personalized, interconnected systems and services. This at query time. Such distributed storage is robust and vision, called _Ambient Intelligence (AmI), is the focal_ point of much ongoing research [1]. The ‘Ambient’ part of (a) **Application** **Application** **Application** AmI refers to the unobtrusiveness of the technology, both **Network Middleware** physically, by embedding it in the environment, and **DBMS** **DBMS** **DBMS** functionally, by making user interaction mostly implicit. **Hardware** **Hardware** **Hardware** The latter will entail, for instance, habit watching [3]. The (b) ‘Intelligence’ part of AmI refers to the way the system **Application** **Application** **Application** **AmbientDB** integrates and relates the data from a wide variety of **Network Middleware** sources to create the perception of intelligence. The **DBMS** **DBMS** **DBMS** **DBMS** sources range from simple sensor nodes to Personal Video **Hardware** **Hardware** **Hardware** **Hardware** Recorders (PVRs) with sophisticated preference-based **Figure 1: Concept of (a) Connected Home, (b)** meta data. Perceived intelligence implies that the AmI **Ambient Intelligence environment** |Application|Col2|Application|Col4|Application| |---|---|---|---|---| |Network Middleware||||| |DBMS||DBMS||DBMS| |Hardware||Hardware||Hardware| |Col1|Application|Col3|Application|Col5|Col6|Col7|Application|Col9| |---|---|---|---|---|---|---|---|---| |AmbientDB||||||||| |||||||||| |Network Middleware||||||||| |DBMS||||DBMS|||DBMS|| |Hardware||||Hardware|||Hardware|| |DBMS|Col2| |---|---| |Hardware|| ----- scalable but data management is not easy anymore. In this paper we discuss a data management approach for the latter strategy and present a prototype system. We first illustrate its requirements using a scenario. In Section 2.2 a technical realization is presented. ### 1.1. Scenario: Music Playlist Generation Consider the problem of playing music for a user that is appropriate to his context and state irrespective of his location. All music owned by this particular user, as well as meta data describing this music is stored in a music database distributed over one or more devices, such as portable music (mp3) players, Smart Phone, PVR and/or PC. To the user, this music collection should appear conceptually as a single collection. Three autonomous processes are active on this music database. The first aims to improve the profile of the user. The user can actively rate pieces of music, indicating which songs he likes or dislikes. This _explicit rating is_ useful to establish a first draft profile but requires direct user interaction. A more convenient way to improve the profile is implicit rating. The system logs the reaction of the user to different pieces of music in different contexts and the logs are later processed to derive preferences of the user in various contexts. The second process aims to extend the music collection. The profile of the user may be compared to profiles of others, using collaborative filtering [3]. If one user likes many songs another user likes, the first may also like songs unknown to him that the second user likes. The third process recommends music. Based on the context, the user is offered a selection of his total music collection. The context of the user may be derived from many clues, e.g. ambient lighting, time of day and facial expression. If the user leaves the house and takes a portable music store with him, the system assists him in selecting the music subset that will probably meet his musical needs for the particular trip. While away, the portable music device records implicit rating cues and may pick up some additional pieces of music. This system should not be static but able to evolve. If the user buys a watch that senses his skin temperature, the music recommender should discover that by combining this new context information with other data, it could gauge the mood of the user. From that moment on, the recommender takes mood into account. ### 1.2. The idea: database middleware Our challenge of managing a sea of evolving data sources among possibly large and dynamic Device Societies should be addressed in the middleware layer (see Fig. 1b). Middleware services already provide the basic infrastructure for application integration [9], integrating not only data but also a wide variety of context sources [6], providing applications with information about network and device resources, and allowing applications to reconfigure when conditions change [5]. While there have been middleware systems with some data management support [8][9] we aim to raise this to the level of full DBMS functionality. By putting such DBMS functionality inside or on top of the middleware layer, all data sources are virtually merged, shielding applications from the underlying complexities. Applications release their queries and get the results as if accessing a single database system. Events such as the integration of new devices or data sources, or failure of devices, and the creation of new data types, can in great part be handled as schema evolution in this database. Accepting queries in a high-level language that describes what data an application needs instead of how to exactly obtain its results, allows a _query optimizer rather_ than the application programmer to automatically find an efficient solution for executing a query, taking into account indexing structures, system and network load conditions and concurrent requests. Also, this provides _data independence, meaning that the data representation_ can change over time without this breaking the applications using the data. These are the classic database advantages that we are now able to bring into the pervasive computing domain. Finally, this approach enables evolutionary introduction of new functionality. By supplementing middleware for the Connected Home with DBMS functionality, we create a breeding ground for more and more sophisticated and AmI applications. ## 2. AmbientDB: a P2P DBMS The goal of our AmbientDB system [10] is to provide full relational database functionality for standalone operation in autonomous devices that may be mobile and disconnected for long periods of time, while enabling them to cooperate in an ad-hoc way with (many) other AmbientDB devices. Hence, our choice for P2P, as opposed to designs that use a central server. AmbientDB uses ‘abstract’ tables, i.e. applications are ignorant of where data resides. Internally, a table may be private to the node, or distributed over many nodes in the network. The actual content of a distributed table is formed by the union of table partitions in all nodes that are connected at that time. ### 2.1. Key Functionalities Since our work touches upon many sub-fields of database research [7], we highlight the main differences. ----- **Figure 2: Concept of AmbientDB. The application** **on top issues a query to the AmbientDB layer that** **is propagated (dashed line) to all connected peers.** **The query result (solid line) is aggregated along** **the query path and presented to the application.** **The binary tree in the network layer represents the** **network topology.** _Distributed database technology presupposes that the_ collection of participating sites and communication topology is known a priori. AmbientDB does not. _Federated database technology, the current approach to_ heterogeneous schema integration, focuses on statically configured combinations of databases instead of ad-hoc Device Societies. _Mobile database technology_ generally assumes that mobile nodes are (weaker) clients that synchronize with a centralized database server over a narrow channel. Again, AmbientDB does not. Finally, _P2P file sharing systems do support decentralized, ad-hoc_ Device Societies, but allow only simple keyword text search (as opposed to structured database queries). **2.1.1. Self Organization** P2P technologies are able to adapt to changes in the environment and work without central planning. In order to provide efficient indexed lookup into its distributed database tables, AmbientDB makes use of Chord [11]. Chord is a Distributed Hash Table (DHT), a scalable P2P data structure for sharing data among a potentially large collection of nodes, allowing nodes to join and leave without making the network unstable. It uniformly distributes data over all nodes using a hash-function, enabling efficient _O(log(N)) data lookup. To improve_ scalability in situations where some devices are resourcepoor, AmbientDB keeps devices out of Chord to prevent overloading them with data they cannot store or with queries they cannot handle. Upon connection, lowresource nodes transfer their data to a resource-rich neighbor that handles queries on behalf of them. **2.1.2. Query Processing** AmbientDB performs a three level query translation: _(1) abstract algebra: A user query is posed in the_ “abstract global algebra”. This is a standard relational query language, providing the basic operators for selection, join, aggregation and sorting. _(2) concrete algebra: These are concrete strategies for_ resolving the basic relational operators. Typically, each abstract operator has multiple concrete variants. E.g., there is a broadcast-select, that executes a selection operator on a distributed table by flooding the network (broadcast) and collecting all matches. There is also a variant that exploits a Chord DHT index, which may be used if a global index on a table column was defined in the schema. Thus, many different concrete plans may exist for an ‘abstract’ query, and the query optimizer in AmbientDB is used to select a good plan. _(3) dataflow algebra: A very small kernel of basic_ operators is sufficient to implement the concrete algebra. Each concrete operator is mapped onto a _wave-plan that_ consists of a graph of dataflow operators. Next to query processing the dataflow operators provide functionality for splitting and merging data streams. We plan to augment AmbientDB with support for triggers, such that applications can be alerted to interesting events rather than poll the global database with queries [12]. **2.1.3. Synchronization** The aim of traditional (distributed) database technology, to provide strict consistency, is not appropriate for P2P database systems. Algorithms, such as two-phase locking, are too expensive for a large and sparsely connected collection of nodes. Many applications do not need full transactional consistency, but just a notion of final convergence of updates. Also, applications often have effective conflict resolution strategies that exploit application-level semantics. Thus, the challenge for a P2P DBMS is to provide a powerful formalism in which applications can formulate synchronization and conflict resolution strategies. Our first target is to support applications that use rule-based synchronization expressed in a prioritized set of database update queries. **2.1.4. Schema Integration & Evolution** As devices differ in functionality and make, their data differs in semantics and form. We use table-view based schema integration techniques [13] to map local schemata to a global schema. AmbientDB itself does not address the automatic construction of such mappings, but aims at providing the basic functionality for applying, stacking, sharing, evolving and propagating such mappings. Providing support for schema evolution within one schema, e.g. such that old devices can cooperate with newer ones, is often forgotten. We foresee that a global certifying entity keeps track of changes in the various subschemas, maintaining bi-directional mappings between versions. Schema deltas are certified such that one peer ----- may carry it to the next, without need for direct communication with a centralized entity. ### 2.2. Scenario using P2P data management The gist of our approach is that we believe that P2P data management functionality will make it easier to construct Ambient Intelligent applications. To illustrate how we see that happen, let us go back to the problem of managing and navigating music intelligently. The schema created by the music player contains a LOG table where per-user song play counts are kept (see Fig. 3). This is a _distributed table, which means that the_ music application sees the union of all (overlapping) horizontal fragments at all participating devices of that moment as one big table. All devices maintain local playcounts for each (artist,user) combination in this LOG table. The schema specifies an _index on LOG.artist, so_ each LOG entry is replicated in a Chord DHT and distributed over all nodes of the Internet domain, using the Chord hashing scheme (see Fig. 3). This allows to quickly locate users that played a particular artist. **2.2.1. Self Organization Example** The family music collection -- typically in the order of a few thousands of songs -- is distributed among the Device Society owned by family members. Some of these devices may have access to the Internet. The music players with embedded AmbientDBs form a self-organizing P2P network, connecting the nodes in order to share all music content in the "home domain", and a second -possibly huge- P2P network consisting of all music players reachable via the Internet, among which only the meta``` create distributed table create distributed table ``` LizLiz U2U2 77 AA `LOG(user,artist,count,…)LOG(user,artist,count,…)` RobRob U2U2 22 AA `create index` AnnAnn U2U2 55 EE `IDX(user,artist,count)` Liz U2 7 other `on LOG(artist)` Rob U2 2 fields.. Ann blur 9 _global table_ _Chord_ A LizLizRob U2U2U2 772 otherotherfields.. AnnAnn blurblur 99 AA Ann blur 9 RobRob blurblur 44 EE Ann U2 5 LizLiz blurblur 66 CC Rob blur 4 Rob rem 1 Ann U2 5 other Liz blur 6 Rob blur 4 fields.. Ann rem 3 Rob rem 1 E Rob rem 1 Liz rem 8 AnnAnnAnnLiz remremremrem 3338 CCCC C E _Find(‘blur’) :=_ LizLizAnn blurblurrem 663 fields..fields..other _hash(‘blur’)hash(‘blur’)��D_ Liz rem 8 _chord_lookup(D)chord_lookup(D)��E_ **Fig.3: Scenario for sharing music metadata** **between many music players in the Internet** **domain. The distributed table LOG holds artist** **play-counts for each user. One can quickly find** **users that play a certain artist using a Chord index.** information is shared. The home domain may contain some very low-resource devices in terms of CPU and storage (e.g. phone) that are kept out of the Chord DHT. In the Internet domain, the number of on-line nodes maybe large and the number of songs huge. **2.2.2. Schema Evolution Example** In our scenario, the user buys a watch with integrated body thermometer. This watch has Body Area Network (BAN) functionality (e.g. Bluetooth) such that it can communicate with the owner’s phone or mp3 player when these are carried in his pocket. With the temperature meter watch comes an AmbientDB _schema update that e.g._ introduces a new TEMPERATURE table that stores (timestamp, temperature) records, and a data propagation _profile with rules that specify the longevity of its records_ and a propagation strategy. Additionally, on a certified (vendor) site, the user community of the music player may store a trigger update that specifies a (complex) rule that derives a mood from the body temperature curve in conjunction with other personal characteristics stemming from other sources. When this mood indicates appreciation for the current song, an automatic playlist creation process is scheduled, aggregating songs similar to the one currently being played (this query is described in Section 2.2.4). Note that schema updates may propagate in a P2P fashion (from watch to phone, from phone to home PC) or from a central Internet site, in any case though with a certifying mechanism. Also, schema updates may depend on a collection of sub-schema versions being present, such that during the next visit to the central vendor site, when the combination of music and temperature sub-schemas is detected, the user is alerted to the possibility of installing the "music-appreciation-trigger". **2.2.3. Update Propagation Example** The watch has a limited storage capacity, it can hold only a few records. Its synchronization rules, however, make it replicate temperature records to devices in the neighborhood (e.g. your Smart Phone). When it arrives at home, the Smart Phone then propagates these records to the home PC, where a health-monitoring agent might be running that periodically analyzes this log data using data mining techniques. The propagation rules may include a maximum lifetime that causes old records to be automatically deleted after e.g. a number of weeks. **2.2.4. Query Processing Example** The music player generates intelligent playlists, either because the user explicitly chooses an sample artist to generate a similar playlist from, or implicitly when the “music-appreciation-trigger” notices that you like an artist |LLLiiizzz RRRooobbb AAAnnnnnn|UUU222 UUU222 UUU222|777 222 555|AAA AAA EEE| |---|---|---|---| |LLiizz RRoobb AAnnnn|UU22 UU22 bblluurr|77 22 99|ootthheerr ffiieellddss....| |A|Col2|Col3|Col4| |---|---|---|---| |AAAnnnnnn RRRooobbb LLLiiizzz|bbbllluuurrr bbbllluuurrr bbbllluuurrr|999 444 666|AAA EEE CCC| |AAnnnn RRoobb RRoobb|UU22 bblluurr rreemm|55 44 11|ootthheerr ffiieellddss....| |LLiizz RRoobb AAnnnn|UU22 UU22 bblluurr|77 22 99|ootthheerr ffiieellddss....| |---|---|---|---| |AAnnnn RRoobb RRoobb|UU22 bblluurr rreemm|55 44 11|| |LLiizz AAnnnn LLiizz|bblluurr rreemm rreemm|66 33 88|| |ccrreeaattee ddiissttrriibbuutteedd ttaabbllee LLOOGG((uusseerr,,aarrttiisstt,,ccoouunntt,,……)) LLLiiizzz UUU222 777AAA RRRooobbbUUU222 222AAA ccrreeaattee iinnddeexx AAAnnnnnnUUU222 555EEE IIDDXX((uusseerr,,aarrttiisstt,,ccoouunntt)) LLiizz UU22 77 ootthheerr oonn LLOOGG((aarrttiisstt)) RRoobbUU22 22 ffiieellddss.... AAnnnnbblluurr99 gglloobbaall ttaabbllee A LLiizz UU22 77 ootthheerr Chord RRoobbUU22 22 ffiieellddss.... AAAnnnnnnbbbllluuurrr999AAA AAnnnnbblluurr99 RRRooobbbbbbllluuurrr444EEE AAnnnnUU22 55 LLLiiizzz bbbllluuurrr666CCC RRoobbbblluurr44 RRoobbrreemm11 AAnnnnUU22 55 ootthheerr LLiizz bblluurr66 RRoobbbblluurr44 ffiieellddss.... AAnnnnrreemm33 RRRooobbbrrreeemmm111EEE RRoobbrreemm11 LLiizz rreemm88 AAAnnnnnnrrreeemmm333CCC C E LLLiiizzz rrreeemmm888CCC FFiinndd((‘‘bblluurr’’)) ::== LLiizz bblluurr66 ootthheerr hhaasshh((‘‘bblluurr’’))DD AAnnnnrreemm33 ffiieellddss.... LLiizz rreemm88 cchhoorrdd__llooookkuupp((DD))EE|Col2|Col3|Col4| |---|---|---|---| |RRRooobbb AAAnnnnnn LLLiiizzz|rrreeemmm rrreeemmm rrreeemmm|111 333 888|EEE CCC CCC| |LLiizz AAnnnn LLiizz|bblluurr rreemm rreemm|66 33 88|ootthheerr ffiieellddss....| ----- being played. The two database queries below express a simple collaborative filtering method. The first query computes a relevance of other users’ music taste from their play-count of an sample artist. The second query then computes a ranking by multiplying all artist playcounts of all users by the user relevance, and summing this per artist, returning a top-N. We kept this example very simple for presentation purposes, but one can easily refine it, e.g. by increasing the granularity to songs (instead of artists) or making it work with a weighted collection of samples instead of one. The benefit is that for the application programmer, writing this kind of data-intensive applications on large ad-hoc networks is reduced to writing some relatively simple database queries. Database indices and query optimization then make sure that it runs efficiently without the application programmer having to worry about it [10]. ``` % each record has the special field #nodeid % that holds the device ID where it is stored RELEVANT := % query will use Chord index SELECT user, SUM(playcount) AS relevance, #nodeid AS site FROM LOG WHERE artist = ‘normalized artist name’ GROUP BY user ORDER BY relevance DESCENDING LIMIT 5 SELECT L.artist AS artist, sum(L.playcount*R.relevance) AS score FROM LOG L, RELEVANT R WHERE L.user = R.user AND L.#nodeid = R.site GROUP BY artist ORDER BY score DESCENDING LIMIT 25 ## 3. Current Status & Research Challenges ``` We hope to release a first version of AmbientDB early next year. We have focused so far on distributed query processing, and identified three functional levels that all require further research. On the top level, we need more experience with a wider variety of Ambient Intelligent applications to see what exact requirements they impose on a P2P DBMS. Also, if applications are to cooperate seamlessly, they need to operate in a compatible semantic framework. This is an "AI-hard" problem for the general case. We do see possibilities when trusted and standardized mappings are available. The second functional level is P2P data management. While we have a working query processor, it is likely that there are query execution algorithms that exploit the P2P architecture better. Also, loosely consistent or converging transactions as well as a schema mapping infrastructure remain open areas of research. The third functional level is P2P networking. P2P overlay technology often exhibits inefficient usage of physical resources, as these are opaque on the TCP overlay level. This could be improved by dynamic re-configuration of P2P networks, an important middleware research issue [5]. Better adaptation to device and network resources using e.g. slave- and super-nodes could be ways forward here. ## 4. Conclusions Transparent distributed data management is crucial to Ambient Intelligent applications in Device Societies, and the P2P approach with AmbientDB as middleware offers a possible solution. It enables the creation of a high-level application development interface that is flexible and provides data independence, while taking the burden of data management optimization in a dynamic and ad-hoc distributed environment out of the hands of application programmers. The ability of AmbientDB to cope with adding devices and functionality dynamically provides for an evolutionary path for the introduction of Ambient Intelligence, thus alleviating one of the most prominent problems from a systems and marketing point of view. ## 5. References [1] www.philips.com/research/ami [2] www.semiconductors.philips.com/connected_home [3] D. Nichols. Implicit Rating and Filtering, Proc. DELOS Workshop on Filtering and Collaborative Filtering, 1998. [4] G. Montenegro. MNCRS: Industry Specifications for the Mobile NC, IEEE Internet Computing, 1998. [5] M. Roman, F. Kon, R. Campbell. Design and Implementation of Runtime Reflection in Communication Middleware: the dynamicTAO Case. ICDCS Workshop on Middleware, 1999. [6] A. Schmidt, M. Beigl, H.-W. Gellersen. There is more to context than location. Proc. of the Intl. Workshop on Interactive Applications of Mobile Computing, 1998. [7] D. Kossmann. The state of the art in distributed query processing. ACM Computing Surveys, 32(4), 2000. [8] G. Picco, A. Murphy, G.-C. Roman. On Global Virtual Data Structures. In: _Process Coordination and Ubiquitous_ _Computing, D. Marinescu, C. Lee, CRC Press._ [9] J. Carter, A. Ranganathan, S. Susarla. Khazana: An infrastructure for building distributed services. In Proc. Int. Conf. on Distributed Computing Systems (ICDCS’98), 1998. [10] P. Boncz, C. Treijtel. AmbientDB: relational query processing in a P2P network. Proc. Workshop On Databases, Information Systems and P2P Computing (at VLDB’03), 2003. [11] I. Stoica, R. Morris, D. Karger, M. Kaashoek, and H. Balakrishnan. Chord: A scalable peer-to-peer lookup protocol for Internet applications. Proc. SIGCOMM Conf., 2001. [12] J. Widom, S. Ceri: Active Database Systems: Triggers and Rules For Advanced Database Processing. Morgan Kaufmann. [13] A. Halevy: Answering queries using views: A survey. VLDB Journal 10(4): 270-294, 2001. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/PERCOMW.2004.1276932?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/PERCOMW.2004.1276932, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "other-oa", "status": "GREEN", "url": "https://ir.cwi.nl/pub/11118/11118B.pdf" }
2,004
[ "JournalArticle", "Conference" ]
true
2004-03-14T00:00:00
[]
6,993
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Mathematics", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/014f27c5272292f84a182c49ca98f873fba06ae5
[ "Computer Science", "Mathematics" ]
0.845528
Detection of Insider Attacks in Distributed Projected Subgradient Algorithms
014f27c5272292f84a182c49ca98f873fba06ae5
IEEE Transactions on Cognitive Communications and Networking
[ { "authorId": "1702905", "name": "Sissi Xiaoxiao Wu" }, { "authorId": "46439532", "name": "Gangqiang Li" }, { "authorId": "1817363", "name": "Shengli Zhang" }, { "authorId": "144591140", "name": "Xiaohui Lin" } ]
{ "alternate_issns": null, "alternate_names": [ "IEEE Trans Cogn Commun Netw" ], "alternate_urls": [ "https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=6687307" ], "id": "65e58b80-9699-4da6-bd60-929b57b8533d", "issn": "2332-7731", "name": "IEEE Transactions on Cognitive Communications and Networking", "type": "journal", "url": "http://ieeexplore.ieee.org/servlet/opac?punumber=6687307" }
The gossip-based distributed projected subgradient algorithms (DPS) are widely used to solve decentralized optimization problems in various multi-agent applications, while they are generally vulnerable to data injection attacks by internal malicious agents as each agent locally estimates its descent direction without an authorized supervision. In this work, we explore the application of artificial intelligence (AI) technologies to detect internal attacks. We show that a general neural network is particularly suitable for detecting and localizing the malicious agents, as they can effectively explore nonlinear relationship underlying the collected data. Moreover, we propose to adopt one of the state-of-the-art approaches in decentralized federated learning, i.e., a gossip-based collaborative learning protocol, to facilitate training the neural network models by gossip exchanges. This advanced approach is expected to make our model more robust to challenges with insufficient training data, or mismatched test data. In our simulations, a least-squared problem is considered to verify the feasibility and effectiveness of AI-based methods. Simulation results demonstrate that the proposed AI-based methods are beneficial to improve performance of detecting and localizing malicious agents over score-based methods, and the peer-to-peer neural network model is indeed robust to target issues.
## Detection of Insider Attacks in Distributed Projected Subgradient Algorithms #### Sissi Xiaoxiao Wu, Gangqiang Li, Shengli Zhang, and Xiaohui Lin **_Abstract—The gossip-based distributed algorithms are widely_** **used to solve decentralized optimization problems in various** **multi-agent applications, while they are generally vulnerable** **to data injection attacks by internal malicious agents as each** **agent locally estimates its decent direction without an authorized** **supervision. In this work, we explore the application of artificial** **intelligence (AI) technologies to detect internal attacks. We** **show that a general neural network is particularly suitable** **for detecting and localizing the malicious agents, as they can** **effectively explore nonlinear relationship underlying the collected** **data. Moreover, we propose to adopt one of the state-of-art** **approaches in federated learning, i.e., a collaborative peer-to-** **peer machine learning protocol, to facilitate training our neural** **network models by gossip exchanges. This advanced approach** **is expected to make our model more robust to challenges** **with insufficient training data, or mismatched test data. In our** **simulations, a least-squared problem is considered to verify the** **feasibility and effectiveness of AI-based methods. Simulation** **results demonstrate that the proposed AI-based methods are** **beneficial to improve performance of detecting and localizing** **malicious agents over score-based methods, and the peer-to-peer** **neural network model is indeed robust to target issues.** **_Index Terms—Gossip algorithms, distributed projected sub-_** **gradient (DPS), artificial intelligence (AI) technology, internal** **attacks, malicious agents.** I. INTRODUCTION ECENTLY, decentralized optimization algorithm as a popular tool to handle large scale computations has been # R broadly applied in various fields [1], [2]. Typical examples of Internet of Things (IoT) [3], [4], multi-agent systems [5], [6], wireless communications network [7], power grid [8], and federated learning [9]. The design approach in above applications is often refereed as gossip-based optimization problems, wherein interacting agents are randomly selected and exchange information following a point-to-point message passing protocol so as to optimize shared variables. Aiming at a coordinated response, these agents explicitly disclose their estimates (states) to neighboring agents in each iteration, thereby leading to a consistent globally optimal decision [1], [2], [10]. It is well known that gossip-based algorithms are inherently robust to intermittent communication and builtin fault-tolerance to agent failures. They can also provide a degree of privacy in many applications for participating This work is supported by the National Natural Science Foundation of China under Grant 61701315; by Shenzhen Technology R&D Fund JCYJ20170817101149906 and JCYJ20190808120415286; by Shenzhen University Launch Fund 2018018. S. X. Wu, G. Li, S. Zhang and X. Lin are with the College of Electronics and Information Engineering, Shenzhen University, Shenzhen, China. G. Li is the corresponding author. E-mails: ligangqiang2017@email.szu.edu.cn, i i l hli@ d agents without exchanging local user information [11]. Despite many advantages, these gossip-based algorithms, such as the distributed projected subgradient (DPS) algorithm [2], are inherently vulnerable to insider data injection attacks due to the flat architecture, since each agent locally estimates its (sub)gradient without any supervision [12]–[14]. Generally speaking, malicious agents’ (or attackers) attack on decentralized algorithms depends on specific attacking strategies. Attackers may interfere with distributed algorithms by injecting random data that hinders convergence [15]. Especially in an insider attack, the attacker always sends misleading messages to its neighbors to affect the distributed system, resulting in false convergence results [16], [17]. For example, a multi-agent system is forced to converge to the target values of attackers in [18] and an average consensus result is disturbed by coordinated attackers in [19]. The attack model we focus on in this work is that the attacker behaves like stubborn agents [20]. To be more specific, they coordinate and send messages to peers that contain a constant bias [12], [13], [17], [21] and their states can not be changed by other agents. As studied in [18], [19], the network always converges to a final state equal to the bias. This will bring serious security problems to distributed algorithms if the attacker cannot be detected effectively. Thus, a good defense mechanism is needed to protect these algorithms from internal data injection attacks. To detect anomalous behaviors in decentralized optimization, one commonly used approach in the literature is to calculate a dependent score through statistical techniques based on the messages received during the execution of protocol. For instance, in [15], the authors show that the convergence speed of network will slow down when the attacker is present, and design a score to identify potential attacks. In [19], two score-based detection strategies are proposed to protect the randomized average consensus gossip algorithm from malicious nodes. In [22], the authors design a comparison score to search for differences between a node and its neighbors, then adjust update rules to mitigate the impact of data falsification attacks. In [13], the decision score is computed by a temporal difference strategy to detect and isolate attackers. Similar score-based methods are also shown in [23]–[25]. While such methods have reasonable performance, the score design is somewhat ad-hoc and relies heavily on the experts to design sophisticated decision functions, and the detection thresholds of these score-based methods need to be adjusted judiciously. To circumvent the above difficulties, our idea in this work is to utilize the artificial intelligent (AI) technology to approximate more sophisticated decision functions. It is worth mentioning that AI technology has succeed in many applications with ----- the same purpose, including image recognition [26], natural language processing [27], power grid [28], and communications [29]. Furthermore, AI also plays an important role in network security [30], such as anomaly intrusion detection [31], malicious PowerShell [32], distributed denial of service (DDoS) attacks [33] and malicious nodes in communication networks [34], [35]. The main purpose of this work is to apply AI technology to address the problem of detecting and localizing attackers in decentralized gossip based optimization algorithms. While our AI-based methods and training philosophy can be applied to a wide set of multi-agent algorithms and attack scenarios, we focus on testing the approach on a case that has been thoroughly studied in [13], [36], to facilitate the comparison. Concretely, we proposed two AI-based strategies, namely the temporal difference strategy via neural networks (TDNN) and the spatial difference strategy via neural networks (SDNN). We will show that even basic neural network (NN) models exhibit a good ability to extract non-linearity from our training data and thus can well detect and localize attackers, given that 1) the collected training data can well represent the attack model, and 2) training data from all agents can be fully learned at the training center. Unfortunately, collecting good and enough data which perfectly fits the real attack model is usually difficult. First of all, due to the intrinsic of gossp-algorithm, it is difficult and expensive to collect sufficient training samples at each agent. Also, with the emergence of new large-scale distributed agents in the network, sometimes it is hard to upload decentralized data at each agent to a fusion center due to storage and bandwidth limitations [37]. Furthermore, as the insider attacks could occur at any agent in the network, the training data may not cover all the occurrences of the attack. Therefore, some individually trained NN model at each agent may not fit in all insider attack events. A new approach to alleviate these issues is to leverage the decentralized federated learning [38], which utilizes the collaboration of agents to perform data operations inside the network by iterating local computations and mutual interactions. Such a learning architecture can be extremely useful for learning agents with access to only local/private data in a communication constrained environment [39]. Specially, as one of the state-of-the-art approach in decentralized federated learning, gossip learning is very suitable for training NN models from decentralized data sources [40], with the advantages of high scalability and privacy preservation. Thus, we propose a collaborative peer-to-peer learning protocol to help training our NN models by gossip exchanges. Specifically, each agent in the network has a local model with the same architecture, and only relies on local collaboration with neighbors to learn model parameters. It is worth noting that in this process each agent trains the local model by its local data periodically, and then send the local model parameters to its neighbors. It is expected that each agent can learn a local model close to the _global model (i.e., a NN trained by the center, which contains_ training data from all agents), so as to provide robustness in the case of insufficient and mismatched local data. It is also worth mentioning differences between this work and some previous work Previous work [19] aims at the score-based method for securing the gossip-based average consensus algorithm. [35] improves the score-based method by using AI-based methods while it still targeted at an average consensus algorithm. We remark that The inputs for AI model in [35] does not always work for optimization algorithms. [13], [36] provide some preliminary results for protecting optimization algorithms while it only focus on partial neighboring information. This work is the first one which well elaborates AI-based methods for a DPS algorithm using full information from neighboring signal. More importantly, the proposed collaborative learning method is novel and effective to make the defense model more robust to different events of attacks, making our models more practical to multi-agent applications. In summary, the proposed AI-based strategies have following characteristics: 1) they can automatically learn appropriate decision models from the training data, thus reducing the dependence on complicated pre-designed models; 2) they adaptively scale the decision thresholds between 0 and 1, which reduces the difficulty of threshold setting; 3) they improve the performance of detecting and localizing attackers and show good adaptability to different degree of agents. 4) they have strong robustness to the scenarios with insufficient training data, or mismatched training data. Preliminary numerical results demonstrate that the proposed AI-based strategies are conducive to solve the insider attack problem faced by the DPS algorithm. The rest of the paper is organized as follows. In Section II, we describe the decentralized multi-agent system and the attack scheme against the DPS algorithm. In Section III, we review score-based strategies and propose two AI-based defense strategies to detect and locate attackers. Section IV introduces a collaborative peer-to-peer training protocol for NN, dealing with insufficient samples or mismatched samples available on different agents. Simulation results are given in Section V to confirm the effectiveness of the proposed strategies. We conclude this work in section VI. II. SYSTEM MODEL We consider a multi-agent network which can be defined by an undirected graph = ( _,_ ), wherein = 1, _, n_ _G_ _V_ _E_ _V_ _{_ _· · ·_ _}_ represents the set of all agents and represents the _E ⊆V × V_ set of all edges. We define the set of the neighbor nodes of an agent i ∈V by Ni = {vj ∈V : (vi, vj) ∈E}. All the agents in the distributed network follow a gossip-based optimization protocol; see Algorithm 1. That is, in each iteration of information exchange, an agent only directly communicates with its neighbors. We thus define a time-varying graph as (t) := ( _,_ (t)) for the tth iteration and the associated _G_ _V_ _E_ weighted adjacency matrix is denoted by A(t) ∈ R[n][×][n], where [A(t)]ij := Aij(t) = 0 if (vj, vi) /∈E(t). For this network with n agents, we have the following assumption: **Assumption 1. There exists a scalar ζ ∈** (0, 1) such that for _all t_ 1 and i = 1, _, n :_ _≥_ _· · ·_ _• Aij(t) ≥_ _ζ if (i, j) ∈E(t),_ _•_ [�]i[n]=1 _[A][ij][(][t][) = 1][,][ A][ij][(][t][) =][ A][ji][(][t][)][;]_ _The graph (_ (t + ℓ)) is connected for B < _V ∪[B][0]_ _E_ _∞_ ----- **Algorithm 1 The gossip-based optimization protocol** **Input:** Number of instances K, and iterations T . **for k = 1, · · ·, K do** Initial states: x[k]i [(0) =][ β]i[k] _[∀]_ _[i][ ∈V]_ **for t = 1, · · ·, T do** Uniformly wake up a random agent i _•_ _∈V_ _• Agent i selects agent j ∈Ni with probability Pij_ _• The trustworthy agents i, j ∈Vt update the states_ according to the rules in (2). The malicious agents follow the attack scheme to _•_ keep their original states, as seen in (3). **end for** **end for** The goal of these agents is cooperatively to solve the following optimization problem: min **_x_** _[f]_ [(][x][) := 1]n _n_ � _fi(x) s.t. x ∈_ _X ._ (1) _i=1_ where X ⊆ R[d] is a closed convex set common to all agents and fi : R[d] _→_ R is a local objective function of agent i. Herein, fi is convex and not necessarily differentiable function which is only known to agent i. In this setting, we denote the optimal value of problem (1) by f _[⋆]. A decentralized solution to estimate f_ _[⋆]_ of this problem is the DPS algorithm [2]. In this algorithm, each agent locally updates the decide variable by fusing the estimates from its neighbors and then take the subgradient of this function at the updated decide variable to be the decent direction for the current iteration. To be more specific, when applied this algorithm to solve problem (1), it performs the following iterations: **_x¯i(t) =_** _n_ � _Aij(t)xj(t) ._ _j=1_ (2) the set of malicious agents (attackers) Vm, as seen in Fig. 1. We have V = Vt∪Vm and n = |Vt|+|Vm|. In our attack model, attackers are defined as agents whose estimates (or states) can not be affected by other agents, and those coordinated attackers try to drag the trustworthy agents to their desired value. If i ∈Vt, a trustworthy agent will perform the rules in (2). Otherwise, an attacker j ∈Vm will update its state with the following rule: **_xj(t) = α + rj(t), ∀_** _j ∈Vm ._ (3) where α is the target value of attackers, rj(t) is an artificial noise generated by attackers to confuse the trustworthy agents. If there are more than one attacker in the network, we assume that they will coordinate with each other to converge to the desired value α. Meanwhile, to disguise attacks, they will independently degenerate artificial noise rj(t) which decays exponentially with time, i.e., limt→∞ _∥rj(t)∥_ = 0 for all _j ∈Vm._ For the time varying network, let E(Vt; t) be the edge set of the subgraph of (t) with only the trustworthy agents in _G_ _Vt. The following assumption is needed to ensure a successful_ attack on the DPS algorithm: **Assumption 2. There exists B1, B2 < ∞** _such that for all_ _t ≥_ 1, 1) the composite sub-graph (Vt, ∪[t]ℓ[+]=[B]t+1[1] _[E][(][V][t][;][ ℓ][))][ is]_ _connected; 2) there exists a pair i ∈Vt, j ∈Vm with (i, j) ∈_ _E(t) ∪_ _. . . ∪E(t + B2 −_ 1). Based on this assumption, we have the following fact: **Fact 2. [13] Under Assumptions 1 and 2. If ∥∇[ˆ]** _fi(x)∥≤_ _C2_ _for some C2 and for all x ∈_ _X, and γ(t) →_ 0, we have: lim _t→∞_ [max]i∈Vt _[∥][x][i][(][t][)][ −]_ **_[α][∥]_** [= 0][ .] This fact implies that in this attack scheme, the attackers will succeed in steering the final states. This was also proved in our previous work [13]. III. DETECTION AND LOCALIZATION STRATEGIES The DPS algorithm runs in a fully decentralized fashion in trustworthy agents i ∈Vt. The neighborhood detection (ND) task and neighborhood localization (NL) task are then introduced for detecting and localizing attacker. To facilitate our discuss, we consider the following hypotheses. The ND task is defined as follows: _H0[i]_ [:][ N][i] _[∩V][m]_ [=][ ∅][,] No neighbor is an attacker, _H1[i]_ [:][ N][i] _[∩V][m]_ _[̸][=][ ∅][,]_ At least one neighbor is the attacker, (4) where H0[i] [and][ H]1[i] [as two events of agent][ i][ for the ND task.] When event H1[i] [is true at agent][ i][, the second task is to check] if the neighbor j ∈Ni is an attacker. The NL task is defined as follows: _H0[ij]_ [:][ j /][∈V][m][,] Neighbor j is not an attacker, (5) _H1[ij]_ [:][ j][ ∈V][m][,] Neighbor j is an attacker, where H0[ij] [and][ H]1[ij] [are as two events of agent][ i][ for the NL] task If event is true we say that the attacker is localized _H[ij]_ **_xi(t + 1) = PX_** �x¯i(t) − _γ(t) ˆ∇fi�x¯i(t)��_ _._ for t ≥ 1, where Aij(t) is a non-negative weight and γ(t) > 0 is a diminishing stepsize. PX � _·_ � denotes the projection operation onto the set X and _∇[ˆ]_ _fi (¯xi(t)) is a subgradient at_ agent i of the local function fi at x = ¯xi(t). Then, we have the following result: **Fact 1. [2] Under Assumption 1. If ∥∇[ˆ]** _fi(x)∥≤_ _C1 for some_ _C1 and for all x ∈_ _X, and the step size satisfies_ [�]t[∞]=1 _[γ][(][t][) =]_ _∞,_ [�]t[∞]=1 _[γ][2][(][t][)][ <][ ∞][, then for all][ i, j][ ∈V][ we have]_ lim lim _t→∞_ _[f]_ [(][x][i][(][t][)) =][ f][ ⋆] _[and]_ _t→∞_ _[∥][x][i][(][t][)][ −]_ **_[x][j][(][t][)][∥]_** [= 0][ .] The above fact tells that for these convex problems, the DPS method will converge to an optimal solution of problem (1). In the next, we will discuss how this convergence will change when there is attack within the network. _A. Data Injection Attack From Insider_ In this setting, we assume that the set of agents can be _V_ divided into two subsets: the set of trustworthy agents and _V_ ----- Fig. 1. Neighborhood tasks in the attack detection scheme. Each trustworthy agent performs ND and NL tasks independently for isolating attacker from the network. We remark that such hypotheses were also made in previous work [19], [36]. An illustration of the neighborhood detection and localization tasks is shown in Fig. 1. Notice that the NL task is executed only if the event H1[i] [in the ND task is true.] Moreover, once the attacker is localized, trustworthy agents will disconnect from the attacker in the next communication. In this way, it is expected that the network can exclude all the attackers from the network. To proceed our tasks, we run the asynchronous gossip-based optimization algorithm (Algorithm 1) for K instances. We denote **_X[˜]_** _i[k]_ [as the neighborhood state matrix collected by agent] _i in kth instance, i.e., k_ [1, _, K]. The ND and NL tasks_ _∈_ _· · ·_ can be described as follows: **_X˜_** _i[k]_ [:=[][x]i[k][,][ x]1[k][,][ · · ·][,][ x]j[k][,][ · · ·][,][ x][k]|Ni|[]][⊤] _[∀]_ _[j][ ∈N][i][,]_ (6) _H1[i]_ _yi = FND( X[˜]_ _i[1][,][ · · ·][,][ ˜]Xi[K][)]_ ≷ _δ,_ (7) _H0[i]_ _H1[ij]_ _zij = FNL( X[˜]_ _i[1][,][ · · ·][,][ ˜]Xi[K][)]_ ≷ _ϵ._ (8) _H0[ij]_ where x[k]j _[∈]_ [R][d][ is the state vector of agent][ j][ ∈N][i][, which can] be directly obtained by agent i ∈Vt from its neighbors, yi ∈ R is a metric that indicates whether an attacker is present in the neighborhood of agent i, and zij = [zi1, · · ·, zi|Ni|][⊤] _∈_ R[|N][i][|] is the metric vector for localization task. Herein, δ > 0 and _ϵ > 0 are some pre-designed thresholds._ On top of the detection and localization strategies, we have an important assumption about the initial states: **Assumption 3. We have the prior information about the** _expected initial states about the mean of attackers E[x[k]j_ [(0)] =] **_α¯, j ∈Vm and trustworthy agents E[x[k]i_** [(0)] = ¯][β][, i][ ∈V][t][ .] _Moreover, ¯α_ = β[¯] in general. _̸_ Note that this assumption is practical as the attacker always aims at dragging the trustworthy agents to its desired value, which is usually different from the optimal solution. Otherwise, we may not consider it as a meaningful attack. **Remark 1. FND(·) and FNL(·) are statistical decision func-** _tions judiciously designed for ND and NL tasks respectively._ _For each agent i ∈Vt, these decision functions are used to_ _calculate the criterion metrics to identify attackers_ _A. The Score-based Method_ As a remedy to protect these distributed optimization algorithms, such score-based methods have been studied in [18], [19], which stem from statistical techniques. For gossip-based DPS algorithm, a temporal difference strategy (TD) in [13] and a spatial difference strategy (SD) in [36] are proposed to detect and localize the attackers, and these two strategies are reviewed below. _1) Temporal Difference Strategy: Since the expected initial_ states about the mean of attackers and trustworthy agents are different, when t, the network will be mislead by the _→∞_ attackers to E[x[k]j [(][∞][)] = ¯][α][ =][ E][[][x]i[k][(][∞][)]][. This implies that the] difference between the initial state and the steady state can be used to detect anomalies. For each trustworthy agent i ∈Vt, the following score can be evaluated [1]: 1 _ξij :=_ _Kd_ _K_ � **1[⊤][�]x[k]j** [(][T] [)][ −] **_[x]j[k][(0)]�, j ∈Ni._** (9) _k=1_ Herein, T is sufficiently large, d is the state dimension _→∞_ of agents, 1 is an all-one vector. x[k]j [(][T] [)][ and][ x]j[k][(0)][ are] respectively the last and the first state for agent j observed by agent i. To discern the events in ND task, the detection criterion is defined as follow: 1 _yˆi :=_ _|Ni|_ � _j∈Ni_ _H0[i]_ ���ξij − _ξi���_ ≶ _δTD._ (10) _H1[i]_ where ξi = 1/ |Ni| [�]j∈Ni _[ξ][ij][ is the average of neighborhood]_ of agent i. Intuitively, E[ˆyi] = 0 when the event H0[i] [is true,] otherwise E[ˆyi] ̸= 0 when the event H1[i] [is true.][ δ][TD] [is a] pre-designed threshold of the ND task. For the NL task, these two events H1[ij] [and][ H]0[ij] [are checked] by the following criterion: _H1[ij]_ _zˆij := |ξij|_ ≶ _ϵTD, ∀_ _j ∈Ni ._ (11) _H0[ij]_ Herein, ϵTD is a pre-designed threshold used to identify which neighbor is the attacker. Note that E[ˆzij] is close to 0 if an agent j is an attacker, seen in (9). _2) Spatial Difference Strategy: According to (3), attackers_ always try to mislead the network to their desired value, and thus the transient state in the network will also be affected during the attack process. Unlike the TD method that only uses the initial state and steady state, the transient states are considered in the SD method for better performance. We expect that the expected state E[x[k]i [(][t][)][ −] **_[x]j[k][(][t][)]][ between]_** neighbor j and monitoring agent i will behave differently in events H0 and H1, i.e., j ∈Ni and 0 < t < ∞. For the ND task, agent i evaluates the following metrics: **_ϕ[k]ij_** [:=] 1 _yˇi :=_ _|Ni|_ _T_ � _t=0_ � _j∈Ni_ � � **_x[k]j_** [(][t][)][ −] **_[x]i[k][(][t][)]_** _, j ∈Ni._ (12) _K_ � �2 H0[i] **1[⊤]ϕ[k]ij** ≶ _δSD._ (13) _k=1_ _H1[i]_ � 1 _Kd_ 1For each instance k, each agent evaluates ∆j (t) ≜ **_xkj_** [(][t][)][ −] **_[x]j[k][(][t][ −]_** [1)][ at] it ti _t_ d it ll th it ti t bt i � _k(T_ ) _k(0)�_ ----- Fig. 2. The TDNN method at trustworthy agent i: (Left) NN for ND task, (Right) NN for NL task. SDNN shares a similarly structure with TDNN. where � **_x[k]i_** [(][t][) = 1][/][ |N][i] _[∪]_ _[i][|]_ **_x[k]j_** [(][t][)] (14) _j∈{Ni∪i}_ is the neighborhood average of agent i when iterating t in instance k. ϕ[k]ij [is the sum of differences between neighbor] agent j and the neighborhood average x[k]i [(][t][)][ in all iterations.] _δSD is a pre-designed threshold._ For the NL task, we compare the state of neighbor agent j with agent i to check the events in (5). The following criteria are used: �x[k]j [(][t][)][ −] **_[x]i[k][(][t][)]�_** _−_ **_ϕ[k]ii[, j][ ∈N][i][.]_** (15) _K_ � �2 H0[ij] **1[⊤]ϕ[k]ij** ≶ _ϵSD, ∀j ∈Ni._ (16) _k=1_ _H1[ij]_ _chosen agent[2]_ _takes the role as an attacker. Based on Assump-_ _tion 3, we run the asynchronous gossip-based optimization_ _algorithm (Algorithm 1) for K instances and record_ **_X[˜]_** _i[k]_ _[as]_ _the data samples with the ground truth label ‘1’ for event H1[i]_ _where agent i is either in the neighborhood of (next to) the_ _attacker or beyond the neighborhood of (far from) the attacker._ We remark here that **_X˜_** _i[k]_ [is the local data collecting by] agent i which is not allowed to exchange among agents. On the other hand, ground truth label ‘0’ for event H0[i] [can be] easily obtained by running gossip-based algorithm on . We _G_ remark here that how to specifically set the training data collecting process is a challenging problem while it is beyond the scope of this work. Herein, we simply assume that each agent can obtain its own training data with correct labels. Other technique problems about the details of the training process will be included in another work. _1) Temporal Difference Strategy via NN: Armed with train-_ ing data, we propose a method called TDNN, which uses the time difference values as the input of the NN to perform neighborhood tasks, as illustrated in Fig. 2. Based on the metric in (9), the inputs for the two neighborhood tasks are as follows, **_a[0]_** = ˆa[0] = [ξi1, ξi2 · · ·, ξiM ][⊤]. (17) where ξij can be obtained by agent i. For the ND task, the computation process of NN can be described below: **_a[h]_** = σ(W _[h]a[h][−][1]_ + b[h]), _h = 1, ..., n_ 1; (18) _−_ _H1[i]_ _y˜i = g(W_ _[n]a[n][−][1]_ + b[n]), _y˜i_ ≷ _δNN,_ (19) _H0[i]_ where a[0] is the input of NN. σ( ) is the activation function, _·_ _g(_ ) is the sigmoid function defined as g( ) = 1/(1 + e[−][x]), _·_ _·_ **_W_** _[h]_ _∈_ R[L][h][×][L][h][−][1] is the weight matrix between the layer h and layer h _−_ 1, Lh represents the number of neurons in layer _h, b[h]_ _∈_ R[L][h] and a[h] _∈_ R[L][h] are the bias vector and the activation output in the layer h, respectively. ˜yi ∈ R is the expected output, and δNN ∈ [0, 1] is some prescribed threshold for detection task. For the NL task, a similar NN structure is used, except for the number of neurons in the output layer. The design is given as follows, **_aˆ[h]_** = σ( W[ˆ] _[h]aˆ[h][−][1]_ + b[ˆ][h]), _h = 1, ..., n_ 1; (20) _−_ _H1[ij]_ **_z˜i = g( W[ˆ]_** _[n]aˆ[n][−][1]_ + b[ˆ][n]), _z˜ij_ ≷ _ϵNN,_ (21) _H0[ij]_ where ˆa[0] is the input of NN, **_W[ˆ]_** _[h],_ **_b[ˆ][h], and ˆa[h]_** are the weight matrix, bias term, and activation output in NN, respectively, **_z˜i = [˜zi1, · · ·, ˜ziM_** ] ∈ R[M] is the expected output of NL task, and ϵNN ∈ [0, 1] is some prescribed threshold. Notice that the actual output is encoded by one-hot vector during training stage, such as ˜zi = ej if j ∈Vm, see Fig. 2 (Right). 2There could be more than one attackers in the training network while h i l id th i l t **_ϕ[k]ij_** [:=] _T_ � _t=0_ � 1 _zˇij :=_ _Kd_ where ¯ϕ[k]ii [is calculated by agent][ i][ itself, seen in (12).][ ϕ]ij[k] [is the] metric between agent i and agent j, and ϵSD is a pre-designed threshold used to identify the attacker. _B. The AI-based Method_ In fact, the reason why (10), (11), (13) and (16) take effect is that the anomalies will cause the the measured metrics to behave statistically different. In such score-based methods, these decision functions of ND and NL tasks are approximately linear or quadratic functions, which fuse the states obtained by agent i into a scalar score for classification. A natural question that follows is whether there exists more sophisticated nonlinear functions that can better classify those events in the two neighborhood tasks. This is a natural application of AI technology for learning the complex mapping relationships in a classification problem. In the following, we propose to apply NN to handle the ND and NL tasks. Let M = maxi |Ni| be the input dimension of these NNs. Then, NNs can be trained at each monitoring agent in an offline manner using data collected from each agent. To facilitate our approach, we use the following process to collect training data for the AI-based methods: **Assumption 4. Assume that we have set a training data** _collecting process which contains P training network Gp =_ ( ) for p 1 2 _P For each network_ _a randomly_ _V_ _E_ _G_ ----- _2) Spatial Difference Strategy via NN: Both TD and TDNN_ only utilize the initial state and the steady state of agents rather than the transient states, leading to the possibility of losing some key features in the neighborhood tasks. In particular, the neighborhood transient state information is not effectively utilized for extracting key classification features. Therefore, we propose a strategy called SDNN to improve the detection and localization performance by using transient states and NN. As a malicious agent always tries to influence and steer the trustworthy agents away from the true value, we have E[x[k]j [(][t][)][ −] **_[x]i[k][(][t][)][|H]1[ij][]][ ̸][=][ E][[][x]j[k][(][t][)][ −]_** **_[x]i[k][(][t][)][|H]0[ij][]][. Thus, we can]_** compare the state of neighbor agent j and the neighborhood average of agent i over time. The metrics for ND and NL tasks can be described as follows: **_s[k]ij_** [:=] _T_ � _t=0_ �x[k]j [(][t][)][ −] **_[x]i[k][(][t][)]�, j ∈Ni._** (22) **Algorithm 2 Gossip training for AI-based methods** **Input: Pij : probability of exchange, η : learning rate.** **Initialize: W is initialized randomly, i ∈V.** **repeat** _• MERGEMODEL(Wr, Wi) in (28)_ _• Wi ←_ **_Wi −_** _η∇[ˆ]_ _Li(Wi), agent i updates parameters_ _• Agent i sends Wi to agent j ∈Ni with probability Pij_ **until Maximum iteration reached** **function MERGEMODEL(Wr, Wi)** **_Wi ←_** **_Wi(1 −_** _µ) + µWr, µ ∈_ [0, 1]. **end function** and privacy preservation. Moreover, we allow the trustworthy agents to have correctly labeled samples from the ND and NL tasks. For instance, the samples labeled in the current training round can be used to the next training round. In the next, we will see how to share AI-based models between different agents to achieve robust performance in ND and NL tasks. _A. The Distributed Collaborative Training Process_ The goal of collaborative training is that participating agents acting as local learners to train good local models (i.e., NN models) through gossip exchanges. That is, an agent i _∈V_ aims to train a model that performs well with respect to the data points available on other agents. For distributed collaborative training, the standard unconstrained empirical risk minimization problem used in machine learning problems (such as NN) can be described as follows [39]: 1 _χij :=_ _Kd_ _K_ � **1[⊤]s[k]ij[, j][ ∈N][i][.]** (23) _k=1_ where x[k]j [(][t][)][ is the][ t][th states of agent][ j][ at instance][ k][,][ s]ij[k] is the sum of statistical differences between agent j and the neighborhood average of agent i. Note that x[k]i [(][t][)][ has been] defined in (14). Herein, our goal is to accurately detect insider attacks and identify whether the attacker appears in the neighborhood of agent i. The detection structures of SDNN are similar with that for TDNN, as seen in Fig. 2. Therein, we use the following inputs for these NN models of ND and NL tasks: **_a[0]_** = ˆa[0] = [χi1, χi2, · · ·, χiM ][⊤]. (24) IV. COLLABORATIVE LEARNING FOR A ROBUST MODEL In previous sections, we have introduced how to use NN to help detect and localize the insider attackers. Our training data comes from a training data collecting process under Assumption 4 wherein the local data samples **_X[˜]_** _i[k]_ [are collected] by agent i which could be within or beyond the neighborhood of the attacker. Apparently, the optimal train way is to upload all agents’ data to a fusion center and train the model in a centralized manner. However in practice, collecting data from decentralized data sources to the center is hard due to storage and bandwidth limitations. On the other hand, as running a gossip algorithm is time-consuming, it is usually difficult and expensive to collect sufficient data at each agent. For example, the attack could occur far from the monitoring agent while the training data may only contains samples from a neighbor attacker. As the training samples may not well represent the general attack network, some individually trained NN may not fit in all insider attack events. To alleviate these issues, we propose a collaborative peer-to-peer protocol to facilitate training our NN models. Before we go to the details, we recall three assumptions for the proposed collaborative learning process. First, we assume that all agents in the network have equal number of neighbors (this is sort of impractical but we can resolve it later). Also, different agents collect their own training data with the advantages of high scalability where Li(W ) = [1]q �ς∈Di _[ℓ][(][W][, ς][)][. This formulation enables]_ us to state the optimization problem (25) in a distributed manner. This distributed collaborative training problem could be addressed by gossip exchanges [41], [42]. We will detail it as follows min **_W_** _[L][(][W][ ) = min 1]n_ � _Li(W )_ (25) _i∈V_ where W is the parameter of the NN model. Li(·) is a local objective function of agent i, which is defined as the expected loss of the local data set. The local objective is to minimize the expected loss of its local sample _Li(W ) = Eς∼Ii_ [ℓ(W, ς)] (26) where ς is a pair variable, composed of an input and related label, following the unknown probability distribution Ii, which is specific for the sample set received by agent i. ℓ( ) is a loss _·_ function used to quantify the prediction error on ς. Let Di = _{ς1, · · ·, ςq} represents the set of training data on agent i ∈V,_ which contains q samples. Thus, we have D = D1 ∪· · · ∪ _Dn_ to optimize problem (25): min **_W_** _[L][(][W][ ) = min 1]n_ � _i∈V_ � � _ℓ(W, ς)_ (27) _ς∈Di_ � 1 _q_ ----- TABLE I TEST SCENARIOS SETTINGS FOR MISMATCHED DATA Scenario Mean Deviation Initial distribution Fig. 3. An example of tailoring neighbor agents. _B. The Gossip Stochastic Gradient Descent Strategy_ Gossip learning is a method to learn models from fully distributed data without central control [9]. The skeleton of the gossip learning protocol is shown in Algorithm 2. Therein, during the training stage, each agent i has a NN with the same architecture and initializes a local model with the parameter **_Wi. This is then sent to another agent j ∈Ni in the network_** periodically with the probability Pij. Upon receiving a model **_Wr, the agent i merges it with the local model, and updates it_** using the local data set Di. We utilize the stochastic gradient descent (SGD) algorithm to estimate the local parameter Wi [43], as follows **_Wi ←_** **_Wi(1 −_** _µ) + µWr,_ _µ ∈_ [0, 1]. (28) **_Wi ←_** **_Wi −_** _η∇[ˆ]_ _Li(Wi),_ _i ∈V,_ (29) where η and _∇[ˆ]_ _Li(·) are the learning rate and the expected_ gradient of agent i, respectively. µ [0, 1] is a weight _∈_ used to merge the receive model Wr. Herein, MERGEMODEL(Wr, Wi) is a merging process as shown in (28) which is typically achieved by averaging the model with parameters, i.e., µ = 0.5. _C. The Tailor-degree Network_ We have introduced the application of NN in detecting and localizing attacker, and assumed that each normal agent in the network has exactly M neighbors. Inevitably, the communication network is an irregular network, where some agents have a heterogeneous number of neighbor agents. In order to adapt to scenarios with different Degree-|Ni| agents, we tailor our M -input NN to fit into the scenario when a normal agent has |Ni| ̸= M neighbors. Two scenarios are considered in this subsection. In the first scenario, we consider the case of |Ni| > M . The |Ni| neighbors is divided into ⌈|Ni|/M _⌉_ potentially overlapping groups. In ND and NL tasks, each group contains exactly M agents, which can be treated as a standard neighbor set of TDNN and SDNN methods. Thus, these two tasks can be implemented with the unified NN model, as seen in Fig. 3. On the other hand, if we have |Ni| < M, the deficient value in the input vector is replaced by a reference value to fit a Degree-|Ni| agent. For the TDNN method, the input is reconstructed by **_a[0](ˆa[0]) ∈_** R[M] = [ξi1, · · ·, ξi|Ni|, ξii, . . ., ξii][⊤], wherein ξii is the temporal difference value of agent i. For the SDNN method, the deficient value of input vector is replaced by _Xii, and the input is reconstructed by a[0](ˆa[0]) ∈_ R[M] = [χ ; ; χ ; χ ; ; χ ][⊤] when _< M_ _|N |_ S0 0.5 1.0 **_β ∼U_** [0.0, 1.0][d] S1 0.5 0.6 **_β ∼U_** [0.2, 0.8][d] S2 0.5 1.4 **_β ∼U_** [−0.2, 1.2][d] S3 0.7 1.0 **_β ∼U_** [0.2, 1.2][d] S4 0.3 1.0 **_β ∼U_** [−0.2, 0.8][d] _Remark: Typically, the training data used to train the AI-_ based methods is collected by trustworthy agents under a scenario of specific prior information β. In practice, the prior information of the gossip-based DPS optimization protocol will be changed in some particular scenarios, that is, the test data is statistically mismatched with the training data. To further verify the robustness of the AI-based detection and localization models, we generate the test data by keeping the target value of attackers α and changing the mean and the deviation of β. As depicted in Section V, we set β [0, 1][d], _∼U_ then several test scenarios are defined in the TABLE I. V. NUMERICAL RESULTS AND ANALYSIS In this section, numerical results are presented to validate the effectiveness of the proposed AI-based methods in neighborhood tasks. The DPS algorithm runs on a Manhattan network with n = 9 agents, as shown in Fig. 4. In our experiment, an example of the least-square optimization problem is considered; i.e., in (1) we set _n_ � _i=1_ _f_ _[k](x) =_ _n_ � _fi[k][(][x][) =]_ _i=1_ _k_ 2 ��(θi [)][⊤][x][k][ −] _[φ]i[k]��_ _, k = 1, ..., K._ Herein, fi[k] is a utility function at agent i. As shown in Algorithm 1, the DPS algorithm runs in an asynchronous manner such that an agent i randomly selects an agent j with probability [P ]ij = Pij = 1/|Ni|. Thus the expected transition matrix in iteration t can be written as E [A(t)] = **_I −_** 2[1]n **[Σ]** [+][ P][ +]2n[P][ ⊤], where Σ is a diagonal matrix with [Σ]ii = �n _j=1[(][P][ij][ +][ P][ji][)][. In each instance, we set][ d][ = 2][,][ T][ = 2000][,]_ the initialization x[k](0) [0, 1][d], α[k] [ 0.5, 0.5][d] and _∼U_ _∼U_ _−_ **_rj[k][(][t][)][ ∼U][[][−][λ][ˆ][t][,][ ˆ][λ][t][]][, where][ λ][ is the second largest eigenvalue]_** of E[A(t)]. In particular, to serve our purpose we change the function fi[k][(][x][)][ by randomly generating][ θ]i[k] _[∼U]_ [[0][.][5][,][ 2][.][5]][d][,] (x[⋆])[k] _∼U[0, 1][d], and thus we have φ[k]i_ [= (][θ]i[k][)][⊤][(][x][⋆][)][k][.] For the AI-based methods, the feed forward neural networks (FFNN) with three hidden layers is applied to perform the ND and NL tasks with neurons in each hidden layer being 200, 100 and 50 respectively. These NNs are implemented using a modified version of the deep learning toolbox in [44]. Rectified linear unit (ReLU) as the activation function is equipped in all hidden layers and the parameters of NN are jointly optimized through a back propagation method by minimizing the loss function defined on different tasks. To provide neighbor data and ground truth labels for the AI-based methods, we run the DPS algorithm independently in each event (4) starting with a new initial state each time ----- 1 Detection, agent i next to attacker TD K=2, d=1 TD K=1, d=2 TD K=5, d=1 TD K=5, d=2 TDNN K=2, d=1 TDNN K=1, d=2 TDNN K=5, d=1 TDNN K=5, d=2 0 0.2 0.4 0.6 Probability of false alarm Pinf Localization, agent i next to attacker 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 |ection, agent i next|t to attacker|Col3| |---|---|---| |||| |||| |||| |||| |||| |TD K=2, TD K=1,|d=1 d=2|| |TD K=5, TD K=5,|d=1 d=2|| |TDNN K TDNN K|=2, d=1 =1, d=2|| |TDNN K TDNN K|=5, d=1 =5, d=2|| |||| |ocaliza|ation, agent i nex|Col3|xt to attack| |---|---|---|---| ||||| ||||| ||||| ||||| ||||| ||TD K||| |||TD K|=2, d=1| |||TD K TD K|=1, d=2 =5, d=1| |||TD K TDN|=5, d=2 N K=2, d=1| |||TDN TDN|N K=1, d=2 N K=5, d=1| ||TDN|TDN|N K=5, d=2| ||||| 0 0.2 0.4 0.6 Probability of false alarm Pilf Fig. 4. The Manhattan network topology with agent 1 as the attacker. TABLE II TRAINING AND TESTING SETS FOR AI-BASED METHODS GIVEN THATZ AGENT j IS THE ATTACKER. Fig. 5. ROCs of TDNN and TD methods: (Left) ND task, (Right) NL task. where ˆ and ˆ are the estimated event based on AI_H[i]_ _H[ij]_ based methods. Pnd[i] [(][P]ld[ i] [) and][ P][ i]nf [(][P]lf[ i] [) are the probability] of detection and false alarm in the ND (NL) task, respectively. More specifically, those probabilities are calculated as follows Task Event Training set Testing set Label _H0[i]_ 10000 6000 0 _Nnd_ � _I(yi[(][n][)]_ = ˆyi[(][n][)] = 1), (32) _n=1_ _Nnf_ � _I(yi[(][n][)]_ = 0 ∧ _yˆi[(][n][)]_ = 1), (33) _n=1_ _Nld_ � _I(zij[(][n][)]_ = ˆzij[(][n][)] = 1), (34) _n=1_ _Nlf_ � _I(zij[(][n][)]_ = 0 ∧ _zˆij[(][n][)]_ = 1). (35) _n=1_ ND _H1[i]_ [(][j][ ∈N][i][)] 10000 6000 1 _H1[i]_ [(][j /][∈N][i][)] 10000 6000 1 NL _H1[i]_ 10000 6000 **_ej_** As the Manhattan network is symmetric, to obtain the training data under hypothesis H1[i] [with label ‘1’, we need to collect] data at two types of the trustworthy agents: the one that stands at the position “next to” the attacker agent (for example agent 1 is the only attacker and we collect data at agent 2), and the one that stands at the position “far from” the attacker agent (for example agent 1 is the only attacker and we collect data at agent 5). Meanwhile, the training data under hypothesis H0[i] with label ‘0’ is collected at any agent when DPS is running on the Manhattan network free of attacker. We collect data from different scenarios as shown in Table II and fuse them into ND and NL models. Therein, available data is typically split into two sets, a training set and a testing set. As for the ND task, the detection task consists of 30, 000 samples as the training set and 18, 000 samples as the testing set. Herein, within the data under hypothesis H1[i] [, we have][ 10000][ samples collecting] at agent next to the attacker and 10000 samples collecting at agent far from the attacker. For the NL task, the training set and testing set contain 10, 000 and 6, 000 samples respectively. Herein, we encode the ground truth labels of event H1[i] [=] _H1[ij]_ _[∪H]0[ij]_ [by one-hot coding, where the neighbor attacker is] labeled by ‘1’ and the trustworthy agent by ‘0’. Usually, the detection and localization models of AI-based methods are actually classifiers for which NN produces continuous quantities to predict class membership through different thresholds. To make a more comprehensive evaluation of these classifiers, we adopt the probabilities of detection and false alarm for the ND and NL tasks. That is, we define _Pnd[i]_ [= 1] _Nnd_ _Pnf[i]_ [= 1] _Nnf_ _Pld[i]_ [= 1] _Nld_ _Plf[i]_ [= 1] _Nlf_ where Nnd (Nnf ) and Nld (Nlf ) are the number of positive (negative) samples in the ND and NL tasks, respectively, ˆyi[(][n][)] (zˆij[(][n][)][) is the predicted class by ND (NL) classifiers, and][ y]i[(][n][)] (zij[(][n][)][) is the ground-truth class label. Note that][ I][(][·][)][ is an] indicator function that has the value 1 when the predicted class label equal to the ground-truth class label. Based on these probabilities, the detection (or localization) performance can be investigated by the receiver operating characteristic (ROC) [45], for which the probability of detection is plotted on the _Y -axis and the probability of false alarm is plotted on the X-_ axis. It is worth noting that the ROC curves that approach the upper left corner outperform those far from it. _A. Detection and Localization for One attacker_ _Pnd[i]_ [:=][ P] [( ˆ][H][i][ =][H]1[i] _[|H]1[i]_ [)][, P]nf[ i] [:=][ P] [( ˆ][H][i][ =][ H]1[i] _[|H]0[i]_ [)][.] (30) _Pld[i]_ [:=][ P] [( ˆ][H][ij][ =][H]1[ij] 1 [)][, P]lf[ i] [:=][ P] [( ˆ][H][ij][ =][ H]1[ij] 0 [)][.][ (31)] _[|H][ij]_ _[|H][ij]_ In this subsection, we show the detection and localization performance of AI-based methods when the Manhattan network contains only one attacker, seen in Fig. 4. Suppose that agent 1 is the attacker and the monitoring node is agent 2 (or 3, 4 and 7). Then, the NN model is trained and also tested by the data collecting from the agent next to the attacker. In Fig. 5, we study the attacker detection and localization performance of TDNN, where TD in [13] is taken as the benchmark method. ROC curves of ND task is depicted in Fig. 5 (Left), variables K and d in the legend are the number of instances and dimensions used to detect insider attacks ----- 1 Detection, agent i next to attacker 1 0.95 0.9 0 0.05 0.1 SD K=1, d=1 SD K=1, d=2 SD K=2, d=1 SD K=2, d=2 SDNN K=1, d=1 SDNN K=1, d=2 SDNN K=2, d=1 SDNN K=2, d=2 0 0.2 0.4 0.6 Probability of false alarm Pinf Localization, agent i next to attacker 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 SD K=1, d=1 SD K=1, d=2 SD K=2, d=1 SD K=2, d=2 SDNN K=1, d=1 SDNN K=1, d=2 SDNN K=2, d=1 SDNN K=2, d=2 0 0.2 0.4 0.6 Probability of false alarm Pilf 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 |Detection, a|Col2|agent i next|Col4|t to attacker|Col6|Col7| |---|---|---|---|---|---|---| |1||||||| |0.95 0.9 0||||||| |||0.05||||| |||||||| |||||||| |||SD K=1, d= SD K=1, d= SD K=2, d=||1 2 1||| |||SD K=2, d= SDNN K=1,||2 d=1||| |||SDNN K=1, SDNN K=2, SDNN K=2,||d=2 d=1 d=2||| |ocalization,|Col2|Col3|, agent i nex|Col5|xt to attacke|Col7|Col8|Col9| |---|---|---|---|---|---|---|---|---| |||||||||| |1 0.9||||||||| |||||||||| |0|||0.05||0.1|||| |||||||||| ||||SD K=1, d=1 SD K=1, d=2 SD K=2, d=1|||||| |||||||||| ||||SD K=2, d=2 SDNN K=1,||d=1|||| ||||SDNN K=1, SDNN K=2, SDNN K=2,||d=2 d=1 d=2|||| 0 0 0.2 0.4 0.6 Probability of false alarm Pinf |Detection,|Col2|agent i next|to attacker|Col5| |---|---|---|---|---| |||||| ||TD Co TD|NN K=5, d=2 llaborative train NN K=5, d=2|ing|| ||Ind SD|ependent traini NN K=2, d=2|ng|| ||Col|laborative train|ing|| ||SD Ind|NN K=2, d=2 ependent trainin|g|| |||||| |||||| |||||| Fig. 6. ROCs of SDNN and SD methods: (Left) ND task, (Right) NL task. respectively. The localization performance of NL task is shown in Fig. 5 (Right), where we assume that the ND task can completely distinguish between events H0[i] [and][ H]1[i] [without] errors (by an ‘Oracle’). In these plots, it is obvious that both the performance of TDNN and TD will improve significantly as K increases when d is fixed, and vice versa. For the first and second curves in ND and NL tasks, TD has the same performance at K = 2, d = 1 as at K = 1, d = 2, and the same is true for TDNN on the fifth and sixth curves, which is inherent to TD strategy (9). Thus, we may say that either increasing K or increasing d will bring the same improvement over performance. From Fig. 5, it can be seen that TDNN improves significantly over TD in terms of both detection and localization performance, performing good performance when _K = 5, d = 2._ The ROCs of SDNN are shown in Fig. 6, while SD in [36] is selected as a benchmark. It can be seen from the plots that both SDNN and SD already provide good detection and localization performance when K = 2, which is better than that of TDNN and TD methods. This result implies that transient states can indeed provide more information to identify the attacker, as spatial methods (SDNN and SD) leverage the entire dynamic information while the temporal methods (TDNN and TD) only utilize the first and last states. Also in this case, the attacker detection and localization performance of SDNN and SD will improve significantly as K increases when d is fixed. When K is fixed, the performance of SDNN and SD slightly improved as d increases. For the ND task in Fig. 6 (Left), the detection performance of SDNN and SD are close to each other, they show excellent performance under the same feature processing condition, seen in (12) and (22). Nevertheless, SDNN has a drastic advantage over SD method in NL task and can completely distinguish the neighboring attacker at K = 2, seen in Fig. 6 (Right). Fig. 7. Comparison between independent training and collaborative training based on matched data. Models are trained on sufficient “next to” data then tested on “next to” data. 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 |Col1|ND task|Col3|Col4| |---|---|---|---| ||||| ||||| ||||| ||||| ||||| ||||| ||||| |Collabo|rative training||| |Collabo|rative training||| ||||| ||||| 0 0.2 0.4 0.6 Probability of false alarm Pinf _B. Performance of the Collaborative Learning_ In this subsection, we show how to utilize the collaborative learning protocol to train a robust model for accommodating more attack events. Specifically, we consider the performance comparison of independent training and collaborative training. Herein, independent training means that each agent trains its model based on its local data where “next to” data refers to Fig. 8. Comparison between independent training and collaborative training for mismatched data. The plots are SDNN with K = 1, d = 2. samples collected at the agent next to an attacker while “far from” data refers to samples collected at the agent far from an attacker. Moreover, “next to” model refers as the independent model trained on “next to” data while “far from” model refers as the independent model trained on “far from” data. We then consider two extreme cases to verify the collaborative learning. For Case 1, we assume that in the independent training process the monitoring agent only collects a very small amount of local “next to” data, which is not enough to complete a meaningful NN model. While in the collaborative training process, the agent update its local model by merging the received neighboring models. We then test the two training methods on “next to” data for ND and NL tasks. In Fig. 7, the dashed and solid lines are the performance for the independent training and the collaborative training, respectively. It is clear that with insufficient samples, the agent in independent learning performs poorly on both ND and NL tasks, while the collaborative training enables the agent to learn models from its neighbors, which greatly improves detection and localization performances. It is expected that a similar result also holds for “far from” cases. For Case 2, we first consider the scenario that each agent has sufficient “next to” data (“far from” data) to train the ----- 1 Localization, agent i next to attacker 1 Localization, agent i next to attacker 1 1 Detection, agent i next to attacker TD p=0 TD p=1 TD p=2 TDNN p=0 TDNN p=1 TDNN p=2 0 0.2 0.4 0.6 Probability of false alarm Pinf Detection, agent i next to attacker TD ~ U[0.0, 1.0] TD ~ U[0.2, 0.8] TD ~ U[-0.2, 1.2] TD ~ U[0.2, 1.2] TD ~ U[-0.2, 0.8] TDNN ~ U[0.0, 1.0] TDNN ~ U[0.2, 0.8] TDNN ~ U[-0.2, 1.2] TDNN ~ U[0.2, 1.2] TDNN ~ U[-0.2, 0.8] 0 0.2 0.4 0.6 Probability of false alarm Pinf |Detection, a|agent i next|t to attacker| |---|---|---| |||| |||| |||| ||TD p=0 TD p=1|| ||TD p=2 TDNN p TDNN p TDNN p|=0 =1 =2| |||| |||| |ocalization,|, agent i nex|xt to attack| |---|---|---| |||| |||| |||| ||TD p=0 TD p=1 TD p=2|| ||TDNN p TDNN p TDNN p|=0 =1 =2| |||| |||| |Detection,|Col2|agent i next|Col4|t to attacker|Col6| |---|---|---|---|---|---| ||||||| ||||||| ||||||| |||TD ~ U[0.0 TD ~ U[0.2 TD ~ U[-0.||, 1.0], 0.8] 2, 1.2]|| |||TD ~ U[0.2 TD ~ U[-0. TDNN ~ U TDNN ~ U||, 1.2] 2, 0.8] [0.0, 1.0] [0.2, 0.8]|| |||TDNN TDNN|~ U ~ U|[-0.2, 1.2] [0.2, 1.2]|| ||||||| |||TDNN|~ U|[-0.2, 0.8]|| ||||||| |ocalization, agent|Col2|t i next to attack| |---|---|---| |||| |||| |||| ||TD TD|~ U[0.0, 1.0] ~ U[0.2, 0.8]| ||TD TD TD TDN|~ U[-0.2, 1.2] ~ U[0.2, 1.2] ~ U[-0.2, 0.8] N ~ U[0.0, 1.0]| ||TDN TDN|N ~ U[0.2, 0.8] N ~ U[-0.2, 1.2]| ||TDN TDN|N ~ U[0.2, 1.2] N ~ U[-0.2, 0.8]| 0 0.2 0.4 0.6 Probability of false alarm Pilf 0 0.2 0.4 0.6 Probability of false alarm Pilf 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Fig. 9. ROCs of TDNN and TD with different deficient size: (Left) ND task, (Right) NL task. p = M −|Ni| means that the number of deficient inputs. Localization, agent i next to attacker 1 1 Detection, agent i next to attacker 1 0.95 0.9 0 0.02 0.04 0.06 SD p=0 SD p=1 SD p=2 SDNN p=0 SDNN p=1 SDNN p=2 0 0.2 0.4 0.6 Probability of false alarm Pinf |Detection, a|Col2|Col3|agent i next|Col5|t to attacker|Col7|Col8| |---|---|---|---|---|---|---|---| |1|||||||| |0.95|||||||| ||||||||| |0.9 0|||||||| ||||0.02 0.||||| ||||||||| ||||||||| ||||||||| ||||SD p=0||||| ||||SD p=1 SD p=2||||| ||||||||| ||||SDNN p= SDNN p=||0 1||| ||||||||| ||||SDNN p=||2||| ||||||||| ||||||||| ||||||||| |ocalization, agen|Col2|nt i next to attacke|Col4| |---|---|---|---| ||||| ||||| ||||| ||||| ||||| |SD||p=0|| ||SD|p=0|| ||SD SD|p=1 p=2|| ||SD SD|NN p=0 NN p=1|| |SD|SD|NN p=2|| ||||| 0 0.5 1 Probability of false alarm Pilf 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Fig. 10. ROCs of SDNN and SD with different deficient size: (Left) ND task, (Right) NL task. p = M −|Ni| means that the number of deficient inputs. NN model. We then test the independent/collaborative training models on the “next to” data (“far from” data), which is matched the training data. The red (blue) dashed and solid lines in Fig. 8 (Left) represent their ROC results, which imply that the collaborative learning model will converge to the independent learning model. On the other hand, we further test the mismatch case of the training data and testing data. That is, we use the “next to” data to test the “far from” model, and vice versa. Interestingly, Fig. 8 (Right) shows that the collaborative learning model has a significant improvement over the independent model, as it learns both the characteristics of the “next to” model and “far from” model. These results demonstrate the advantage of the collaborative learning on the robustness of the model. Therefore, when there are enough samples, collaborative learning also has strong competitiveness compared with independent learning. Fig. 11. ROCs of TDNN and TD for the mismatch model: (Left) ND task, (Right) NL task. α[k] _∼U_ [−0.5, 0.5][d]. Each entry of x[k](0) is distributed as legended for testing data. as the target example, and we have M = 4 and |Ni| = {2, 3}. The attack scenario of m = 1, c = 1 is applied to verify our proposed method, and the performance in p = 0 is taken as the baseline. We choose agent 2 as the test agent whose neighbors are agents 1, 3, 5 and 8, i.e., p = 0. When p = 1, we cut off the connection between agents 2 and 3, then the neighbors of the test agent are agents 1, 5, and 8. Next when p = 2, we further cut off the connection between agents 2 and 5, leaving only agents 1 and 8 as neighbors of the test agent 2. Note that the parameters for the AI-based methods are the same as those in previous subsection V-A, and that the testing data is generated from a modified Manhattan network with p = 1 and _p = 2._ Fig. 9 shows the attacker detection and localization performance of TDNN and TD methods with K = 5, d = 2. It can be seen that the performance of TDNN and TD in ND and NL tasks will not fluctuate significantly with the increase of p. However, TDNN has more stable detection and localization performance than TD. In Fig. 10, we shows the performance of SDNN and SD methods with K = 2, d = 2. As p increases, SD has good detection performance, but its localization performance slightly decreases. Obviously, the results in both ND and NL tasks show that in our setting p does not have a significant effect on the performance of SDNN. SDNN still can provide stable performance for detecting and localizing attackers. These results suggest that the proposed AI-based models may fit well with irregular degree networks. _C. Performance for Different Degree-|Ni| agents_ In this subsection, we discuss the scenario where the communication network is an irregular network. We assume that the number of mismatched inputs not matching the unified model is p = M −|Ni|. According to the scheme described in subsection IV-C, we test the detection and localization performance of AI-based methods when p = 0. To set up _̸_ this simulation the Manhattan network topology is selected _D. Robustness test when data is mismatched with prior infor-_ _mation_ Furthermore, we test the robustness of AI-based methods when the prior information is inconsistent with the actual environment. The prior information of test scenarios are present in subsection IV-C. We consider one attacker in the Manhattan network and train the parameters of AI-based methods in the scenario with α[k] [ 0.5, 0.5][d] and β [0.0, 1.0][d], and _∼U_ _−_ _∼U_ the test data is generated by other scenarios in TABLE I. In Fig. 11, we show the detection and localization performance of TDNN and TD methods in different test scenarios when K 5 d 2 Specifically we generate the test data ----- 1 Detection, agent i next to attacker 1 Localization, agent i next to attacker 1 Localization, agent i next to attacker 1 Detection, agent i next to attacker TD m=1, c=1 TD m=2, c=1 TD m=2, c=2 TD m=5, c=1 TD m=5, c=2 TD m=5, c=3 TDNN m=1, c=1 TDNN m=2, c=1 TDNN m=2, c=2 TDNN m=5, c=1 TDNN m=5, c=2 TDNN m=5, c=3 0 0.2 0.4 0.6 Probability of false alarm Pinf |Detection, a|Col2|agent i next|Col4|Col5|t to attacker|Col7| |---|---|---|---|---|---|---| |||||||| |||||||| |||||||| |||||||| ||SD SD|K=1, ~ U[0. K=1, ~ U[0.|||0, 1.0] 2, 0.8]|| ||SD|K=1,|~ U[-0||.2, 1.2]|| ||SD SD|K=1, ~ U[0. K=1, ~ U[-0|||2, 1.2] .2, 0.8]|| ||SDN SDN|N K=1, N K=1,||~ ~|U[0.0, 1.0] U[0.2, 0.8]|| ||SDN SDN|N K=1, N K=1,||~ ~|U[-0.2, 1.2] U[0.2, 1.2]|| |SDN||N K=1, ~|||U[-0.2, 0.8]|| |ocalization,|, agent i nex|Col3|Col4|xt to attack| |---|---|---|---|---| |||||| |||||| |||||| |||||| |SD K SD K|=1, ~ U[0. =1, ~ U[0.|||0, 1.0] 2, 0.8]| |SD K|=1,|~ U[-0||.2, 1.2]| |SD K SD K|=1, =1,|~ U[0. ~ U[-0||2, 1.2] .2, 0.8]| |SDNN SDNN|K=1, K=1,||~ ~|U[0.0, 1.0] U[0.2, 0.8]| |SDNN SDNN|K=1, K=1,||~ ~|U[-0.2, 1.2] U[0.2, 1.2]| |SDNN|K=1, ~|||U[-0.2, 0.8]| |Detection, a|Col2|agent i next|t to attacker|Col5| |---|---|---|---|---| |||||| |||||| |||||| |||TD m=1, c= TD m=2, c=|1 1|| |||TD m=2, c=|2|| |||||| |||TD m=5, c= TD m=5, c=|1 2|| |||||| |||TD m=5, c= TDNN m=1|3, c=1|| |||||| |||TDNN m=2 TDNN m=2|, c=1, c=2|| |||||| |||TDNN m=5 TDNN m=5|, c=1, c=2|| |||TDNN m=5|, c=3|| |ocalization,|Col2|, agent i nex|xt to attack|Col5| |---|---|---|---|---| |||||| |||||| |||||| |||TD m=1, c=|1|| |||TD m=2, c= TD m=2, c=|1 2|| |||||| |||TD m=5, c= TD m=5, c=|1 2|| |||||| |||TD m=5, c= TDNN m=1,|3 c=1|| |||||| |||TDNN m=2, TDNN m=2,|c=1 c=2|| |||||| |||TDNN m=5, TDNN m=5,|c=1 c=2|| |||TDNN m=5,|c=3|| 0 0.2 0.4 0.6 Probability of false alarm Pilf 0 0.2 0.4 0.6 Probability of false alarm Pilf 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 Probability of false alarm Pinf 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Fig. 12. ROCs of SDNN and SD for the mismatch model: (Left) ND task, (Right) NL task. α[k] _∼U_ [−0.5, 0.5][d]. Each entry of x[k](0) is distributed as legended for testing data. Fig. 14. ROCs for multiple attackers of TDNN and TD: (Left) ND task, (Right) NL task. m is the number of attackers in the Manhattan network, and _c is the number of attackers in the testing agent’s neighborhood._ Localization, agent i next to attacker 1 Localization, agent i next to attacker 1 1 Detection, agent i next to attacker 1 0.95 0.9 0 0.02 0.04 0.06 SD K=2, ~ U[0.0, 1.0] SD K=2, ~ U[0.2, 0.8] SD K=2, ~ U[-0.2, 1.2] SD K=2, ~ U[0.2, 1.2] SD K=2, ~ U[-0.2, 0.8] SDNN K=2, ~ U[0.0, 1.0] SDNN K=2, ~ U[0.2, 0.8] SDNN K=2, ~ U[-0.2, 1.2] SDNN K=2, ~ U[0.2, 1.2] SDNN K=2, ~ U[-0.2, 0.8] |Detection, a|Col2|Col3|agent i next|Col5|Col6|Col7|t to attacker|Col9|Col10| |---|---|---|---|---|---|---|---|---|---| |1|||||||||| ||||||||||| |0.95 0.9|||||||||| |0|||0.02 0.04||||0.06||| |SD K|||=2, ~ U[0.||||0, 1.0]||| ||SD K SD K||=2, =2,||~ U[0. ~ U[-0||2, 0.8] .2, 1.2]||| ||SD K SD K SDN SDN SDN||=2, ~ U[0. =2, ~ U[-0 N K=2, ~ U N K=2, ~ U N K=2, ~ U||||2, 1.2] .2, 0.8] [0.0, 1.0] [0.2, 0.8] [-0.2, 1.2]||| |ocalization,|Col2|, agent i nex|Col4|xt to attack| |---|---|---|---|---| |||||| |1||||| |0.95 0.9||||| |0||0.02 0||.04| |SD K||=2, ~ U[0||.0, 1.0]| |SD K SD K||=2, =2,|~ U[0 ~ U[-|.2, 0.8] 0.2, 1.2]| |SD K SD K SDN SDN SDN||=2, ~ U[0 =2, ~ U[- N K=2, ~ N K=2, ~ N K=2, ~||.2, 1.2] 0.2, 0.8] U[0.0, 1.0] U[0.2, 0.8] U[-0.2, 1.2]| |ocalization,|Col2|, agent i nex|xt to attacke|Col5| |---|---|---|---|---| |||||| |||||| |||||| |||SD m=1,c=|1|| |||||| |||SD m=2,c= SD m=2,c=|1 2|| |||SD m=5,c= SD m=5,c=|1 2|| |||SD m=5,c= SDNN m=1 SDNN m=2 SDNN m=2 SDNN m=5|3,c=1,c=1,c=2,c=1|| 0 0.2 0.4 0.6 Probability of false alarm Pilf 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 Probability of false alarm Pinf 0 0.2 0.4 0.6 Probability of false alarm Pilf 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Fig. 13. ROCs of SDNN and SD for the mismatch model: (Left) ND task, (Right) NL task. α[k] _∼U_ [−0.5, 0.5][d]. Each entry of x[k](0) is distributed as legended for testing data. for the second and third curves by changing the deviation of **_β to β_** [0.2, 0.8][d] and β [ 0.2, 1.2][d]. The results _∼U_ _∼U_ _−_ indicate that the performances of TDNN and TD methods deteriorate when the deviation of β increases, and improve when the deviation of β decreases. While in the fourth and fifth curves, we change the mean of β to β [0.2, 1.2][d] _∼U_ and β [ 0.2, 0.8][d]. As can be seen that the performances _∼U_ _−_ of TDNN and TD will improve when [E[α] − E[β]] increases and deteriorate when the gap decreases. Meanwhile, TDNN performs better than TD in both ND and NL tasks. In addition, Fig. 12 and Fig. 13 respectively show the detection and localization performance of SDNN and SD methods when _K = 1, d = 2 and K = 2, d = 2. In these plots, the_ ROC curves of SDNN and SD follow the same trends as those in Fig. 11. It is worth mentioning that SDNN still shows good detection and localization performance despite the mismatching of the training data and testing data. Specifically, in Fig. 11, Fig. 12 and Fig. 13. _E. Detection and Localization for Multiple Attackers_ Then, we investigated the performance of TDNN and SDNN with the case of multiple attackers. Note that the parameters of AI-based methods are the same as those in subsection V-A. In Fig 14 and 15 we set agents 1 _m_ as the attackers _{_ _}_ Fig. 15. ROCs for multiple attackers of SDNN and SD: (Left) ND task, (Right) NL task. m is the number of attackers in the Manhattan network, and _c is the number of attackers in the testing agent’s neighborhood._ when considering a scenario with m attackers in the Manhattan network. The legend in these plots with ‘m and c’ indicates that there are m attackers in the Manhattan network, and c attackers are in the neighborhood of the monitoring agent. In this case, the same α[k] is shared by all cooperating attackers, but the noise is random and independent among each attacker. In Fig. 14, we show the ROC curves of TDNN and TD methods when K = 5, d = 2. Obviously, both the detection and localization performance of TDNN and TD methods fluctuate obviously in different m and c. We notice that the total number of attackers (m) has only a slight impact on the detection performance of TDNN, which can be seen from the sixth (m = 1, c = 1), seventh (m = 2, c = 1) and ninth curves (m = 5, c = 1). It shows that the detection performance for TDNN depends on the number of attacking neighbors. As for NL task, we observe that TDNN exhibits similar performance in different attack scenarios. Nevertheless, the proposed TDNN method outperforms TD and has good performance in the case of multiple attackers. For SDNN and SD methods, the detection and localization performance are shown in Fig. 15 with K = 2, d = 2. From the Fig. 15 (Left), SDNN and SD still show good detection performance in the case of multiple attackers. We notice that the detection performance of SDNN is a slightly better than SD when ----- VI. CONCLUSION 1 Detection, agent i next to attacker Manhattan K=2, d=2 Small-world K=2, d=2 Manhattan K=5, d=2 Small-world K=5, d=2 0 0.2 0.4 0.6 Probability of false alarm Pinf Localization, agent i next to attacker 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 |Detection, a|Col2|agent i next|t to attacker|Col5| |---|---|---|---|---| |||||| |||||| |||||| |||||| ||M S|anhattan K mall-world K|=2, d=2 =2, d=2|| ||M S|anhattan K mall-world K|=5, d=2 =5, d=2|| |||||| |||||| |||||| |||||| |ocalization|n, agent i ne|ext to attacke|Col4| |---|---|---|---| ||||| ||||| ||||| ||||| |M|anhattan|K=2, d=2|| |S M|mall-world anhattan|K=2, d=2 K=5, d=2|| |S|mall-world|K=5, d=2|| ||||| ||||| ||||| 0 0.2 0.4 0.6 Probability of false alarm Pilf Fig. 16. ROCs of TDNN for the small world network: (Left) ND task, (Right) NL task. Solid lines show the average detection and localization performance in the small world network, and the parameters of TDNN are trained by the Manhattan network. This work is dedicated to the detection of insider attacks in the DPS algorithm through AI technology. We have proposed two AI-based defense strategies (TDNN and SDNN) for securing the gossip-based DPS algorithm. Unlike the traditional score-based methods, this work utilizes NN to learn the complex mapping relationships in this classification problem, thus reducing the design difficulty of the attacker detector. To circumvent the mismatch of the training data and the actual network attack, we propose a federated learning approach to learn a local model close to the global model using training data from all agents. Experiment results demonstrate that the proposed AI-based methods have good detection and localization performance in different attack scenarios. They also have good adaptability to different degree of agent, and have strong robustness to the inconsistency of prior information with the actual environment. Therefore, it is convinced that the proposed AI-based defense strategies have a high potential for practical applications in the DPS algorithm. As a future work, it would be interesting to try the AI-based methods on more complicated attack models and other decentralized algorithms. REFERENCES Localization, agent i next to attacker 1 1 Detection, agent i next to attacker Manhattan K=1, d=2 Small-world K=1, d=2 Manhattan K=2, d=2 Small-world K=2, d=2 0 0.2 0.4 0.6 Probability of false alarm Pinf |Detection, a|Col2|agent i next|t to attacker|Col5| |---|---|---|---|---| |||||| |||||| |||||| |||||| ||M|anhattan K|=1, d=2|| ||Sm M|all-world K anhattan K|=1, d=2 =2, d=2|| |Sm||all-world K|=2, d=2|| |||||| |||||| |||||| |ocalization|n, agent i ne|ext to attacke|Col4| |---|---|---|---| ||||| ||||| ||||| ||||| ||Manhattan Small-world|K=1, d=2 K=1, d=2|| ||Manhattan Small-world|K=2, d=2 K=2, d=2|| ||||| ||||| ||||| ||||| 0 0.2 0.4 0.6 Probability of false alarm Pilf 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Fig. 17. ROCs of SDNN for the small world network: (Left) ND task, (Right) NL task. Solid lines show the average detection and localization performance in the small world network, and the parameters of SDNN are trained by the Manhattan network. _m = 5, c = 3. While in the NL task, SDNN has excellent_ localization performance and outperforms SD in all attack scenarios. _F. Performance Test In a Small World Network_ In addition, we also test the detection and localization performance of AI-based methods in a small world network. We consider a small world network with 20 agents. The average degree is set to be 8 and the rewiring probability is set to be 0.2. We assume that among all the nodes, agents 3, 10 and 17 are the attackers. The AI-based models are trained by the training data from Manhattan network as those in V-A, while the test data are collected from the small world network. We only consider to monitor at the agents next to the attacker. Note that herein the degree-mismatch problem is solved by the proposed method in Section IV-C. We therefore show the performance of TDNN and SDNN in Fig. 16 and Fig. 17 respectively. In these plots, the solid lines show the average detection and localization performance in the small world network. We notice that AI-based methods also exhibit good detection and localization performance in a small world network. These results further illustrate the potential of the proposed defense strategies [1] A. Nedic and A. Ozdaglar, “Distributed subgradient methods for multiagent optimization,” IEEE Transactions on Automatic Control, vol. 54, no. 1, pp. 48–61, 2009. [2] A. Nedic, A. Ozdaglar, and P. A. Parrilo, “Constrained consensus and optimization in multi-agent networks,” IEEE Transactions on Automatic _Control, vol. 55, no. 4, pp. 922–938, 2010._ [3] G. Hattab and D. Cabric, “Distributed wideband sensing-based architecture for unlicensed massive iot communications,” IEEE Transactions on _Cognitive Communications and Networking, vol. 5, no. 3, pp. 819–834,_ 2019. [4] B. V. Philip, T. Alpcan, J. Jin, and M. Palaniswami, “Distributed realtime iot for autonomous vehicles,” IEEE Transactions on Industrial _Informatics, vol. 15, no. 2, pp. 1131–1140, 2019._ [5] S. X. Wu, H.-T. Wai, L. Li, and A. Scaglione, “A review of distributed algorithms for principal component analysis,” Proceedings of the IEEE, vol. 106, no. 8, pp. 1321–1340, 2018. [6] C. Zhang and Y. Wang, “Enabling privacy-preservation in decentralized optimization,” IEEE Transactions on Control of Network Systems, vol. 6, no. 2, pp. 679–689, 2018. [7] D. Gesbert, S. G. Kiani, A. Gjendemsjo, and G. E. Oien, “Adaptation, coordination, and distributed resource allocation in interference-limited wireless networks,” Proceedings of the IEEE, vol. 95, no. 12, pp. 2393– 2409, 2007. [8] G. B. Giannakis, V. Kekatos, N. Gatsis, S.-J. Kim, H. Zhu, and B. F. Wollenberg, “Monitoring and optimization for power grids: A signal processing perspective,” IEEE Signal Processing Magazine, vol. 30, no. 5, pp. 107–128, 2013. [9] I. Heged˝us, G. Danner, and M. Jelasity, “Gossip learning as a decentralized alternative to federated learning,” in IFIP International Conference _on Distributed Applications and Interoperable Systems. Springer, 2019,_ pp. 74–90. [10] J. Tsitsiklis, “Problems in decentralized decision making and computation,” Ph.D. dissertation, Dept. of Electrical Engineering and Computer Science, M.I.T., Boston, MA, 1984. [11] S. S. Ram, A. Nedi´c, and V. V. Veeravalli, “Distributed stochastic subgradient projection algorithms for convex optimization,” Journal of _optimization theory and applications, vol. 147, no. 3, pp. 516–545, 2010._ [12] S. Sundaram and B. Gharesifard, “Consensus-based distributed optimization with malicious nodes,” in 2015 53rd Annual Allerton Con_ference on Communication, Control, and Computing (Allerton)._ IEEE, 2015, pp. 244–249. [13] S. X. Wu, H.-T. Wai, A. Scaglione, A. Nedi´c, and A. Leshem, “Data injection attack on decentralized optimization,” in 2018 IEEE International _Conference on Acoustics, Speech and Signal Processing (ICASSP)._ IEEE 2018 3644 3648 ----- [14] G. Li, X. Wu, S. Zhang, H.-T. Wai, and A. Scaglione, “Detecting and localizing adversarial nodes using neural networks,” in 2018 IEEE _19th International Workshop on Signal Processing Advances in Wireless_ _Communications (SPAWC) (IEEE SPAWC 2018), Kalamata, Greece, Jun._ 2018. [15] Q. Yan, M. Li, T. Jiang, W. Lou, and Y. T. Hou, “Vulnerability and protection for distributed consensus-based spectrum sensing in cognitive radio networks,” in INFOCOM, 2012 Proceedings IEEE. IEEE, 2012, pp. 900–908. [16] C. Zhao, J. He, and J. Chen, “Resilient consensus with mobile detectors against malicious attacks,” IEEE Transactions on Signal and Information _Processing over Networks, vol. 4, no. 1, pp. 60–69, 2017._ [17] S. Sundaram and B. Gharesifard, “Distributed optimization under adversarial nodes,” IEEE Transactions on Automatic Control, vol. 64, no. 3, pp. 1063–1076, 2018. [18] R. Gentz, H.-T. Wai, A. Scaglione, and A. Leshem, “Detection of datainjection attacks in decentralized learning,” Asilomar Conf, 2015. [19] R. Gentz, S. X. Wu, H. T. Wai, A. Scaglione, and A. Leshem, “Data injection attacks in randomized gossiping,” IEEE Transactions on Signal _and Information Processing Over Networks, vol. 2, no. 4, pp. 523–538,_ 2016. [20] M. Mobilia, “Does a single zealot affect an infinite group of voters ?” _Physical Review Letters, July 2003._ [21] B. Kailkhura, S. Brahma, and P. K. Varshney, “Data falsification attacks on consensus-based detection systems,” IEEE Transactions on Signal _and Information Processing over Networks, vol. 3, no. 1, pp. 145–158,_ 2017. [22] ——, “Consensus based detection in the presence of data falsification attacks,” IEEE Transactions on Signal Processing, vol. PP, no. 99, 2015. [23] O. Shalom, A. Leshem, A. Scaglione, and A. Nedi´c, “Detection of data injection attacks on decentralized statistical estimation,” in 2018 IEEE _International Conference on the Science of Electrical Engineering in_ _Israel (ICSEE)._ IEEE, 2018, pp. 1–5. [24] N. Ravi and A. Scaglione, “Detection and isolation of adversaries in decentralized optimization for non-strongly convex objectives,” IFAC_PapersOnLine, vol. 52, no. 20, pp. 381–386, 2019._ [25] S. Patel, V. Khatana, G. Saraswat, and M. V. Salapaka, “Distributed detection of malicious attacks on consensus algorithms with applications in power networks,” in Preprint. Golden, CO: National Renewable _Energy Laboratory. NREL/CP-5D00-76848, 2020._ [26] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in CVPR. IEEE Computer Society, 2016, pp. 770–778. [27] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv _preprint arXiv:1810.04805, 2018._ [28] H. Wang, J. Ruan, G. Wang, B. Zhou, Y. Liu, X. Fu, and J. Peng, “Deep learning-based interval state estimation of ac smart grids against sparse cyber attacks,” IEEE Transactions on Industrial Informatics, vol. 14, no. 11, pp. 4766–4778, 2018. [29] Y. Yu, T. Wang, and S. C. Liew, “Deep-reinforcement learning multiple access for heterogeneous wireless networks,” IEEE Journal on Selected _Areas in Communications, vol. 37, no. 6, pp. 1277–1290, 2019._ [30] W. Wu, R. Li, G. Xie, J. An, Y. Bai, J. Zhou, and K. Li, “A survey of intrusion detection for in-vehicle networks,” IEEE Transactions on _Intelligent Transportation Systems, vol. 21, no. 3, pp. 919–933, 2019._ [31] A. Doboli, “Discovery of malicious nodes in wireless sensor networks using neural predictors,” Wseas Transactions on Computers Research, vol. 2, 2007. [32] G. Rusak, A. Al-Dujaili, and U.-M. O’Reilly, “Ast-based deep learning for detecting malicious powershell,” in Proceedings of the 2018 ACM _SIGSAC Conference on Computer and Communications Security, 2018,_ pp. 2276–2278. [33] O. Rahman, M. A. G. Quraishi, and C.-H. Lung, “Ddos attacks detection and mitigation in sdn using machine learning,” in 2019 IEEE World _Congress on Services (SERVICES), vol. 2642._ IEEE, 2019, pp. 184– 189. [34] G. Li, S. X. Wu, S. Zhang, H.-T. Wai, and A. Scaglione, “Detecting and localizing adversarial nodes usig neural networks,” in 2018 IEEE _19th International Workshop on Signal Processing Advances in Wireless_ _Communications (SPAWC)._ IEEE, 2018, pp. 1–5. [35] G. Li, S. X. Wu, S. Zhang, and Q. Li, “Neural networks-aided insider attack detection for the average consensus algorithm,” IEEE Access, vol. 8, pp. 51 871–51 883, 2020. [36] ——, “Detect insider attacks using cnn in decentralized optimization,” in _ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech_ _d Si_ _l P_ _i_ _(ICASSP)_ IEEE 2020 8758 8762 [37] L. Giaretta and S. Girdzijauskas, “Gossip learning: off the beaten path,”[ˇ] in 2019 IEEE International Conference on Big Data (Big Data). IEEE, 2019, pp. 1117–1124. [38] S. Savazzi, M. Nicoli, and V. Rampa, “Federated learning with cooperating devices: A consensus approach for massive iot networks,” IEEE _Internet of Things Journal, vol. 7, no. 5, pp. 4641–4654, 2020._ [39] Z. Jiang, A. Balu, C. Hegde, and S. Sarkar, “Collaborative deep learning in fixed topology networks,” in Advances in Neural Information _Processing Systems (NIPS), 2017, pp. 5904–5914._ [40] R. Orm´andi, I. Heged˝us, and M. Jelasity, “Gossip learning with linear models on fully distributed data,” Concurrency and Computation: _Practice and Experience, vol. 25, no. 4, pp. 556–571, 2013._ [41] M. Blot, D. Picard, N. Thome, and M. Cord, “Distributed optimization for deep learning with gossip exchange,” Neurocomputing, vol. 330, pp. 287–296, 2019. [42] J. Daily, A. Vishnu, C. Siegel, T. Warfel, and V. Amatya, “Gossipgrad: Scalable deep learning using gossip communication based asynchronous gradient descent,” arXiv preprint arXiv:1803.05880, 2018. [43] M. Blot, D. Picard, and M. Cord, “Gosgd: Distributed optimization for deep learning with gossip exchange,” arXiv preprint arXiv:1804.01852, 2018. [44] R. B. Palm, “Prediction as a candidate for learning deep hierarchical models of data,” Technical University of Denmark, vol. 5, 2012. [45] T. Fawcett, “An introduction to roc analysis,” Pattern recognition letters, vol. 27, no. 8, pp. 861–874, 2006. -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2101.06917, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "http://arxiv.org/pdf/2101.06917" }
2,021
[ "JournalArticle" ]
true
2021-01-18T00:00:00
[ { "paperId": "edfb119b453a7ee7055eeab7e407c4334ae57ed7", "title": "Distributed Detection of Malicious Attacks on Consensus Algorithms with Applications in Power Networks" }, { "paperId": "c7c75508eca75e7a2c29724f5b735c24c0d3f72e", "title": "Detect Insider Attacks Using CNN in Decentralized Optimization" }, { "paperId": "2d8db2937d3015d8ddc5ab16dfb9628c2d393430", "title": "Neural Networks-Aided Insider Attack Detection for the Average Consensus Algorithm" }, { "paperId": "f97424b7026d344ccab301da4da81ef292a89aee", "title": "A Survey of Intrusion Detection for In-Vehicle Networks" }, { "paperId": "0d7e26c623068f7119878f12f5ee1a49b20b9c9d", "title": "Federated Learning With Cooperating Devices: A Consensus Approach for Massive IoT Networks" }, { "paperId": "0ef9776f1abd2111c3d89101fd71a26273f3b7ae", "title": "Gossip Learning: Off the Beaten Path" }, { "paperId": "aa551505e92f565e2985f478f9cb01442b58a6b2", "title": "Detection and Isolation of Adversaries in Decentralized Optimization for Non-Strongly Convex Objectives" }, { "paperId": "56c27a7b93cd8fc372f8459d151ee589e752b8e0", "title": "DDoS Attacks Detection and Mitigation in SDN Using Machine Learning" }, { "paperId": "5940b6a654300ed7bd6e495407b2002e5b33ba98", "title": "Gossip Learning as a Decentralized Alternative to Federated Learning" }, { "paperId": "2de3693d22b2ae489ccafd2a447f06c0dcb58bf8", "title": "Enabling Privacy-Preservation in Decentralized Optimization" }, { "paperId": "900a44e087e5ea69ddaea883fdc4e7cb46b1ad90", "title": "Distributed Real-Time IoT for Autonomous Vehicles" }, { "paperId": "b44144f12c522b7cdf2bff9f5a2be23ce53058c4", "title": "Distributed Wideband Sensing-Based Architecture for Unlicensed Massive IoT Communications" }, { "paperId": "989c107a890690634b4b87b7e5e2d35f1484ba9a", "title": "Detection of Data Injection Attacks on Decentralized Statistical Estimation" }, { "paperId": "77369f12dd131a755129b1b5b923b1d479eff5db", "title": "AST-Based Deep Learning for Detecting Malicious PowerShell" }, { "paperId": "bd2c2165197ac4ee13cc8a692e46703c43b37623", "title": "Data Injection Attack on Decentralized Optimization" }, { "paperId": "5bcd7726beba0d40cbaf218ff57f9ffdea2b14d4", "title": "A Review of Distributed Algorithms for Principal Component Analysis" }, { "paperId": "c4b3439af61021703b6f589b0a6f1380bc51439f", "title": "Detecting and Localizing Adversarial Nodes Usig Neural Networks" }, { "paperId": "dcd206eb55da934f996b7ab5740f2a3630e1501e", "title": "GoSGD: Distributed Optimization for Deep Learning with Gossip Exchange" }, { "paperId": "a7e9f6c55c1118c9947c6ef63bddd11764b85d33", "title": "GossipGraD: Scalable Deep Learning using Gossip Communication based Asynchronous Gradient Descent" }, { "paperId": "b6691a340b287047bf1e56cbb3224113a2791d89", "title": "Resilient Consensus with Mobile Detectors Against Malicious Attacks" }, { "paperId": "1ee1ab9ea9e79a3de7f5a1b8e0dcc882d9e8b5de", "title": "Deep Learning-Based Interval State Estimation of AC Smart Grids Against Sparse Cyber Attacks" }, { "paperId": "953c6f0c1ae5311e1bcc523fbba2de57d751c248", "title": "Deep-Reinforcement Learning Multiple Access for Heterogeneous Wireless Networks" }, { "paperId": "4d886c571c0849fda73a2d24e944d59fd37bcf9c", "title": "Collaborative Deep Learning in Fixed Topology Networks" }, { "paperId": "ac80c839dc4b47062af7db9ccf71b8c406bfab3f", "title": "Distributed Optimization Under Adversarial Nodes" }, { "paperId": "2c03df8b48bf3fa39054345bafabfeff15bfd11d", "title": "Deep Residual Learning for Image Recognition" }, { "paperId": "45532b07468f95ccd1509e7783a0ebb6d046320e", "title": "Detection of data injection attacks in decentralized learning" }, { "paperId": "2916496933fe0f4e35ad074bd53e63fe0393cfee", "title": "Consensus-based distributed optimization with malicious nodes" }, { "paperId": "26ae65513b8c5d93a881fb04a39213f85a3227be", "title": "Data Falsification Attacks on Consensus-Based Detection Systems" }, { "paperId": "a45c0fb9ab7f9933656c4e1b45616625455d6c03", "title": "Monitoring and Optimization for Power Grids: A Signal Processing Perspective" }, { "paperId": "06818ca61e095700b021debb4bf0b693e602f4e2", "title": "Vulnerability and protection for distributed consensus-based spectrum sensing in cognitive radio networks" }, { "paperId": "cbca96cd533e2f0f7097277e16a597d8bd073158", "title": "Gossip learning with linear models on fully distributed data" }, { "paperId": "92a203af1a25201dc09148c63adf7bbb3ca56112", "title": "Distributed Subgradient Methods for Multi-Agent Optimization" }, { "paperId": "f6c3b74e5e1d89b56da324edc87958659c05c8bf", "title": "Distributed Stochastic Subgradient Projection Algorithms for Convex Optimization" }, { "paperId": "eca39922b65b91c53c195f0ea6c1380850f2db19", "title": "Constrained Consensus and Optimization in Multi-Agent Networks" }, { "paperId": "d40ee5dd758c525dfb9932d726bb4e844b7b8478", "title": "An introduction to ROC analysis" }, { "paperId": "7ca531a49c8f9a53cedc58a90ddfe826809db75d", "title": "Does a single zealot affect an infinite group of voters?" }, { "paperId": "49d003dec54e01141d885df0fd41a23427a82cab", "title": "Problems in decentralized decision making and computation" }, { "paperId": "df2b0e26d0599ce3e70df8a9da02e51594e0e992", "title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding" }, { "paperId": "e6c08dd8f02b292b17f70f3094f7321149657bd5", "title": "Data Injection Attacks in Randomized Gossiping" }, { "paperId": null, "title": "Consensus based detection in the presence of data falsification attacks" }, { "paperId": "7c616fe341381a4866135042dbb565d2eda415c3", "title": "Prediction as a candidate for learning deep hierarchical models of data" }, { "paperId": "dc28a23b92d9c2f6fdf46bbac47f0c99c8318401", "title": "Discovery of Malicious Nodes in Wireless Sensor Networks Using Neural Predictors" }, { "paperId": null, "title": "“Adaptation, coordination, and distributed resource allocation in interference-limited wireless networks,”" } ]
26,407
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/014fdd56682f5f3e91789e83114aad81f5aa5670
[ "Computer Science", "Medicine" ]
0.851831
Adaptive Square-Shaped Trajectory-Based Service Location Protocol in Wireless Sensor Networks
014fdd56682f5f3e91789e83114aad81f5aa5670
Italian National Conference on Sensors
[ { "authorId": "8782316", "name": "Hwa-Jung Lim" }, { "authorId": "66152655", "name": "Joahyoung Lee" }, { "authorId": "70342309", "name": "Heonguil Lee" } ]
{ "alternate_issns": null, "alternate_names": [ "SENSORS", "IEEE Sens", "Ital National Conf Sens", "IEEE Sensors", "Sensors" ], "alternate_urls": [ "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-142001", "http://www.mdpi.com/journal/sensors", "https://www.mdpi.com/journal/sensors" ], "id": "3dbf084c-ef47-4b74-9919-047b40704538", "issn": "1424-8220", "name": "Italian National Conference on Sensors", "type": "conference", "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-142001" }
In this paper we propose an adaptive square-shaped trajectory (ASST)-based service location method to ensure load scalability in wireless sensor networks. This first establishes a square-shaped trajectory over the nodes that surround a target point computed by the hash function and any user can access it, using the hash. Both the width and the size of the trajectory are dynamically adjustable, depending on the number of queries made to the service information on the trajectory. The number of sensor nodes on the trajectory varies in proportion to the changing trajectory shape, allowing high loads to be distributed around the hot spot area.
_Sensors_ **2010, 10, 4497-4520; doi:10.3390/s100504497** _Article_ **OPEN ACCESS** # sensors **ISSN 1424-8220** www.mdpi.com/journal/sensors ## Adaptive Square-Shaped Trajectory-Based Service Location Protocol in Wireless Sensor Networks **Hwa-Jung Lim, Joa-Hyoung Lee and Heon-Guil Lee *** Dept. of Computer Science and Engineering, Kangwon National University, Chuncheon Gangwondo, 200-701, Korea; E-Mails: jinnie4u@kangwon.ac.kr (J.-H.L.); hjlim@kangwon.ac.kr (H.-J.L.) - Author to whom correspondence should be addressed; E-Mail: hglee@kangwon.ac.kr; Tel.: +82-01-4914-0107; Fax: +82-33-252-6390. _Received: 22 March 2010; in revised form: 8 April 2010 / Accepted: 14 April 2010 /_ _Published: 30 April 2010_ **Abstract:** In this paper we propose an adaptive square-shaped trajectory (ASST)-based service location method to ensure load scalability in wireless sensor networks. This first establishes a square-shaped trajectory over the nodes that surround a target point computed by the hash function and any user can access it, using the hash. Both the width and the size of the trajectory are dynamically adjustable, depending on the number of queries made to the service information on the trajectory. The number of sensor nodes on the trajectory varies in proportion to the changing trajectory shape, allowing high loads to be distributed around the hot spot area. **Keywords: service location; trajectory; data replication; load scalability; robust** **1. Introduction** Advances in wireless networking have set new paradigms in computing, including pervasive computing based on a large-scale wireless sensor network. A wireless sensor network, a type of ad hoc network (MANets), is designed to be an infrastructure-less, unattended, and rapidly-deployable network. A fundamental issue in wireless sensor network environments is the efficient location of the required service in the network. The service location protocol is imperative to the design of a wireless sensor network because each network node lacks prior knowledge of the service available in the network [1-7]. ----- _Sensors 2010, 10_ **4498** Service location in wireless sensor networks is a challenging problem for several reasons. First, due to a lack of infrastructure, there are no well-known servers in a pre-defined network structure. Second, energy scarcity in a network node in a wireless network necessitates the design of new service location protocols that are qualitatively different from those designed for the wired network. Third, in many cases, wireless networks may scale up to thousands of nodes, rendering the location problem even more challenging [8-20]. In pervasive computing, users receive information regarding the environment in real-time; therefore, the sensor network, which is the foundation of pervasive computing, should enable real-time access. In particular, service information is very time critical in pervasive computing; therefore, the service location protocol for the wireless sensor network should provide high accessibility to service information [21-23]. The easiest way to provide high accessibility is to periodically broadcast (flood) service information to the entire network. This method entails major energy consumption, but it is simple and some protocols use this approach. To reduce the overhead associated with broadcasting, some protocols restrict the flooding area by forwarding packets in a specific direction, as cross shape or restricted regions. These schemes could reduce the broadcasting overhead but still require unnecessary replications if the service information is not popular. Load scalability is the ease with which a distributed system can expand and contract its resource pool to accommodate heavier or lighter loads; it is the ease with which a system or component can be modified, added, or removed to accommodate a changing load. Service location protocols should rapidly provide service information with a large number of users. Therefore, load scalability is an important metric for a service location protocol [24-40]. In this paper, we propose an adaptive square-shaped trajectory (ASST)-based service location method, which is a novel self-configuring, scalable, energy efficient, and robust service location protocol. ASST is based on Geographic Hash Table (GHT) and Trajectory Based Forwarding (TBF). GHT maps the geographic position of a sensor network field to a hash table. In GHT, the sensor node closest to the position where is computed by hash function is responsible for a set of key and data [6,8,9,11]. ASST stores service information in groups of sensor nodes, called a trajectory. A node wishing to publish (advertise) service information obtains a position through the hash function, and it then uses geographic-aided routing such as GPSR (Greedy Perimeter Stateless Routing) to store service information to the trajectory surrounding the hashed position, as in GHT [6,7]. ASST uses TBF to form a trajectory storing the service information. Replication between nodes in the trajectory reduces the network load on a node because queries from users are distributed to several nodes in the trajectory. To further distribute the network load, ASST adjusts the range and size of the trajectory in proportion to the frequency of user queries. In the next section, we review related work. Section 3 describes ASST, and Section 4 provides performance evaluation. We conclude the paper in Section 5. **2. Related Work** Conventional solutions related to this paper can be classified into the following two approaches: Data Storage architecture in a wireless sensor network and Service Location protocols in an _ad hoc_ network, as shown in Table 1 [4,5,15,22,34,37,40]. ----- _Sensors 2010, 10_ **4499** **Table 1. Classification of Data Storage Schemes and Service Location Protocols.** **Data Storage Scheme** **Service Location Protocol** **Wired Networks** External Storage Service Directory Internal Storage Flooding Data Centric Storage FMMS **Wireless Networks** GHT GCLP In the early days of sensor networks, the data sensed by sensor nodes were collected by a base station and stored externally. Here, external storage means that the node providing the storage space is located separately from the sensor networks. For users on a wired network seeking to access and use the sensed data, there are no problems associated with the external storage. However, if users are mobile, it is difficult to access the external storage on the wired network. To address this problem, internal storage architectures are proposed. In internal storage, each sensor node saves the sensed data to its local storage. Users obtain the data by directly querying the sensor network. Internal storage architecture can reduce the data collection overhead; however, users have to query the entire network to find the data. Data-centric storage architecture provides fast data dissemination by storing the data on the basis of its name. Data-centric storage is an enhanced version of data-centric routing. A data-centric routing scheme presented firstly is directed diffusion. Directed diffusion uses flooding to advertise the interests from sinks to sources throughout the network [4,13]. GHT is a type of data-centric storage architecture. GHT is based on the Distributed Hash Table (DHT) that is the results of research efforts on peer-to-peer (P2P) computing networks. GHT was proposed for data-centric storage with geographic information in a sensor network. GHT is a geographic hash table system that hashes keys into geographic points and stores the key-value set at the sensor node closest to the hashed point. GHT uses geographic perimeter routing to identify a packet home node. GHT provides fast access to the data in the sensor nodes but does not take account of availability and scalability. In GHT, only the home node responds to user queries. This causes a concentration of network load and reduces the energy of the home node [6,22]. _Ad-hoc networks and DHT share key characteristics in terms of self organization, decentralization,_ redundancy requirements, and limited infrastructure. However, node mobility and the continually changing physical topology pose a special challenge to scalability and the design of a DHT for mobile ad-hoc network. Using DHT over wireless sensor networks has gained a lot of attention in the research arena recently. In wireless sensor network, the most important issue in routing is to gather the routed information coming from sensor nodes to the sink node regardless of the identity of the donating node. The problem in this context is to locate efficiently the sensor node, which holds the data item with the minimum number of intermediate nodes to save network energy. A ScatterPastry platform based on Pastry DHT as an overlay routing platform for distributed applications over wireless sensor network using Scatterweb nodes, a real-world wireless sensor platform was proposed in [27]. A topology-based distributed hash tables (T-DHT) as an infrastructure |Table 1. Classific|cation of Data Storage Schemes and|d Service Location Protocols.| |---|---|---| ||Data Storage Scheme|Service Location Protocol| |Wired Networks|External Storage|Service Directory| |Wireless Networks|Internal Storage|Flooding| ||Data Centric Storage|FMMS| ||GHT|GCLP| ----- _Sensors 2010, 10_ **4500** for data-centric storage, information processing, and routing in ad hoc and sensor networks was introduced in [28]. T-DHTs do not rely on location information and work even in the presence of voids in the network. Using a virtual coordinate system, we construct a distributed hash table which is strongly oriented to the underlying network topology. The mobile hash-table (MHT) [29] addresses this challenge by mapping a data item to a path through the environment. In contrast to existing DHT, MHT does not to maintain routing tables and thereby can be used in networks with highly dynamic topologies. Thus, in mobile environments it stores data items with low maintenance overhead on the moving nodes and allows the MHT to scale up to several ten thousands of nodes. In [30], the appropriateness of using DHT routing paths for service placement in an SBON was evaluated, when aiming to minimize network usage. SBONs are one approach to implementing largescale stream processing systems. A fundamental consideration in an SBON is that of service placement, which determines the physical location of in-network processing services or operators, in such a way that network resources are used efficiently. Service placement consists of two components: node discovery, which selects a candidate set of nodes on which services might be placed, and node selection, which chooses the particular node to host a service. By viewing the placement problem as the composition of these two processes we can trade-off quality and efficiency between them. For this, two DHT-based algorithms for node discovery, which use either the union or intersection of DHT routing paths in the Stream-based overlay networks (SBON), was considered and compared their performance to other techniques. In [31] a GHT based service discovery protocol including the mechanism that constructs topology-aware overlay networks in wireless sensor network was proposed. It does not require a central lookup server and does not rely on multicast or flooding. A Similarity Search Algorithm (SSA) for efficiently processing similarity search queries was proposed in [32]. A data-centric storage structure based on the concept of Hilbert curve and DHT was presented and then an algorithm designed for efficiently probing the most similar data item for the sensor network was proposed. A dynamic geographic hash table for data-centric storage in wireless sensor networks was proposed in [33]. Unbalanced resource utilization problem was addressed by proposing a dynamic GHT solution that relies on two schemes—a temporal-based geographic hash table to achieve overall load balancing among sensor nodes over time and a location selection scheme based on node contribution potential to proactively adapt the system to network dynamics. An effective hotspot storage management schemes to solve the hotspot storage problem in GHT was proposed in [35]. The scheme included the cover-up and multi-threshold mechanisms. The coverup mechanism can adjust to another storage node dynamically when a storage node is full, while the multi-threshold mechanism can spread the data into several storage nodes for load balancing of the sensor nodes. Increasing Ray Search (IRS), an energy efficient and scalable search protocol, and k-IRS, an enhanced variant of IRS based on the GHT was proposed in [36]. The priority of IRS is energy efficiency and sacrifices latency whereas k-IRS is configurable in terms of energy-latency trade-off and this flexibility makes it applicable to varied application scenarios. The basic principle of these protocols is to route the search packet along a set of trajectories called rays that maximizes the likelihood of discovering the target information by consuming least amount of energy. The classical protocols for service location in wired networks rely on a central server called a service directory. The central server advertises its information periodically. Users who want to use a ----- _Sensors 2010, 10_ **4501** service connect to the service directory in order to obtain a service description. In wired networks, the network topologies hardly change; therefore, users can access the service directory anytime. However, the central server cannot be used in a wireless network because the network topology in a wireless network frequently changes. The simplest form of service location is global flooding in the network; however, flooding does not scale well. To overcome the weakness of flooding, restricted flooding techniques are developed, such as the Facilitating Match-Making Service (FMMS) and Geography-based Content Location Protocol (GCLP) [5,15]. In FMMS, a service provider advertises the service in a cross-shaped trajectory along the network, as shown in Figure 1. **Figure 1. Service location in FMMS.** Service provider P sends a service advertisement packet in four directions, and the packet is forwarded until it reaches the boundary of the network forming the publish trajectory. A user C who wants to use the service also propagates the query packet in four directions, and the query packet is forwarded until it reaches the boundary of the network packet forming subscribe trajectory similar to advertisement trajectory. The nodes that belong to both trajectories (publish trajectory and subscribe trajectory) reply to user C with the service descriptions. GCLP reduces the query forwarding overhead by stopping the propagation of a query packet when the subscribe trajectory crosses the publish trajectory. FMMS and GCLP reduce the flooding overhead by limiting the flooding to four trajectories; however, both protocols are unable to provide scalability because they do not consider the query amounts, such that the trajectory always has the same size. ----- _Sensors 2010, 10_ **4502** The distance-sensitive service discovery problem in wireless sensor and actor networks was formalized, and a novel localized algorithm, iMesh was proposed in [38]. Unlike existing solutions, iMesh uses no global computation and generates constant per-node storage load. In iMesh, new service providers (i.e., actors) publish their location information in four directions, updating an information mesh such as GCLP. Information propagation for relatively remote services is restricted by a blocking rule, which also updates the mesh structure. Based on an extension rule, nodes along mesh edges may further advertise newly arrived relatively near service by backward distance-limited transmissions, replacing previously closer service location. The final information mesh is a planar structure constituted by the information propagation paths. It stores locations of all the service providers and serves as service directory. Service consumers (i.e., sensors) conduct a lookup process restricted within their home mesh cells to discover nearby services. We analytically study the properties of iMesh including construction cost and distance sensitivity over a static network model. A service discovery protocol based on hierarchical grid architecture in an ad hoc network which enhanced the GCLP was proposed in [39]. The geographical area was divided into a 2D logical hierarchical grid and the information of available services was registered to a specific location along a predefined trajectory. To enhance resource availability and effective discovery of GCLP, each grid cell selects a directory to cache available services. This work utilizes the transmitting trajectory to improve the efficiency of registration and discovery. First, the service provider registers a service along the proposed register trajectory. The requestor then discovers the service along the discovery trajectory to acquire the service information. **3. ASST** _3.1. Basic Concept_ This section presents the basic design of ASST that is based on the following assumptions: (1) a vast field is covered by a large number of homogeneous sensor nodes that communicate with each other through short-range radios. (2) Each sensor node is aware of its own location and uses geographic routing such as GPSR to accomplish long distance delivery. (3) Service information means the service description that describes the characteristic of the service, as shown in Figure 2. (4) The storage space on the sensor node is sufficiently large to save the service information. Recently, sensor nodes with a very large memory space, such as RISE (RIverside SEnsor) with 1 GB of flash memory, have been developed so that sensor nodes can store a large amount of service information and occupy very little space [17]. ----- _Sensors 2010, 10_ **4503** **Figure 2. Example of service description and service location in a wireless network.** (A)Example of Service Description (B) Service Location in Wireless Network Services have descriptions that include a service name and a service provider ID. Figure 2A shows an example of the service description of a printer. The service description can differ from applications to application and mainly depends on the types of services. Users wishing to use services provided by a wireless sensor network need to obtain the service description. The simple way to obtain the service description is by query flooding the entire network, as shown in Figure 2B. In the figure, a user with a laptop wants to print documents. Query packets with the service name that the user wants to use are broadcasted to all the nodes in the network until they reach the service provider matching the requirement in the query packet. The user connects to the printer server on the basis of the service ----- _Sensors 2010, 10_ **4504** description reply sent by the service provider. The flooding is easy and simple, but it may require a long time to discover the required service provider. Furthermore, the flooding could cause a considerable amount of energy consumption on the sensor nodes as a result of broadcast storming. Therefore, efficient service location protocols that provide service information quickly and with low network load are required in a wireless sensor network. ASST is based on GHT and extends the DHT. Service descriptions that consist of a service name, a service provider ID, _etc., are stored in a distributed manner using an algorithm based on DHT. The_ specific service description is stored in repository nodes that correspond to the hash value of the service name. Geographic routing delivers the information for the service descriptions to and from the repository nodes. When a service provider wants to advertise its service description information on the network, the target point Q(Tx,Ty) hash key corresponding to the service name is first obtained using a well-known hash function such as MD5 and SHA. Once the target point Q(Tx, Ty) hash key is obtained, the service provider finds the repository nodes surrounding the hash value and stores the service description in those nodes. For example, in Figure 3, a node provides a printer service and wants to advertise its service to the network. First, the node finds the repository nodes for the printer by calculating the hash key for the printer service. If the printer’s hash key is (11, 28), nodes around (11, 28) become the repository nodes for the printer. A user wanting to use a printer finds the repository nodes for the printer by calculating the hash key of the printer service, same as the service provider. **Figure 3. Example of ASST.** (A) Service Advertisement ----- _Sensors 2010, 10_ **4505** **Figure 3.** _Cont._ (B) Service Discovery ASST provides a general architecture for service discovery. Both the store and the query operations can be viewed as a general insertion and lookup, and the key can be looked-up in the hash table. Different applications can have different insertion and lookup characteristics. Therefore, different policies should be applied that are based on the characteristics of application. The popularity of data (equivalent to the data query frequency) is one of the most important characteristics of the lookup operation. Different data in an application may have different popularities, and this could lead to an unbalanced load distribution in the existing GHT. ASST transforms the cross shaped trajectory into the square shaped trajectory to control the number of nodes on the trajectory in proportion of popularity of data as shown in Figure 4. In the cross shaped trajectory, the publish trajectories are forwarded toward the boundary of sensor network field and the shape is fixed regardless of popularity of data. Therefore, it is possible some nodes on the trajectory not to receive any subscription trajectory (query packet) if the popularity if data is very low. Moreover, forwarding the publish/subscription trajectory toward the boundary of sensor network field can be overhead when a sensor network is deployed in large area. On the other hand, the shape of square trajectory can be increased or decreased easily and thus can be said more flexible than cross shaped trajectory. ASST distributes the query processing load to several replicated nodes, this is called trajectory zone. Trajectory zone in ASST is a square-shaped region with replica nodes for the data. A sensor node within the trajectory zone is called a repository node, which is responsible for the data storage and query response. To provide high scalability and accessibility for the data, ASST adjusts the range of the trajectory in proportion to the data’s popularity; this is referred to as dynamic trajectory. ----- _Sensors 2010, 10_ **4506** **Figure 4. Cross Shaped Trajectory and Square Shaped Trajectory.** _3.2. Square-Shaped Trajectory_ As in the case of many distributed hash table systems, ASST provides a _(key, value)-based_ associative memory. Services are named with keys. Both the storage of the service and its retrieval are performed using these keys. Any naming scheme that distinguishes the services that users of the sensor network wish to distinctly identify will suffice in ASST [6]. ASST supports the following two operations: **Put(k; v) stores v (the observed data) according to the key k, the name of the service.** **Get(k) retrieves whatever stored value is associated with key k.** First, the backbone node Rn in the trajectory Tr obtains data from the data packet sent by source S, and it then forwards the data packet to the next backbone node along edge E and toward vertex V in a counterclockwise direction, as shown in Figure 5. By overhearing the data packet forwarded by the backbone node, replica nodes around the backbone node obtain the data and store it. Source node S frequently sends data packets to update the data on the repository node. Each repository node keeps a timer for the data it stores. If the timer expires before any update packet is received from the source, then the data is discarded. To prevent query slippage in the trajectory, we use the method proposed in [9]. The method divides the network into a virtual grid and ensures that the trajectory is constructed through the sequential cells with sensor nodes. The grid line in Figure 5 shows the virtual grid in the network; the trajectory is formed by forwarding the data packet through the sequential cells. In [9], the trajectory was cross shaped; the trajectory in this paper is square shaped. However, the process of trajectory formation is the same for both, only the order differs. ----- _Sensors 2010, 10_ **4507** **Figure 5. Trajectory formation.** Trajectory Tr is formed away from the target point Q with a distance H (h, h1, h2,… hn) computed in proportion to the data popularity (the same as the query frequency) p by the source node S. In ASST, source nodes only need to know the target point Q and the distance h to form the trajectory Tr. A node that receives a data packet from source S could check whether or not it is involved in trajectory Tr. Trajectory Tr is shaped with vertex V, edge E, and boundary B. Figure 6 shows the elements and formation process of the trajectory Tr. Vertex V of trajectory Tr is a corner of the square with a distance H(h, h1, h2,… hn) from the target point Q (Tx, Ty). The value of vertex V is one of (Tx – h, Ty – h), (Tx + h, Ty – h ), (Tx – h, Ty + h), and (Tx + h, Ty + h). Edge E of trajectory Tr is a line connecting two vertexes V. Trajectory Tr has a boundary B at each side of edge E with width W. Trajectory Tr is a set of data-centric storage nodes, called repository node Rn. Repository node Rn consists of a backbone node b, which lies on the edge E of trajectory Tr, and replica nodes u, which lie between boundary B and edge E. Source S sends a data packet toward the target point. The header of the data packet contains the target point Q (Tx, Ty), range distance h, Source ID, and the Data ID. Each node between the source node and the target point computes the vertex V and edge E for the data packet. If the node is on the edge or vertex of the trajectory’s square, then the node becomes the backbone node and starts to form trajectory Tr. ----- _Sensors 2010, 10_ **4508** **Figure 6. Square-Shaped Trajectory Formation.** _3.3. Dynamic Trajectory_ To provide scalability for a large sensor network, ASST dynamically adjusts the square range of the trajectory in proportion to the query frequency for the data. If the number of queries for the specific data increases, the range of trajectory Tr also increases. On the other hand, if the number of queries for the data decreases, the range of trajectory Tr decreases. This is known as Dynamic Trajectory. To decide the distance H which is the range of trajectory Tr, ASST computes the data popularity (query frequency) p for the data. Initially, we gather the number of total queries on all repository nodes Rn. This information can be gathered through the summation of the number of queries at each node and by piggybacking this on the data packet. The repository node feeds back the query information by returning the data packet to the source node. First, the backbone node b in trajectory Tr gathers query counts from the replica nodes around it, and then it piggybacks the total query counts, which is a summation of the query counts, for both itself and the replica nodes. Next, the backbone node that receives the data packet forwarded from the previous backbone node, adds both its query counts and the replica node’s query counts to the query counts from the data packet, and then it forwards this to the next backbone node. When the data packet returns to the first backbone node, the data packet is returned to source S, which computes a new distance H. Figure 7 shows the Dynamic Trajectory in ASST. Dynamic Trajectory starts with a minimum distance h1. The source node S sends a data packet with a distance h1. At the next update, the source S sends a data packet with a distance h1 and computes a new distance with the query frequency. The new distance is not applied immediately but at the time of the next update. If the query frequency is increased over the threshold, the distance h is also increased to h2. At distance h2, if the query ----- _Sensors 2010, 10_ **4509** frequency is decreased, the distance is decreased to h1; if the query frequency is increased again, the distance becomes h3. When the distance is increased or decreased, a new trajectory is generated with the new repository nodes. We do not need to manually remove the old trajectory because each repository node has an update timer. If the data is not updated until the timer is expired, the node discards the data and this leads to deformation of the trajectory. Table 2 shows the procedures of dynamic trajectory in ASST. **Figure 7. Dynamic Trajectory Development.** **Table** **2. Dynamic ASST procedure.** _procedure_ _put(_ _k,_ _v_ ) _TP(Tx,Ty)_  _Hash(k)_ _QF_  _get_ _ _Query_ _ _Freq(k)_ _h_  _QF_  _QT_ 2 _Vertex[4]_  {(Tx  _h,Ty_  _h),_ (Tx  _h,Ty_  _h),_ (Tx  _h,Ty_  _h),_ (Tx  _h,Ty_  _h)}_ _Min_ _ _Dis_  _MAX_ _ _DIS_ _for_ _i_  1 _to_ 4 _do_ _dis_  _dis_ tan _ce(my_ _ _P,Vertex[i])_ _if_ _dis_  _Min_ _ _Dis_ _then_ _Min_ _ _dis_  _dis_ _T_ arg _et_ _Vertex  _Vertex[i]_ _Direction_  _i_ _end_ _if_ _end_ _for_ _Send(US,_ _k,_ _v,TP,_ _h,Vertex,T_ arg _et_ _Vertex, _Direction,_ _ET_ ) _return_ ----- _Sensors 2010, 10_ **4510** **Table 2.** _Cont._ _procedure_ _receiveVertex(k,_ _v,TP,_ _h,Vertex,T_ arg _et_ _Vertex, _Direction,_ _ET_ ) _if_ _Us_  _my_ _US _then_ _return_ _my_ _ _US_  _US_ _my_ _ _k_  _k_ _my_ _ _v_  _v_ _my_ _ _ET_  _ET_ _set_ _ _Timer(my_ _ _ET_ ) _if_ _Direction_  _EAST_ _then_ _Direction_  _NORTH_ _else_ _if_ _Direction_  _NORTH_ _then_ _Direction_  _WEST_ _else_ _if_ _Direction_  _WEST_ _then_ _Direction_  _SOUTH_ _else_ _if_ _Direction_  _SOUTH_ _then_ _Direction_  _EAST_ _end_ _if_ _T_ arg _et_ _Vertex  _Vetext[Direction]_ _GN_  _Grid_ _ _Number(Direction)_ _Neighbor[]_  _Neighbor_ _ _Set[GN_ ] _Backbone_  min _dis(TP,_ _h,_ _Neighbor)_ _Send(US,_ _k,_ _v,TP,_ _h,Vertex,T_ arg _et_ _ _Vertex,_ _Direction,_ _Backbon,_ _GN_ ) _return_ _procedure_ _receiveRe_ _pository(k,_ _v,TP,_ _h,Vertex,T_ arg _et_ _Vertex, _Direction,_ _ET_, _Backbone,GN_ ) _if_ _US_  _my_ _US _then_ _return_ _end_ _if_ _if_ _my_ _ _GN_  _GN_ _then_ _my_ _US  _US_ _my_ _ _k_  _k_ _my_ _ _v_  _v_ _myET_  _ET_ _set_ _ _Timer(ET_ ) _end_ _if_ _if_ _Backbone_  _my_ _ _ID_ _then_ _T_ arg _et_ _Vertex  _Vertex[Direction]_ _Neighbor[]_  _Neighbor_ _ _Set[Direction]_ _Backbone_  min _dis(TP,_ _h,_ _Neighbor)_ ----- _Sensors 2010, 10_ **4511** **Table 2.** _Cont._ _end_ _if_ _Send_ (US, _k,_ _v,TP,_ _h,Vertex,T_ arg _et_ _Vertex, _Direction,_ _Backbone)_ _return_ k : key of value v : value TP : Target Point QF : Query Frequency my_P : position of source US : Update Sequence ET : Expire Time GN : Next Grid Number _3.4. Analysis_ 3.4.1. Robustness Sensor networks consist of small fragile sensor nodes with limited resources such as computing power, energy, and network bandwidth. Sensor nodes can easily fail as a result of a physical attack or through energy dissipation. Therefore, sensor nodes should be protected, and systems running on sensor nodes should provide failure tolerance. The service location system should also be able to rapidly provide service information to the user, even if there are node failures in the network. ASST provides failure tolerance with a virtual grid when ASST forms a trajectory and the sensor node forwards the query packet, as shown in Figure 8. **Figure 8. Failure Avoidance and Recovery.** ----- _Sensors 2010, 10_ **4512** **Figure 8.** _Cont._ A virtual grid on the trajectory consists of a backbone node and several replica nodes; other nodes in the virtual grid can respond to a query in the event of node failure, as long as at least one node is alive in the grid. If all the nodes in the grid fail, the query is forwarded to other grids by geographic routing, thus avoiding the failed area. Both the trajectory formation and query forwarding are based on the virtual grid; therefore, the query cannot miss the trajectory, as shown in [9]. When the service provider updates the trajectory, the trajectory is recovered by making a detour around the failed area. This technique can be extended for multiple grid failures. 3.4.2. Load scalability To be load scalable, a scheme has to offer uniform processing times irrespective of the loads. The service location protocol should be scalable to accommodate the variable popularity of a service. ASST increases the distance H of the trajectory in proportion to the query frequency in order to provide a uniform query processing time to the user. The number of virtual blocks on the trajectory 4(2H ) with a distance H is _NUMVB_  _W_, and the number of queries that one virtual block can process during T is _NUM_ _proc_  _T_ tSNproc _Avg_, where tproc is the processing time of one query on the sensor node and SNAvg is the average number of sensor nodes in a virtual block. When the total number of queries during T is NUMquery, the number of virtual block that requires to process the NUMquery is _NUM_ _reqVB_  _NUMNUMqueryproc_ and thus the distance Hquery is _H_ _query_  _W8NUMNUMprocquery_ . The required processing time for NUMquery is _treq_  _NUM_ _query_  _t_ _proc_ . The query processing time per node is: ----- _Sensors 2010, 10_ **4513** _ttotal_  _SNtreqtotal_  _SNNUMAvg_ _query_ _NUM_ _t_ _procVB_  _NUMSN_ _Avg_ _query_ 8HWqueryt _proc_  _NUM8_ _queryW8NUMNUMt_ _proc_ _SN_ _Avg_  _W_  _NUM_ _query_  _t_ _proc_  _NUM_ _query_  _t_ _proc_  _T_ _SN_ _Avg_  _NUMNUMqueryproc_ _SN_ _Avg_  _TNUM_ _SNqueryAvg_ _query_ _proc_ _t_ _proc_ Therefore, there is no delay time on the repository node. 3.4.3. Time and message For the data transmissions from a child sensor node to a parent node, CSMA is used to reserve the channel. We assume that each node knows the number of contending neighbor nodes (m) and contends the channel with the optimal probability p = 1/m. The probability that one contending node wins the channel is psucc = (1 − 1/m)[m−1]. Since the number of slots needed until the successful reservation is a geometric random variable, the average number of contending slots (ACS) is given by: _ACS_ 1  1( 1 )M 1 _M_ where M is the Average number of Neighbor nodes. One node on the grid included to the Trajectory Zone has to forward the packet to next virtual block and therefore, the required time slots for Trajectory Zone (STZ) is given by: 1 8H _STZ_  _ACS_  _NUM_ _VB_  1 _M_ 1  _W_ 1( ) _M_ The main part of procedure in Table 2 is finding the next node which is closest to the trajectory Edge and the time for this routine is in proportion to the number of nodes in a virtual zone, which is in proportion to the width of Trajectory Zone (W). Sorting with distance and finding a node minimum distance requires O(n log n) time. The required time for inter communication between backbone nodes is in proportion to the width of Trajectory Zone (W) and the distance H as shown above and thus can be also O(n log n). As a consequence, the total time complexity is O(n log n). The number of message for Trajectory Zone formation is in proportion to the number of virtual block in Trajectory Zone since every backbone node in virtual blocks has to forward the formation message to next virtual block and thus the number of message for Trajectory Zone formation is given by O(N). **4. Performance Evaluation** In Section 3, we proposed ASST, a new mechanism for data dissemination based on DHT and TBF. In this section, we evaluate the performance of our proposed mechanism in ns-2 simulations. Ns-2 supports detailed simulation of mobile, wireless networks. Our simulation uses an 802.11 radio with a 30 m radio range, rather than the 250 m radio range of IEEE-compliant hardware; this choice is similar ----- _Sensors 2010, 10_ **4514** to that made in the evaluation of GHT with a 40 m radio range. ASST was implemented on the basis of GPSR. 900 sensor nodes were deployed in a 600 m × 600 m field in grid topology. The distance between sensor nodes was 20 m so that each sensor node has 8 neighboring nodes. The sensor nodes at the border line of the field were set as query nodes, frequently sending queries. The target point was set at the center of the field. Each sensor node consumes 0.5 w of power when it sends and 0.2 w of power when it receives. Table 3 provides details of the simulation parameters. RXThresh is a receive threshold, CPThresh capture threshold, and CSThresh carrier-sensing Threshold. **Table 3. Simulation parameters.** **Parameter** **Value** **Parameter** **Value** Radio Propagation Shadowing Routing Protocol GPSR MAC 802.11 Idle Power 0.1 w Energy Queue DropTail Rx Power 0.2 w Consumption Antenna OmniAntenna Tx Power 0.5 w Network size 600 m × 600 m Nodes 900 nodes CPThresh 10.0 Packet size 60 bytes Radio CSThresh 1.559e–11 X 0 Antenna Parameter RXThresh 3.652e–10 Y 0 Parameter Pt 2.5872e–4 Z 1.5 Scalability, as a property of systems, is generally difficult to define; however, it is essential to define the specific requirements for scalability based on the dimensions that are deemed important. In telecommunications and software engineering, scalability is a desirable property of a system, a network, or a process; it indicates the ease with which the system can either handle growing amounts of work, or can be readily enlarged. For example, it can refer to the capability of a system to increase total throughput under an increased load. An algorithm, design, networking protocol, program, or other system is said to be scalable if it is suitably efficient and practical when applied to large situations (e.g., a large input data set or a large number of participating nodes in the case of a distributed system). If the design fails when the quantity increases then it does not scale. In particular, load scalability means the ability of a distributed system to expand easily and to contract its resource pool to accommodate heavier or lighter loads. Alternatively, it is the ease with which a system or component can be modified, added, or removed to accommodate a changing load. In this paper, we use load scalability as an evaluation metric. We evaluate the performance of ASST with a diverse query frequency. We compare ASST with GHT and GCLP. To evaluate the scalability of ASST, we varied the query frequency from 0.5 queries per second to 100 queries per second. To reflect query processing, each query has a query processing time of 100 ms. A replica node receiving the query packet has to wait for 100 ms before responding. To provide load scalability, each query has to maintain the response time of 100 ms. The query was uniformly distributed among sensor nodes. |Col1|Col2|Table 3. Sim|mulation parameters.|Col5|Col6|Col7| |---|---|---|---|---|---|---| |Parameter||Value|Parameter|||Value| |Radio Propagation||Shadowing|Routing Protocol|||GPSR| |MAC||802.11|Energy Consumption||Idle Power|0.1 w| |Queue||DropTail|||Rx Power|0.2 w| |Antenna||OmniAntenna|||Tx Power|0.5 w| |Network size||600 m × 600 m|Nodes|||900 nodes| |Radio Parameter|CPThresh|10.0|Packet size|||60 bytes| ||CSThresh|1.559e–11|Antenna Parameter|X||0| ||RXThresh|3.652e–10||Y||0| ||Pt|2.5872e–4||Z||1.5| ----- _Sensors 2010, 10_ **4515** **Figure 9. Number of repository nodes and the number of queries received per repository node.** (A) Number of repository nodes (B) Number of queries received per repository node Figure 9 shows the number of repository nodes and the number of queries received per repository node given several query frequencies. GHT saves information into very few repository nodes—home nodes and neighboring nodes surrounding the target point—and GCLP uses many repository nodes forming a trajectory from end-to-end in the network. Neither GHT nor GCLP take account of the query frequency; therefore, the number of repository nodes is fixed in spite of the increased query frequency. This leads to an increase in the number of queries received per node. On the other hand, ASST varies the size of the trajectory in proportion to the query frequency, so that the number of repository nodes in ASST is increased in proportion to the increased query frequency. Therefore, the number of received queries per node in ASST is smaller than that in GHT and GCLP. GHT can be regarded as minimum replication, and GCLP can be regarded as maximum replication. It should be noted that the number of repository nodes in GCLP and ASST is similar when the query frequency is high, but the number of received queries per node in GCLP is higher than that in ASST due to the two ----- _Sensors 2010, 10_ **4516** matching points in GCLP. In GCLP, both the publish trajectory and the subscribe trajectory form a cross, making two matching points. Therefore, the number of received queries is higher than that in ASST. The number of queries per node can affect the response time. Figure 10 shows the response time with an increased query frequency. When the query frequency is low, GCLP shows a shorter response time than ASST or GHT because GCLP has more repository nodes than GHT or ASST. Moreover, GCLP forms a trajectory crossing the network so that GCLP occupies a larger area than GHT or ASST, and this leads to a reduction in the hop counts between the repository and query nodes. However, as the query frequency increases, the response time also increases. When the query frequency is very high, the response time is longer than that in ASST. This is caused by the larger number of received queries. However, GHT and ASST have the same number of repository nodes and the same hop counts between the repository and query nodes when the query frequency is low, so that the response times of both is the same. However, when the query frequency is increased, the response time of GHT is increased drastically while the response time of ASST is almost uniform. This is due to the difference between GHT and ASST in terms of the number of received queries per node. GHT always has a fixed number of repository nodes so that the number of received queries per node increases, and this leads to an increased response time. On the other hand, ASST increases the number of repository nodes in proportion to the query frequency so that the number of received queries per node and response time are almost uniform. **Figure 10. Response time with increased query frequency.** 120 100 **Figure 10.** 700 500 400 GCLP GHT 80 60 600 500 300 200 40 20 100 0 |GCLP ASST GHT query frequency|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16|Col17|Col18|Col19|Col20|Col21|Col22|Col23|Col24|Col25|Col26|Col27|Col28|Col29|Col30|Col31|Col32|Col33|Col34|Col35|Col36|Col37|Col38|Col39|Col40| |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| ||||||||||||||||||||||||||||||||||||||||| 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0 11.0 12.0 13.0 14.0 **query frequency per second** 0 1.0 4.0 5.0 6.0 Replicating data in several sensor nodes requires energy costs for packet transmission. More nodes participate in replication, and more energy is consumed because energy consumption is increased whenever a node transmits a packet. Figure 11 shows the energy consumption for repository construction. As mentioned earlier, the number of repository nodes in GHT and GCLP is fixed in spite 2.0 3.0 ----- _Sensors 2010, 10_ **4517** of query frequency change. GHT maintains a minimum number of repository nodes around the target point so that the energy consumption is lower than that with GCLP or ASST. In contrast, GCLP maintains repository nodes between the network boundaries in a cross shape so that energy consumption is higher than that under GHT or ASST. On the other hand, ASST varies the number of repository nodes in proportion to the query frequency so that the energy consumption is low when the query frequency is low, and the energy consumption is increased in proportion to the query frequency. The energy consumption is similar with GHT when the query frequency is low, and the energy consumption is increased towards the GCLP as the query frequency increases. **Figure 11. Energy consumption for storage construction.** **5. Conclusions** In this paper, we have proposed a new energy efficient and scalable data dissemination method, i.e., ASST, based on DHT and TBF. ASST is a type of data-centric storage system for sensor networks. ASST provides fast access to the data stored in a sensor network by using the distributed hash function. In ASST, the source node stores data at the nodes forming a trajectory around the target position computed by the hash function. A client sends a query packet to the position computed by the hash function that is the same as the source node. By storing data at several repository nodes, ASST distributes the network load caused by query packets. ASST also provides scalability for a large sensor network. When the query frequency is increased, the range of the trajectory is also increased to reduce the network load at the repository node. By adjusting the range of the trajectory in proportion to the query frequency, ASST can ensure a fast response time for multiple clients and reduce the network load assigned to a sensor node. Such capabilities result in an increased life span for the sensor network. The main focus of ASST is providing load scalability with constant response time by adjusting the width of trajectory in proportion to the query frequency under uniform distribution of query, however the query distribution might not be uniform over network and thus increasing or decreasing four edges with same distance H can cause load unbalance under non-uniform distribution of query. In the future ----- _Sensors 2010, 10_ **4518** work, we will consider the adjustment of distance H of four edges in trajectory zone in separated manner under with under non-uniform distribution of query. **References** 1. Akyildiz, I.F.; Su, W.; Sankarasubramaniam, Y.; Cayirci, E. A Survey on Sensor Networks. IEEE _Commun. Mag._ **2002, 102-114.** 2. Henn, H.; Hepper, S.; Rindtorff, K.; Schack, T.; Burkhardt, J. Pervasive Computing: Technology _and Architecture of Mobile Internet Applications, 1st ed.; Addison-Wesley Professional: Boston,_ MA, USA, 2002. 3. Niculescu, D. Communication Paradigms for Sensor Networks. _IEEE Commun. Mag. 2005, 43,_ 116-122. 4. Seada, K.; Helmy, A. Rendezvous Regions: A Scalable Architecture for Service Location and Data-Centric Storage in Large-Scale Wireless Networks. In _proceedings of_ _18th International_ _Parallel and Distributed Processing Symposium (IPDPS'04), Santa Fe, New Mexico, April 26–_ 30, 2004. 5. Tchakarov, J.B.; Vaidya, N.H. Efficient Content Location in Wireless Ad Hoc Networks. _IEEE_ _International Conference on Mobile Data Management, Berkeley, CA, USA, January 19–22,_ 2004; p. 74. 6. Ratnasamy, S.; Karp, B.; Yin, L.; Yu, F; Estrin, D.; Govindan, R.; Shenker, S. GHT: A Geographic Hash-table for Data centric Storage in Sensornets. In Proceedings of the First ACM International _Workshop on Wireless Sensor Networks and Applications (WSNA), Atlanta, GA, USA, September_ 28, 2002; pp.78-87. 7. Karp, B.; Kung, H.T. _GPSR: Greedy Perimeter Stateless Routing for Wireless Networks;_ A870044; MobiCom: Houston, TX, USA, 2000. 8. Capone, A.; Pizziniaco, L.; Filippini, I. A SiFT: An Efficient Method for Trajectory Based Forwarding. In _Proceedings of_ _Wireless Communication Systems 2nd International Symposium,_ Siena, Italy, September 5–7, 2005; pp. 135-139. 9. Tscha, Y.; Caglayan, M.U. Query Slipping Prevention for Trajectory-based Publishing and Subscribing in Wireless Sensor Networks. Comp. Commun. **2006, 29, 1979-1991.** 10. Machado, M.; Goussevskaia, O.; Mini, R.A.F.; Rezende, C.G.; Loureiro, A.A.F.; Mateus, G.R.; Nogueira, J.M.S. Data Dissemination in Autonomic Wireless Sensor Networks. IEEE J. Sel. Area. _Commun._ **2005, 23, 2305-2319.** 11. Akkaya, K.; Younis, M. A survey on routing protocols for wireless sensor networks. Ad Hoc Netw. **2005, 3, 325-349.** 12. Liu, X.; Huang, Q.; Zhang, Y. _Combs, Needles, Haystacks: Balancing Push and Pull for_ _Discovery in Large-scale Sensor Networks; ACM Press: New York, NY, USA, 2004; pp. 122-133._ 13. Intanagonwiwat, C.; Govindan R.; Estrin, D. Directed Diffusion: A Scalable and Robust Communication Paradigm for Sensor Networks, submitted to ACM MobiCom 2000. 14. GridwiseTech. Principles of Scalable Systems—Describing Technologies/layers that Are Components of Scalable System. Available online: http://www.gridwisetech.com/scalability_ expert (accessed November 2007). ----- _Sensors 2010, 10_ **4519** 15. Aydin, I.; Shen, C.C. Facilitating Match-making Service in Ad Hoc and Sensor Networks Using Pseudo Quorum. In _Proceedings of the 11th IEEE International Conference on Computer_ _Communications and Networks (ICCCN), Miami, FL, USA, October 14–16, 2002._ 16. Sailhan, F.; Issarny, V. Scalable service discovery for MANET. In _Proceedings of the 3rd IEEE_ _International Conference on Pervasive Computing and Communications, Kauai Island, HI, USA,_ March 8–12, 2005. 17. Mitra, A.; Banerjee, A.; Najjar, W.; Zeinalipour-Yazti, D.; Kalogeraki, V.; Gunopulos, D. HighPerformance Low Power Sensor Platforms Featuring Gigabyte Scale Storage. In Proceedings of _IEEE/ACM 3rd International Workshop on Measurement, Modelling, and Performance Analysis_ _of Wireless Sensor Networks, San Diego, CA, USA, July 21, 2005._ 18. Jurczyk, P.; Xiong, L.; Sunderam, V. DObjects: Enabling Distributed Data Services for Metacomputing Platforms. ICCS. Krakow, Poland, June 23-25. 2008; pp.136-145. 19. Yuksel, M.; Pradhan, R.; Kalyanaraman, S. An Implementation Framework for Trajectory-based Forwarding in Ad-Hoc Networks. In Proceedings of IEEE ICC, June 20–24, 2004; pp. 4062-4066. 20. Park, S.; Lee, D.; Lee, E.; Yu, F.; Choi, Y.; Kim, S-H. A Communication Architecture to Reflect User Mobility Issue in Wireless Sensor Fields. In Proceedings of IEEE Communications Society _Subject Matter Experts for publication in the WCNC, Kowloon, HongKong, March 11-15, 2007;_ pp.3376-3381. 21. Krishnamachari, B.; Ahn, J. Optimizing Data Replication for Expanding Ring-based Queries in Wireless Sensor Networks. USC Computer Engineering Technical Report; CENG-05-14, 2005. 22. Shenker, S.; Ratnasamy, S.; Karp, B; Govindan, R.; Estrin, D. Data-Centric Storage in Sensornets. _ACM SIGCOMM, Comp. Commun. Rev._ **2003, 33, 137-142.** 23. Ball, R. Content Sharing for Mobile Devices. CoRR **2008, abs/0809.4395.** 24. Gjermundrød, H.; Bakken, D.E.; Hauser, C.H.; Bose, A. GridStat: A Flexible QoS-Managed Data Dissemination Framework for the Power Grid. IEEE Trans. Power Delivery **2009, 24, 136-143.** 25. Liang, S.H.L. A New Fully Decentralized Scalable Peer-to-Peer GIS Architecture. In Proceedings _of ISPRS Congress, Beijing, China, July 3–11, 2008; p. 687._ 26. Bondi, A.B. Characteristics of scalability and their impact on performance. In Proceedings of the _2nd International Workshop on Software and Performance, Ottawa, Canada, September 17–20,_ 2000; pp. 195-203. 27. Al-Mamou, A.; Labiod Houda. ScatterPastry: An Overlay Routing Using a DHT over Wireless Sensor Networks. In Proceedings of _The 2007 International Conference on Intelligent Pervasive_ _Computing (IPC 2007), Jeju Island, Korea, October 11–13, 2007; pp.274-279._ 28. Landsiedel O.; Lehmann K.; Wehrle K. T-DHT: Topology-based Distributed Hash Tables. _In_ _Proceedings of the Fifth IEEE International Conference on Peer-to-Peer Computing (P2P'05),_ Konstanz, Germany, August 31–September 2, 2005; pp. 143-144. 29. Landsiedel, O.; Gotz, S.; Wehrle, K. Towards Scalable Mobility in Distributed Hash Tables. _In_ _Proceedings of the Sixth IEEE International Conference on Peer-to-Peer Computing, Cambridge,_ UK, October 2–4, 2006; pp. 203-209. 30. Pietzuch, P.; Shneidman, J.; Ledlie, J.; Welsh, M.; Seltzer, M. Roussopoulos, M. Evaluating DHTBased Service Placement for Stream-Based Overlays. In Proceedings of the 4th International _Workshop on Peer-to-Peer Systems (IPTPS'05), Ithaca, NY, USA, February 24–25, 2005._ ----- _Sensors 2010, 10_ **4520** 31. Jung, J.; Lee, S.; Kim, N.; Yoon, H. Efficient Service Discovery Mechanism for Wireless Sensor Networks. Comput. Commun. **2008, 31, 3292-3298.** 32. Su, I.; Chung, Y.; Lee, C. Finding Similar Answers in Data-Centric Sensor Networks. _IEEE_ _International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (Sutc_ _2008), Taichung, Taiwan, June 11–13, 2008. pp. 217-224._ 33. Le, T.; Yu, W.; Bai, X.; Xuan, D. A Dynamic Geographic Hash Table for Data Centric Storage in Sensor Networks. In Proceeding of IEEE Wireless Communications and Networking Conference _(WCNC 2006), Las Vegas, NV, USA, April 3–6, 2006._ 34. Liao W.; Wu W.; A cover-up scheme for data-centric storage in wireless sensor networks, _In_ _Proceeding of IEEE Symposium on Computers and Communications (ISCC 2007), Aveiro,_ Portugal, July 1-4. 2007. 35. Liao W.; Wu W. Effective Hotspot Storage Management Schemes in Wireless Sensor Networks. _Comput. Commun._ **2008, 31, 2131-2141.** 36. Rachuri, K.; Murthy, S. Energy Efficient and Scalable Search in Dense Wireless Sensor Networks. _IEEE Trans. Comput._ **2009, 58, 812-826.** 37. Satoshi, M.; Kazutoshi, F.; Hideki, S. Implementation Methodology of Geographical Overlay Network Suitable for Ubiquitous Sensing Environment. J. Inform. Process. **2008, 16, 80-92.** 38. Li, X.; Santoro, N.; Stojmenovic, I. Localized Distance-Sensitive Service Discovery in Wireless Sensor and Actor Networks. IEEE Trans. Comput. **2009, 58, 1275-1288.** 39. Tsai, H.; Chen, T.; Chu, C. Service Discovery in Mobile Ad Hoc Networks Based on Grid. IEEE _Trans. Veh. Technol. 2009, 58, 1528-1545._ 40. Mann, C.; Baldwin, R.; Kharoufeh, J.; Mullins, B. A Trajectory-Based Selective Broadcast Query Protocol for Large-Scale, High-Density Wireless Sensor Networks. _Telecommun. Syst.: Model._ _Anal. Des. Manag._ **2007, 35, 67-86.** © 2010 by the authors; licensee MDPI, Basel, Switzerland. This article is an open-access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/). -----
{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC3292128, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/1424-8220/10/5/4497/pdf?version=1403312442" }
2,010
[ "JournalArticle" ]
true
2010-04-30T00:00:00
[ { "paperId": "b8d9b10cf54629364523ec065e6307ab87f7d4f0", "title": "iRun: Horizontal and Vertical Shape of a Region-Based Graph Compression" }, { "paperId": "e44581f9aed1c309f9b7bb7054b7cba6a40e77ed", "title": "A survey on routing protocols in Wireless Sensor Networks" }, { "paperId": "a8220c1e619befd4ee02a9175c9e7f16281abfac", "title": "Localized Distance-Sensitive Service Discovery in Wireless Sensor and Actor Networks" }, { "paperId": "845206025b38ef09e9e0e5629ec29140cbfa723f", "title": "Energy Efficient and Scalable Search in Dense Wireless Sensor Networks" }, { "paperId": "5f7d282fd35b1391e35d2a722fb6cfe7e20169b1", "title": "Service Discovery in Mobile Ad Hoc Networks Based on Grid" }, { "paperId": "d7503f1126d6df93d1f418d89fe7f5b52bb0ecad", "title": "Content Sharing for Mobile Devices" }, { "paperId": "1cfa82527fee869bcaddc060a02cbb6a2fb9a541", "title": "Efficient service discovery mechanism for wireless sensor networks" }, { "paperId": "0355405fbadfba27d7f682bbae6777c5878c2269", "title": "DObjects: enabling distributed data services for metacomputing platforms" }, { "paperId": "dc694cd4ca181f01a737bb17d1ae78bd7e342eab", "title": "Finding Similar Answers in Data-Centric Sensor Networks" }, { "paperId": "4a63f0a098bdccb905e6202c9b70f4a0f493a0f9", "title": "Effective hotspot storage management schemes in wireless sensor networks" }, { "paperId": "8a82225d733ed2fb1c1ff0e06150f6e281654f3c", "title": "ScatterPastry: An Overlay Routing Using a DHT over Wireless Sensor Networks" }, { "paperId": "c289d2fa550ff33457655749b5cfc91b3cc7714a", "title": "A Cover-Up Scheme for Data-Centric Storage in Wireless Sensor Networks" }, { "paperId": "00e810580940e3568ab2e473b847d79db414a8bb", "title": "A trajectory-based selective broadcast query protocol for large-scale, high-density wireless sensor networks" }, { "paperId": "c53b063a0ff982012a7cc7d81153b6ca6856e445", "title": "A Communication Architecture to Reflect User Mobility Issue in Wireless Sensor Fields" }, { "paperId": "a9bf09db95ea70d4f29d965d9cd92f7e2c639c43", "title": "Towards Scalable Mobility in Distributed Hash Tables" }, { "paperId": "20587e2ec5e5c296c86c5cf21a2a4677273ebd6c", "title": "Query slipping prevention for trajectory-based publishing and subscribing in wireless sensor networks" }, { "paperId": "4626c35153566f4e6b1f00eb1513c9de74f5a60c", "title": "A dynamic geographic hash table for data-centric storage in sensor networks" }, { "paperId": "8b17c8876351d26a6ed55799d196dfcaf84799e7", "title": "Optimizing Data Replication for Expanding Ring-based Queries in Wireless Sensor Networks" }, { "paperId": "7a3d67136402eac90e59e5e6a545bf75eafec029", "title": "A SiFT: an efficient method for trajectory based forwarding" }, { "paperId": "f3ce8b8f56427ff3cb7b253d1938ac7170de0ebf", "title": "Data dissemination in autonomic wireless sensor networks" }, { "paperId": "ab3812744dcea0d6f2dcac994ac88254c9fe00c4", "title": "T-DHT: topology-based distributed hash tables" }, { "paperId": "7c2af5ebe723580e90348d3bf1eaf31fe7bef828", "title": "Service discovery in mobile ad hoc networks" }, { "paperId": "014d24556957a61ba4a3b6f01066a8cbdf6547c6", "title": "Scalable Service Discovery for MANET" }, { "paperId": "841a78677c613d063aa930245cf4d6277d396d04", "title": "Communication paradigms for sensor networks" }, { "paperId": "665578c35c379446c5d22e2e12ef57983517ce29", "title": "Evaluating DHT-Based Service Placement for Stream-Based Overlays" }, { "paperId": "b8e81f7def2f9bd1207e41e49c6b456c81ea0609", "title": "Combs, needles, haystacks: balancing push and pull for discovery in large-scale sensor networks" }, { "paperId": "ff96b433d5b5ef9dec64f78ab914c798e527f811", "title": "Efficient content location in wireless ad hoc networks" }, { "paperId": "c791c71e27541abfd6d61f60e50d7978511c02fb", "title": "An implementation framework for trajectory-based routing in ad-hoc networks" }, { "paperId": "eb0361300b9786312caa54a4661df5dec376a216", "title": "Rendezvous regions: a scalable architecture for service location and data-centric storage in large-scale wireless networks" }, { "paperId": "1799995b17df1d9266bed56d7c4902392d513863", "title": "Facilitating match-making service in ad hoc and sensor networks using pseudo quorum" }, { "paperId": "19d9d5e14d4d3f3e4ef66e00cdbc11ff6a93e9d5", "title": "GHT: a geographic hash table for data-centric storage" }, { "paperId": "c3ed671bf7c9bb71c9b791309c11f2b8f3ecf204", "title": "Sensor Networks" }, { "paperId": "6f2b8a31d0139bc93ec64da1a44d8be584d3dbd2", "title": "Pervasive Computing: Technology and Architecture of Mobile Internet Applications" }, { "paperId": "cfa353b0652b3d1799a91ff454c9d8e90c792c10", "title": "Characteristics of scalability and their impact on performance" }, { "paperId": "829cf8d5540342b6a7fd190befd0874e8e0e3b20", "title": "Directed diffusion: a scalable and robust communication paradigm for sensor networks" }, { "paperId": "ef9d0802dad4d6e8fb39a5fc14841651736d6503", "title": "GPSR: greedy perimeter stateless routing for wireless networks" }, { "paperId": "dbc6af4eb4088158990e737440c7724193c2b1ab", "title": "GridStat: A Flexible QoS-Managed Data Dissemination Framework for the Power Grid" }, { "paperId": "f65b4d6e874d22e6c6d4eab8bdd7dd3fdad05651", "title": "A NEW FULLY DECENTRALIZED SCALABLE PEER-TO-PEER GIS" }, { "paperId": "45af9fe862a92406249cd7d91f6e0e46a67b078d", "title": "Implementation Methodology of Geographical Overlay Network Suitable for Ubiquitous Sensing Environment" }, { "paperId": "f2687e43bfee9958963fbc4d19b799a4504e1edb", "title": "High Performance , Low Power Sensor Platforms Featuring Gigabyte Scale Storage" }, { "paperId": "3e8cbf0a19e997d93afd3b3aa89ea54ea728a98f", "title": "Data-centric storage in sensornets" }, { "paperId": "798f3053268dbfc1c6df40fee61ac862ef713d1f", "title": "An Implementation Framework for Trajectory-Based Forwarding in Ad-Hoc Networks" }, { "paperId": null, "title": "Principles of Scalable Systems — Describing Technologies / layers that Are Components of Scalable System" }, { "paperId": null, "title": "This article is an open-access article distributed under the terms and conditions of the Creative Commons Attribution license" }, { "paperId": null, "title": "Principles of Scalable Systems—Describing Technologies/layers that Are Components of Scalable System Available online" } ]
14,644
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0153cde1ef613a5261c93354a43fb137d3bbf2a4
[]
0.879905
A Comparative Study of Consensus Mechanisms in Blockchain for IoT Networks
0153cde1ef613a5261c93354a43fb137d3bbf2a4
Electronics
[ { "authorId": "2183437602", "name": "Zachary Auhl" }, { "authorId": "9219862", "name": "N. Chilamkurti" }, { "authorId": "51132514", "name": "Rabei Alhadad" }, { "authorId": "2145540588", "name": "Will Heyne" } ]
{ "alternate_issns": [ "2079-9292", "0883-4989" ], "alternate_names": null, "alternate_urls": [ "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-247562", "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-247562", "https://www.mdpi.com/journal/electronics" ], "id": "ccd8e532-73c6-414f-bc91-271bbb2933e2", "issn": "1450-5843", "name": "Electronics", "type": "journal", "url": "http://www.electronics.etfbl.net/" }
The consensus mechanism is a core component of Blockchain technology, allowing thousands of nodes to agree on a single and consistent view of the Blockchain. A carefully selected consensus mechanism can provide attributes such as fault tolerance and immutability to an application. The Internet of Things (IoT) is a use case that can take advantage of these unique Blockchain properties. IoT devices are commonly implemented in sensitive domains such as health, smart cities, and supply chains. Resilience and data integrity are important for these domains, as failures and malicious data tampering could be detrimental to the systems that rely on these IoT devices. Additionally, Blockchains are well suited for decentralised networks and networks with high churn rates. A difficulty involved with applying Blockchain technology to the IoT is the lack of computational resources. This means that traditional consensus mechanisms like Proof of Work (PoW) are unsuitable. In this paper, we will compare several popular consensus mechanisms using a set of criteria, with the aim of understanding which consensus mechanisms are suitable for deployment in the IoT, and what trade-offs are required. We show that there are opportunities for both PoW and PoS to be implemented in the IoT, with purpose-made IoT consensus mechanisms like PoSCS and Microchain. Our analysis shows that Microchain and PoSCS have characteristics that are well suited for IoT consensus.
# electronics _Article_ ## A Comparative Study of Consensus Mechanisms in Blockchain for IoT Networks **Zachary Auhl** **[1,]*, Naveen Chilamkurti** **[1]** **, Rabei Alhadad** **[1]** **and Will Heyne** **[2]** 1 Cybersecurity Innovation Node, La Trobe University, Melbourne, VIC 3086, Australia 2 BAE Systems, Adelaide, SA 5000, Australia ***** Correspondence: z.auhl@latrobe.edu.au **Citation: Auhl, Z.; Chilamkurti, N.;** Alhadad, R.; Heyne, W. A Comparative Study of Consensus Mechanisms in Blockchain for IoT Networks. Electronics 2022, 11, 2694. [https://doi.org/10.3390/](https://doi.org/10.3390/electronics11172694) [electronics11172694](https://doi.org/10.3390/electronics11172694) Academic Editor: Asma Khatoon Received: 18 July 2022 Accepted: 22 August 2022 Published: 27 August 2022 **Publisher’s Note: MDPI stays neutral** with regard to jurisdictional claims in published maps and institutional affil iations. **Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **Abstract: The consensus mechanism is a core component of Blockchain technology, allowing thou-** sands of nodes to agree on a single and consistent view of the Blockchain. A carefully selected consensus mechanism can provide attributes such as fault tolerance and immutability to an application. The Internet of Things (IoT) is a use case that can take advantage of these unique Blockchain properties. IoT devices are commonly implemented in sensitive domains such as health, smart cities, and supply chains. Resilience and data integrity are important for these domains, as failures and malicious data tampering could be detrimental to the systems that rely on these IoT devices. Additionally, Blockchains are well suited for decentralised networks and networks with high churn rates. A difficulty involved with applying Blockchain technology to the IoT is the lack of computational resources. This means that traditional consensus mechanisms like Proof of Work (PoW) are unsuitable. In this paper, we will compare several popular consensus mechanisms using a set of criteria, with the aim of understanding which consensus mechanisms are suitable for deployment in the IoT, and what trade-offs are required. We show that there are opportunities for both PoW and PoS to be implemented in the IoT, with purpose-made IoT consensus mechanisms like PoSCS and Microchain. Our analysis shows that Microchain and PoSCS have characteristics that are well suited for IoT consensus. **Keywords: consensus; IoT; Blockchain** **1. Introduction** Blockchains are cryptographically linked distributed ledgers that are known for storing the transaction history of the Bitcoin network. Bitcoin adopted the Blockchain as it has two important properties: tamper-evidence and the triple-entry ledger. As each block in the Blockchain is cryptographically linked to the previous block, attempts to tamper with the Blockchain will invalidate the blocks cryptographic link. This means malicious parties cannot arbitrarily alter the history of the Blockchain. Triple-entry accounting refers to the distributed characteristics of a Blockchain. Instead of relying on two parties to provide evidence of their activities, transactions on the Blockchain are transmitted to the whole network, allowing anyone to validate all the transactions on the Blockchain. Often paired with a Blockchain, is a consensus mechanism. Consensus mechanisms allow Blockchains to converge on network-wide agreement on the state of the Blockchain, meaning that nodes all agree on the same history of the ledger. Consensus on the Bitcoin’s Blockchain relies on two mechanisms, Proof of Work (PoW), and the Longest Chain Rule (LcR). PoW is a cryptographic puzzle that miners on the Bitcoin network attempt to solve. This provides miners with a financial incentive to support the network, and prevents Sybil attacks. The LcR is a fork resolution tool, that manages competing histories of the Blockchain, and converges the network back onto a single state in cases where the Blockchain forks. Recently, there has been interest in applying Blockchain technology to the Internet of Things (IoT). Specifically, consensus mechanisms have been modified to be less resource intensive, ----- _Electronics 2022, 11, 2694_ 2 of 23 and more suitable for deployment in the IoT, with consensus mechanisms such as the Credit-Based PoW (CBPoW) and Proof of Supply Chain Share (PoSCS). This paper will examine the suitability of permissionless Blockchains for the IoT and the trade-offs required, especially for resource limited IoT devices. _1.1. Outline_ We begin the paper by discussing the criteria. We use these criteria to compare the consensus mechanisms discussed in the paper. Next, we briefly discuss the fundamental properties of Blockchains in the background section. The second half of the paper is focused on analysing consensus mechanisms. We start by covering consensus mechanisms commonly found in blockchains, such as PoW and PoS; then, we discuss four newer consensus mechanisms from the literature, specifically designed for the IoT. Finally, we compare the discussed consensus mechanisms using our proposed criteria. We pay close attention to properties that positively and negatively impact the critical characteristics of IoT devices. Finally, we conclude with consensus recommendations for the IoT and outline future work. _1.2. Contributions_ In this paper, we provide the following contributions: - Analysis of PoW and PoS consensus mechanisms and their usability in the IoT. - Analysis of four novel consensus mechanisms from the literature, specifically designed for the IoT. - A comparison between the mentioned consensus mechanisms, with clear criteria to show their suitability for the IoT. **2. Criteria for Blockchain Consensus** The IoT environment is incredibly diverse, involving a wide range of hardware solutions and software solutions paired with a stringent set of requirements, involving power consumption, storage, and computational capabilities. IoT devices work in dynamic environments, where sensors generate data constantly, coming online and offline depending on their power requirements, and expected to work in ad hoc networks. Due to this flexibility, IoT devices have seen widespread usage in applications such as smart cities [1] supply chains [2–5] and healthcare [6]. The consensus mechanism is a critical part of most Blockchain deployments, but the choice becomes even more important when working with Blockchain deployments targeted towards the IoT. All Blockchains come with trade-offs; there is no such thing as a ’perfect’ Blockchain. Some are more resourceintensive, some are faster, and others are more centralised. To compare the Blockchains discussed throughout this paper, we will define a set of requirements, to better understand their usability and impact in IoT environments. 1. Processor Usage: How are IoT devices going to agree on the content and order of the Blockchain? Historically, Blockchains made use of PoW to decide this. However, PoW is well-known for being computationally expensive, and environmentally destructive, and has seen waning interest in newer consensus mechanisms. Extending the battery life of IoT devices, and maintaining an acceptable processor utilisation, will be an important factor when selecting a consensus mechanism. 2. Security: Blockchain implementations may provide stronger security guarantees, when compared to traditional IoT networks with central points of failure (coordinators, controllers, cloud networks etc.). Many Blockchains suffer from potential attacks when the number of malicious nodes reaches 51% and 33% depending on the consensus mechanism. While there is no longer a single point of failure for a malicious actor to target, Blockchain specific attacks can still compromise the IoT network. 3. Decentralisation: A choice for most Blockchains which is not binary and operates on a sliding scale. Increasing the network’s decentralisation further diversifies Blockchain storage and decision-making, but usually impacts the speed and scalability of the net ----- _Electronics 2022, 11, 2694_ 3 of 23 work. Decreasing decentralisation has an inverse effect, which reduces the diversity in the consensus process, and prioritises scalability and speed. 4. Storage: A factor that needs to be considered for security and the decentralisation aspects of a Blockchain. If all nodes on the network store a full copy of the Blockchain, they can independently verify transactions, and help new nodes bootstrap their Blockchains. IoT devices generally do not have the capacity to store hundreds of gigabytes worth of Blockchain data, so a compromise that still allows for security, and potentially decentralisation needs to be made. 5. Transactions Per Second (TPS): Another trade-off occurs between decentralisation and speed. The more nodes that participate in consensus, the higher the latency to make a decision, which results in generally lower speed, but higher decentralisation. A lower number of nodes participating in consensus could lead to increased transaction throughput and lower block times, which is generally desirable for IoT devices. **3. Background** Consensus mechanisms were traditionally deployed to maintain critical control systems, such as those aboard commercial airlines. Aboard an airline, the consensus mechanism is used to coordinate multiple control systems, keeps the system operational even in partial failure, and keeps hundreds of passengers on an aircraft safe [7]. Consensus mechanisms, such as Paxos, have also been widely adopted by Google and Amazon in their distributed systems. Google’s projects such as Spanner [8], which provide a distributed, replicated database system, and Mesa [9], which provides distributed data warehousing. A consensus mechanism defines a set of rules or protocols a group of systems needs to abide by, in order to make a decision. Let us use Bitcoin as an example. Part of Bitcoin’s consensus mechanism lets users running full nodes agree on Bitcoin’s Blockchain history. Each nonfaulty participant running a full node, and enforcing Bitcoin’s consensus rules, will check transactions for issues like: a user spending Bitcoin they do not own, a user trying to print Bitcoin out of thin air, or a miner creating a block and rewarding themselves with thousands of Bitcoin [10]. Rather than a central entity enforcing these rules, every full node on the Bitcoin network is enforcing these rules. One of the most famous examples of distributed agreement was published by Lamport et al. titled the “Byzantine Generals Problem” [11]. This paper includes a thought experiment involving a set of generals planning an attack on an enemy city. If a treacherous commander or lieutenant attempts to deceive their peers by sending incorrect orders, this could lead to disaster for the army. If parts of the army attack, while other parts of the army retreat, the battle cannot be won. The paper proves that in order to deal with b Byzantine nodes, there must be at least 3b+1 nodes on the network [11]. Byzantine actors are capable of colluding to deceiving other users in our system, which breaks consensus by simple majority. To accommodate for Byzantine actors and to maintain safety and liveness guarantees, b < 1/3 of the total nodes on the network, so that consensus cannot be split by a 50/50 vote on our network by colluding nodes [12]. _Types of Blockchains_ Blockchains can be generalised into two categories: permissionless, and permissioned. Permissionless, like Bitcoin and Ethereum, are public Blockchains that anyone can participate in. Users are free to create transactions on these networks, interact with smart contracts, and are free to propose blocks if they participate in mining. Users on permissionless Blockchains are generally pseudonymous, which means users wallets are associated with certain public addresses, but not necessarily linked to a person’s name. Permissioned Blockchains are more restrictive on how users can interact with the Blockchain, and generally have different use cases. Examples of permissioned Blockchains include R3 [13] and J.P. Morgan’s Quorum [14], which are both targeted at the financial ----- _Electronics 2022, 11, 2694_ 4 of 23 industry. These Blockchains may allow the public to join, but are generally invite only, and have a stricter set of rules. Private Blockchains can be thought of as an extension to permissioned Blockchains. The difference is permissioned Blockchains may be publicly accessible, as long as a user meets some sort of criteria. Private Blockchains are generally not accessible to the public, and are almost always invite only [15]. In Section 4, we will discuss PoW Blockchains, and their usability in the IoT. **4. Proof of Work** PoW is a consensus mechanism used by Bitcoin, and has been forked, modified and copied by many other cryptocurrencies [16]. At its core, most PoW implementations involve solving a cryptographic puzzle with certain parameters, and the first machine to solve this problem is rewarded. Figure 1 shows a diagram of this process. More specifically, miners are searching for a nonce (a random number), that can be hashed together with the block header, to produce a block hash, with a specific number of starting zeros [17]. The first miner on the network to produce a hash with these specific requirements, is given the block reward as payment for their service. Bitcoin’s consensus mechanism involves the interaction between two important components, PoW, and the Longest Chain Rule (LcR), otherwise known as Nakamoto Consensus [10]. PoW provides two important features: a mechanism that provides a financial incentive for mining, and a way of preventing Sybil attacks. Bitcoin’s second component, LcR, exists due to a trade-off Bitcoin made to maintain consensus guarantees. Bitcoin had to compromise either safety or liveness to guarantee consensus under Byzantine actors, and in an asynchronous network environment, with Bitcoin choosing safety [18]. Bitcoin is unable to provide strong safety, which means Bitcoin cannot guarantee that every node on the network will have identical copies of the Blockchain [19]. With the ability for forks to occur, Bitcoin requires the LcR as a fork resolution tool. In the case the network is split, the LcR states that the fork with the most aggregated computational work, is the correct Blockchain [20]. **Figure 1. The process of mining on PoW Blockchains.** Proof of work is still widely used by several cryprocurrencies such as Ethereum, Litecoin, and Monero with slight differences. Ethereum uses a modified version of the SHA3 hashing algorithm called Keccak-256 [21], Litecoin uses the Scrypt hash function [22], and Monero uses an evolution of the CryptoNote hash function called CryptoNight [23]. With these two rules we have a system that rewards miners, stops Sybil attacks, and can reconcile forks in a trustless and decentralised manner. With hundreds of billions of dollars at stake, Bitcoin is yet to be hacked catastrophically, and, in the words of Andreas ----- _Electronics 2022, 11, 2694_ 5 of 23 [Antonopoulos, Bitcoin has become the “sewer rat” of Blockchains (https://aantonop.com/](https://aantonop.com/bubble-boy-and-the-sewer-rat/) [bubble-boy-and-the-sewer-rat/ (accessed on 20 August 2022)).](https://aantonop.com/bubble-boy-and-the-sewer-rat/) _4.1. Credit-Based PoW (CBPoW)_ Huang et al. proposes a credit-based PoW system that’s suitable to run on IoT devices [24]. The authors created a consensus mechanism that dynamically adjusts a device’s PoW difficulty depending on their adherence to the consensus rules. A node’s total score can be calculated by taking the sum of their positive score along with their negative score as shown in Figure 2. A positive score can be increased by following the consensus mechanism, while their negative score grows by disobeying consensus. The paper focuses on two specific attacks that could lower a client’s score: lazy tips and double spending. Lazy tips is an issue that specifically effects Directed Acyclic Graphs (DAGs), where a malicious actor will avoid confirming recent transactions by building on top of old, preexisting transactions. This can be detrimental to the network, as honest nodes may not have their new transactions approved. Huang et al. also penalises users for attempting to spend their tokens twice. If nodes are found to be acting maliciously, a penalty function that takes the sum of their malicious transactions, over a certain period of time, multiplied by a punishment coefficient, is used to penalise their credit amount. As nodes are required to confirm two previous transactions before submitting their own, a low PoW credit will make adding a transaction time intensive, and computationally expensive. CBPoW uses a tiered node network, where lite nodes are responsible for collecting data and broadcasting transactions, and full nodes are responsible for maintaining the tangle. **Figure 2. CBPoW Consensus.** _4.2. Proof of Elapsed Work and Luck (PoEWAL)_ PoEWAL is a consensus mechanism with similar traits to Bitcoin, but has been modified to be mineable on resource constrained devices [25]. PoEWAL still requires devices to solve a cryptographic puzzle, however, rather than devices searching for the matching nonce, miners just need to mine for a short period of time. This heavily reduces the power and computational load on IoT devices. Once the mining time in a round elapses, miners will compare their hash values used to solve the computational puzzle. The node, whose ----- _Electronics 2022, 11, 2694_ 6 of 23 hash value has the highest number of consecutive zeros, has the right to produce a block for the round. In the case that two miners propose hashes with the same number of consecutive zeros, Huang et al. proposes a fork resolution tool called Proof of Luck. Proof of Luck compares the two hashes values with equal consecutive zeros, then selects the node whose hash value was the lowest to propose a block as presented in Figure 3. PoEWAL is able to enforce these highly synced time limits on their consensus mechanism as the authors assume that their IoT devices will have synced clocks. PoEWAL also implements a dynamic difficulty level depending on the number of collisions. If collisions happen at a regular frequency, the difficulty level will be raised in an attempt to lower the number of collisions. **Figure 3. PoEWAL Consensus.** **5. Proof of Stake** Proof of Stake (PoS) was presented in a paper written by Sunny King and Scott Nadal in 2012 [26]. King et al. proposed that the age of a cryptocurrencies coin, known as the coinage, could be used to develop an alternative consensus mechanism to PoW. The authors propose a system where PoW mints the initial supply of coins on the network, and then slowly diminishes the mining rewards to lower the reliance on PoW. Sunny King went on to create Peercoin (PPC) a fork of Bitcoin in 2013. Peercoin implemented an initial PoW coin distribution. The proposed consensus mechanism also used coin-age to stop wealthy users from hoarding staking rewards, and checkpoints to deny changes to the Blockchain after a certain point [24]. Rather than finding a nonce, a node is selected to mine the next block using a pseudorandom lottery. The larger the node’s stake of coins in proportion to the rest of the network, the higher the chance of being selected to mine a block [16]. Similarly, to PoW, the header is hashed, but rather than spending large amounts of electricity, constantly hashing different nonces, PoS does one calculation. If the coin age > blockhash/target, the node can create a new valid block. Figure 4 shows a similar mechanism, but displays the more sophisticated PoS kernel as a replacement to the coin age. If not, the node waits til the next round to check if it meets the criteria to produce a block [27]. PoS proved popular due to its minimal hardware requirements and reduced energy usage compared to PoW. Cryptocurrencies such as Algorand, Cardano, and PIVX are examples of cryptocurrencies that have adopted PoS as their consensus mechanism. Algorand [28] co-found by cryptographer Silvio Micali, Cardano, one of the largest smart contract platforms by market cap [29], and PIVX, that allows users to run ’masternodes’, nodes which provide extra security and functionality for the network [6]. ----- _Electronics 2022, 11, 2694_ 7 of 23 **Figure 4. PoS Consensus.** _5.1. Byzantine Agreement Protocol (BAP)_ Algorand is a cryptocurrency co-found by Silvio Micali, and uses a Verifiable Random Function (VRF) to power its consensus mechanism, the Byzantine Agreement Protocol [28]. Nodes on the network can choose to participate in consensus by computing an evaluation function. The Decentralised Random Beacon (DRB) allows nodes to agree on a VRF and to collaboratively create one new output of the VRF every round. A VRF in this context means a commitment to a deterministic, pseudorandom value. In particular, the VRF outputs are unbiased due to their pseudorandom qualities [30]. On the Algorand network, a user has their own unique secret key, and a ’magic seed’ known to nodes on the Algorand network. The evaluation function returns an output string (which is used to select committee members) and a proof to verify the output. Next the output string is checked to see if it falls between a range [0, user stake] where user stake is the proportion of the coins a user has staked compared to the total coins staked. If the output string falls in this range, the user is selected to join the committee for the current round [28]. The VRF also acts as a lottery to select leaders to propose blocks to the committee. If most of the committee is honest, and a node proposes a valid block, a block can be certified and added to the Blockchain. This process is shown in Figure 5. ----- _Electronics 2022, 11, 2694_ 8 of 23 **Figure 5. BAP Consensus.** _5.2. Dfinity_ Dfinity Launched on the 18 December 2020 and is positioning itself as the “Internet Computer”. Dfinity is creating a Blockchain that can host various online services, such as those provided by Amazon AWS and Google Cloud. The Dfinity consensus mechanism is split into 4 segments [31], summarised here, and depicted in detail in Figure 6. 1. Identities and Registry: Used to register clients to the network, each client has a permanent pseudonymous identity. This is a form of Sybil protection to defend against malicious users flooding the network with fake identities. 2. Random Beacon: built on top of a Verifiable Random Function (VRF) that allows registered clients on the network to generate and agree upon random numbers. Dfinity uses an optimised implementation of the Boneh–Lynn–Shacham (BLS) signature scheme, which they have used to solve the last actor problem. This problem involves ----- _Electronics 2022, 11, 2694_ 9 of 23 the last actor in the protocol knowing the random value for the next round, and having the ability to abort the protocol 3. Blockchain and Fork Resolution: This segment implements the Probabilistic Slot Protocol (PSP) that is used to rank the clients for a particular round according to the output from the Random Beacon. This rank is used to assign a weight to block proposers, with a higher rank resulting in a better chance of being selected to create a block. PSP offers instantaneous ranking and a deterministic block time. 4. Notarisation and near-instant finality: Block notarisation is Dfinity’s technique to provide near-instant finality, that is, network-wide and irreversible agreement on a new block. Dfinity takes advantage of the BLS threshold signatures [32], Random Beacon, and the client ranking system to achieve this. **Figure 6. Dinifity Consensus.** _5.3. Ethereum PoS Consensus_ Ethereum is a Blockchain that was originally conceived by Vitalik Buterin in 2013 [33], and extended upon in 2014 by Gavin Wood to define Ethereum’s smart contract functionality [34]. Ethereum launched in 2015, with the code name ‘Frontier’. At the time of writing, Ethereum is the second largest cryptocurrency in terms of market cap. Ethereum is a Blockchain known for decentralised applications, commonly called ‘DApps’. The Ethereum DApp ecosystem is diverse, with a range of widely used DApps in areas such as finance [35,36], gaming [37] and prediction markets [38]. Currently, Ethereum is using PoW as its consensus mechanism, but plans to move to PoS in an upgrade commonly ----- _Electronics 2022, 11, 2694_ 10 of 23 known as ‘Eth2’. The Eth2 upgrade is underway, and being rolled out progressively. The roll out is planned to occur in 3 stages, with each stage implementing several changes. Stage 1, launches the Beacon chain, which will introduce PoS, stage 2 will focus on merging the Ethereum PoW chain, and the Beacon chain, and step 3 will focus on scalability with the implementation of sharding [39]. Consensus on Eth2 is going to involve two components, the Greedy Heaviest Observed Subtree (GHOST), which will act as a fork resolution tool, and Casper the Friendly Finality Gadget (CFFG), which will finalise the decisions that GHOST makes [40]. The GHOST protocol compromises on its safety, which means that it is possible to switch between different forks, with different chain heights. However, as GHOST has liveness guarantees, blocks can continue to be added to the Ethereum Blockchain even when the chain is under attack. The CFFG protocol finalises the blocks that are added to the chain, and as CFFG favours safety over liveness, the protocol’s decisions are final. CFFG has similar properties to the Practical Byzantine Fault Tolerance (PBFT) consensus mechanism, in that both protocols use justification rounds and finalisation rounds to come to consensus [41]. CFFG also employs a method to batch justification and finalisation messages, which increases Ethereum’s potential scalability. While the Ethereum network is functioning normally, GHOST will provide a fork resolution process, then CFFG will finalise the decision and add the block to Ethereum’s Blockchain. However, in the event the network is under attack, or there is an issue that causes many nodes to go offline, GHOST will continue to function, and blocks will still be added to the Blockchain, but will not be finalised. Once the attack subsides, CFFG will start working again, and will finalise the blocks that GHOST has proposed, and add them to the Blockchain if they are valid. CFFG and GHOST cover each other’s weaknesses, and allow a consensus mechanism with both safety, and liveness guarantees [42]. A partial snapshot of Ethereum’s PoS implementaion is shown in Figure 7. **Figure 7. Ethereum PoS Consensus.** ----- _Electronics 2022, 11, 2694_ 11 of 23 _5.4. Microchain_ Microchain proposes a lightweight consensus mechanism aimed at the IoT [43]. Microchain’s consensus mechanism has similar properties to PoS, where a number of validators are selected to join a committee, and from the committee a node is selected to produce a block. The purpose of the committee is to select a pseudorandom subset of the network, to avoid biased or malicious block producers. Microchain also uses a committee called a ‘Dynasty’, where eligible validators are selected to join the committee. Microchain’s consensus mechanism is broken down into two major components: Proof of Credit (PoC), and Voting based Chain Finality (VCF). PoC is a PoS mechanism that uses a credit weight to increase the chance a particular node has of producing a block as depicted in Figure 8. Given the distribution of credits in a particular Dynasty, nodes that have a higher credit weight have a larger chance of being selected to produce a block. The VCF is a fork resolution tool, and is also responsible for extending the chain by adding new blocks, and protecting the Blockchain from malicious or accidental reorganising by adding checkpoints. The consensus mechanism proposed by Xu et al. leverages a VRF to power its slot selection (that is, the process of picking nodes to join a Dynasty). Microchain makes an assumption that networks are synchronous, and is able to provide two guarantees: persistence and liveness. Persistence guarantees that all the users agree on the same history of the Blockchain, and if one honest node finds a transaction to be finalised, all honest nodes see the transaction as final. Liveness guarantees that a valid transaction submitted by an honest node will eventually be added to a new block. **Figure 8. Microchain Consensus.** _5.5. Proof of Supply Chain Share (PoSCS)_ PoSCS is a consensus mechanism proposed by Tsang et al. targeted towards the Perishable Food Supply Chain (PFSC) [44]. The project uses a framework that incorporates an IoT network to manage monitoring and communication, a Blockchain to manage the data of food through the life cycle of the supply chain, and a database to archive supply chain information. The authors quickly point out that PoW is not suitable for the IoT due to the computationally expensive mining process. The authors propose a consensus mechanism like PoS, but replace the need for a currency, with a reputation system. Each node ----- _Electronics 2022, 11, 2694_ 12 of 23 participating in consensus has four components that determine its reputation: Influence Factor (INF), Interest Factor (INT), Devotion Factor (DEV), and Satisfaction Factor (SAT). These factors can then be weighted using three strategies: the interest-first strategy, moderate-strategy, and devotion-first strategy. These weights prevent the consensus mechanism from favouring participants who attempt to maximise a single factor. Lastly, the shipment volume, considers the ingoing and outgoing volume a particular party is moving on the supply chain network. This process is summarised in Figure 9. These factors and weights are used to pseudorandomly select a block producer, who will be required to forge a block. Block forgers are also required to do a small amount of PoW mining, which allows the block creation time to be controlled. Rather than all the nodes participating in PoSCSs consensus mechanism, only the block forger is required to mine. The architecture of the system uses a hybrid approach, combining Blockchain and the cloud. The Blockchain is used to record the data about a particular object in the supply chain in tandem with a traditional database. Once the object has completed its journey through the supply chain, the object is removed from the storage of the IoT devices, and remains archived in the cloud. **Figure 9. PoSCS Consensus.** _5.6. Tendermint_ A flexible consensus mechanism in the Byzantine Fault Tolerance (BFT) family that can be configured to work in either public, private, or permissioned networks [16]. Tendermint can be configured as a public consensus mechanism with PoS, or as a permissioned/private Blockchain with predetermined validator nodes. Tendermint’s consensus mechanism uses a voting mechanism that has three steps: proposal, prevote and precommit. The proposal message is used by a proposer to suggest a particular value or state, while the prevote and precommit messages allow other nodes to vote on the proposal [45]. Tendermint uses a locking mechanism to guarantee consensus if the number of malicious nodes on the network does not surpass one-third of the total participants [12]. This locking mechanism uses the term ‘polka’, which checks that two-thirds of the prevotes are for a single block. If a validator tries to publish a block without a polka, it is considered malicious behaviour as shown in the ‘Commit’ phase in Figure 10. Cosmos is an example of a Blockchain that ----- _Electronics 2022, 11, 2694_ 13 of 23 is leveraging the Tendermint consensus mechanism. Cosmos is a multichain Blockchain, which allows many independent Blockchains, called zones, to run in parallel, with the ability to communicate through a central Blockchain, called the hub. The native token on the Cosmos Blockchain is called Atom. **Figure 10. Tendermint Consensus.** **6. Alternative Consensus Mechanisms** The consensus mechanisms PoW and PoS are well-known and widely used in Blockchains and cryptocurrencies. In this section, well cover 3 consensus mechanisms that deviate from purely PoW and PoS. PoC creates consensus using hard drive capacity, PoI heavily integrates a reputation system into its consensus mechanism. Section 6 concludes by covering hybrid consensus mechanisms. _6.1. Proof of Capacity_ A consensus mechanism that focuses on hard drive capacity, rather than mining with graphics cards or ASICs (Application-specific integrated circuit). Proof of Capacity saw its first use in the cryptocurrency BurstCoin. Mining on BurstCoin has two phases, plotting, and mining. Plotting involves hashing a list of nonce values, and then storing them on a hard drive. BustCoin uses the Shabal hashing algorithm, which is harder to hash than Bitcoins SHA256. Rather than discarding the hashes like in Bitcoin, they are bundled together into scoops (a pair of hashes), and stored on the nodes hard drive. In the mining phase, miners calculate a scoop number, and they use that scoop number to create a deadline value [46]. A node’s deadline value will vary depending on the hashes that have been calculated, and will represent a time limit in seconds. A node that is able to calculate ----- _Electronics 2022, 11, 2694_ 14 of 23 a deadline with the lowest time, is given the right to produce a block. An outline of this porocess can be found in Figure 11. BurstCoin has since rebranded to Signum, and has changed to a hybrid consensus mechanism called PoC+. PoC+ still requires a commitment of hard drive space, with miners now having the option to stake their Signa coins, which will increase their chance of mining a block. **Figure 11. PoC Consensus.** _6.2. Proof of Importance_ Proof of Importance (PoI) is a consensus mechanism originally proposed by the New Economy Movement (NEM). PoI shares similarities to PoS, where nodes are required to lock up a certain about of coins. However, rather than just keeping a node running like in the case of PoS, PoI has some extra requirements to encourage network usage and calculate a wallet’s importance [47] as scene in Figure 12. To be selected for the importance calculation, NEM wallets must have a minimum of 10,000 coins vested for a certain period. An importance score can also be increased by using the NEM network and sending transactions. Safeguards have been put in place against loop attacks, which involves sending coins between accounts controlled by a single actor, to boost their importance [48]. NEM has added a mechanism that heavily weights the importance of an account sending NEM, and minimal weight to an account that sends many coins, but receives most or all of their NEM back. Even if an account were to attempt the loop attack, they gain a minor increase in their importance score (<10%) but gain very little monetarily, as the extra money that receive from their higher importance, is lost in transaction fees attempting to boost their importance [48]. The name of NEM’s native Blockchain token is XEM. ----- _Electronics 2022, 11, 2694_ 15 of 23 **Figure 12. PoI Consensus.** _6.3. Hybrid Consensus (PoW/PoS)_ A number of cryptocurrencies have taken alternative approaches to consensus, by combining elements from PoW and PoS. Decred is a cryptocurrency that saw the flaws in PoW (double-spend problem) and the issues in PoS (nothing at stake) and decided to create a hybrid consensus mechanism to mitigate these problems as shown in Figure 13. Miners on the Decred network are still used to produce blocks, but are unable to add blocks directly to the Blockchain. Instead, miners propose their blocks to a network of PoS nodes who purchase tickets as their stake [49]. If a PoS node is pseudorandomly selected from this pool of tickets, they will be required to validate the block and add it to Decred’s Blockchain, as shown in figure Figure 13. These improvements stop miners from creating private chains and adds a checkpoint system that stops large parts of the Blockchain from being reorganised in the event of an attack. The cryptocurrency Horizen also uses a Hybrid consensus mechanism. Horizen leverages a network of PoW miners, to solve a cryptographic puzzle. Horizen full nodes are still given a reward for running honestly but are not part of the consensus process. Horizen’s ‘Secure Nodes’ provide a more secure version of standard full nodes which are found in other cryptocurrencies. Horizen requires their Secure Nodes to use TLS encryption, hold a small number of tokens, and maintain a full copy of the Blockchain. Secure Node users are compensated with part of the block reward, providing a financial incentive to support the network [50]. In 2018, [Horizen was a victim of a 51% attack (https://www.coindesk.com/markets/2018/06/08](https://www.coindesk.com/markets/2018/06/08/blockchains-once-feared-51-attack-is-now-becoming-regular/) [/blockchains-once-feared-51-attack-is-now-becoming-regular/ (accessed on 20 August](https://www.coindesk.com/markets/2018/06/08/blockchains-once-feared-51-attack-is-now-becoming-regular/) 2022)), and decided to make a modification to their consensus mechanism to make future attacks more difficult. Horizen added a delay function, that penalises miners for keeping their private Blockchain hidden from the network. Malicious miners will be required to continue mining their Blockchain for a certain number of blocks, according to the delay function, rather than having honest nodes instantly adopt their modified Blockchain once it is made public [51]. This makes 51% attacks require more time to execute, and require more electricity, when comparing attacks to Horizen’s original implementation of the LcR. The consensus mechanisms discussed in Sections 4–6 have been conveniently summarised into 3 tables. Table 1 discusses common properties of consensus mechanisms, such as block time, Transactions Per Second (TPS), and adversary tolerance. Table 2 specficially discusses the consensus mechanisms designed for IoT devices (PoSCS, Microchain, PoEWAL, and CBPoW) in more detail. Table 3 compares the discussed consensus mechnaisms against our criteria that was defined in Section 2, and allocates each consensus mechanisms a rating. ----- _Electronics 2022, 11, 2694_ 16 of 23 **Figure 13. Decred’s hybrid PoW/PoS consensus mechanism.** **Table 1. Overview of all the consensus mechanisms mentioned in Sections 4–6.** **Adversary** **Consensus** **Blockchain** **Block Time** **TPS** **L2 Network** **Reference** **Tolerance** PoW PoS Bitcoin 10 min 7 Lightening [52,53] Litecoin 2.5 min 56 Network [54] Monero 2 min Variable None [55] <51% Side Chains, Ethereum 12–14 s 15 [56] Rollups Horizen 2.5 min N/A Side Chains [50] CBPoW Variable 500+ None [24] PoEWAL Variable 25 None [25] Ethereum (PoS) 12 s TBD <51% TBD [40] Off-chain Algorand 4.5 s 1000 <33% [57] Contracts Dfinity Variable Variable <33% [31] Cosmos 6 s 1000+ <33% [58] PIVX 60 s 173 <51% None [55,59] Microchain 9 s 230+ <33% [43] PoSCS Variable Variable <51% [44] Lightening PoW + PoS Decred 5 min 14 <51% [49] Network PoC BurstCoin 4 min 80+ <50% [60] None PoI NEM 1 min 4000 <51% [48] A Blockchains layer 1 network is the Blockchains primary network, where transactions are created on-chain. Transactions on layer 2 are created off-chain, and are often compressed and posted on a Blockchains layer 1 network to increase scalability. Note: Transactions Per Second (TPS)—Only considers a Blockchains layer 1 network. ----- _Electronics 2022, 11, 2694_ 17 of 23 **Table 2. Overview of the IoT specific consensus mechanisms mentioned in Sections 4 and 5.** **Consensus** **Similar to** **Decentralised** **Features** **Apps** **Drawbacks** **Reference** Reputation PoSCS PoS No Supply Chains Cloud Reliance [44] System Crypto Synchronous Microchain PoS Partially IoT Blockchain [43] Sortition Networks Time-limited PoEWAL PoW IoT Dapps Synced Clocks [25] Partially PoW DAG CBPoW PoW Credit System Industrial IoT [24] Coordinator **Table 3. Consensus mechanism suitability for IoT devices, measured against our criteria defined in** Section 2. **Processor** **Consensus** **Security** **Decentralisation** **Storage** **TPS** **Suitable?** **Reference** **Usage** PoW High High High High Low No [16] PoS Medium High Medium High Variable Partially [16] PoW + PoS High High High High Low No [59] PoC Low High High High Low No [60] PoI Low High High High High Partially [48] PoSCS Low High Low Low Variable Partially [44] CBPoW Low High Medium Low Medium Yes [24] PoEWAL Low High High High low Partially [25] Microchain Medium High Medium High Medium Yes [43] Storage refers to the internal memory needed to store the Blockchain on IoT devices. TPS Refers to the transactions per second of the consensus mechanism. less than 100 TPS is low, 100–1000 TPS is considered medium, and 1000+ is considered high TPS. **7. Analysis** Before starting the analysis, we will discuss the three trade-offs Blockchains commonly make. The Blockchain trilemma commonly effects the design choices of consensus mechanisms, and also has consequences related to IoT devices. After, we will discuss each consensus mechanism, and talk about their suitability for the IoT. _7.1. Blockchain Trilemma_ Blockchains have three important properties: security, decentralisation, and scalability. Many Blockchains are only able to pick two of these properties, while having to compromise on the third [61]. The Blockchain trilemma was originally coined by Ethereum’s creator, Vitalik Buterin. Buterin explains that, using these three properties, that simple (meaning with no advanced techniques, such as sharding [62]) Blockchains can broadly be placed into three categories: traditional Blockchains, high Transaction Per Second (TPS) Blockchains, and multichain Blockchains. Bitcoin and Ethereum (pre-PoS Ethereum) are examples of traditional Blockchains. Bitcoin and Ethereum highly value decentralisation, and security, at the expense of scalability [63]. Blockchains that prioritise speed, and security, generally have a limited amount of nodes participating in consensus. Blockchains with Delegated Proof of Stake (DPoS) such as EoS [64] and TRON [65] are examples of Blockchains that prioritise performance. These Blockchains are able to process more transactions per second, but are more prone to centralisation due to a smaller number of nodes participating in consensus [66]. Blockchains such as COSMOS [58] and Avalanche [67] are two examples of multi-chain Blockchains. From the trilemma triangle, these sorts of Blockchains generally prioritise scalability and decentralisation. Buterin suggests that multichain Blockchains may not be able to pro ----- _Electronics 2022, 11, 2694_ 18 of 23 vide certain security guarantees, when implementing more advanced techniques, such as sharding [61]. _7.2. Proof of Work Suitability_ PoW can be quickly discarded from the list of suitable consensus mechmanisms for IoT devices. PoW is extemely energy intensive [65], processor intensive, and requires specalised mining hardware. PoW is not suitable for the IoT. _7.3. Proof of Stake Suitability_ PoS was given a suitability score of partial, and could potentially be implemented for the IoT. PoS has more desirable qualities than PoW for deployment in the IoT, as the energy and processor intensive nature required for consensus has been removed. However, PoS still has challenges for IoT usage: 1. Cryptography such as VRFs and the BLS signature scheme can be processor intensive, and may become problematic when working with resource constrained devices, especially as the network grows. Ethereum (PoS) Figure 7 and Dfinity Figure 6 are well-known to use these cryptographic functions in their consensus implementations. 2. PoS Blockchains are based on monetary concepts that involve a currency, which may not be suitable for some IoT applications. Arbitrary tokens could replace a monetary system, but this still comes with economic issues (who generates the tokens, how are tokens distributed, is the token amount capped etc.). 3. The transaction throughput of the network may not be adequate depending on the IoT use case, the PoS TPS varies dramatically from just over 100 TPS, to well over 1000 TPS as shown in Table 1. Selecting a particular PoS implementation will become important, as different PoS implementations have varying performance capabilities. Due to these issues, PoS may be suitable for the IoT, but only under particular circumstances, such as where a more decentralisaed network is required. _7.4. Hybrid Consensus (PoW/PoS) Suitability_ PoW/PoS hybrid consensus mechanisms are an alternative solution when attempting to solve the 51% attack problem and nothing at stake problem using a novel two consensus solution. Unfortunately, having PoW as a prerequisite brings all the same problems faced by PoW and its IoT suitability. Theoretically, the PoW porition of this hybrid consensus mechanism could be done offsite by powerful ASICS. The PoS portion could then be delegated to the IoT devices. We believe a configuration like the one mentioned above would be excessively complicated for consensus in the IoT and do not recommend Hybrid PoW/PoS for the IoT. _7.5. Proof of Capacity Suitability_ PoC is another novel consensus mechanism that uses the concept of hard drive capacity to create consensus. Local storage on IoT devices is limited and would not be available for use in consensus. PoC is not suitable for use in the IoT. _7.6. Proof of Importance Suitability_ This consensus mechanism takes concepts from PoS, and combines them with an importance mechanism. The higher a node’s importance score, combined with the nodes total amount of staked coins, calculates their chance of being selected to mint a block. PoI satisfies many criteria in Table 1, which make it well suited for the IoT. Adoption of PoI consensus in Blockchains is limited, as NEM is the only Blockchain to implement the PoI consensus mechanism as seen on CoinMarketCap [68]. PoI is referenced in the literature, in surveys such as [16,69] giving a general description of the flow of the consensus mechanism. However, the authors are not aware of any papers or references that validate these claims and support the figures in Table 1. Until papers are developed that independently ----- _Electronics 2022, 11, 2694_ 19 of 23 verify the characteristics of the PoI, we can only partially recommend the PoI consensus mechanism for the IoT. _7.7. PoSCS Suitability_ PoSCS is a consensus mechanism that also uses elements of PoS. PoSCS uses a staking system, but replaces a monetary system with a reputation system based on how a node interacts with the supply chain. PoSCS also relies on the cloud to archive parts of the Blockchain that are not required to be stored on IoT devices, one of the few consensus mechanisms discussed here that addresses this problem. Based on the results of PoSCS [44], the transaction throughput may not be high enough for some IoT use cases, and the reliance on the cloud may not be suitable for some implementations. For these reasons, we think PoSCS is partially suitable, and may be usable for some IoT implementations. _7.8. CBPoW Suitability_ One of the two consensus mechanisms that adapts PoW for the IoT. CBPoW dynamically adjusts the PoW mining difficulty for nodes, and actively punishes misbehaving nodes by making their PoW mining difficulty very high, to the point where mining is infesable. CBPoW also replaces a traditional Blockchain with a DAG, which has the ability to prune itself to reduce the sizde of the Blockchain stored locally on the device. This consensus mechanism performes well according to Table 1 with a throughput of 500 TPS. CBPoW has one central point of failure in current implementations (the coordinator, in the case of the IoTA DAG [70]) shown in Table 2. CBPoW has a number of characteristics that are suitable for the IoT including: a DAG Blockchain structure which has the ability to reduce its size, a lightweight PoW consensus mechanism and moderately high transaction throughput. Due to these 3 features that are favorable for IoT devices, we have labeled CBPoW as suitable in Table 3. _7.9. PoEWAL Suitability_ PoEWAL consensus mechanism is also based on a modified version of PoW. In PoEWAL, the PoW mining process is time limited. Devices have a short amount to mine a block, significantly reducing power and processor usage. PoEWAL also makes the assumption that devices all have synced clocks, which is probably a reasonable assumption for IoT devices on a wireless sensor network collecting time series data. This requirement may unsuitable for some implementations, as IoT devices use commodity parts and are generally more susceptible to drifting out of sync [71]. PoEWAL has two limitations: low transaction throughput and reliance on sycnhronised clocks. Due to these limitations, we have labeled PoEWAL as only partially suitable on Table 3. _7.10. Microchain Suitability_ Microchain has adapted concepts from PoS and made them more suitable for IoT usage. Nodes use credit amounts, rather than a monetary system, and Microchain has a block selection process to accommodate this change in PoS. Microchain also uses VRF cryptography to power its consensus mechanism, which may incur high processor usage on IoT devices as the network grows. Microchain also made some assumptions around network environments (synchronous networks in Table 2), which makes this consensus mechanism unusable for public Blockchains. Microchain is also shown to have reasonable performance, managing 230 TPS in their tests. Microchain has a number of features that make it suitable for IoT devices: - A PoS implementation not dependent on a monetary system - A moderately fast transaction trhougput - Processor usage which remains low on controlled private networks Due to these features which favour IoT devices, Microchain has been labeled as suitable in Table 3. ----- _Electronics 2022, 11, 2694_ 20 of 23 **8. Conclusions** In this paper, we discussed resource constrained IoT devices, and the limitations of the current consensus mechanisms on the IoT. We started the discussion by defining a criterion to rank the consensus mechanisms against, including speed, security, decentralisation, and others. The discussion continued by individually describing a number of consensus mechanisms and defining their general flow through a series of figures. We described well-known consensus mechanisms such as PoW and PoS, but also discuss 4 IoT consensus mechanisms purpose built for the IoT (CBPoW, Microchain, PoEWAL, and PoSCS). These IoT focused consensus mechanisms make modifications to the already existing PoW and PoS consensus, but remove the need for energy inefficient mining and monetary systems respectfully. In our analysis, discuss the advantages and disadvantages of each consensus mechanism, and their suitability for the IoT. The results from our analysis show that Microchain and CBPoW are suitable for the IoT. Microchain shows that it has suitable performace for the IoT in private environments, and CBPoW addresses problems around local Blockchain storage on IoT devices. Microchain and CBPoW have been labeled suitable in our analysis. Four other consensus mechanisms where also labeled partially suitable, including: PoEWAL, PoSCS, PoI and PoS. Some of the partially recommended consensus mechanisms addressed issues with monetary systems, computational overhead, and storage requirements. However, other issues where introduced, such as cloud reliance, synchronised clocks and poor and unverified performance claims. _Future Work_ The current trend around consensus research in academia, and in industry, has been focused on making mechanisms that are lightweight enough for low powered devices, and reducing the carbon footprint of the mining process. Our future research will look more closely into novel consensus mechanisms. Specifically, those that can meet the unique requirements of an IoT device, and mechanisms that can be deployed to meet business specific goals in both private, and potentially public operating environments. **Author Contributions: Conceptualization, Z.A. and N.C.; investigation, Z.A.; writing—original draft** preparation, Z.A.; writing—review and editing, N.C., R.A., W.H.; supervision, N.C., R.A. All authors have read and agreed to the published version of the manuscript. **Funding: This work was supported by the SmartSat C.R.C., whose activities are funded by the** Australian Government’s C.R.C. Program. **Acknowledgments: The authors would like to thank SmartSat CRC for providing scholarship fund-** ing, and BAE Systems Australia, our industry collaborator. **Conflicts of Interest: The authors declare no conflict of interest.** **References** 1. Singh, S.; Sharma, P.K.; Yoon, B.; Shojafar, M.; Cho, G.H.; Ra, I.H. Convergence of blockchain and artificial intelligence in IoT [network for the sustainable smart city. Sustain. Cities Soc. 2020, 63, 102364. [CrossRef]](http://doi.org/10.1016/j.scs.2020.102364) 2. [Min, H. Blockchain technology for enhancing supply chain resilience. Bus. Horizons 2019, 62, 35–45. [CrossRef]](http://dx.doi.org/10.1016/j.bushor.2018.08.012) 3. Dujak, D.; Sajter, D. Blockchain Applications in Supply Chain. In SMART Supply Network; Springer: Berlin/Heidelberg, Germany, 2019; pp. 21–46. 4. Sternberg, H.S.; Hofmann, E.; Roeck, D. The Struggle is Real: Insights from a Supply Chain Blockchain Case. J. Bus. Logist. 2021, _[42, 71–87. [CrossRef]](http://dx.doi.org/10.1111/jbl.12240)_ 5. Casado-Vara, R.; Prieto, J.; Prieta, F.D.L.; Corchado, J.M. How blockchain improves the supply chain: Case study alimentary [supply chain. Procedia Comput. Sci. 2018, 134, 393–398. [CrossRef]](http://dx.doi.org/10.1016/j.procs.2018.07.193) 6. Werner, R.; Lawrenz, S.; Rausch, A. Blockchain Analysis Tool of a Cryptocurrency. In PervasiveHealth: Pervasive Computing _Technologies for Healthcare; Springer: Berlin/Heidelberg, Germany, 2020; pp. 80–84._ 7. Hopkins, A.L.; Lala, J.H.; Smith, T.B. The Evolution of Fault Tolerant Computing at the Charles Stark Draper Laboratory, 1955–85; Springer: Berlin/Heidelberg, Germany, 1987; pp. 121–140. ----- _Electronics 2022, 11, 2694_ 21 of 23 8. Corbett, J.C.; Dean, J.; Epstein, M.; Fikes, A.; Frost, C.; Furman, J.J.; Ghemawat, S.; Gubarev, A.; Heiser, C.; Hochschild, P.; et al. [Spanner: Google’s Globally-Distributed Database. ACM Trans. Comput. Syst. (TOCS) 2013, 31, 1–21. [CrossRef]](http://dx.doi.org/10.1145/2491245) 9. Gupta, A.; Yang, F.; Govig, J.; Kirsch, A.; Chan, K.; Lai, K.; Wu, S.; Dhoot, S.; Kumar, A.; Agiwal, A.; et al. Mesa: Geo-Replicated, [Near Real-Time, Scalable Data Warehousing. Proc. Vldb Endow. 2014, 7, 1259–1270. [CrossRef]](http://dx.doi.org/10.14778/2732977.2732999) 10. Nakamoto, S. Bitcoin: A Peer-to-Peer Electronic Cash System; Portal Unicamp: Campinas, Brazil, 2008. 11. [Lamport, L.; Shostak, R.; Pease, M. The Byzantine generals problem. ACM Trans. Program. Lang. Syst. 2019, 4, 382–401. [CrossRef]](http://dx.doi.org/10.1145/357172.357176) 12. Buchman, E. Tendermint: Byzantine Fault Tolerance in the Age of Blockchains. Ph.D. Thesis, The University of Guelph, Guelph, ON, Canada, 2016. 13. Helliar, C.V.; Crawford, L.; Rocca, L.; Teodori, C.; Veneziani, M. Permissionless and permissioned blockchain diffusion. Int. J. Inf. _[Manag. 2020, 54, 102136. [CrossRef]](http://dx.doi.org/10.1016/j.ijinfomgt.2020.102136)_ 14. Polge, J.; Robert, J.; Traon, Y.L. Permissioned blockchain frameworks in the industry: A comparison. ICT Express 2021, 7, 229–233. [[CrossRef]](http://dx.doi.org/10.1016/j.icte.2020.09.002) 15. Yang, R.; Wakefield, R.; Lyu, S.; Jayasuriya, S.; Han, F.; Yi, X.; Yang, X.; Amarasinghe, G.; Chen, S. Public and private blockchain [in construction business process and information integration. Autom. Constr. 2020, 118, 103276. [CrossRef]](http://dx.doi.org/10.1016/j.autcon.2020.103276) 16. Salimitari, M.; Chatterjee, M.; Fallah, Y.P. A survey on consensus methods in blockchain for resource-constrained IoT networks. _[Internet Things 2020, 11, 100212. [CrossRef]](http://dx.doi.org/10.1016/j.iot.2020.100212)_ 17. Conti, M.; Sandeep, K.E.; Lal, C.; Ruj, S. A survey on security and privacy issues of bitcoin. IEEE Commun. Surv. Tutorials 2018, _[20, 3416–3452. [CrossRef]](http://dx.doi.org/10.1109/COMST.2018.2842460)_ 18. Anceaume, E.; Lajoie-Mazenc, T.; Ludinard, R.; Sericola, B. Safety analysis of Bitcoin improvement proposals. In Proceedings of the 2016 IEEE 15th International Symposium on Network Computing and Applications, NCA 2016, Boston, MA, USA, 31 October–2 November 2016; pp. 318–325. 19. Decker, C.; Seidel, J.; Wattenhofer, R. Bitcoin meets strong consistency. In Proceedings of the 17th International Conference on Distributed Computing and Networking, Singapore, 4–7 January 2016. 20. Shi, E. Analysis of deterministic longest-chain protocols. In Proceedings of the 2019 IEEE 32nd Computer Security Foundations Symposium (CSF), Hoboken, NJ, USA, 25–28 June 2019; pp. 122–135. 21. Antonopoulos, A.M.; Wood, G. Mastering Ethereum: Building Smart Contracts and Dapps; O’Reilly Media: Sebastopol, CA, USA, 2018. 22. Alwen, J.; Chen, B.; Pietrzak, K.; Reyzin, L.; Tessaro, S. Scrypt is maximally memory-hard. In Lecture Notes in Computer Science _(Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Springer: Berlin/Heidelberg, Germany,_ 2017; Volume 10212 LNCS, pp. 33–62. 23. Jamtel, E.L. Swimming in the Monero pools. In Proceedings of the 11th International Conference on IT Security Incident Management and IT Forensics, IMF 2018, Hamburg, Germany, 7–9 May 2018; pp. 110–114. 24. Huang, J.; Kong, L.; Chen, G.; Wu, M.Y.; Liu, X.; Zeng, P. Towards secure industrial iot: Blockchain system with credit-based [consensus mechanism. IEEE Trans. Ind. Inform. 2019, 15, 3680–3689. [CrossRef]](http://dx.doi.org/10.1109/TII.2019.2903342) 25. Raghav; Andola, N.; Venkatesan, S.; Verma, S. PoEWAL: A lightweight consensus mechanism for blockchain in IoT. Pervasive _[Mob. Comput. 2020, 69, 101291. [CrossRef]](http://dx.doi.org/10.1016/j.pmcj.2020.101291)_ 26. [King, S.; Nadal, S. Ppcoin: Peer-to-Peer Crypto-Currency with Proof-of-Stake. Available online: https://bitcoin.peryaudo.org/](https://bitcoin.peryaudo.org/vendor/peercoin-paper.pdf) [vendor/peercoin-paper.pdf (accessed on 20 August 2022).](https://bitcoin.peryaudo.org/vendor/peercoin-paper.pdf) 27. [Zhang, S.; Lee, J.H. Analysis of the main consensus protocols of blockchain. ICT Express 2020, 6, 93–97. [CrossRef]](http://dx.doi.org/10.1016/j.icte.2019.08.001) 28. Gilad, Y.; Hemo, R.; Micali, S.; Vlachos, G.; Zeldovich, N. Algorand: Scaling Byzantine Agreements for Cryptocurrencies. In Proceedings of the 26th Symposium on Operating Systems Principles, Shanghai, China, 28–31 October 2017. 29. Badertscher, C.; Gaži, P.; Kiayias, A.; Russell, A.; Zikas, V. Ouroboros genesis: Composable proof-of-stake blockchains with dynamic availability. In Proceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security, Toronto, ON, Canada, 15–19 October 2018; pp. 913–930. 30. Galindo, D.; Liu, J.; Ordean, M.; Wong, J.M. Fully Distributed Verifiable Random Functions and their Application to Decentralised Random Beacons. IACR Cryptol. EPrint Arch. 2020, 2020, 96. 31. Hanke, T.; Movahedi, M.; Williams, D. DFINITY Technology Overview Series, Consensus System. arXiv 2018, arXiv:1805.04548. 32. Boneh, D.; Lynn, B.; Shacham, H. Short Signatures from the Weil Pairing. In Lecture Notes in Computer Science; Springer: Berlin/Heidelberg, Germany, 2001; Volume 2248, pp. 514–532. 33. Buterin, V. A Next-Generation Smart Contract and Decentralized Application Platform; Ethereum Foundation: Bern, Switzerland, 2014. 34. Wood, G. Ethereum: A Secure Decentralised Generalised Transaction Ledger; Ethereum Foundation: Bern, Switzerland, 2022. 35. [Angeris, G.; Kao, H.T.; Chiang, R.; Noyes, C.; Chitra, T. An analysis of Uniswap markets. Cryptoecon. Syst. 2019, 1. [CrossRef]](http://dx.doi.org/10.21428/58320208.c9738e64) 36. Schär, F. Decentralized Finance: On Blockchain- and Smart Contract-Based Financial Markets. Fed. Reserve Bank St. Louis Rev. **[2021, 103, 153–174. [CrossRef]](http://dx.doi.org/10.20955/r.103.153-74)** 37. Scholten, O.J.; Gerard, N.; Hughes, J.; Deterding, S.; Drachen, A.; Walker, J.A.; Zendle, D. Ethereum Crypto-Games: Mechanics, Prevalence and Gambling Similarities. In Proceedings of the Annual Symposium on Computer-Human Interaction in Play, Barcelona, Spain, 22–25 October 2019; Association for Computing Machinery: New York, NY, USA, 2019. 38. Peterson, J.; Krug, J.; Zoltu, M.; Williams, A.K.; Alexander, S. Augur: A decentralized oracle and prediction market platform (v2. 0). arXiv 2021, arXiv:1501.01042. ----- _Electronics 2022, 11, 2694_ 22 of 23 39. Mohanty, D. Ethereum: What Lies Ahead. In Ethereum for Architects and Developers; Apress: New York, NY, USA, 2018; pp. 245–258. 40. Buterin, V.; Hernandez, D.; Kamphefner, T.; Pham, K.; Qiao, Z.; Ryan, D.; Sin, J.; Wang, Y.; Zhang, Y.X. Combining GHOST and Casper. arXiv 2020, arXiv:2003.03052. 41. Buterin, V.; Griffith, V. Casper the Friendly Finality Gadget. arXiv 2017, arXiv:1710.09437. 42. [Beekhuizen, C. Validated, Staking on Eth2: #2—Two Ghosts in a Trench Coat. Available online: https://blog.ethereum.org/2020](https://blog.ethereum.org/2020/02/12/validated-staking-on-eth2-2-two-ghosts-in-a-trench-coat) [/02/12/validated-staking-on-eth2-2-two-ghosts-in-a-trench-coat (accessed on 20 August 2022).](https://blog.ethereum.org/2020/02/12/validated-staking-on-eth2-2-two-ghosts-in-a-trench-coat) 43. Xu, R.; Chen, Y.; Blasch, E.; Chen, G. Microchain: A Hybrid Consensus Mechanism for Lightweight Distributed Ledger for IoT. _arXiv 2019, arXiv:1909.10948._ 44. Tsang, Y.P.; Choy, K.L.; Wu, C.H.; Ho, G.T.S.; Lam, H.Y. Blockchain-Driven IoT for Food Traceability with an Integrated Consensus [Mechanism. IEEE Access 2019, 7, 129000–129017. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2940227) 45. Buchman, E.; Kwon, J.; Tendermint, Z.M. The latest gossip on BFT consensus. arXiv 2018, arXiv:1807.04938. 46. Andrey, A.; Petr, C. Review of Existing Consensus Algorithms Blockchain. In Proceedings of the 2019 IEEE International Conference Quality Management, Transport and Information Security, Information Technologies IT and QM and IS 2019, Sochi, Russia, 23–27 September 2019; pp. 124–127. 47. Wen, Y.; Lu, F.; Liu, Y.; Cong, P.; Huang, X. Blockchain Consensus Mechanisms and Their Applications in IoT: A Literature Survey. In Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Springer: Berlin/Heidelberg, Germany, 2020; Volume 12454 LNCS, pp. 564–579. 48. [Nem Foundation. NEM Technical Reference. Available online: https://www.cryptoground.com/nem-white-paper (accessed on](https://www.cryptoground.com/nem-white-paper) 20 August 2022). 49. [Jepson, C. DTB001: Decred Technical Brief. Available online: https://decred.org/dtb001.pdf (accessed on 20 August 2022).](https://decred.org/dtb001.pdf) 50. [Viglione, R.; Versluis, R.; Lippencott, J. Zen White Paper. Available online: https://www.horizen.io/assets/files/Zen-White-](https://www.horizen.io/assets/files/Zen-White-Paper.pdf) [Paper.pdf (accessed on 20 August 2022).](https://www.horizen.io/assets/files/Zen-White-Paper.pdf) 51. Garoffolo, A.; Viglione, R. Sidechains: Decoupled Consensus between Chains. arXiv 2018, arXiv:1812.05441. 52. Bhaskar, N.D.; Chuen, D.L.K. Bitcoin Mining Technology. In Handbook of Digital Currency: Bitcoin, Innovation, Financial Instruments, _and Big Data; Academic Press: Cambridge, MA, USA, 2015; pp. 45–65._ 53. Gobel, J.; Krzesinski, A.E. Increased block size and Bitcoin blockchain dynamics. In Proceedings of the 2017 27th International Telecommunication Networks and Applications Conference, ITNAC 2017, Melbourne, Australia, 22–24 November 2017; pp. 1–6. 54. Boshuis, S.; Braam, T.; Marchena, A.P.; Jansen, S. The Effect of Generic Strategies on Software Ecosystem Health: The Case of Cryptocurrency Ecosystems. In Proceedings of the 2018 IEEE/ACM 1st International Workshop on Software Health (SoHeal), Gothenburg, Sweden, 27 May 2018. 55. [Lee, J.H. Rise of Anonymous Cryptocurrencies: Brief Introduction. IEEE Consum. Electron. Mag. 2019, 8, 20–25. [CrossRef]](http://dx.doi.org/10.1109/MCE.2019.2923927) 56. Bach, L.M.; Mihaljevic, B.; Zagar, M. Comparative analysis of blockchain consensus algorithms. In Proceedings of the 2018 41st International Convention on Information and Communication Technology, Electronics and Microelectronics, Opatija, Croatia, 21–25 May 2018; pp. 1545–1550. 57. Esgin, M.F.; Kuchta, V.; Sakzad, A.; Steinfeld, R.; Zhang, Z.; Sun, S.; Chu, S. Practical Post-quantum Few-Time Verifiable Random Function with Applications to Algorand. In Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial _Intelligence and Lecture Notes in Bioinformatics); Springer: Berlin/Heidelberg, Germany, 2021; Volume 12675 LNCS, pp. 560–578._ 58. [Kwon, J.; Buchman, E. Cosmos Whitepaper. Available online: https://v1.cosmos.network/resources/whitepaper (accessed on](https://v1.cosmos.network/resources/whitepaper) 20 August 2022). 59. Baudlet, M.; Fall, D.; Taenaka, Y.; Kadobayashi, Y. The Best of Both Worlds: A New Composite Framework Leveraging PoS and PoW for Blockchain Security and Governance. In Proceedings of the 2020 2nd Conference on Blockchain Research and Applications for Innovative Networks and Services, BRAINS 2020, Paris, France, 28–30 September 2020; pp. 17–24. 60. [Gauld, S.; von Ancoina, F.; Stadler, R. The Burst Dymaxion. Available online: https://www.allcryptowhitepapers.com/wp-](https://www.allcryptowhitepapers.com/wp-content/uploads/2018/05/Burst-Coin-whitepaper.pdf) [content/uploads/2018/05/Burst-Coin-whitepaper.pdf (accessed on 20 August 2022).](https://www.allcryptowhitepapers.com/wp-content/uploads/2018/05/Burst-Coin-whitepaper.pdf) 61. [Buterin, V. Why Sharding Is Great: Demystifying the Technical Properties. Available online: https://vitalik.ca/general/2021/04/](https://vitalik.ca/general/2021/04/07/sharding.html) [07/sharding.html (accessed on 20 August 2022).](https://vitalik.ca/general/2021/04/07/sharding.html) 62. Zamani, M.; Movahedi, M.; Raykova, M. Rapidchain: Scaling blockchain via full sharding. In Proceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security, Toronto, ON, Canada, 15–19 October 2018; pp. 931–948. 63. Kiayias, A.; Panagiotakos, G. Speed-Security Tradeoffs in Blockchain Protocols. IACR Cryptol. EPrint Arch. 2015, 2015, 1019. 64. Xu, B.; Luthra, D.; Cole, Z.; Blakely, N. EOS: An Architectural, Performance, and Economic Analysis. Available online: [https://blog.bitmex.com/wp-content/uploads/2018/11/eos-test-report.pdf (accessed on 20 August 2022).](https://blog.bitmex.com/wp-content/uploads/2018/11/eos-test-report.pdf) 65. Li, J.; Li, N.; Peng, J.; Cui, H.; Wu, Z. Energy consumption of cryptocurrency mining: A study of electricity consumption in [mining cryptocurrencies. Energy 2019, 168, 160–168. [CrossRef]](http://dx.doi.org/10.1016/j.energy.2018.11.046) 66. Al-Saqaf, W.; Seidler, N. Blockchain technology for social impact: Opportunities and challenges ahead. J. Cyber Policy 2017, _[2, 338–354. [CrossRef]](http://dx.doi.org/10.1080/23738871.2017.1400084)_ 67. Tanana, D. Avalanche blockchain protocol for distributed computing security. In Proceedings of the 2019 IEEE International Black Sea Conference on Communications and Networking (BlackSeaCom), Sochi, Russia, 3–6 June 2019. 68. [CoinMarketCap. Top POI Tokens by Market Capitalization. Available online: https://coinmarketcap.com/view/poi/ (accessed](https://coinmarketcap.com/view/poi/) on 20 August 2022). ----- _Electronics 2022, 11, 2694_ 23 of 23 69. Zhang, P.; Schmidt, D.C.; White, J.; Dubey, A. Chapter Seven—Consensus mechanisms and information security technologies. _Adv. Comput. 2019, 15, 181–209._ 70. Silvano, W.F.; Marcelino, R. Iota Tangle: A cryptocurrency to communicate Internet-of-Things data. Future Gener. Comput. Syst. **[2020, 112, 307–319. [CrossRef]](http://dx.doi.org/10.1016/j.future.2020.05.047)** 71. Tirado-Andrés, F.; Rozas, A.; Araujo, A. A Methodology for Choosing Time Synchronization Strategies for Wireless IoT Networks. _[Sensors 2019, 19, 3476. [CrossRef] [PubMed]](http://dx.doi.org/10.3390/s19163476)_ -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/electronics11172694?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/electronics11172694, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/2079-9292/11/17/2694/pdf?version=1662026333" }
2,022
[]
true
2022-08-27T00:00:00
[]
17,890
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01540651818ded5880dbfa4a6aecf665efd8a00a
[ "Computer Science" ]
0.854752
Policies and Security Aspects For Distributed Scientific Laboratories
01540651818ded5880dbfa4a6aecf665efd8a00a
IFIP International Information Security Conference
[ { "authorId": "1751943", "name": "N. Dessì" }, { "authorId": "1702365", "name": "M. Fugini" }, { "authorId": "122042482", "name": "R. Balachandar" } ]
{ "alternate_issns": null, "alternate_names": [ "Symp Edge Comput", "Symposium on Edge Computing", "Inf Secur", "Information Security", "IFIP Int Inf Secur Conf", "SEC" ], "alternate_urls": [ "https://acm-ieee-sec.org" ], "id": "c60f6e64-434b-4c82-931d-faefcb362d4c", "issn": null, "name": "IFIP International Information Security Conference", "type": "conference", "url": "http://www.tc11.uni-frankfurt.de/" }
null
# Policies and Security Aspects For Distributed Scientific Laboratories Nicoletta Dess´ı, Maria Grazia Fugini, R. A. Balachandar **Abstract Web Services and the Grid allow distributed research teams to form dy-** namic, multi-institutional virtual organizations sharing high performance computing resources, large scale data sets and instruments for solving computationally intensive scientific applications, thereby forming Virtual Laboratories. This paper aims at exploring security issues of such distributed scientific laboratories and tries to extend security mechanisms by defining a general approach in which a security policy is used both to provide and regulate access to scientific services. In particular, we consider how security policies specified in XACML and WS-Policy can support the requirements of secure data and resource sharing in a scientific experiment. A framework is given where security policies are stated by the different participants in the experiment, providing a Policy Management system. A prototype implementation of the proposed framework is presented. ## 1 Introduction Web Services (WS) and the Grid have revolutionized the capacity to share information and services across organizations that execute scientific experiments in a wide range of disciplines in science and engineering (including biology, astronomy, highenergy physics, and so on) by allowing geographically distributed teams to form dynamic, multi-institutional virtual organizations whose members use shared community tools and private resources to collaborate on solutions to common problems. Since WS have been recognized as the logical architecture for the organization of Nicoletta Dess´ı and R. A. Balachandar Dipartimento di Matematica e Informatica, Universit´a degli Studi di Cagliari, Via Ospedale 72, 09124 Cagliari, Italy, e-mail: dessi@unica.it, balsonra@yahoo.co.in Maria Grazia Fugini Dipartimento di Elettronica e Informazione, Politecnico di Milano, piazza L. da Vinci 32, 20133 Milano, Italy, e-mail: fugini@elet.polimi.it _Please use the following format when citing this chapter:_ ----- Grid services, they can enable the formation of Virtual Laboratories, which are not simply concerned with file exchange, but also with direct access to computers, software, and other resources, as required by a dynamic collaboration paradigm among organizations [6]. As the community of researchers begins to use Virtual Laboratories, exploiting Grid capabilities [16], the definition of secure collaborative environments for the next generation of the science process will need further potentialities. In order to extend common security mechanisms such as certification, authorization or cryptography. These new functions include, for example, the definition and the enforcement of policies in place for single Virtual Laboratories in accordance with dynamically formed Virtual Organizations (VOs), and the integration of different local policies, in order to make the resources available to the VO members, who deploy their own services in the VO environment. These considerations motivate the approach that we propose in this paper, whose aim is to explore the security of environments supporting the execution of scientific experiments in a Virtual Laboratory. Specifically, the paper elaborates on extending usual control access mechanism by defining a general approach in which security policies are expressed and enforced to regulate _resource sharing and service provisioning. In detail, the paper proposes a reference_ framework for secure collaboration where security policies can be formulated in order to regulate access to scientific services and to their provisioning. Since each Virtual Laboratory has a set of local security policies, we examine how these polices can be expressed and enforced such that the allocation process of resources to a distributed experiment is made aware of security implications. As a sample application of the proposed approach, some implementation hints are presented for distributed experiments that incorporate security policies. This paper is structured as follows. Section 2 reviews related work. Section 3 addresses requirements to be considered when security policies for experiments are defined. Section 4 presents our reference framework for Virtual Laboratories, with emphasis on security issues. Section 5 details our approach to Policy Management, giving a component architecture and providing implementation aspects. Finally, Section 6 contains the conclusions. ## 2 Related Work A Virtual Laboratory for e-Science can be viewed as a cooperative System where WS are dynamically composed in complex processes (experiments) and executed at different organizations. WS security [19] is assuming more and more relevance since WS handle users’ private information. WS-Trust [9] describes a framework for managing, assessing and establishing trust relationships for WS secure interoperation. In WS-based systems, security is often enforced through security services [20], for which new specifications have been developed to embed such security services in the typical distributed and WS-based elements, considering also security policies [18]. Examples are the SOAP header [19], the Security Assertion Markup Language (SAML) [12], XML Signature [4] and XML Encryption [14]. WS-Security [3] ap ----- plies XML security technologies to SOAP messages with XML elements. Based on SOAP e-Services, [8] proposes an access control system, while XACML (XML Access Control Markup Language) [2] allows fine-grained access control policies to be expressed in XML. However, all these mechanisms prove useful in specifying specific aspects of security, but need to be selected first, and integrated later, into a uniform framework addressing all issues regarding e-collaboration. Policies, as an increasingly popular approach to dynamic adjustability of applica tions, require an appropriate policy representation and the design and development of a policy management framework. Considering that security policies should be part of WS representations, [19] and [10] specify the Web Services Policy Framework (WS-Policy). Policy-based management is supported by standards organizations, such as the Internet Engineering Task Force (IETF). The IETF framework [13] defines a policy-based management architecture, as the basis for other efforts at designing policy architectures. Existing technology for the Grid (e.g., see [11]) allows scientists to develop project results and to deploy them for ongoing operational use, but only within a restricted community. However, security is still implemented as a separate subsystem of the Grid, making the allocation decisions oblivious of the security implications. Lack of security [20] may adversely impact future investment in e-Science capabilities. The e-Science Core Programme initiated a Security Taskforce (STF) [http://www.nesc.ac.uk/teams/stf/], developing a Security Policy for e-Science (http://www.nesc.ac.uk/teams/stf/links/), while an authorization model for multipolicies is presented in [17]. An approach combining Grid and WS for e-Science is presented in [5, 1]. Authorizations in distributed workflows executed with their own distinctive ac cess control policies and models has been tackled in [7]; security is handled through alarms and exceptions. In [15] access control for workflows is described explicitly addressing cooperation. However, decentralization of workflow execution is not explicitly addressed nor security policies handling is specifically tackled. ## 3 Basic Security Aspects for Virtual Laboratories At least for certain domains, scientific experiments are cooperative processes that operate on, and manipulate, data sources and physical devices, whose tasks can be decomposed and made executable as (granular) services individually. Workflows express appropriate modeling of the experiment as a set of components that need to be mapped to distinct services and support open, scalable, and cooperative environments for scientific experiments [5]. We denote such scientific environments as Virtual Laboratories (VLs) or eLabs. Each VL node (or eNode) is responsible for offering services and for setting the rules under which the service can be accessed by other eNodes through service invocation. Usually, the execution of an experiment involves multiple eNodes interacting to offer or to ask for services. Services correspond to different functionalities ----- that encapsulate problem solving and data processing capabilities. Services can be designed to use of VOs resources while the network infrastructure promotes the exploitation of distributed resources in a transparent manner. This offers good opportunities for achieving an open, scalable and cooperative environment. We classify services in: - Vertical services, that include components for a range of scientific domains, in cluding various software applications. - Horizontal services, that provide adaptive user interfaces, plug-and-play collab orative work components, interoperability functions, transaction co-ordination, and security. Vertical services expose interfaces that convey information about specific application functions. Their interfaces are implemented from within the component embedding them and are assembled in a workflow that globally expresses the experiment _model. Horizontal services allow for easier, more dynamic and automated eNode_ integration and for more precise run-time integration of remote services. They are designed to facilitate collaboration. A VO member plans a complex scientific experiment by repeatedly choosing a sequence of services and including these services in a workflow. He can wait for the fulfilment of a specific workflow and/or choose the next service to invoke on the basis of the returned information. The workflow execution may require the collaboration of various services spread over different VLs whose owners must be confident that users accessing their software or data respect fine-grained access restrictions controlling the varying levels of access to the resource that a user may be granted for. For example, a service may require commodity processors or may have a limited choice of input data (possibly requiring a specific file-format or database access). Similarly, a scientist executing a service on a remote eNode must trust the administrator of the eNode to deliver a timely and accurate result (and possibly proprietary data sent to the site). This requires the extension of security aspects related to resource sharing to those related to service sharing. However, security is still currently unsupported in an integrated way by any of the available WS technologies, nor a standard method to enforce Grid security is defined. Moreover, security policy requirements have to be considered. The approach of this paper regards the definition of the basic aspects to be tackled when extending WS and Grid security infrastructures to VLs environments. ## 4 A Reference Framework for Virtual Laboratories Based on what illustrated so far, we now introduce some basic modeling elements for the context of VLs security, by defining as an actor each relevant subject capable of executing experiments supported by networked resources, which we consider as _objects. In detail:_ ----- - Subjects execute activities and request access to information, services, and tools. Among subjects we mention the remote user of a WS/Grid enabled application, which would generally be composed of a large, distributed and dynamic population of resources. Subjects may also include organizations, servers and applications acting on behalf of users. In this paper, we consider only trusted groups which are not requested to exchange security tokens or credentials during a scientific experiment, since they know and trust each other, and received authentication and authorization to access resources when first joining the VL. - Objects are the targets of laboratory activities. Services are considered as objects. Methods are also regarded as objects, which can be grouped together to form experiments. Fine-grained access control would thus be required over input and output parameters, methods, WS and groupings among WS (to form a process) and among WS and other applications (e.g., legacy software or device control software). Other objects are the server hosting the WS, an IP address, or the URI of a WS. Internal data, kept in a database and other objects accessed by the WS, should also be considered as part of the list of objects to be managed. - Actions that can be performed are various, depending on the type of subject issu ing a request. Remote users or applications would generally be allowed to execute a WS method, or access a server hosting a number of WS objects or an application. Rights corresponding to actions such as place experiment, or view results, update data could be granted. The identification of subjects and objects in a scientific environment defines a framework for secure collaboration based on the idea of integrating components that control the workflow execution through a set of specific security components. Such framework, depicted in Fig. 1 comprises components (diamonds), their specific activities (ovals) and specific security aspects (double-border boxes). The framework elements are as follows: **Process Manager - Each process manager supervises the execution of the work-** flow modeling the scientific experiment. It is responsible for the transformation of the abstract workflow into a concrete plan whose components are the executions of specific tasks/tools and/or actual accesses to data repositories. This planning process can be performed in cooperation with a service broker, acting as a mediator, in that it supplies, at run time, the description and location of useful resources and services. **Task Manager - This is in charge of executing a particular set of activities which are** instances of the workflow plan. It is also responsible for collaborating with others components for managing the service execution. In fact, execution involves contacting data sources and components and requesting the appropriate execution steps. **Service Manager - This supervises the successful completion of each task request.** In case of failure, the service manager takes appropriate repair actions. Repair may involve either restarting the task execution or re-entering the configuration component in order to explore alternative ways of instantiating the task execution to avoid service failures, e.g., due to a security attack or service misuse. In that case, the service flow can be rerouted to other services able to provide substitute functionalities, thus allowing redo or retry operations on tasks that were abnormally ended before ----- rerouting. Moreover, this component waits until completion of the task request, and notifies to the task manager the end of the activity. **Policy Manager - This component supports and updates the resource provision pol-** icy that regulates the flow of information through the applications and the network, and across organizational boundaries, to meet the security requirements established by members who are in charge of deploying their own services under their own policies that assert privileges and /or constraints on resource and services utilization. **Fig. 1 Security Aspects and Related Components of a Virtual Laboratory** Two major concerns in this framework are: structural and dynamic concerns, and security concerns. i) Structural and dynamic concerns deal with the execution of a scientific experiment in a VL and incorporate controls on vertical services. ii) _Security concerns refer to horizontal services supporting privileges and constraints_ on the of VL resources, and may differ from user to user for each individual service. The sequel of the paper presents how these policies can be implemented and how fine-grained constraints can be defined in the VL to gain restricted different access levels to services according to a policy that is fully decided by software owners themselves. ## 5 Policy Management Policy management in VLs, as the ability to support an access control policy in accordance with the resource access control goals, should support dynamically changing decentralized policies, policy administration and integrated policy enforcement. A typical policy management system would include two components, namely the _Policy Enforcement Point (PEP), and the Policy Decision Point (PDP), as shown in_ ----- Fig. 2. The PEP is the logical entity, or location within a server, responsible for enforcing policies with respect to authentication of subscribers, authorization to access and services, accounting and mobility, and other requirements. The PEP is used to ensure that the policy is respected before the user is granted access the WS resource. The PDP is a location where an access decision is formed, as a result of evaluating the user’s policy attributes, the requested operation, and the requested resource, in the light of applicable policies. The policy attributes may relate to authorization and authentication. They may also refer to the attributes related to Quality of Service (QoS), or to service implementation details, such as transport protocol used, and security algorithms implemented. The PEP and the PDP components may be either distributed or resident on the same server. In our VL, access control works as follows. A user who wants to perform an experiment submits a request to the appropriate resource(s) involved in the experiments through a set of invocations to WS providers. The Policy Manager (see Fig. 2) located in each of the requested resources, implements the PEP and the PDP to take the access decision about the user access request. The PEP wraps up an access request based on the user’s security attributes or credentials, on the requested resource, and on the action the user wants to perform on the resource. It then forwards this request to the PDP, which checks the request against the resource policy and determines whether the access can be granted. **Fig. 2 Policy Management System** There is no standard way of implementing the PDP and PEP components; they shall either be located in a single machine or be distributed in the different machines depending on the convenience of the Grid Administrator and of the resource provider. The Policy Manager (see Fig. 2) has the ability to recognize rules from the WS requestor and provider of relevant sources, and is able to correctly combine applicable access rules to return a proper, enforceable access decision. Generally, policies are defined for access to a single resource; hence, the PEP and the PDP can be contained in a single eNode or be distributed. VL resources may ----- be part of more than one application and therefore there should be a defined access _control service. Further, these resources can be used contemporaneously by different_ applications with different associated policies; hence they will be processed by the applicable Policy Managers. In that case, the applications have their own PEP and _PDP, which control user access to the applications. Further, the Policy Manager_ must be able to recognize the policy attributes related to access control, as well as, the information related to QoS. In the following subsection, we describe the implementation methodology employed for the Policy Manager and the standard specification used to express the access policy requirements for a resource. The described access control mechanisms of the Policy Manager can be imple mented using XACML, which includes both a policy language and an access control decision request/response language (both encoded in XML). The policy language is used to describe general access control requirements, and has standard extension points for defining new functions, data types, combining logic, etc. The request/response language allows queries on whether a given action should be allowed, and the interpretation of the result. The response always includes an answer about whether the request should be allowed using one of four values: Permit, Deny, Indeterminate (in case of error or required values missing, that so a decision cannot be made) or Not Applicable (the request can’t be answered by this service). A Policy represents a single access control policy, expressed through a set of Rules. Each XACML policy document contains exactly one Policy or a PolicySet, that contains other policies or a reference to policy locations. For example, consider a scenario where a user wants to access and read a web page available in a resource. The XACML representation of this request in the PEP is as follows: _Request_ < _Subject_ < _Attribute AttributeId = ”urn : oasis : names : tc : xacml : 1_ 0 : subject : subject − _id”_ < . _DataType = ”urn : oasis : names : tc : xacml : 1_ 0 : data − _type : r fc822Name”_ . _AttributeValue_ _www_ _unica_ _it_ _AttributeValue_ < - . . < / _Attribute_ < / _Subject_ < / _Resource_ < _AttributeAttributeId = ”urn : oasis : names : tc : xacml : 1_ 0 : resource : resource − _id”_ < . _DataType = ”http :_ _www_ _w3_ _org_ 2001 _XMLSchema#anyURI”_ // . . / / _AttributeValue_ _http :_ _webmail_ _dsf_ _unica_ _it_ _userGuide gLite_ _html_ _AttributeValue_ < - // . . . / . < / _Attribute_ < / _Resource_ < / _Action_ < _AttributeAttributeId = ”urn : oasis : names : tc : xacml : 1_ 0 : action : action − _id”_ < . _DataType = ”http :_ _www_ _w3_ _org_ 2001 _XMLSchema#string”_ // . . / / _AttributeValue_ _read_ _AttributeValue_ < - < / _Attribute_ < / ----- _Action_ < / _Request_ < / The PEP submits this request form to the PDP component which checks this request against the policy of the resource hosting the intended web page. For example, the following policy states that the ”developers” group is allowed to read the resource (i.e., the Web Page): _RuleRuleId = ”ReadRule”E f fect = ”Permit”_ < _Target_ < _Subjects_ < _AnySubject_ < / > _Subjects_ < / _Resources_ < _AnyResource_ < / > _Resources_ < / _Actions_ < _Action_ < _ActionMatchMatchId = ”urn : oasis : names : tc : xacml : 1_ 0 : function : string − _equal”_ < . _AttributeValue_ < _DataType = ”http :_ _www_ _w3_ _org_ 2001 _XMLSchema#string”_ _read_ _AttributeValue_ // . . / / - < / _ActionAttributeDesignatorDataType = ”http :_ _www_ _w3_ _org_ 2001 _XMLSchema#string”_ < // . . / / _AttributeId = ”urn : oasis : names : tc : xacml : 1_ 0 : action : action − _id”_ . / > _ActionMatch_ < / _Action_ < / _Actions_ < / _Target_ < / _ConditionFunctionId = ”urn : oasis : names : tc : xacml : 1_ 0 : function : string − _equal”_ < . _ApplyFunctionId = ”urn : oasis : names : tc : xacml : 1_ 0 : function : string − _one_ − _and −_ _only”_ < . _SubjectAttributeDesignatorDataType = ”http :_ _www_ _w3_ _org_ 2001 _XMLSchema#string”_ < // . . / / _AttributeId = ”group”_ / > _Apply_ < / _AttributeValue_ < _DataType = ”http :_ _www_ _w3_ _org_ 2001 _XMLSchema#string”_ _developers_ _AttributeValue_ // . . / / - < / _Condition_ < / _Rule_ < / The PDP checks this policy against the request and determines whether the read request can be allowed for the web page. It then forms a XACML response and forwards it to the PEP which eventually allows the user to read the page. The implementation of XACML provides a programming interface to read, evaluate and validate XACML policies. It can also be used to develop the Policy Manager con ----- taining the PEP and the PDP, and performs most of the functionalities of the Policy Manager. We can create a PEP which interacts with a PDP by creating requests and interpreting the related responses. A PEP typically interacts in an applicationspecific manner and there is currently no standard way to send XACML requests to an online PDP. Hence, we need to include code for both PEP and PDP in the same application. For instance, the following code snippet will create an XACML request and pass the same to the PDP. _RequestCtxrequest = newRequestCtx(subjects_ _resourceAttrs_ _actionAttrs_ ,,, _environmentAttrs);_ _ResponseCtxresponse = pdp_ _evaluate(request);_ . The XACML based Policy Manager can recognize policy attributes related to authentication and authorization. Hence, they can be used only for implementing access control mechanisms. However, such authorization policies do not express the capabilities, requirements, and general characteristics of entities (i.e., users and resources) in an XML WS-based system and there are some more attributes, different from the access control attributes, that need to be examined before accessing a WS. For instance, one may need to negotiate QoS characteristics of the service, or privacy policies and also the kind of security mechanism used in the WS. Unfortunately, XACML does not provide the grammar and syntax required to express these policies. For this aspects, we use WS-policy specifications which provide a flexible and extensible grammar for expressing various aspects of policy attributes, such as the used authentication scheme, the selected transport protocol, the algorithm suite, and so on. For example, the following specification represents the policy for the algorithm suite required for cryptographic operations with symmetric or asymmetric key based security tokens (it is also possible to include timestamps to the policy specifications to prevent any misuse of the policies). _wsp : Policy_ < _xmlns : sp = ”http :_ _schemas_ _xmlsoap_ _org_ _ws_ 2005 07 _securitypolicy”_ // . . / / / / _xmlns : wsp = ”http :_ _schemas_ _xmlsoap_ _org_ _ws_ 2004 09 _policy”_ // . . / / / / _wsp : ExactlyOne_ < _sp : Basic256Rsa15_ < / > _sp : TripleDesRsa15_ < / > _wsp : ExactlyOne_ _wsp : All_ < / >< _sp : IncludeTimestamp_ < / > _wsp : All_ < / _wsp : Policy_ < / The Apache implementation of WS-Policy provides versatile APIs for program matic access to WS-Policies. Under this approach, we can implement a policy matching mechanism to negotiate security attributes, and other QoS attributes, be ----- fore actual access to the WS. Moreover, WS-policy APIs are a flexible tool to read, compare and verify the attributes present in WS-Policies. For instance, the following code snippet shall be used for creating a Policy Reader object to access a WS-Policy (here Policy A) and to compare this object with another policy (Policy B): _PolicyReaderreader =_ _PolicyFactory_ _getPolicyReader(PolicyFactory_ _DOM POLICY READER);_ . . _PolicyReaderreader =_ _PolicyFactory_ _getPolicyReader(PolicyFactory_ _DOM POLICY READER);_ . . _FileInputStreamPolicy A = newFileInputStream(”ResA_ _xml”);_ . _PolicypolicyA = reader_ _readPolicy(Policy A);_ . _FileInputStreamPolicy B = newFileInputStream(”ResB_ _xml”);_ . _PolicypolicyB = reader_ _readPolicy(Policy B);_ . _Booleanresult = PolicyComparator2_ _compare(Policy A_ _Policy B)_ ., Through the combination of XACML and WS-Policy specifications, we can implement a full fledged Policy Management system for WS to manage authorization policies on resources as well as policies related to security and other QoS aspects. However, this Policy Management system cannot be used as such in Grid environments, considering the very nature of jobs and resources in the Grid. In fact, in the Grid, there are computationally intensive resources, such as clusters, that can either host an experiment as a service, or allow jobs to be executed in it. Hence, the policy requirements in this environment will be different from those of WS environments. For example, suppose that a resource wants to contribute up to (but not more than) 200MB of its memory for job execution in the Grid. To express such policy, currently existing policy languages do not offer enough grammar and syntax. Hence, we suggest to extend the existing policy language schema to include policies regarding elements typical of Grid Services, such as bandwidth information, memory, CPU cycle, etc. For our prototype implementation, we consider three attributes namely the memory, CPU cycle and the available nodes in the cluster resource and a schema is developed with these attributes. The APIs of the WS-Policy implementation are modified accordingly, to deal with this schema and be able to perform operations such as compare, read, normalize, and so on. The schema that includes the attributes related to a Grid resource, and its usage in WS-Policy is as follows: _xs : schema_ < _targetNamespace = ”http :_ _unica_ _it_ _gridpolicy_ _xsd”_ // . / . _xmlns : tns = ”http :_ _unica_ _it_ _gridpolicy_ _xsd”_ // . / . _xmlns : xs = ”http :_ _www_ _w3_ _org_ 2001 _XMLSchema”_ // . . / / _elementFormDe fault = ”qualified”_ _blockDe fault = ”#all”_ ----- _xs : elementname = ”Mem”type = ”tns : OperatorContentType”_ < / > _xs : elementname = ”ProcessorSpeed”type = ”tns : OperatorContentType”_ < / > _xs : elementname = ”DiskSpace”type = ”tns : OperatorContentType”_ < / > The following WS-Policy uses this schema to represent the capabilities and policy information of a Grid resource: _wsp : Policyxmlns : sp = ”http :_ _schemas_ _xmlsoap_ _org_ _ws_ 2005 07 _securitypolicy”_ // . . / / / / _xmlns : wsp = ”http :_ _schemas_ _xmlsoap_ _org_ _ws_ 2004 09 _policy”_ // . . / / / / _xmlns : cs = ”http :_ _schemas_ _mit_ _edu_ _cs”_ _wsp : ExactlyOne_ _wsp : All_ // . . / >< >< _cs : Mem_ 1024 _cs : Mem_ < - < / _cs : ProcessorSpeed_ 2GHz _cs : ProcessorSpeed_ < - < / _wsp : All_ < / _wsp : All_ _sp : Basic256Rsa15_ _sp : TripleDesRsa15_ _wsp : ExactlyOne_ _wsp : All_ < >< / >< / >< / < _wsp : ExactlyOne_ < / _wsp : Policy_ < / Through this policy, the Grid resource wants to advertise that it can allocate no more than 1GB of its free memory to Grid job execution, and that it is able to provide 2GHz of its processor speed. This policy information can be read and compared with other policies using the WS-Policy implementation libraries. This prototype implementation modifies the WS-Policy specification to deal with a larger number of attributes. To implement these issues in a real time dynamic environment, an extensive survey of Grid resource usage policies and their representation in a WS-policy schema are needed. Our future research will investigate the development of a Policy Management system working for both WS and Grid environments. ## 6 Implementation Hints The illustrated framework has been the basis for developing a prototype VL which, in an initial validation stage, has been used to test secure cooperation from the perspective of one scientific server only, for which a Security Server has been implemented, containing security functions deployed as Security WS. The prototype (see Fig. 3) is built on top of Taverna[1], a workflow composer that allows designers to map the initial abstract workflow into a detailed plan. Each Taverna workflow consists of a set of components, called Processors, each with a name, a set of inputs and a set of outputs. The aim of a Processor is to define an inputs-to-outputs transformation. Vertical services can be installed by adding to Taverna new plug-in processors that can operate alone or can be connected with data and workflows through control links. When a workflow is executed and the execution reaches a security Proces 1 Taverna is available in the myGrid open source E-science environment http://www.mygrid.org.uk/ ----- sor, an associated invocation task is called that invokes a specific horizontal service implementing security mechanisms. The Scufl workbench included in MyGrid provides a view for composition and execution of processors. The internal structure of a VL includes four components: a Security Server, a Front-End, a Back-End, a Workflow Editor. The Security Server exposes various functionalities aimed at data privacy and security both in the pole and during the interaction among poles. It manages User Authentication, Validity check of Security Contracts, Trust Levels, Cryptographic Functions, and Security Levels. The Security Server service communicates with the front-end scientific services by sending them the local Security Levels and the list of remote poles offering a specific resource. User authentications occurs through insertion of a secret code by the user requesting the execution of a protected workflow. The Front-end of the scientific pole is a set of WS that can be invoked by a workflow editor, after negotiation. These WS interact with the Security Server, from which they require information related to the local pole access policy. The Front-end includes services that do not hold their own resource trust level, but rather inherit the clearance level of the user executing the WS. However, the Front-end service receives, at creation time, a threshold security level, reflecting the quality and sensitiveness of the service. **Fig. 3 Security Components Implementation Architecture** The Back-end of a scientific pole is constituted by the local resources of the sci entific pole, e.g., calculus procedures or data stored in databases. All the resources in the Back-end are exposed as WS, and can be invoked by a remote Virtual Laboratory. Each resource has its own Resource Service Level assigned by an administrator. The applied policy is ”no read up, no write down”. The invocations of the Back-end services are protected via SSL. Finally, the scientific workflow is defined ----- using the Taverna workflow editor of MyGrid [2]. Upon proper negotiation of security contracts, a download of the workflow modifier tool and the encryption/decryption module from the provider pole is required. The modifier tool modifies the scientific workflow, by adding crypt and decrypt activities and the input data related to access codes of services. The crypt/decrypt module implements cryptographic functions on exchanged data (we use AES). These editors are designed to be used by scientists teams, generally co-ordinated by a Chief Scientist. However, a workflow is not associated to a whole, given global Security Level, but rather each service of the workflow has an associated Security Level depending on the qualification of the user requiring the service. ## 7 Concluding Remarks This paper has highlighted the requirements that should be considered when access control policies of Virtual Laboratories are written. To allow an access control policy to be flexible and dynamic, it can no longer be a high-level specification, but must become a dynamic specification that allows real-time access control administration of WS and the Grid resources. To this aim, we have presented the security requirements of a cooperative environment for executing scientific experiments. Namely, we have illustrated XACML policy specifications, and the use of the WSPolicy to define scientific resource sharing requirements needed to securely activate a collaboration in experiments with negotiating of QoS policy attributes. A security framework and a prototype environment have been presented, with the purpose of providing a uniform view of Grid service policies for a dynamic environment where a set of nodes cooperate to perform a scientific experiment. Currently there exists no standardized access control for virtual applications implemented with WS on the Grid. We plan to extend the requirements presented in this paper and define a formal security model and architecture for WS and Grid enabled scientific applications. The model will be based on the security policy languages used in this paper, independently of specific technologies and configuration models. This should ensure industry-wide adoption by vendors and organizations alike to allow crossorganization business integration. Interoperation requires a standard-based solution. In fact, a Virtual Laboratory, created with WS and the Grid, where scientific relationships may frequently change, requires a highly flexible, but robust security framework, based on approval and universal acceptance of standards. This would allow business partners to avoid interoperability problems among their disparate applications and maintain a security context to allow interoperation. **Acknowledgements This paper has been partially supported by the Italian TEKNE Project.** 2 Taverna, and other e-Science management tools, are freely available on the Internet, but to ensure encryption, decryption and server authentication capabilities they require additional features. ----- ## References 1. Amigoni F., Fugini M.G., Liberati D., ”Design and Execution of Distributed Experiments”, Proc. 9th International Conference on Enterprise Information Systems, (ICEIS’07), Madeira, June 2007 2. Anderson A. et. al., XACML 1.0 Specification, http://www.oasis-open.org/committees/tc home.php?wg abbrev=xacml, 2003 3. Atkinson B. et al., Web Services Security (WS-Security), 2002, Version 1.0 April 5, 2002, http://www.verisign.com/wss/wss.pdf 4. Bartel M., Boyer J., Fox B., LaMacchia B. and Simon, XML Signatures, http://www.w3.org/TR/2002/REC-xmldsig-core-20020212/E 5. Bosin A., Dess N., Fugini M.G., Liberati D., Pes B., ”Supporting Distributed Experiments in Cooperative Environments”, in Business Process Management, Springer-Verlag Bussler C., Haller A. (Eds.), vol. 25, 2006, pp. 281 - 292 6. Camarinha-Matos L.M., Silveri I., Afsarmanesh H., and Oliveira A.I., ”Towards a Frame work for Creation of Dynamic Virtual Organizations”, in Collaborative Networks and Their Breeding Environments, Springer, Boston Volume 186/2005, 2005, pp. 69-80 7. Casati F., Castano S., Fugini M.G., ”Managing Workflow Authorization Constraints Through Active Database Technology”, Journal of Information Systems Frontiers, Special Issue on Workflow Automation and Business Process Integration, 2002 8. Damiani E., De Capitani di Vimercati S., Paraboschi S., Samarati P., ”Fine Grained Access Control for SOAP E-Services”, in Proc. of the Tenth International World Wide Web Conference, Hong Kong, China, May 1-5, 2001. 9. Della-Libera G. et al., Web Services Trust Language (WS-Trust), available at http://www.ibm.com/developerworks/library/ws-trust/index.html 10. Della-Libera G., et al, ”Web Services Security Policy Language (WS-SecurityPolicy,” July 2005. (See http://www.oasis-en.org/committees/download.php/16569/) 11. Foster, I. 2006. ”Service-Oriented Science: Scaling e-Science Impact”, Proceedings of the 2006 IEEE/WIC/ACM International Conference on Web intelligence, 2006 12. Hallam-Baker P., Hodges J., Maler E., McLaren C., Irving R., SAML 1.0 Specification, http://www.oasis-open.org/committees/tc home.php?wg abbrev=security, 2003 13. IETF Policy Framework Working Group, A framework for policy-based admission control, available at http://www.ietf.org/rfc/rfc2753.txt, 2003 14. ImamuraT., Dillaway B., Simon E., XML Encryption, http://www.w3.org/TR/xmlenc-core/ 15. Jiang H., Lu S., ”Access Control for Workflow Environment: The RTFW Model”, in Com puter Supported Cooperative Work in Design III, LNCS Springer Berlin / Heidelberg, Volume 4402/2007, 2007, pp. 619-626 16. Kim K.H., Buyya R., ”Policy-based Resource Allocation in Hierarchical Virtual Organiza tions for Global Grids”, 18th International Symposium on Computer Architecture and High Performance Computing (SBAC-PAD’06), 2006, pp. 36-46 17. Lang B., Foster I., Siebenlist F., Ananthakrishnan R., Freeman T., ”A Multipolicy Authoriza tion Framework for Grid Security,” Fifth IEEE International Symposium on Network Computing and Applications (NCA’06), 2006, pp. 269-272 18. Mohammad A.,Chen A.,Wang G. W., Changzhou C., Santiago R., ”A Multi-Layer Security Enabled Quality of Service (QoS) Management Architecture”, in Enterprise Distributed Object Computing Conference, 2007 (EDOC 2007) Oct. 2007, pp.423-423 19. Nadalin A., C. Kaler, P. Hallam-Baker, R. Monzillo (Eds.) Web Services Security, available at http://docs.oasis-open.org/wss/2004/01/oasis-200401-wss-soap-message-security-1.0.pdf 20. Welch V., Siebenlist F., Foster I., Bresnahan J., Czajkowski K., Gawor J., Kesselman C., Meder S., Pearlman L., Tuecke S., ”Security for Grid Services”, Proc. 12th IEEE International Symposium on High Performance Distributed Computing, 22-24 June 2003, pp. 48- 57 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-0-387-09699-5_15?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-0-387-09699-5_15, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://link.springer.com/content/pdf/10.1007/978-0-387-09699-5_15.pdf" }
2,008
[ "JournalArticle" ]
true
2008-09-07T00:00:00
[]
9,527
en
[ { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0154a7e5dfd13d425e75b44ccbbb87043c308356
[]
0.942783
Analysis of Consensus Mechanisms: PoW and PoS
0154a7e5dfd13d425e75b44ccbbb87043c308356
Applied and Computational Engineering
[ { "authorId": "73529960", "name": "Jia Wenxuan" } ]
{ "alternate_issns": null, "alternate_names": [ "Appl Comput Eng" ], "alternate_urls": null, "id": "38ef5a81-0fad-4de7-abc2-0fb847d3ece7", "issn": "2755-2721", "name": "Applied and Computational Engineering", "type": "journal", "url": null }
Along with the trend of Bitcoin blockchain, the concept behind all virtual currency has be-come popular in the study of the Internet. This essay mainly researches two kinds of common consensus mechanisms for the current blockchain network and looks forward to the future development of the technologys usage in daily life. This research aims to overview the two most common consensus mechanisms in the construction of blockchain. By reviewing re-sources from other research, an explanation of the goal of the consensus method, the ad-vantages, and disadvantages of each approach and the future development of these two meth-ods are summarized and developed. The result of the review explains the shifts of mature vir-tual currencies from proof of work to proof of stake and advises what mechanism should be used at starting stage and why a shift is necessary for proof of stake currencies.
Proceedings of the 2023 International Conference on Software Engineering and Machine Learning DOI: 10.54254/2755-2721/8/20230187 # Analysis of consensus mechanisms: PoW and PoS **Jiang Wenxuan** Ningbo Xiaoshi High School, Ningbo, Zhejiang, China, 315000 vincentjiang0207@163.com **Abstract. Along with the trend of Bitcoin blockchain, the concept behind all virtual currency** has become popular in the study of the Internet. This essay mainly researches two kinds of common consensus mechanisms for the current blockchain network and looks forward to the future development of the technology’s usage in daily life. This research aims to overview the two most common consensus mechanisms in the construction of blockchain. By reviewing resources from other research, an explanation of the goal of the consensus method, the advantages, and disadvantages of each approach and the future development of these two methods are summarized and developed. The result of the review explains the shifts of mature virtual currencies from proof of work to proof of stake and advises what mechanism should be used at starting stage and why a shift is necessary for proof of stake currencies. **Keywords: PoW, PoS, consensus mechanism, blockchain, virtual currency.** **1. Introduction** Just like what Satoshi Nakamoto said in his paper, “A purely peer-to-peer version of electronic cash would allow online payments to be sent directly from one party to another without going through a financial institution” [1]. Bitcoin is a kind of electronic cash in which there is no trusted third party required. As a kind of currency which is going to lead the world’s development in the future, the importance of researching and overviewing the current mechanism development is significant. In detail, this paper focused on the advantages and disadvantages of proof of work and proof of stake, known as PoW and PoS. The review of the structure of the algorithm and document from the developers is the source used by this report. By summarizing the feature of the two methods, both the long-term and short-term influences are discussed, and risks of exchanges and starting stages’ problems are also evaluated in the paper. The method of research is by overview mechanism of both proves and summarize their features. Furthermore, the situation is stimulated to evaluate the advantages and disadvantages of a specific currency development or usage stage. As a strong foundation, the report, which is a consensus mechanism analysis, could help researchers or organizations to make a better choice when constructing their new currency in the future. It would help other resources to summarize the new mechanism’s advantages and disadvantages. **2. Basic knowledge of blockchain and consensus mechanism** Blockchain is a chain list which connects blocks. In each block, numerous things are packaged, such as all trade data and a new key for the next block. The most important feature of a chain list is a connection to the next block requires addressing information in the block, and thus the problem of double pay and © 2023 The Authors. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0 (https://creativecommons.org/licenses/by/4.0/). ----- Proceedings of the 2023 International Conference on Software Engineering and Machine Learning DOI: 10.54254/2755-2721/8/20230187 modification of the contents of the block easily. The original blockchain is invented by Satoshi Nakamoto, and the product of this algorithm is Bitcoin, known as BTC. A schematic diagram for the original blockchain is shown in figure 1. In this schematic diagram, the inheritance of the blockchain is displayed, which is the connection between the first transaction and the following transaction that needs the previous key to verify the signature. **Figure 1. Schematic diagram of blockchain [1].** The idea of blockchain is essential for a currency without a third party. It provides the currency with numerous features different feature. However, the concern remains. In this paper, how to make a new block is the focus, and different consensus mechanisms would be researched. “Blockchain technology, the consensus mechanism determines the security, scalability, and decentralization of Blockchain” [2]. As the mechanism is introduced, the production of a new block is not defined properly. If a new block would be produced easily, too many branches of a blockchain would destroy the currency since people do not know where to write their trade and no branch would be really overall. Tons of merge processes are required and always required. Blockchain has many outstanding features, such as immutability, security, speed, and consensus as Lashkari mentioned in his review, and the consensus mechanism is the source of this feature [3]. For the consensus mechanism, the main idea is to add some obstacles for creating a new block and the difficulty would be verified easily. As the description shows, a one-way function is required for this obstacle. With this mechanism, the block produced would be much more reasonable and more people would trade on it and tend to add a new block after it, which is also known as consensus. Furthermore, a different way of proof would be possible if the current owner does not only determine the validity by the computer power. Another kind of consensus mechanism is based on the stake (currency) that the block creator owes and the standard of creating a new block is the currency they held (often, the time of holding is also considered an important factor). **3. Proof of work** For the proof of work, the main idea of this mechanism is to set a difficult task which would be difficult for everyone in every block, and thus it is difficult to add a new block. Two widely used approaches of proof of work are one-way function and space-time exam. _3.1. Hash applied to bitcoin_ For the hash, the most widely known example is bitcoin. It is the basis of countless different currencies. Bitcoin creates a fundament for all other algorithms using the hash. Although other, more practical ways have been developed, it is still critical for us to learn how a simple hash function works for blockchain. ----- Proceedings of the 2023 International Conference on Software Engineering and Machine Learning DOI: 10.54254/2755-2721/8/20230187 The main idea is to find a difficult problem and start to modify it. In this case, using the hash function secure hash algorithm 256-bit, to produce a number with numerous zeros at the first of the answer. This issue is relatively random, and no trend could be obtained. So, a huge number of trials, which means computing power, is required to make a new block. When the zero number increases, the difficulty of the concern increases as well. The change in difficulties would help to limit the rate of block produced. However, solving a problem would not avoid the modification concern, so the real rule of new block development is based on the data in the previous block. Every time when the miner wanted to produce a new block, their hash value needs to be added with the previous value of all the trades in the block, the signature of the miner who make the last block and the last block’s hash. If someone would like to modify the previous trade records, these people need to redo the hash calculation. Moreover, changing the content in the block means the miner must change all the following blocks. Using this kind of algorithm, the blockchain would be produced. Choosing bitcoin as an example is great since its process is easy to understand and straightforward. In other cases, some currencies would use a different hash function or a different way of validation, and some coins even use MD5 messagedigest algorithm as their one-way function (it has been proven to be easy to crash). This algorithm would be the mode widely known approach. _3.2. Storage applied for proof of the capacity_ POC means proof of capacity. It is different from the hash approach mentioned above. The examination takes place by storage ability. The developer of this mechanism claims it would not cost a lot of power compared to bitcoin. However, the end of POC causes a gigantic rise in hard disk prices. Most of the POC would also include time as another factor. Every time when a miner claim they have this ability of storage, the examiner request would be sent, and the miner needs to rent the request. A verifier can check if a prover has stored the data they committed to over space and over a period. _3.3. Advantages of proof of work_ The advantages of PoW are that it is a strong verification for every people, and it could be used at the start phase of a currency. Everyone would verify easily because it relied solely on a one-way function or simple same message. A hash function would be difficult to obtain a certain value, but easy to calculate a value. Moreover, in the process of building blocks, the total amount of currency is not enough at first. The consensus mechanism cannot be based on stake. For the PoW, it would reward the first group of miners with some currencies. It would encourage them to build this blockchain more and get more rewards due to the increase in the value of coins. No matter what time it is, the PoW mechanism would be useful for building blocks, since it depends solely on solving problems, instead of proving a huge amount of currency owned. Furthermore, this method of construction of new blocks could prevent forking because of limited computing power for every miner. As explained by Cointelegraph, “A miner would have to split their computational resources between the two sides of the fork in order to support both blockchains. As a result, through an economic incentive, proof-of-work systems naturally prevent constant forking and urge the miners to pick the side that does not wish to harm the network” [5]. _3.4. Disadvantages of proof of work_ As the currency developed more, plenty of disadvantages appear. When the difficulties of producing a new block increase to a high level, a huge number of resources, such as computers and electricity, would be wasted for calculating a simple hash function. And as the currency developed, even more resources would be required. According to Bonheur, “Powerful computers inherently consume a lot of energy. Furthermore, these machines require effective heat management or cooling system to remain operational [and prevent overheating, as well as associated damages to hardware components due to internal heat](https://www.profolus.com/topics/causes-of-overheating-in-electronic-devices/) build-up” [6]. Moreover, when someone wants to attack a kind of currency, a 51% attack is possible for some currency at the start. Since not plenty of miners are focusing on this blockchain, the attackers would use ----- Proceedings of the 2023 International Conference on Software Engineering and Machine Learning DOI: 10.54254/2755-2721/8/20230187 51% of their computing power to create a whole new branch and replace the original chain. It is possible to happen if someone really aimed to attack. So, it would increase the probability of being attacked by excessive computing power. **4. Proof of stake** The basic concept of proof of stake is that instead of using computer power as the indicator of the capability to build a new block, PoS is aiming for using a stake to prove the validity of a new block. In this way, relatively less computing power is required to maintain a currency and more problems emerge. _4.1. Ethereum 2.0_ It is widely known that Ethereum, known as ETH, is another widely used currency which uses PoW to create new blocks on the blockchain. However, in the next version of the ETH, its developer decided to shift to PoS. Before this transition, a huge amount of energy usage is required for bitcoin mining and ETH mining. For the further development of this currency, a shift is needed. For ETH 2.0, PoS has been used. In principle, PoS is a new consensus mechanism for ETH which “This staked ETH then acts as collateral that can be destroyed if the validator behaves dishonestly or lazily. The validator is then responsible for checking that new blocks propagated over the network are valid and occasionally creating and propagating new blocks themselves” [4]. As the schematic diagram in figure 2 shown, a part of the stake of the block creator would be used to verify the new block’s validity, and the stake decision chooses whether it would be the main branch or not to be selected. **Figure 2.** Schematic diagram of PoS used by ETH 2.0 [4]. It asks a group of validators to deposit some ETH into a contract and run software programs. The validator is responsible for creating a new block and sending it out to other people in the network. Furthermore, a feature of finality is introduced, which means the block’s content cannot be changed unless a lot of ETHs are burnt. Only two-thirds of the people agree the process would take place. _4.2. Advantages_ PoS mechanism only requires currency to prove the stakeholder’s ability to maintain a block, instead of using plenty of resources to verify their ability. In most cases, the PoS chain of a currency is developing a lot faster than the PoW chain, and thus, the speed of payment is faster. According to Gehmlich, “The proof-of-stake solves scalability issues that have been a thorn in the flesh in the proof-of-work consensus mechanism. PoS facilitates faster transactions since blocks are approved faster as there’s no need to solve complex mathematical equations. Since no physical machines or mining farms requiring ample energy supplies are needed to generate consensus, there is better scalability” [7]. ----- Proceedings of the 2023 International Conference on Software Engineering and Machine Learning DOI: 10.54254/2755-2721/8/20230187 For another reason, the people who take part in PoS are more likely to contribute more to the currency. Because they hold a lot of this kind of currency, improving in value of these coins would be beneficial for them, which means they want to and would work on it. For them, they have a good reason to maintain a new block nicely and fast. It is essential to reward the people who want to improve this currency. For the PoW, some miners might just complete the work and sell the coins. Moreover, it is less likely that a coin-used PoS would experience a 51% attack. Since it is possible that the attacker uses a lot of computer power to attack a coin, the 51% attack for a coin based on PoW is possible when it is relative, not popular. However, no matter what kind of coin, it is much more difficult for an attacker to control 51% of the coin. In some algorithms, change blocks even require the attacker to have two-thirds of the total currency. It is much more difficult and if you really have that number of coins, the people will not want to attack. _4.3. Disadvantages_ One of the disadvantages is that in a normal PoS community, more stake means more power in policymaking, but that is not always the case. It is usual for more people to invest some of their money into electric currency such as bitcoin and ETH, but they would not create their own electric wallet. Instead, they would only open an account on a platform and the platform would help them purchase and sell all the coins. For the concept of stake equal to power, people who owe the coins do not have the right to vote since the platform helps them to keep them, which means some massive cooperation might make several bad decisions to make these coins bad. And it is also difficult for most people to understand and try to use a hash code as their account. In another case, if the coins just start developing, PoS is not likely to work. Not numerous people have coins, and they could only buy them from other people or from mining. It is possible that only the introducer of the coin would maintain the chain and the reward cannot be distributed thus the influence of the coins cannot increase. As a result, no one would spend more on this coin. So, it is inappropriate to use PoS at the beginning of a new electric currency. Furthermore, proof of stake might cause the problem of centralization again. According to Chandler [8], since there is no limit on how much crypto a single validator could stake, a huge validator might act like a bank and control the currency. **5. Future development of consensus mechanisms** Nowadays, more and more consensus mechanisms are being developed. Combined networks such as "an improved network for blockchain is proposed to combine different blockchain networks together. It uses the POU consensus mechanism to improve the network environment, which consists of Proof of Stake Entrance, Hash Net Verification and Delegated Parliament” [9]. This kind of mixed network requires a different level of explaining and evaluating since it would combine the effects, too. Moreover, different methods from basic are developed, such as Proof-of-Stake with Delegated Ownership (DPoS) Blockchain-based Consensus and Fault Tolerance in the Byzantine Style (BFT) [10]. In the future, blockchain and electric currency will be used more and more. Furthermore, the certain difference would appear while developing. The content in each block would increase and the frequency of requests and receive would increase noticeably. To cope with this increase in demand for trade, better algorithms would be developed and applied to the currencies. Moreover, the ability to be updated is also being considered. Instead of having a different currency, a future developer should be able to update existing currency and blockchain. It could reduce the cost of transfer a lot. In the future, the blockchain could be used for smart contracts and create a global country. It is possible to see that during the usage of the electric currency, the concept of country and world could change. In mete verse, more currency system needs to be established. Moreover, NTF has appeared for a long time, and this merchandise could be stored on a blockchain. The more the title represents all kinds of possibilities. ----- Proceedings of the 2023 International Conference on Software Engineering and Machine Learning DOI: 10.54254/2755-2721/8/20230187 **6. Conclusion** In conclusion, the features of the PoW and PoS are varying, and all comes down to a simple fact: PoW is better for starting and PoS could be used for further development. It is certain that PoS will be the mainstream of blockchain in the future, but the contribution of the PoW should not be forgotten. The lack of method analysis is one of the improvements that the report could improve. Moreover, practical examples of risk for PoW and PoS could be added to illustrate these advantages and disadvantages in much more clarity and detail. These improvements could help readers to think about the current virtual currency’s development and provide important resources for them to make choices in the mechanism’s use. **References** [1] Bitcoin: A Peer-to-Peer Electronic Cash System Satoshi Nakamoto Institute. (n.d.). Retrieved [March 7, 2023, from https://nakamotoinstitute.org/bitcoin/](https://nakamotoinstitute.org/bitcoin/) [2] Liu, Z., Liu, W., Zhang, Y., Xu, G., & Yu, H. 2019. Overview of Blockchain Consensus Mecha nisms. _Journal_ _of_ _Cryptologic_ _Research,_ **6,** 395–432. [https://doi.org/10.13868/j.cnki.jcr.000311](https://doi.org/10.13868/j.cnki.jcr.000311) [3] Lashkari, B., & Musilek, P. 2021. A Comprehensive Review of Blockchain Consensus Mecha [nisms. IEEE Access, 9, 43620–43652. https://doi.org/10.1109/ACCESS.2021.3065880](https://doi.org/10.1109/ACCESS.2021.3065880) [4] [Proof-of-stake (PoS). (n.d.). Ethereum.Org. Retrieved March 7, 2023, from https://ethereum.org](https://ethereum.org/) [5] Proof-of-stake vs. proof-of-work: Pros, cons, and differences explained. (n.d.). Cointelegraph. [Retrieved March 12, 2023, from https://cointelegraph.com/blockchain-for-beginners/proof-](https://cointelegraph.com/blockchain-for-beginners/proof-of-stake-vs-proof-of-work:-differences-explained) [of-stake-vs-proof-of-work:-differences-explained](https://cointelegraph.com/blockchain-for-beginners/proof-of-stake-vs-proof-of-work:-differences-explained) [6] Bonheur, K. (2021, September 15). PoW: Advantages and Disadvantages of Proof-of-Work. [Profolus. https://www.profolus.com/topics/pow-advantages-and-disadvantages-of-proof-of-](https://www.profolus.com/topics/pow-advantages-and-disadvantages-of-proof-of-work/) work/ [7] Gehmlich, B. (2022, October 10). Pros and Cons of Proof of Stake for Ethereum Blockchain [Security. Gigster. https://gigster.com/pros-and-cons-of-pos-for-ethereum-security/](https://gigster.com/pros-and-cons-of-pos-for-ethereum-security/) [8] Chandler, S. (n.d.). Proof of stake vs. proof of work: Key differences between these methods of verifying cryptocurrency transactions. Business Insider. Retrieved March 12, 2023, from [https://www.businessinsider.com/personal-finance/proof-of-stake-vs-proof-of-work](https://www.businessinsider.com/personal-finance/proof-of-stake-vs-proof-of-work) [9] Guo, H., Zheng, H., Xu, K., Kong, X., Liu, J., Liu, F., & Gai, K. 2018. An Improved Consensus Mechanism for Blockchain. In M. Qiu (Ed.), Smart Blockchain **11373, 129–138. Springer In-** [ternational Publishing. https://doi.org/10.1007/978-3-030-05764-0_14](https://doi.org/10.1007/978-3-030-05764-0_14) [10] Blockchain Consensus Algorithms: What and How? Blockchain Certification Programs CBCA. [(n.d.). Retrieved March 7, 2023, from https://www.cbcamerica.org/blockchain-insights/block-](https://www.cbcamerica.org/blockchain-insights/blockchain-consensus-algorithms-what-and-how) [chain-consensus-algorithms-what-and-how](https://www.cbcamerica.org/blockchain-insights/blockchain-consensus-algorithms-what-and-how) -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.54254/2755-2721/8/20230187?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.54254/2755-2721/8/20230187, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "HYBRID", "url": "https://www.ewadirect.com/proceedings/ace/article/view/2511/pdf" }
2,023
[ "JournalArticle", "Review" ]
true
2023-08-01T00:00:00
[]
5,304
en
[ { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0155fe1751b05dda3dfde6e95745a5fc3e0c6e95
[]
0.833873
Cloud-Assisted Private Set Intersection via Multi-Key Fully Homomorphic Encryption
0155fe1751b05dda3dfde6e95745a5fc3e0c6e95
Mathematics
[ { "authorId": "3236877", "name": "Cunqun Fan" }, { "authorId": "2201198555", "name": "Peiheng Jia" }, { "authorId": "8766367", "name": "Manyun Lin" }, { "authorId": "2110972903", "name": "Lan Wei" }, { "authorId": "2075394776", "name": "Peng Guo" }, { "authorId": "3134309", "name": "Xiangang Zhao" }, { "authorId": "2154487593", "name": "Ximeng Liu" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-283014", "https://www.mdpi.com/journal/mathematics" ], "id": "6175efe8-6f8e-4cbe-8cee-d154f4e78627", "issn": "2227-7390", "name": "Mathematics", "type": null, "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-283014" }
With the development of cloud computing and big data, secure multi-party computation, which can collaborate with multiple parties to deal with a large number of transactions, plays an important role in protecting privacy. Private set intersection (PSI), a form of multi-party secure computation, is a formidable cryptographic technique that allows the sender and the receiver to calculate their intersection and not reveal any more information. As the data volume increases and more application scenarios emerge, PSI with multiple participants is increasingly needed. Homomorphic encryption is an encryption algorithm designed to perform a mathematical-style operation on encrypted data, where the decryption result of the operation is the same as the result calculated using unencrypted data. In this paper, we present a cloud-assisted multi-key PSI (CMPSI) system that uses fully homomorphic encryption over the torus (TFHE) encryption scheme to encrypt the data of the participants and that uses a cloud server to assist the computation. Specifically, we design some TFHE-based secure computation protocols and build a single cloud server-based private set intersection system that can support multiple users. Moreover, security analysis and performance evaluation show that our system is feasible. The scheme has a smaller communication overhead compared to existing schemes.
# mathematics _Article_ ## Cloud-Assisted Private Set Intersection via Multi-Key Fully Homomorphic Encryption **Cunqun Fan** **[1,2], Peiheng Jia** **[3], Manyun Lin** **[1,2], Lan Wei** **[1,2,][∗], Peng Guo** **[1,2], Xiangang Zhao** **[1,2]** **and Ximeng Liu** **[4]** 1 Key Laboratory of Radiometric Calibration and Validation for Environmental Satellites, National Satellite Meteorological Center (National Center for Space Weather), China Meteorological Administration, Beijing 100081, China 2 Innovation Center for FengYun Meteorological Satellite (FYSIC), Beijing 100081, China 3 School of Mathematics and Computer Science, Shanxi Normal University, Taiyuan 030031, China 4 College of Computer and Data Science, Fuzhou University, Fuzhou 350108, China ***** Correspondence: weilan@cma.cn **Citation: Fan, C.; Jia, P.; Lin, M.;** Wei, L.; Guo, P.; Zhao, X.; Liu, X. Cloud-Assisted Private Set Intersection via Multi-Key Fully Homomorphic Encryption. _Mathematics 2023, 11, 1784._ [https://doi.org/10.3390/](https://doi.org/10.3390/math11081784) [math11081784](https://doi.org/10.3390/math11081784) Academic Editor: Antanas Cenys Received: 21 March 2023 Revised: 4 April 2023 Accepted: 4 April 2023 Published: 8 April 2023 **Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **Abstract: With the development of cloud computing and big data, secure multi-party computation,** which can collaborate with multiple parties to deal with a large number of transactions, plays an important role in protecting privacy. Private set intersection (PSI), a form of multi-party secure computation, is a formidable cryptographic technique that allows the sender and the receiver to calculate their intersection and not reveal any more information. As the data volume increases and more application scenarios emerge, PSI with multiple participants is increasingly needed. Homomorphic encryption is an encryption algorithm designed to perform a mathematical-style operation on encrypted data, where the decryption result of the operation is the same as the result calculated using unencrypted data. In this paper, we present a cloud-assisted multi-key PSI (CMPSI) system that uses fully homomorphic encryption over the torus (TFHE) encryption scheme to encrypt the data of the participants and that uses a cloud server to assist the computation. Specifically, we design some TFHE-based secure computation protocols and build a single cloud server-based private set intersection system that can support multiple users. Moreover, security analysis and performance evaluation show that our system is feasible. The scheme has a smaller communication overhead compared to existing schemes. **Keywords: private set intersection; homomorphic encryption; multi-key TFHE; cloud computing;** privacy protection **MSC: 68U99; 68U99; 68T09; 68Q06** **1. Introduction** With the rapid growth of data in the Internet era, the demand for data storage and computing capacity in various fields far exceeds the capacity of their own devices. To solve this problem, cloud computing has been proposed. Cloud computing is generally defined as an internet-based computing method. In this way, the shared software and hardware information and resources can be provided to various terminals and other devices of the computer as required. Cloud computing technology can transmit various information to the Internet and store and calculate data, and users can view the calculation results and data information. However, current security issues in the context of cloud computing are more prominent [1]. Data security issues in cloud computing mainly include storage data security, computing data security, and transmission data security. When users store data on the cloud server, the cloud server will obtain the users’ data first, but the abnormal use of malicious users can also cause a risk of data leakage. In the process of cloud server computing, the cloud server will know the calculation results and additional data. This information that should only be known by users also has a risk of leakage. In addition, ----- _Mathematics 2023, 11, 1784_ 2 of 20 data theft can easily occur during data transmission, and user data can show problems of theft and tampering [2]. Private set intersection (PSI), as an interactive encryption protocol, calculates the intersection of two data owners’ data and returns it to one of them. We generally refer to the party receiving the data as the receiver and the party receiving nothing as the sender. It is important and necessary to protect the privacy of the set in computing, especially when the information in the set is important private information such as the customer transaction information of a bank or the address book of a user. With the concerted efforts of many researchers, PSI technology has developed rapidly, and more and more efficient solutions have been proposed [3–15]. After several years of development, PSI technology has been applied to the fields of internet of vehicles [16], profile matching [17], and private contact search [18]. In the current situation where the data volume is large and scattered in the hands of different participants, PSI technology can well balance the relationship between privacy and information sharing. Leveraging the storage and computing power of cloud servers allows PSI protocols to compute larger datasets, but current cloud-assisted PSI schemes suffer from information leakage [19] or large communication overhead [20]. Fully homomorphic encryption (FHE) refers to the computation of data that has been homomorphically encrypted, and the computed decryption result is the same as that obtained by the same computation for unencrypted data. The concept of FHE has been proposed as early as the late 1970s, but it has only started to develop rapidly in the last two decades. The development of fully homomorphic encryption is generally divided into three stages. In 2009, the first generation of fully homomorphic encryption started to develop, and Gentry constructed the first fully homomorphic encryption scheme [21]. The scheme first constructs a somewhat homomorphic encryption (SHE) scheme that can homomorphically compute circuits of a certain depth, then compresses and decrypts the circuits and performs bootstrapping operations in an orderly manner, and finally obtains a scheme that can homomorphically compute arbitrary circuits. The second generation of fully homomorphic encryption schemes arose in 2011 when Brakerski and Vaikuntanathan implemented FHE for the first time under the LWE assumption using linearization and modulo conversion [22] and implemented FHE under the RLWE assumption [23]. These schemes do not require compression and decryption circuits, and the security and efficiency are greatly improved. In 2013, the third generation of fully homomorphic encryption schemes was born, and Gentry et al. for the first time designed a fully homomorphic encryption scheme, Gentry–Sahai–Waters (GSW), that does not require the computation of a key using the approximate eigenvector technique [24]. There are two broad categories of fully homomorphic algorithms, the BGV [25] scheme proposed by Professor Brakerski of Stanford University, Research Fellow Gentry of IBM, and Professor Vaikuntanathan of the University of Toronto, and the GSW [24] scheme proposed by Gentry of IBM, Sahai of the University of California and Waters of the University of Austin. Fully homomorphic encryption over toru (TFHE) [26] is an improvement of the GSW scheme with higher efficiency. TFHE can accomplish fast comparisons, supports arbitrary boolean circuits, and allows fast bootstrapping to reduce the noise due to ciphertext computation. In previous studies, the BGV scheme has been used to focus on the unbalanced privacy aggregation scenario [27–29]. Unlike previous works, this paper uses the TFHE encryption scheme for the first time to implement privacy-seeking protocol based on cloud computing. At a high level, our contributions can be summarized as follows: - We have designed a series of security sub-protocols for the MKTFHE cryptosystem, including some basic circuit gate operations and security comparison protocols. - We have built a cloud-assisted multi-key private set intersection (CMPSI) system based on a single cloud server. Our system can prevent collusion attacks between servers and participants. - We strictly prove the security of the proposed CMPSI system under the semi-honest model. ----- _Mathematics 2023, 11, 1784_ 3 of 20 - We have conducted extensive experimental evaluation on the performance of the scheme, which proves that our scheme has greatly reduced the communication cost of the participants. The rest of the paper is organized as follows. In Section 2, we describe the related work of private set intersection. In Section 3, we provide the preliminaries. Section 4 details the system model, threat model, and design goals. Section 5 elaborates on the cryptographic protocol for the private set intersection. Section 6 analyzes the security of our proposed protocols. Section 7 conducts a series of experimental comparisons. Finally, Section 8 concludes this paper. **2. Related Work** PSI was first proposed by Freedman et al. [30], who transformed the element comparison problem into the polynomial root problem and realized PSI through multiplicative homomorphic encryption. However, when the polynomial order is large, it will lead to a costly exponential computation of the homomorphic encryption. In recent years, many researchers have intensively studied the PSI problem, and many PSI protocols with high efficiency and low communication overhead have emerged. PSI computing protocols are mainly divided into two categories according to whether there is a third party, namely, the traditional PSI computing protocol based on public key encryption, obfuscation circuit [31–33] and inadvertent transmission [34] technology and the cloudassisted PSI computing protocol that uses cloud servers to complete computing. Traditional PSI computing protocols rely on a series of basic cryptography technologies for computing. These basic cryptography technologies are mainly divided into PSI based on public key encryption mechanism, PSI based on obfuscation circuit, and PSI based on inadvertent transmission. The PSI calculation protocol proposed by Freedman et al. [34] is based on the public key encryption mechanism. This scheme represents the elements in the set as the roots of polynomials and uses polynomials to calculate the intersection. However, the cost of calculation will become large with the increase in the order of polynomials. Hazay et al. also improved the article [30] and adopted the bit commitment protocol to prevent the scenario of inconsistent input data on the server [35], so that the PSI protocol can be applied to the protocol of malicious adversaries. In 2012, Huang et al. first proposed PSI computing protocols based on obfuscated circuits [36], which are Bitwise-AND (BWA), Pairwise-Comparisons (PWC), and Sort-Compare-Suffle (SCS) protocols. In 2013, the PSI protocol proposed by Dong et al. [37] used OT technology for the first time. The author used OT technology to ensure the security of the protocol. Pinkas et al. [38] proposed a new PSI protocol based on Hash and random OT protocols and optimized the SCS protocol in [36]. The computational efficiency of the protocol was greatly improved, and the complexity of the algorithm was also reduced. Based on the article [34], Freedman et al. further optimized and improved their scheme in 2014 [39]. Specifically, the scheme uses different hash functions for the client and server when mapping the set elements. In 2018, Pinkas et al. realized PSI based on unintentional pseudorandom function [40] through the circuit. In 2020, Pinkas et al. [12] constructed a PSI protocol with malicious security based on the protocols [41] in the literature. The traditional PSI does not need the assistance of a third party, but in the application, the participants are generally resource-constrained users, who are insufficient in providing sufficient data storage and computing power. With the development of cloud computing, the PSI protocol based on cloud servers began to develop. The cloud-assisted PSI scheme provides a new optimization method for the existing PSI scheme by the excellent storage and computing capabilities of the cloud server. The cloud-assisted PSI uses the third-party cloud computing framework to complete the calculation and uses the storage and computing resources of the cloud server to enable the protocol to calculate large-scale datasets. Kerschbaum [42] implemented the anti-collusion outsourcing PSI protocol through two single functions, but the method has the risk of brute force cracking. Then, Kerschbaum [43] proposed another kind of cloudassisted PSI using bloom filter and homomorphic encryption. Liu et al. [19] proposed a ----- _Mathematics 2023, 11, 1784_ 4 of 20 relatively simple PSI protocol, but it can disclose the cardinality of set intersection. Abadi et al. [44] implemented the PSI protocol using homomorphic encryption and polynomial interpolation in 2015. This protocol outsources the collection of clients to a third-party server to perform infinite PSI operations. Based on this work, a verifiable cloud outsourcing PSI protocol [45] is proposed to ensure the privacy and integrity of data. Ali et al. [46] proposed an attribute-based private set intersection scheme. The cloud server can calculate the corresponding access rights of the participants. The PSI protocol based on the cloud server can use the computing and storage capabilities of the cloud server, but it has produced the privacy disclosure problem of data outsourcing, and the excessive cost of users in the operation of the protocol is another problem that needs to be solved. Table 1 shows the comparison between our scheme and the existing scheme. **Table 1. Comparison with existing schemes.** **CMPSI** [46] [47] [48] [19] [42] [43] The year 2023 2020 2019 2014 2014 2012 2012 Private against the CSP 52 52 52 52 56 52 52 PSI computation authorization 52 52 52 52 56 52 52 Supports multiple user queries 52 52 52 52 56 56 56 Participants can go offline after uploading data 52 52 56 56 56 56 56 CSP can collude with participants 52 56 56 56 56 56 56 **3. Preliminaries** In this section, we first introduce the concept of private set intersection and have an example to better understand the concept. Then, we introduce the cryptosystem MKTFHE used in our system and present the algorithm as an example of a NAND gate. Table 2 lists some of the symbols used in this paper. **Table 2. Notation used.** **Notations** **Definition** _λ_ Security parameter Z Integer set T (R)LWE over the real torus _si_ Private key of participant i (PKi, BKi, KSi) Public key set of participant i [[x]]si Encrypted data x under si **MKHENAND** NAND gate in multi-key TFHE **CMPSI** Cloud-assisted multi-party private set intersection _3.1. Private Set Intersection_ PSI allows two parties holding sets to compare encrypted versions of these sets to compute the intersection. Let the two parties holding the sets be sender X and receiver Y. The sender and receiver hold datasets of size Nx and Ny respectively, each with a number of bits σ. In a basic PSI protocol, receiver Y encrypts its own dataset and sends it to sender _X. For each of Y’s data, sender X calculates the homomorphic product of the difference_ with all of its own terms and sends the result to receiver Y. Y decrypts the result of X’s calculation and obtains the final intersection information. The result of the calculation is sent to the receiver Y. Y decrypts the result of X’s computation and obtains the final intersection information. The basic PSI protocol is shown in Figure 1. ----- _Mathematics 2023, 11, 1784_ 5 of 20 **Figure 1. Basic PSI protocol.** In the scheme of this paper, the storage of data and the computation are performed on the cloud server. We construct a new PSI scheme using fully homomorphic encryption. Both the sender and the receiver encrypt the data locally and then send it to the cloud server. Suppose that the sender has encrypted data a1, . . ., aNX and the receiver has encrypted data b1, . . ., bNY . Both parties send their encrypted data to the cloud server. On the cloud server, for each data bi of the receiver, ci = ∏0<j≤Nx �bi − _aj�_ is computed. ci is a Boolean value that represents whether the data bi of the receiver are in the sender X or not. Figure 2 shows the handshake model of this scheme. **Figure 2. Handshake model.** _3.2. MKTFHE Cryptosystem_ Homomorphic encryption is the computation of the encrypted data to obtain the encrypted computational result, and the result of the decryption of the obtained encryption result is the same as the result obtained by performing the same operations on the unencrypted plaintext. Fully homomorphic encryption [24,25] is a homomorphic encryption that can satisfy both additive and multiplicative operations. Fully homomorphic encryption over the toru (TFHE) [26] is a type of fully homomorphic encryption that can accomplish fast comparisons and support operations on arbitrary Boolean circuits.TFHE differs from other FHE schemes in that it can be fast bootstrapping to reduce noise during ciphertext operations. In this paper, we use multi-key TFHE [49] to meet the needs of our system. MKTFHE is a multi-key version of TFHE that can compute Boolean circuits on ciphertexts encrypted under different keys, and then performs bootstrapped to refresh the noise as ----- _Mathematics 2023, 11, 1784_ 6 of 20 each binary gate is computed. However, the MKTFHE library only implements multi-key homomorphic NAND gates, which cannot meet the needs of our system. The following describes the five components of MKTFHE and gives an example of the homomorphic computation process with a multi-key homomorphic NAND gate. 1. **Setup(1[λ]): Takes as input the security parameter λ and returns the public parameter** **pp[MKTFHE].** (a) Run LWE.Setup(1[λ]) to generate the LWE parameter pp[LWE] = (n, χ, α, B[′], d[′]). In the LWE parameters, n is the dimension of the LWE secret, χ is the key distribution of the LWE secret, α is the error rate, B[′] is the decomposition basis, and d[′] is the dimension of the key transformation gadget vector. We use the � key-switching gadget vector g[′] = _B[′−][1], . . ., B[′−][d][′]_ [�]. (b) Run RLWE.Setup(1[λ]) to generate the RLWE parameter pp[RLWE] = (N, ψ, B, d, a). We define N as the dimension of RLWE secret (a power of 2), ψ as the distribution of RLWE secret over R and with error rate α, B ≥ 2 as an integer base, � decomposition dimension d, and gadget vector g = _B[−][1], . . ., B[−][d][�]. a is a_ uniformly distributed sample over distribution T[d]. (c) Returns the generated public parameter pp[MKTFHE] = �pp[LWE], pp[RLWE][�]. 2. **KeyGen(pp[MKTFHE]): Each participant generates its keys independently. Take the** public parameter pp[MKTFHE] as input and return the key si and the public key set (PKi, BKi, KSi). (a) Generate the LWE secret si ← **LWE.KeyGen(). This step is only for sampling** the key from distribution χ. (b) Run (zi, bi) ← **RLWE.KeyGen(), and set the public key to PKi = bi. Sample** _z from distribution ψ, and then, set z = (1, z). Take an error vector e from_ _Dα[d]_ [and calculate the public key][ b][ =][ −][z][ ·][ a][ +][ e][(][mod][1][)][. For][ z]i [=][ z]1,0 [+][ z]i,1[X][ +] . . . + zi,N−1X[N][−][1], note zi[∗] [= (][z][i][,0][,][ −][z][i][,][N][−][1][, . . .,][ z][i][,1][)][ ∈] [Z][N][.] (c) For j ∈ [n], generate �di,j, Fi,j� _←_ **RLWE.UniEnc�si,j, zi�, this step is to encrypt** the LWE secret using the RLWE secret. In addition, set the bootstrap key to **BKi =** ��di,j, Fi,j��j∈[n][. Taking a random value][ r][ from][ ψ][, one can think of][ d] as the LWE key s under the encryption of the random value r and F as the random value r under the encryption of the RLWE key z. (d) Generate a key conversion key KS ← **LWE.KSGen�zi[∗][,][ s][i]�, capable of convert-** ing an LWE ciphertext corresponding to t ∈ Z[N] into another LWE ciphertext for the same message under s ∈ Z[N] encryption. (e) Returns key si, a triple (PKi, BKi, KSi) of public keys, public key, bootstrap key and key transformation key, respectively. 3. **Enc(m): The data m to be encrypted are taken as input, and return TLWE ciphertext** [[m]] = (b, a) ∈ T[n][+][1] satisfies b + ⟨a, s⟩≈ [1]4 _[m][(][mod1][)][.]_ (a) Using standard LWE encryption, uniformly sample from T[n] to obtain a as the mask and sample from Dα to obtain e as the error. (b) Output ciphertext [[m]] = (b, a) ∈ T[n][+][1], where b + ⟨a, s⟩≈ [1]4 _[m][(][mod1][)][.]_ 4. **Dec([[m]], {si}i∈[k]): Takes as input the TLWE ciphertext [[m]] = (b, a1, . . ., ak) ∈** T[kn][+][1] with a set of keys (s1, . . ., sk) and returns the decrypted message m which minimizes _| b + ∑i[k]=1[⟨][a][i][,][ s][i][⟩−]_ [1]4 _[m][ |][.]_ (a) Input [[m]] = (b, a1, . . ., ak) ∈ T[kn][+][1] with a set of keys (s1, . . ., sk). (b) Returns the bit m ∈{0, 1} that minimizes | b + ∑i[k]=1[⟨][a][i][,][ s][i][⟩−] [1]4 _[m][ |][.]_ 5. **NAND([[m1]], [[m2]], {(PKi, BKi, KSi)}i∈[k]): Takes two TLWE ciphertexts and the** public key as input. Expand [[m1]] ∈ T[k][1][n][+][1] and [[m2]] ∈ T[k][1][n][+][1] to [[m1[′] []]][,][ [[][m]2[′] []]][ ∈] [T][kn][+][1] and evaluate the gate homomorphically on encrypted bits. Then the algorithm evalu ----- _Mathematics 2023, 11, 1784_ 7 of 20 ates the decryption circuit of the TLWE ciphertext and execute the multi-key switching algorithm. Finally, returning the TLWE ciphertext of the same message under joint key encryption. (a) Given two ciphertexts [[m1]] ∈ T[k][1][n][+][1] and [[m2]] ∈ T[k][1][n][+][1], let k be the number of participants, associated with either [[m1]] or [[m2]]. For a public key set, PKi = bi represents the public key, BKi = ��di,j, Fi,j��j∈[n] [represents the bootstrap key,] and KSi represents the key transformation key of the j-th participant. Expand ciphertext [[m1]] and [[m2]] to [[m1]][′], [[m2]][′] _∈_ T[kn][+][1], i.e., the same message under joint key s = (s1, . . ., sk) ∈ Z[kn] encryption. The process of expansion is the process of rearrangement, and 0 is put into the empty slot. Using the expanded ciphertext to perform the calculations. Only the calculation of NAND gate is supported in the document. (b) Use the Mux gate to implement the main calculation, for i ∈ [k], let ˜ai = �a˜i,j�j∈[n][.] For _i_ _∈_ [k] and _j_ _∈_ [n], recursively compute [[c]] _←_ [[c]] + RLWE.Prod�[[c]] · X[a][˜][ij] _−_ [[c]], �di,j, Fi,j�, {bl}l∈[k]�, where **RLWE.Prod�[[c]], (di, Fi),** �bj�j∈[k]� is a hybrid product algorithm that multi plies a single encrypted ciphertext (di, Fi) by a multi-key RLWE ciphertext [[c]]. (c) For [[c]] = (c0, c1, . . ., ck) ∈ _T[k][+][1], let b[∗]_ be a constant term of c0 and for i ∈ [k], let ai[∗] [be a vector of coefficients of][ c][i][. Compute the LWE ciphertext][ [[][m][]]][∗] [=] �b[∗], a1[∗][, . . .,][ a][∗]k � _∈_ Tkn+1. Finally a multi-key key conversion algorithm is exe � � cuted and returns the ciphertext [[m]][′′] _←_ **LWE.MKSwitch** [[m]][∗], {KSi}i∈[k], � � where LWE.MKSwitch [[m]][∗], {KSi}i∈[k] inputs the expanded ciphertext and a series of key conversion keys, returning the ciphertext of the same message under joint key encryption. **4. System Model and Design Goal** _4.1. Problem Formulation_ Suppose the receiver Y has a dataset TY, and Y wants to know their intersection with other data owners but does not want to expose more information. The data owners encrypt their datasets separately and send them to the cloud server. The cloud server can store this encrypted information but cannot decrypt it. Data receiver Y encrypts its data and uploads it to the cloud server, which executes privacy intersection and obtains the intersection information of dataset TY with other datasets. The cloud server computes and returns the cryptographic result to receiver Y. Y decrypts the intersection result and obtains the intersection information. Note that each data owner including the data receiver has their separate key to encrypt the data. _4.2. System Model_ In Section 3.1, we mention the flow of the basic PSI protocol, in which the sender interacts directly with the receiver for information. Unlike the basic PSI protocol, our system consists of four entities, which are Parameter Generation Center (PGC), Cloud Server (CS), Data Receiver (DR), and Data Owners (DOs). DO owns its own dataset and is able to let other participants obtain information about the intersection of the dataset but does not want to expose more information. DR wants to query the intersection of its own dataset with the dataset of other participants and does not want to expose more information. Specifically, PGC is responsible for generating public parameters in the system and sending them to other entities. CS can store a large amount of data and has excellent computing resources. DR needs to query the intersection. DOs provide their encrypted data to CS. Note that in our system, the data owners can be multiple participants. The general model of our private set intersection system is shown in Figure 3. ----- _Mathematics 2023, 11, 1784_ 8 of 20 **Figure 3. System model.** 1. PGC: PGC generates public parameters for our system and sends them to each entity involved in the computation (See 1 ). _⃝_ 2. CS: CS has huge storage resources to store the encrypted data of the participating parties. At the same time, CS has large enough computing power to satisfy the intersection of the datasets of the participating parties. 3. DR: DR generates its own private key and public key set using public parameters, encrypts its own data using the private key and sends it to CS (See 3 ), and receives _⃝_ the computation results sent by CS (See 4 ). _⃝_ 4. DOs: Each DO generates its own private key and public key set using public parameters, encrypts its own dataset using the private key, and sends it to CS (See 2 ). _⃝_ Please note that in our system, the participants do not need to be online all the time. Since CS can store the encrypted data, the DOs can go offline after they send their encrypted data to CS. Similarly, DR can be offline after sending data until CS returns the calculation results. In our scheme, DO can be used as DR for frequent item set queries, and the DR can query the intersection information with multiple DOs to achieve multi-user query. _4.3. Threat Model_ In our system model, the participating entities are curious but honest individuals. Curious means that the server and the participants try to use existing resources and data to obtain the data of other participants and are curious about the data of other entities; honest means that the server and the participants do not falsify the experimental data and follow the developed protocols to complete the computation. A is the active adversary we introduce to obtain the real data from other entities. Specifically, desires to obtain the _A_ real data of DOs and DR. We assume that adversary has the following capabilities. _A_ 1. can obtain all the data that passes through the public channel. _A_ 2. may collude with CS. Try to obtain the original values of the encrypted data _A_ uploaded by DOs and DR. 3. may be a DR used to obtain its dataset information, the cryptographic query results _A_ returned by the CS, and the encryption and decryption capabilities of the DR. 4. may be a DO used to obtain its dataset information and encryption and decryption _A_ capabilities. ----- _Mathematics 2023, 11, 1784_ 9 of 20 Note that in our threat model, the attacking adversary can be a DR. Since the joint _A_ key of multiple participants must be used in decryption to decrypt the computed result of CS, the final decryption result is also not available when A has only the key of DR. Unlike existing schemes when the attacking adversary A is a CSP, A can collude with DR or DO. In our scheme, decryption requires the keys of all participants to perform; thus, CSP colluding with some DR or DO still cannot decrypt the computation results. _4.4. Design Goal_ According to the system model and threat model proposed above, the design objectives of this paper are as follows. 1. Data privacy: the original data of DR and the query intersection result as well as the original dataset of DOs cannot be revealed to adversary . _A_ 2. Calculation accuracy: The accuracy of the calculation results of the system cannot be reduced compared with other methods. 3. Low overhead: The time and upload overhead of the calculation cannot be too large compared with other methods. 4. Offline participant: The participant should be able to go offline after encrypting the data and uploading it to ensure the scalability of the system. **5. Cloud-Assisted Multi-Party Private Set Intersection** In this section, we first introduce the initialization of the system. Then, we design the secure computing sub-protocol based on MKTFHE. Finally, we describe our private set intersection scheme. _5.1. System Initialization_ Our system can satisfy the DR to query the information of its intersection with multiple participants, and we assume that there is a DR and n DOs. First, PGC generates public parameters for each participant and the cloud server and sends the public parameters to CS, DR, and n DOs. Then, each entity that receives the public parameters generates its own public key set (PK, BK, KS) and private key s based on the public parameters. _5.2. Security Protocol Design_ In this paper, four secure computation protocols are proposed to help complete the privacy-seeking intersection, which is a secure AND gate computation protocol (SCAND), secure OR gate computation protocol (SCOR), secure XNOR computation protocol (SCXNOR), and secure comparison protocol (SCP). 5.2.1. Secure AND Gate Computation Protocol We implement the AND operation between two MKLwe samples. We implement the addition between multi-key Lwe samples (MKlweAddTo) to implement this secure computation protocol. Suppose CS has two MKLwe samples ca and cb: initialize an intermediate sample temp, add ca and cb using MKlweAddTo twice, and finally return the result to res (Algorithm 1). **Algorithm 1 Secure AND gate computation protocol (SCAND).** **Input: MKLwe Sample ca, cb.** **Output: MKLwe Sample res.** 1: CS initializes temp using the public parameter pp to hold the intermediate variable LWE sample. 2: AndConst = modSwitchToTorus32( 1, 8) _−_ 3: temp **MKlweNoiselessTrivial(AndConst, pp)** _←_ 4: temp **MKlweAddTo(temp + ca)** _←_ 5: res **MKlweAddTo(temp + cb)** _←_ ----- _Mathematics 2023, 11, 1784_ 10 of 20 5.2.2. Secure OR Gate Computation Protocol We implement the OR operation between two MKLwe samples. As with SCAND above, we use the addition MKlweAddTo between multi-key Lwe samples to implement this secure computation protocol. Suppose CS has two MKLwe samples ca and cb, initialize an intermediate sample temp, add ca and cb using MKlweAddTo twice respectively, and finally, return the result to res to obtain the result of the OR gate operation between ca and _cb (Algorithm 2)._ **Algorithm 2 Secure OR gate computation protocol (SCOR).** **Input: MKLwe Sample ca, cb.** **Output: MKLwe Sample res.** 1: CS initializes temp using the public parameter pp to hold the intermediate variable LWE sample. 2: ORConst = modSwitchToTorus32(1, 8) 3: temp **MKlweNoiselessTrivial(ORConst, pp)** _←_ 4: temp **MKlweAddTo(temp + ca)** _←_ 5: res **MKlweAddTo(temp + cb)** _←_ 5.2.3. Secure XNOR Gate Computation Protocol We implement the XNOR operation between two MKLwe samples. We implement this secure computation protocol using the addition and multiplication of multi-key Lwe samples MKlweAddMulTo. Suppose CS has two MKLwe samples ca and cb: initialize an intermediate sample temp, add 2 ∗ _ca and 2 ∗_ _cb using MKlweAddMulTo twice, return the_ result to temp to obtain the XOR gate operation result of ca and cb, and use the multi-key homomorphic NOT gate SCNOT once to obtain the XNOR gate operation result. Note that in the cryptographic scheme we use, MKTFHE, the computation of the NOT gate does not require bootstrapping operations; thus, the computation overhead is very small (Algorithm 3). **Algorithm 3 Secure XNOR gate computation protocol (SCXNOR).** **Input: MKLwe Sample ca, cb.** **Output: MKLwe Sample res.** 1: CS initializes temp using the public parameter pp to hold the intermediate variable LWE sample. 2: XNORConst = modSwitchToTorus32(1, 8) 3: temp **MKlweNoiselessTrivial(XNORConst, pp)** _←_ 4: temp **MKlweAddMulTo(temp + 2** _ca)_ _←_ _∗_ 5: temp **MKlweAddMulTo(temp + 2** _cb)_ _←_ _∗_ 6: res ← **SCNOT(temp)** 5.2.4. Secure Comparison Protocol **SCP is important in our protocol and is used to determine whether the two input** ciphertext vectors are equal or not. Suppose DR has its own encrypted data [[x]]sDR = ([[x1]]sDR, . . ., [[xn]]sDR ) sent to CS and DO has its own encrypted data [[y]]sDO = ([[y1]]sDO, . . ., [[yn]]sDO ) also sent to CS, where sDR and sDO are the private keys of DR and DO, respectively. For each of [[x]]sDR = ([[x1]]sDR, . . ., [[xn]]sDR ) and [[y]]sDO = ([[y1]]sDO, . . ., [[yn]]sDO ), the protocol performs SCXNOR and SCAND protocols to finally obtain a ciphertext with a Boolean value (Algorithm 4). ----- _Mathematics 2023, 11, 1784_ 11 of 20 **Algorithm 4 Secure Comparison Protocol (SCP).** **Input: Encrypted** data vectors [[x]]sDR = ([[x1]]sDR, . . ., [[xn]]sDR ),[[y]]sDO = ([[y1]]sDO, . . ., [[yn]]sDO ). **Output: Encrypted Boolean values [[z]]si** . 1: CS initializes the intermediate data vector [[v]] = ([[v1]], . . ., [[vn]]) using the public parameter pp. 2: for k = 0 to n 1 do _−_ 3: [[vk]]si [[xk]]sDR XNOR[[yk]]sDO _←_ 4: [[z]]si [[vk]]si AND[[z]]si _←_ 5: end for _5.3. Private Set Intersection_ CMPSI is performed by CS, DR, and DOs working together. Now DR wants to obtain the intersection information of their dataset and DOs dataset. First DOs encrypt their dataset with their own private key _sDO,_ send the encrypted dataset � � _ADO =_ [[a1]]sDO, [[a2]]sDO, . . ., [[am]]sDO with the public key set (PKsDO, BKsDO, KSsDO ) to CS, and then they can go offline. DR encrypts the dataset with its own private key � � _sDR and then sends the encrypted dataset BDR =_ [[b1]]sDR, [[b2]]sDR, . . ., [[an]]sDR with its public key set (PKsDR, BKsDR, KSsDR ) to CS, and then, it can be offline until CS completes the calculation. CS receives the encrypted dataset sent by DOs and DR, saves the data, and performs the secure computation in a secure environment. Finally, DR receives the encryption result calculated by CS and decrypts it using the joint key to obtain the intersec � � tion. Let there be m items in the encrypted dataset ADO = [[a1]]sDO, [[a2]]sDO, . . ., [[am]]sDO of DOs with k Boolean values in each item, and n items in the encrypted dataset BDR = � � [[b1]]sDR, [[b2]]sDR, . . ., [[bn]]sDR of DR with k Boolean values in each item. S1(DOs): Each DO encrypts its dataset using its own key sDO generated by the public parameter pp issued by PGC and sends it to CS. CS stores the encrypted dataset of all � � DOs, and for item i of dataset ADO = [[a1]]sDO, [[a2]]sDO, . . ., [[am]]sDO, we have [[ai]]sDO = ([[a1]]sDO, . . ., [[ak]]sDO ). S2(DR): DR uses the public parameter pp to generate its own key sDR to encrypt its dataset and sends it to CS. CS uses DR’s encrypted database for secure computation and has � � [[bj]]sDR = ([[b1]]sDR, . . ., [[bk]]sDR ) for item j of dataset BDR = [[b1]]sDR, [[b2]]sDR, . . ., [[bn]]sDR . � � S3(CS): CS receives the encrypted data message BDR = [[b1]]sDR, [[b2]]sDR, . . ., [[bn]]sDR � � from DR and the encrypted data message ADO = [[a1]]sDO, [[a2]]sDO, . . ., [[am]]sDO from DO. For j ∈{1, 2, . . ., n} and i ∈{1, 2, . . ., m}, each item [[bj]]sDR = ([[b1]]sDR, . . ., [[bk]]sDR ) in � � _BDR_ = [[b1]]sDR, [[b2]]sDR, . . ., [[bn]]sDR performs **SCP** with each item � � [[ai]]sDO = ([[a1]]sDO, . . ., [[ak]]sDO ) in ADO = [[a1]]sDO, [[a2]]sDO, . . ., [[am]]sDO, i.e., **SCP([[ai]]sDO**, [[bj]]sDR ). The result [[gi]]s = ([[g1]]s, . . ., [[gk]]s) is obtained as the result of whether the current [[bj]]sDR = ([[b1]]sDR, . . ., [[bk]]sDR ) is the same as each item in ADO = � � [[a1]]sDO, [[a2]]sDO, . . ., [[am]]sDO . S4(CS): For each computed [[gi]]s = ([[g1]]s, . . ., [[gk]]s), CS runs SCOR to obtain [[cj]]s = ([[c1]]s, . . ., [[ck]]s). [[cj]]s = ([[c1]]s, . . ., [[ck]]s) is a cryptographic Boolean value indicating � � whether each item in _BDR_ = [[b1]]sDR, [[b2]]sDR, . . ., [[bn]]sDR exists in � � _ADO =_ [[a1]]sDO, [[a2]]sDO, . . ., [[am]]sDO . A value of 1 means it exists and 0 means it does not. S5 (CS): For j ∈{1, 2, . . ., n}, execute S4, and send the calculated result C = {[[c1]]s, [[c2]]s, . . ., [[cn]]s} to DR. ----- _Mathematics 2023, 11, 1784_ 12 of 20 S6 (DR): Receive the calculation result from C = {[[c1]]s, [[c2]]s, . . ., [[cn]]s} sent by CS and decrypt it using the joint key to obtain the result. Please note that in our PSI scheme, the dense state computation is performed by FHE cryptography. All the calculations are performed on the cloud server, and the data on the cloud server are all cryptographic data, so that the privacy of the participants is protected. During the calculation process, the DR does not obtain any information other than its own information and the query result. The DOs do not obtain any information other than their own information and do not expose their information to other participants. The result of the CS calculation is in cryptographic form and cannot be decrypted by the participants except by the DR, which protects the privacy of the calculation result. **6. Security Analysis** In this section, we prove that our scheme is secure under a semi-honest model. We will prove the security of the MKTFHE cryptosystem, SCAND, SCOR, SCXOR, SCP and PSI schemes separately. We first present the security of the semi-honest model below. **Definition 1 (Security of the semi-honest model). According to protocol π, let ai be the input** _of participant Pi and bi be the output of Pi. REALi[Π][(][π][)][ is the viewpoint of][ P][i][ when protocol][ π][ is]_ _actually executed. IDEALi[Π][(][π][)][ is the viewpoint of][ P][i][, simulated by][ a][i][ and][ b][i][, executed in the ideal]_ _world of protocol π. If REALi[Π][(][π][)][ is computationally indistinguishable from][ IDEAL]i[Π][(][π][)][, then]_ _protocol π is secure in the semi-fair model [50]._ Note that in our protocols, the execution image usually consists of the exchanged data and the information that can be computed from these data. It follows from Definition (1) that when proving the security of these protocols, the image we simulate should be indistinguishable from the actual execution image when we compute it. _6.1. Security of MKTFHE Cryptosystem_ Privacy of LWE Assumption: The j-th component Kj of a key-switching key KS = _{Kj}j∈[N] from t ∈_ Z[N] to s ∈ Z[N] is generated by adding tj · g[′] to the first column of the T[d][′][×][(][n][+][1][)] matrix, the rows of which are instances of LWE under the secret s. Therefore, KS **LWE.KSGen(t, s) is computationally indistinguishable from a uniform distribu-** _←_ tion over (T[d][′][×][(][n][+][1][)])[N] where LWE assumes a parameter of (n, χ, β) and s is sampled according to χ. Privacy of RLWE Assumption: Under the assumption that the parameter is (N, ψ, α), a uniform distribution over T[d][×][5] is computationally indistinguishable from the distribution _D0 = {(a, b, d, F) : pp[RLWE]_ _←_ **RLWE · Setup�1[λ][�], (z, b) ←** **RLWE.KeyGen(), (d, F) ←** **RLWE · UniEnc(µ, z)} for any µ ∈** _R. We consider the following distribution: First, we_ transform F = [f0 | f1] and (b, a) into independent uniform distributions of T[d][×][2] using the RLWE assumption of a secret z. Therefore, D0 is indistinguishable from D1 in terms of � calculation. _D1_ = _{(a, b, d, F)_ : **a, b** _←_ _U_ _T[d][�],_ � � **F ←** _U_ _T[d][×][2][�], r_ _←_ _ψ, e1 ←_ _Dα[d][,][ d][ =][ r][ ·][ a][ +][ µ][ ·][ g][ +][ e]1_ (mod1) .Then, d is made uniformly distributed using the RLWE assumption with a secret of r. Therefore, D1 is indistinguish � � � able from the distribution D2. D2 = (a, b, d, F) : a, b, d ← _U_ _T[d][�], F ←_ _U_ _T[d][×][2][��]. Since_ _D2 is independent from µ, our RLWE scheme is semantically private._ In summary, under the (R)LWE assumption, our cryptosystem is semantically private; thus, we can appropriately choose parameters pp[LWE] and pp[RLWE] to achieve a security level of at least λ-bit. _6.2. Security of Secure Computing Protocols_ In this section, we demonstrate the security of our secure computing subprotocols, including SCAND, SCOR, SCXOR and SCP. ----- _Mathematics 2023, 11, 1784_ 13 of 20 **Theorem 1. The SCAND proposed is secure under the semi-honest model.** **Proof of Theorem 1. We use REALCS[Π]** [(][SC][AND][)][ to denote the execution view in the real world] of the of CS, where it is specified as REALCS[Π] [(][SC][AND][)] = [[ca]], [[cb]], [[AndConst]], [[temp]], [[res]] . [[AndConst]] is obtained from [[ 1]] and [[8]] by _{_ _}_ _−_ _modSwitchToTorus32. [[temp]] is obtained from [[AndConst]] and [[ca]] by MKlweAddTo and_ **MKlweNoiselessTrival.** We assume that IDEALCS[Π] [(][SC][AND][)] = [[ca[′]]], [[cb[′]]], [[temp[′]]], [[res[′]]], [[AndConst[′]]] is the execution view of the simulation in the ideal _{_ _}_ world, where [[ca[′]]], [[cb[′]]], [[temp[′]]], [[res[′]]] and [[AndConst[′]]] are chosen randomly from T[n][+][1]. The semantic privacy of our encryption scheme makes [[ca]], [[cb]], [[temp]] and [[AndConst]] computationally indistinguishable from [[ca[′]]], [[cb[′]]], [[temp[′]]] and [[AndConst[′]]] respectively. In addition, [[res]] is computationally indistinguishable from [[temp[′]]] and [[AndConst[′]]] respectively. Thus, it can be concluded that REALCS[Π] [(][SC][AND][)][ and][ IDEAL]CS[Π] [(][SC][AND][)][ are com-] putationally indistinguishable. We can obtain that SCAND is secure under the semi-honest model. **Theorem 2. The SCOR proposed is secure under the semi-honest model.** **Proof of Theorem 2. We use REALCS[Π]** [(][SC][OR][)][ to denote the execution view in the real world] of CS, where it is specified as REALCS[Π] [(][SC][OR][) =][ {][[[][ca][]]][,][ [[][cb][]]][,][ [[][temp][]]][,][ [[][ORConst][]]][,][ [[][res][]]][}][.] [[ORConst]] is obtained from [[1]] and [[8]] by modSwitchToTorus32. [[temp]] is obtained from [[ca]] and [[ORConst]] by MKlweNoiselessTrivial and MKlweAddTo. [[res]] is obtained from [[temp]] and [[cb]] by MKlweAddTo. We assume that IDEALCS[Π] [(][SC][OR][) =] [[ca[′]]], [[cb[′]]], [[temp[′]]], [[ORConst[′]]], [[res[′]]] is the execution view of the simulation in the _{_ _}_ ideal world, where [[ca[′]]], [[cb[′]]], [[temp[′]]], [[ORConst[′]]] and [[res[′]]] are chosen randomly from T[n][+][1]. The semantic privacy of our encryption scheme makes [[ca]] and [[cb]] computationally indistinguishable from [[ca[′]]], [[cb[′]]], [[temp[′]]] and [[ORConst[′]]], respectively. In addition, [[res]] is computationally indistinguishable from [[ca[′]]], [[cb[′]]], [[temp[′]]] and [[ORConst[′]]] respectively. Thus, it can be concluded that REALCS[Π] [(][SC][OR][)][ and][ IDEAL]CS[Π] [(][SC][OR][)][ are computationally] indistinguishable. We can obtain that SCOR is secure under the semi-honest model. **Theorem 3. The SCXOR proposed is secure under the semi-honest model.** **Proof of Theorem 3. Since the design ideas of SCAND and SCOR are similar, we can prove** the theorem based on Theorem (1). **Theorem 4. The SCP proposed is secure under the semi-honest model.** **Proof of Theorem 4. We use REALCS[Π]** [(][SCP][)][ to denote the execution view in the real world] of the CS, where it is specified as REALCS[Π] [(][SCP][) =][ {][([[][x][]]][,][ [[][y][]])][,][ [[][z][]]][}][.][ [[][x][]]][ and][ [[][y][]]][ are the] encrypted data vectors. [[z]] is the result of determining whether the encrypted data vectors [[x]] and [[y]] are equal. [[z]] is a random number between 0 and 1 in the ciphertext. We assume that IDEALCS[Π] [(][SCP][) =][ {][([[][x][′][]]][,][ [[][y][′][]])][,][ [[][z][′][]]][}][ is the execution view of the simulation in] the ideal world, where the encrypted data in both [[x[′]]] and [[y[′]]] are chosen randomly from T[n][+][1]. [[z[′]]] are chosen randomly from T[n][+][1]. The semantic privacy of our encryption scheme makes [[x]] and [[y]] computationally indistinguishable from [[x[′]]] and [[y[′]]], respectively. In addition, [[z[′]]] takes 0 or 1 with equal probability. [[z]] are computationally indistinguishable from [[z[′]]], respectively. Thus, it can be concluded that REALCS[Π] [(][SCP][)][ and][ IDEAL]CS[Π] [(][SCP][)] are computationally indistinguishable. We can obtain that SCP is secure under the semihonest model. _6.3. Security of CMPSI_ **Theorem 5. The CMPSI proposed is secure under the semi-honest model, and the security of** _encrypted data, mining results, and query data can be guaranteed._ ----- _Mathematics 2023, 11, 1784_ 14 of 20 **Proof of Theorem 5. We can use the above method to prove that our proposed CMPSI is** secure under the semi-honest model. In S1, CS obtains the encrypted dataset from DOs. In S2, CS obtains the encrypted dataset from DR. From Section 6.1, our cryptosystem is semantically secure, and the semi-honest CS cannot distinguish these messages from the random values of T[n][+][1]. In S3, SCP is executed to obtain the intersection information of the encryption of individual items in the dataset. Since SCP is secure in our system, it can be confirmed that the protocol in S3 is secure. In S4, SCOR is used to obtain the final encryption result. Since SCOR is secure in our system, the protocol in S4 is secure. In S5 and S6, the execution of S4 is repeated, the DR receives the message and decrypts it using the joint key, and the protocol is secure from the security of MKTFHE. **Theorem 6. The CMPSI proposed is able to resist man-in-the-middle attacks.** **Proof of Theorem 6. As shown in Figure 4, the participants represent the DR and DOs in** our scenario. Under normal conditions, the participants can communicate with the CS, and Figure 4a shows the communication under normal conditions. The man-in-the-middle attack changes the original communication channel and can access the communication data between the participant and the cloud server, and Figure 4b shows the impact of the man-in-the-middle attack on the communication. We will prove that our model is resistant to man-in-the-middle attacks in three ways. First, DO encrypts its own dataset TDO into [[TDO]]sDO using its own key sDO and then sends [[TDO]]sDO to CS. Intermediary A obtains [[TDO]]sDO through the new channel, but A does not have DO’s key, and it is known from the security of MKTFHE that A cannot decrypt [[TDO]]sDO . Thus, our model can resist the man-in-the-middle attack during the data transmission from DO to CS. Second, DR wants to obtain the intersection information and sends the encrypted data [[TDR]]sDR to CS. Intermediary A obtains [[TDR]]sDR through the illegal channel. By the security of MKTFHE, _A does not have sDR and cannot obtain TDR from [[TDR]]sDR_ . Thus, our model can resist the man-in-the-middle attack from DR to CS man-in-the-middle attack during the data transfer. Finally, CS needs to return the computed intersection information [[TDR∩DO]]s to DR. The middleman A obtains the information [[TDR∩DO]]s, and it is known from the security of **MKTFHE that A does not have the key to obtain TDR∩DO. Thus, our model can resist the** man-in-the-middle attack during the data transmission from CS to DR. (a) (b) **Figure 4. Man-in-the-middle attack; (a) normal communication; (b) post-attack communication.** _6.4. Security Services_ According to the above proof of CMPSI security, Table 3 shows the security services provided by the scheme and a demonstration from our model of how the method provides each of these functions. ----- _Mathematics 2023, 11, 1784_ 15 of 20 **Table 3. Security services provided.** **Security Services** **Definition** **Proof** Confidentiality Network information is not dis- In our system, DO uses its own key sDO to encrypt its own dataset closed to non-authorized users, enti- _TDO into [[TDO]]sDO and then sends [[TDO]]sDO to CS. An unauthorized_ ties, or processes. user A illegally obtains [[TDO]]sDO, and according to the security of the MKTFHE cryptosystem in Section 6.1, it is known that without the key, sDO cannot perform decryption. Therefore, unauthorized illegal users A cannot obtain the information of DO’s dataset TDO. Integrity Information is transmitted, ex- In our system, DR encrypts the dataset TDR as [[TDR]]sDR using the changed, stored, and processed in key sDR and sends [[TDR]]sDR to CS. Attacker A obtains the dataset such a way that it remains uncor- [[TDR]]sDR through the intermediate channel, and according to the rupted or unmodified, that it is not definition of the semi-honest model in Section 4.3, A does not modify lost, and that it cannot be changed or corrupt the data, and CS can obtain the dataset [[TDR]]sDR intact. without authorization. Availability Assurance that information is avail- In our system, DR is the legal user. When DR wants to obtain the able to authorized users, i.e., assur- intersection information of its dataset TDR, DR sends [[TDR]]sDR to CS, ance that legitimate users can use the and CS sends the computed intersection result [[TDR∩DO]]s to DR. The required information when needed. legitimate user DR can obtain the required data when needed, which proves the usability of our system. Non-repudiation The two parties of information ex- In our system, DO sends its encrypted dataset [[TDO]]sDO to CS. Acchange cannot deny that they send cording to the definition of the semi-honest model in Section 4.3, or receive information in the ex- DO will not deny that the dataset [[TDO]]sDO is its data, proving the change process. non-repudiation of our system. **7. Performance Analysis** In this section, we evaluate the time overhead and communication overhead of our proposed scheme. The experimental parameters we used [51] are shown in Table 4 below. According to one study [52], the parameters we use reach a privacy level of at least 110 bits, which is a common reference in this field. **Table 4. Parameter sets.** **LWE-n** **LWE-α** **LWE-B[′]** **LWE- d[′]** **RLWE-N** **RLWE-β** **RLWE-B** **RLWE-d** 560 3.05 10[−][5] 2[2] 8 1024 3.72 10[−][9] 2[9] 3 _×_ _×_ The test environment used for our experiments was as follows: a 2.30 GHz Intel (R) Core(TM) i5-8300H Dell laptop. The programming language we used was C++, and our system was based on the MKTFHE library. First, we tested the efficiency of the security subprotocols separately. Then, we tested the communication overhead of our scheme and compared it with existing schemes. Finally, we tested our scheme. _7.1. Experiments on Security Computing Protocols_ Our secure subprotocol experiments were performed using the MKTFHE library [(https://github.com/ilachill/MK-TFHE) (1 February 2023). MKTFHE is a proof-of-concept](https://github.com/ilachill/MK-TFHE) implementation of a multi-key version of TFHE. The code is written on top of the TFHE [library (https://tfhe.github.io/tfhe/) (1 February 2023). The computation of secure NAND](https://tfhe.github.io/tfhe/) gates is given in the MKTFHE library. In the MKTFHE-based implementation, our goal is to implement the MKLwe sample addition and multiplication operations as a way to implement the other circuit gates needed in our scheme in addition to the NADN gate. We first performed experiments on single circuit gates, including experiments on secure AND gate computation protocol, secure OR gate computation protocol, and secure XNOR computation protocol, and the experimental results are shown in Table 5. We compared these with NAND gates and found that the efficiency of individual gate computation is close. ----- _Mathematics 2023, 11, 1784_ 16 of 20 **Table 5. Experimental results for single circuit gates.** **Gate Circuit** **Key Generation Time (s)** **FFT Conversion Time (s)** **Bootstrapping Time (s)** _AND_ 1.973 0.039 0.226 _NAND_ 1.982 0.038 0.227 _OR_ 1.956 0.040 0.227 _XNOR_ 1.975 0.039 0.220 Then, as shown in Table 6, we tested the experimental time overhead of SCP for k = 8, 16, and 32, where k is the bits of data. The results show that the time overhead of the SCP protocol is linearly related to the number of bits of input. **Table 6. Running time of SCP.** **_k_** **8** **16** **32** Running time (s) 3.52 7.17 14.20 _7.2. Overhead Evaluation_ In our scenario, DOs and DRs are resource-constrained users; thus, it is important to have a smaller communication overhead. In our scheme, each participant uses their key to encrypt the data and uploads it to the cloud server; thus, the total communication overhead is related to the total data size. We tested the communication overhead of our scheme on datasets with aggregate sizes of 2[8], 2[12], 2[16], and 2[20]. We compared our scheme with the scheme based on RSA [53] and the scheme based on pseudorandom permutation (PRP) [48]. As shown in Figure 5, our scheme is significantly superior to the privacy intersection scheme based on RSA. For the server-assisted scheme with limited security [48], the communication cost of our scheme is also lower. Our experimental results are the average of ten experiments. **Figure 5. Communication overhead.** Our scheme is based on the underlying PSI protocol, and the computation of the ciphertext is performed directly on the cloud server. To the best of our knowledge, our proposed scheme is the first scheme that uses MKTFHE to achieve the ideal PSI, and the time overhead of the scheme is a very important metric. For users with limited resources, ----- _Mathematics 2023, 11, 1784_ 17 of 20 low overhead in the process of data encryption and decryption is necessary. We tested the time cost of using encryption and decryption and the size of ciphertext on datasets with sizes of 2[8], 2[12], 2[16] and 2[20]. Table 7 shows that for DOs and DR with limited resources, the cost of our scheme in data encryption and decryption is very small. Finally, we tested the computing cost of the cloud server. In the experiment, we used data from 16, 32, and 64 bit systems to test the performance of our proposed scheme. Table 8 shows our experimental results. The results show that the time cost of the scheme is linearly related to the size of the dataset and the number of bits of data. Please note that the cloud has excellent computing power, so that the efficiency of the solution can be faster in actual use. **Table 7. Cost during encryption.** **2[8]** **2[12]** **2[16]** **2[20]** Encryption time (ms) 13.5 208.2 3162.2 47,987.1 Cipher size (kb) 3.5 57.3 917.5 15,083.6 **Table 8. Cloud computing time (min).** **Data Set Size** **16bit** **32bit** **64bit** 2[2] 0.51 0.99 1.98 2[4] 8.47 16.02 31.70 2[6] 137.68 273.83 547.66 **8. Conclusions** In this paper, we proposed CMPSI, a cloud-assisted private set intersection via multi-key fully homomorphic encryption, which allows the participants to outsource the encrypted data to cloud servers for storage and computation. We also designed some MKTFHE-based secure computing protocols to complete the design of our system. We analytically demonstrated the security of our scheme under a semi-honest model. Through experiments, we tested the performance of our proposed scheme and proved that our scheme has less communication overhead by comparing it with existing schemes. We also proved the feasibility of the scheme. As future research work, we plan to apply our proposed MKTFHE to a wider range of areas, such as association rule mining systems in large shopping malls. In addition, we will improve our framework to handle more complex computations and further improve the performance of our system. **Author Contributions: Conceptualization, C.F.; Methodology, X.L.; Software, C.F. and P.J.; Validation,** P.J.; Formal analysis, M.L.; Investigation, M.L. and P.G.; Data curation, X.Z.; Writing—original draft, X.Z.; Writing—review & editing, L.W. and X.L.; Visualization, P.G.; Supervision, L.W. All authors have read and agreed to the published version of the manuscript. **Funding: This work was funded by the National Key Technology Research and Development** Program of China (grant nos. 2021YFB3901000 and 2021YFB3901005); the Civil Aerospace Technology Advance Research Project of China (D040405); the Application Pilot Plan of Fengyun Satellite (FYAPP-2021.0501). **Data Availability Statement: Not applicable.** **Conflicts of Interest: The authors declare no conflict of interest.** ----- _Mathematics 2023, 11, 1784_ 18 of 20 **Abbreviations** The following abbreviations are used in this manuscript: PSI Private set intersection CMPSI Cloud-assisted multi-key private set intersection TFHE Fully homomorphic encryption over toru MKTFHE Multi-key fully homomorphic encryption over toru **References** 1. [Abdulsalam, Y.S.; Hedabou, M. Security and privacy in cloud computing: technical review. Future Internet 2022, 14, 11. [CrossRef]](http://doi.org/10.3390/fi14010011) 2. Aburukba, R.; Kaddoura, Y.; Hiba, M. Cloud Computing Infrastructure Security: Challenges and Solutions. In Proceedings of the 2022 International Symposium on Networks, Computers and Communications (ISNCC), Shenzhen, China, 19–22 July 2022; pp. 1–7. 3. Shao, Z.; Bo, Y. Private set intersection via public key encryption with keywords search. Secur. Commun. Netw. 2015, 8, 396–402. [[CrossRef]](http://dx.doi.org/10.1002/sec.988) 4. Shi, R.H.; Mu, Y.; Zhong, H.; Cui, J.; Zhang, S. An efficient quantum scheme for Private Set Intersection. Quantum Inf. Process. **[2016, 15, 363–371. [CrossRef]](http://dx.doi.org/10.1007/s11128-015-1165-z)** 5. Yang, X.; Luo, X.; Xu, A.W.; Zhang, S. Improved outsourced private set intersection protocol based on polynomial interpolation. _[Concurr. Comput. Pract. Exp. 2018, 30, e4329. [CrossRef]](http://dx.doi.org/10.1002/cpe.4329)_ 6. Tajima, A.; Sato, H.; Yamana, H. Outsourced Private Set Intersection Cardinality with Fully Homomorphic Encryption. In Proceedings of the 2018 6th International Conference on Multimedia Computing and Systems (ICMCS), Rabat, Morocco, 10–12 May 2018. 7. Ruan, O.; Huang, X.; Mao, H. An efficient private set intersection protocol for the cloud computing environments. In Proceedings of the 2020 IEEE 6th International Conference on Big Data Security on Cloud (BigDataSecurity), Baltimore, MD, USA, 25–27 May 2020; pp. 254–259. 8. Jiang, Y.; Wei, J.; Pan, J. Publicly Verifiable Private Set Intersection from Homomorphic Encryption. In Proceedings of the Security and Privacy in Social Networks and Big Data: 8th International Symposium, SocialSec 2022, Xi’an, China, 16–18 October 2022; pp. 117–137. 9. Debnath, S.K.; Kundu, N.; Choudhury, T. Efficient post-quantum private set-intersection protocol. Int. J. Inf. Comput. Secur. 2022, _[17, 405–423. [CrossRef]](http://dx.doi.org/10.1504/IJICS.2022.122381)_ 10. Wang, Q.; Zhou, F.; Xu, J.; Peng, S. Tag-based verifiable delegated set intersection over outsourced private datasets. IEEE Trans. _[Cloud Comput. 2020, 10, 1201–1214. [CrossRef]](http://dx.doi.org/10.1109/TCC.2020.2968320)_ 11. Pinkas, B.; Rosulek, M.; Trieu, N.; Yanai, A. SpOT-light: lightweight private set intersection from sparse OT extension. In Proceedings of the Advances in Cryptology—CRYPTO 2019: 39th Annual International Cryptology Conference, Santa Barbara, CA, USA, 18–22 August 2019; pp. 401–431. 12. Pinkas, B.; Rosulek, M.; Trieu, N.; Yanai, A. PSI from PaXoS: fast, malicious private set intersection. In Proceedings of the Advances in Cryptology—EUROCRYPT 2020: 39th Annual International Conference on the Theory and Applications of Cryptographic Techniques, Zagreb, Croatia, 10–14 May 2020; pp. 739–767. 13. Chase, M.; Miao, P. Private set intersection in the internet setting from lightweight oblivious PRF. In Proceedings of the Advances in Cryptology—CRYPTO 2020: 40th Annual International Cryptology Conference, CRYPTO 2020, Santa Barbara, CA, USA, 17–21 August 2020; pp. 34–63. 14. Rindal, P.; Schoppmann, P. VOLE-PSI: fast OPRF and circuit-psi from vector-ole. In Proceedings of the Advances in Cryptology— EUROCRYPT 2021: 40th Annual International Conference on the Theory and Applications of Cryptographic Techniques, Zagreb, Croatia, 17–21 October 2021; pp. 901–930. 15. Shi, R.H.; Li, Y.F. Quantum private set intersection cardinality protocol with application to privacy-preserving condition query. _[IEEE Trans. Circuits Syst. Regul. Pap. 2022, 69, 2399–2411. [CrossRef]](http://dx.doi.org/10.1109/TCSI.2022.3152591)_ 16. Zhou, Q.; Zeng, Z.; Wang, K.; Chen, M. Privacy Protection Scheme for the Internet of Vehicles Based on Private Set Intersection. _[Cryptography 2022, 6, 64. [CrossRef]](http://dx.doi.org/10.3390/cryptography6040064)_ 17. Qian, Y.; Xia, X.; Shen, J. A profile matching scheme based on private set intersection for cyber-physical-social systems. In Proceedings of the 2021 IEEE Conference on Dependable and Secure Computing (DSC), Aizuwakamatsu, Japan, 30 January–2 February 2021; pp. 1–5. 18. Demmler, D.; Rindal, P.; Rosulek, M.; Trieu, N. PIR-PSI: Scaling Private Contact Discovery; Cryptology ePrint: Archive, CA, USA, 2018. 19. Liu, F.; Ng, W.K.; Zhang, W.; Han, S.; et al. Encrypted set intersection protocol for outsourced datasets. In Proceedings of the 2014 IEEE International Conference on Cloud Engineering, Boston, MA, USA, 11–14 March 2014; pp. 135–140. 20. De Cristofaro, E.; Tsudik, G. Practical private set intersection protocols with linear complexity. In Proceedings of the Financial Cryptography and Data Security: 14th International Conference, FC 2010, Tenerife, Canary Islands, 25–28 January 2010; pp. 143–159. 21. Gentry, C. Fully homomorphic encryption using ideal lattices. In Proceedings of the 41st Annual ACM Symposium on Theory of Computing, Bethesda, MA, USA, 31 May–2 June 2009; pp. 169–178. ----- _Mathematics 2023, 11, 1784_ 19 of 20 22. Brakerski, Z.; Perlman, R. Lattice-based fully dynamic multi-key FHE with short ciphertexts. In Proceedings of the Advances in Cryptology—CRYPTO 2016: 36th Annual International Cryptology Conference, Santa Barbara, CA, USA, 14–18 August 2016; pp. 190–213. 23. López-Alt, A.; Tromer, E.; Vaikuntanathan, V. On-the-fly multiparty computation on the cloud via multikey fully homomorphic encryption. In Proceedings of the 44th Annual ACM Symposium on Theory of Computing, New York, NY, USA, 20–22 May 2012; pp. 1219–1234. 24. Gentry, C.; Sahai, A.; Waters, B. Homomorphic encryption from learning with errors: Conceptually-simpler, asymptotically-faster, attribute-based. In Proceedings of the Annual Cryptology Conference, Barbara, CA, USA, 18–22 August 2013; pp. 75–92. 25. Brakerski, Z.; Gentry, C.; Vaikuntanathan, V. (Leveled) fully homomorphic encryption without bootstrapping. ACM Trans. _[Comput. Theory (TOCT) 2014, 6, 1–36. [CrossRef]](http://dx.doi.org/10.1145/2633600)_ 26. Chillotti, I.; Gama, N.; Georgieva, M.; Izabachene, M. Faster fully homomorphic encryption: Bootstrapping in less than 0.1 seconds. In Proceedings of the International Conference on the Theory And Application of Cryptology and Information Security, Taipei, Taiwan, 5–9 December 2016; pp. 3–33. 27. Chen, H.; Laine, K.; Rindal, P. Fast private set intersection from homomorphic encryption. In Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, Dallas, TX, USA, 30 October–3 November 2017; pp. 1243–1255. 28. Chen, H.; Huang, Z.; Laine, K.; Rindal, P. Labeled PSI from fully homomorphic encryption with malicious security. In Proceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security, Toronto, ON, Canada, 15–19 October 2018; pp. 1223–1237. 29. Cong, K.; Moreno, R.C.; da Gama, M.B.; Dai, W.; Iliashenko, I.; Laine, K.; Rosenberg, M. Labeled PSI from homomorphic encryption with reduced computation and communication. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, Copenhagen, Denmark, 15–19 November 2021; pp. 1135–1150. 30. Freedman, M.J.; Nissim, K.; Pinkas, B. Efficient private matching and set intersection. In Proceedings of the Advances in Cryptology-EUROCRYPT 2004: International Conference on the Theory and Applications of Cryptographic Techniques, Interlaken, Switzerland, 2–6 May 2004; pp. 1–19. 31. Yao, A.C. Protocols for secure computations. In Proceedings of the 23rd Annual Symposium on Foundations of Computer Science (SFCS 1982), Chicago, IL, USA, 3–5 November 1982; pp. 160–164. 32. Micali, S.; Goldreich, O.; Wigderson, A. How to play any mental game. In Proceedings of the 19th ACM Symposium on Theory of Computing, New York, NY, USA, 1 January 1987; pp. 218–229. 33. Kolesnikov, V. Gate evaluation secret sharing and secure one-round two-party computation. In Proceedings of the Advances in Cryptology-ASIACRYPT 2005: 11th International Conference on the Theory and Application of Cryptology and Information Security, Chennai, India, 4–8 December 2005; pp. 136–155. 34. [Even, S.; Goldreich, O.; Lempel, A. A randomized protocol for signing contracts. Commun. ACM 1985, 28, 637–647. [CrossRef]](http://dx.doi.org/10.1145/3812.3818) 35. Hazay, C.; Nissim, K. Efficient Set Operations in the Presence of Malicious Adversaries. In Proceedings of the Public Key Cryptography, Paris, France, 26–28 May 2010; Volume 6056; pp. 312–331. 36. Huang, Y.; Evans, D.; Katz, J. Private set intersection: Are garbled circuits better than custom protocols? In Proceedings of the NDSS, San Diego, CA, USA, 5–8 February 2012. 37. Dong, C.; Chen, L.; Wen, Z. When private set intersection meets big data: An efficient and scalable protocol. In Proceedings of the 2013 ACM SIGSAC Conference on Computer & Communications Security, Berlin, Germany, 4–8 November 2013; pp. 789–800. 38. Pinkas, B.; Schneider, T.; Zohner, M. Faster Private Set Intersection based on OT Extension (Full Version). In Proceedings of the USENIX Security Symposium, San Diego, CA, USA, 20–22 August 2014. 39. Freedman, M.J.; Hazay, C.; Nissim, K.; Pinkas, B. Efficient set intersection with simulation-based security. J. Cryptol. 2016, _[29, 115–155. [CrossRef]](http://dx.doi.org/10.1007/s00145-014-9190-0)_ 40. Pinkas, B.; Schneider, T.; Zohner, M. Scalable private set intersection based on OT extension. ACM Trans. Priv. Secur. (TOPS) 2018, _[21, 1–35. [CrossRef]](http://dx.doi.org/10.1145/3154794)_ 41. Orrù, M.; Orsini, E.; Scholl, P. Actively secure 1-out-of-N OT extension with application to private set intersection. In Proceedings of the Topics in Cryptology–CT-RSA 2017: The Cryptographers’ Track at the RSA Conference 2017, San Francisco, CA, USA, 14–17 February 2017; pp. 381–396. 42. Kerschbaum, F. Collusion-resistant outsourcing of private set intersection. In Proceedings of the 27th Annual ACM Symposium on Applied Computing, Trento, Italy, 25–29 March 2012; pp. 1451–1456. 43. Kerschbaum, F. Outsourced private set intersection using homomorphic encryption. In Proceedings of the Proceedings of the 7th ACM Symposium on Information, Computer and Communications Security, Hong Kong, 7–11 June 2012; pp. 85–86. 44. Abadi, A.; Terzis, S.; Dong, C. O-PSI: delegated private set intersection on outsourced datasets. In Proceedings of the ICT Systems Security and Privacy Protection: 30th IFIP TC 11 International Conference, SEC 2015, Hamburg, Germany, 26–28 May 2015; Proceedings 30; pp. 3–17. 45. Abadi, A.; Terzis, S.; Dong, C. VD-PSI: verifiable delegated private set intersection on outsourced private datasets. In Proceedings of the Financial Cryptography and Data Security: 20th International Conference, FC 2016, Christ Church, Barbados, 22–26 February 2016; Revised Selected Papers 20; pp. 149–168. 46. Ali, M.; Mohajeri, J.; Sadeghi, M.R.; Liu, X. Attribute-based fine-grained access control for outscored private set intersection [computation. Inf. Sci. 2020, 536, 222–243. [CrossRef]](http://dx.doi.org/10.1016/j.ins.2020.05.041) ----- _Mathematics 2023, 11, 1784_ 20 of 20 47. Abadi, A.; Terzis, S.; Metere, R.; Dong, C. Efficient Delegated Private Set Intersection on Outsourced Private Datasets. IEEE Trans. _[Dependable Secur. Comput. 2019, 16, 608–624. [CrossRef]](http://dx.doi.org/10.1109/TDSC.2017.2708710)_ 48. Kamara, S.; Mohassel, P.; Raykova, M.; Sadeghian, S. Scaling private set intersection to billion-element sets. In Proceedings of the Financial Cryptography and Data Security: 18th International Conference, FC 2014, Christ Church, Barbados, 3–7 March 2014; Revised Selected Papers 18; pp. 195–215. 49. Chen, H.; Chillotti, I.; Song, Y. Multi-key homomorphic encryption from TFHE. In Proceedings of the International Conference on the Theory and Application of Cryptology and Information Security, Kobe, Japan, 8–12 December 2019; pp. 446–472. 50. Oded, G. Foundations of Cryptography: Volume 2, Basic Applications; Cambridge University Press: Cambridge, MA, USA, 2009. 51. Pradel, G.; Mitchell, C. Privacy-Preserving Biometric Matching Using Homomorphic Encryption. In Proceedings of the 2021 IEEE 20th International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom), Shenyang, China, 18–20 August 2021; pp. 494–505. 52. [Albrecht, M.R.; Player, R.; Scott, S. On the concrete hardness of learning with errors. J. Math. Cryptol. 2015, 9, 169–203. [CrossRef]](http://dx.doi.org/10.1515/jmc-2015-0016) 53. Ciampi, M.; Orlandi, C. Combining private set-intersection with secure two-party computation. In Proceedings of the Security and Cryptography for Networks: 11th International Conference, SCN 2018, Amalfi, Italy, 5–7 September 2018; pp. 464–482. **Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual** author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/math11081784?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/math11081784, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/2227-7390/11/8/1784/pdf?version=1680946668" }
2,023
[ "JournalArticle" ]
true
2023-04-08T00:00:00
[ { "paperId": "5c2d23341c3462eef259c9829fa8b0f208e7b171", "title": "Privacy Protection Scheme for the Internet of Vehicles Based on Private Set Intersection" }, { "paperId": "fd2f276b3392c877e535fd8645ae9144f127dca7", "title": "Quantum Private Set Intersection Cardinality Protocol With Application to Privacy-Preserving Condition Query" }, { "paperId": "0315646aef872dd04595a6d81cfedc3634bc64d9", "title": "Tag-Based Verifiable Delegated Set Intersection Over Outsourced Private Datasets" }, { "paperId": "14320c540d0ef53b9b87d5673ad7ce4d89b12dd4", "title": "Security and Privacy in Cloud Computing: Technical Review" }, { "paperId": "2eddb162b782692d46f71f4118ba2f6a315506d7", "title": "Labeled PSI from Homomorphic Encryption with Reduced Computation and Communication" }, { "paperId": "90aa26fe3b9b34a9a4afab67cf358f25e4f7db61", "title": "Attribute-based fine-grained access control for outscored private set intersection computation" }, { "paperId": "efba98ce884d7ff8341f963651f6999e93ca5187", "title": "Efficient Delegated Private Set Intersection on Outsourced Private Datasets" }, { "paperId": "2b51ea49ac326f4a37550c15988ab2415712bb48", "title": "Labeled PSI from Fully Homomorphic Encryption with Malicious Security" }, { "paperId": "424e9f80879e3f085673df8ca7559d4a20142c9f", "title": "PIR-PSI: Scaling Private Contact Discovery" }, { "paperId": "6767847f52f8a1c48bd59aeedf2eefecd87e061e", "title": "Improved outsourced private set intersection protocol based on polynomial interpolation" }, { "paperId": "a9d07270be6e48448ef17b348f3455d76ea1d68f", "title": "Scalable Private Set Intersection Based on OT Extension" }, { "paperId": "eccb9301284339575275c3788e71d36b5908c9ea", "title": "Fast Private Set Intersection from Homomorphic Encryption" }, { "paperId": "76b62d5ce49829dc34855a2b573007cf2b195469", "title": "An efficient quantum scheme for Private Set Intersection" }, { "paperId": "3365f31550db350ce54b7ccfd9dff3cb7715185c", "title": "On the concrete hardness of Learning with Errors" }, { "paperId": "d1db76412435907060cdf9d05b13d2f5e2ae27c5", "title": "Private set intersection via public key encryption with keywords search" }, { "paperId": "517d88b6ab20c6041876ace670199054a1509f22", "title": "(Leveled) Fully Homomorphic Encryption without Bootstrapping" }, { "paperId": "c320a52959ed7bdbde14775338ce867b97697601", "title": "When private set intersection meets big data: an efficient and scalable protocol" }, { "paperId": "c3f095c11102196896ed10ed27b9adb2e5bca68f", "title": "On-the-fly multiparty computation on the cloud via multikey fully homomorphic encryption" }, { "paperId": "2c9ebc00a3e4b78c2754c3e084e9af3c5ac569f8", "title": "Collusion-resistant outsourcing of private set intersection" }, { "paperId": "df2473061df11b76cebb7400c50246d0b354390c", "title": "How to play ANY mental game" }, { "paperId": "f2c4398e489bed6cd2ac00492c762f6b112aa7bc", "title": "A randomized protocol for signing contracts" }, { "paperId": "0856b3ce6350e5cbfe10a7b5881bc67c105708e9", "title": "Efficient post-quantum private set-intersection protocol" }, { "paperId": "9f17818d52ddbb13998097a87964a14731b8849d", "title": "Efficient Set Intersection with Simulation-Based Security" } ]
20,396
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Engineering", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01595e97cdf2cb27afbab014d58eb69ab9ffe030
[ "Computer Science", "Engineering" ]
0.879541
Low Complexity Approaches for End-to-End Latency Prediction
01595e97cdf2cb27afbab014d58eb69ab9ffe030
International Conference on Computing Communication and Networking Technologies
[ { "authorId": "2164874701", "name": "Pierre Larrenie" }, { "authorId": "1907649", "name": "J. Bercher" }, { "authorId": "2076500", "name": "O. Venard" }, { "authorId": "2541806", "name": "Iyad Lahsen Cherif" } ]
{ "alternate_issns": null, "alternate_names": [ "Int Conf Comput Commun Netw Technol", "International Conference on Computing, Communication and Networking Technologies", "ICCCNT" ], "alternate_urls": null, "id": "47c47088-7640-441b-85ba-375fe8f5c3a4", "issn": null, "name": "International Conference on Computing Communication and Networking Technologies", "type": "conference", "url": null }
Software Defined Networks have opened the door to statistical and AI-based techniques to improve efficiency of networking. Especially to ensure a certain Quality of Service (QoS) for specific applications by routing packets with awareness on content nature (VoIP, video, files, etc.) and its needs (latency, bandwidth, etc.) to use efficiently resources of a network.Predicting various Key Performance Indicators (KPIs) at any level may handle such problems while preserving network bandwidth.The question addressed in this work is the design of efficient and low-cost algorithms for KPI prediction, implementable at the local level. We focus on end-to-end latency prediction, for which we illustrate our approaches and results on a public dataset from the recent international challenge on GNN [1]. We propose several low complexity, locally implementable approaches, achieving significantly lower wall time both for training and inference, with marginally worse prediction accuracy compared to state-of-the-art global GNN solutions.
# Low Complexity Approaches for End-to-End Latency Prediction ## Pierre Larrenie, Jean-François Bercher, Olivier Venard, Iyad Lahsen-Cherif To cite this version: #### Pierre Larrenie, Jean-François Bercher, Olivier Venard, Iyad Lahsen-Cherif. Low Complexity Ap- proaches for End-to-End Latency Prediction. 2022 13th International Conference on Computing Communication and Networking Technologies (ICCCNT), Oct 2022, Kharagpur, France. pp.1-6, ￿10.1109/ICCCNT54827.2022.9984543￿. ￿hal-03957811￿ ## HAL Id: hal-03957811 https://hal.science/hal-03957811 #### Submitted on 30 Jan 2023 #### HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. #### L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. ----- # LOW COMPLEXITY APPROACHES FOR END-TO-END LATENCY PREDICTION **Pierre Larrenie** Thales SIX & LIGM Université Gustave Eiffel, CNRS Marne-la-Vallée, France ``` pierre.larrenie@esiee.fr ``` **Jean-François Bercher** LIGM Université Gustave Eiffel, CNRS Marne-la-Vallée, France ``` jean-francois.bercher@esiee.fr ``` **Olivier Venard** ESYCOM Université Gustave Eiffel, CNRS Marne-la-Vallée, France ``` olivier.venard@esiee.fr ``` **Iyad Lahsen-Cherif** Institut National des Postes et Télécommunications (INPT) Rabat, Morocco ``` lahsencherif@inpt.ac.ma ### ABSTRACT ``` Software Defined Networks have opened the door to statistical and AI-based techniques to improve efficiency of networking. Especially to ensure a certain Quality of Service (QoS) for specific applications by routing packets with awareness on content nature (VoIP, video, files, etc.) and its needs (latency, bandwidth, etc.) to use efficiently resources of a network. Predicting various Key Performance Indicators (KPIs) at any level may handle such problems while preserving network bandwidth. The question addressed in this work is the design of efficient and low-cost algorithms for KPI prediction, implementable at the local level. We focus on end-to-end latency prediction, for which we illustrate our approaches and results on a public dataset from the recent international challenge on GNN [1]. We propose several low complexity, locally implementable approaches, achieving significantly lower wall time both for training and inference, with marginally worse prediction accuracy compared to state-of-the-art global GNN solutions. **_Keywords KPI Prediction · Machine Learning · General Regression · SDN · Networking · Queuing Theory ·_** GNN ### 1 Introduction Routing while ensuring Quality of Service (QoS) is still a great challenge in any networks. Having powerful ways to transmit data is not sufficient, we must use resources wisely. This is true for wide static networks but even more for mobile networks with dynamic topology. The emergence of Software-Defined Networking (SDN) [2, 3] has made it possible to share data more efficiently between communication layers. Services are able to provide network requirements to routers based on their nature; routers acquire data about network performance, and finally allocate resources to meet these requirements. However, acquiring overall network performance can result in high consumption of network bandwidth for signalization; that is particularly constraining for networks with limited resources like Mobile _Ad-Hoc Networks (MANET)._ 1 **Note: This paper has been accepted for publication at IEEE 13th ICCCNT 2022. ©2022 IEEE. Personal use of this** material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes creating new collective works for resale or ----- _k = T_ _k = T −_ 1 _k = T −_ 2 _k = 0_ A _· · ·_ B C _· · ·_ A _· · ·_ A AGGREGATE C E _· · ·_ F _· · ·_ D A _· · ·_ Figure 1: GNN repeating T Message Passing mechanisms: message propagation and aggregation (inspired _from [7])_ We consider network for which we wish to reduce the amount of signalization and perform intelligent routing. In order to limit signalization, a first axis is to be able to estimate some key performance indicators (KPI) from other KPIs. A second point would be to be able to perform this prediction locally, at the node level, rather than a global estimation of the network. Finally, if predictions are to be performed locally, the complexity of the algorithms will need to be low, but still preserve good prediction quality. The question we address is thus the design of efficient and low-cost algorithms for KPI prediction, implementable at the local level. We focus on end-to-end latency prediction, for which we illustrate our approaches and results on a public dataset from the recent international challenge [1]. The best performances of the state-of-the-art are obtained with Graph Neural Networks (GNNs) [4, 5, 1]. Although this is a global method while we favor local methods, we use these performances as a benchmark. We first propose to use standard machine learning regression methods, for which we show that a careful feature engineering and feature selection (based on queue theory and the approach in [6]) allows to obtain near state-of-the-art performances with a very low number of parameters and very low computational cost, with the ability to operate at the link level instead of a whole-graph level. Building on that, we show that it is even possible to obtain similar performances with a single feature and curve-fitting methods. The presentation is structured as follows. In section 2, we first recall the key concepts on GNNs and queues; present some related works in the literature, before introducing the dataset used for the validation of our proposals. In section 3, we present the different approaches proposed, starting with the choice of features for machine learning methods, followed by general curve fitting methods. We then compare in section 4 the performances of these different approaches, in terms of performance as well as in terms of learning time and inference time. Finally, we conclude, discuss the overall results and draw some perspectives. ### 2 Related work and dataset **2.1** **Graph Neural Networks (GNNs)** GNN [7, 8] is a machine learning paradigm that handles non-euclidean data: graphs. A graph is defined as a set of nodes and edges with some properties on its nodes and its edges. The key point in GNNs is the concept of Message Passing: each node of the graph will update its state according to states of its neighborhood by sending and receiving messages transmitted along edges. By repeating this mechanism T times, a node is able to capture states of its T -hop neighborhood as shown in Figure 1. **2.2** **Queue Theory** Queue Theory is a well studied domain and for most of simple queue systems, explicit equations exist [9]. Further, we will refer to queue systems by using their Kendall’s notation. We often take at reference M/M/1 and M/M/1/K for their markovian property, since equations are particularly easy to handle in this case. 2 **Note: This paper has been accepted for publication at IEEE 13th ICCCNT 2022. ©2022 IEEE. Personal use of this** material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes creating new collective works for resale or |A|Col2| |---|---| |Col1|AGGREGATE| |---|---| ||| |A E F|Col2| |---|---| |B C D|Col2| |---|---| |T k = B AGGREGATE C D|T −1 k = A C A E F A|T −2 k · · · · · · · · · · · · · · · · · ·| |---|---|---| ----- However, for more general queue systems such as M/G/1 and M/G/1/K, equations are getting more complex. Whereas closed formulas exist for M/G/1 queues, M/G/1/K queues require to solve an equation system with K + 1 unknowns. Queue systems analysis focus on stable queue, i.e. when the ratio ρ = _µ[λ]_ _[≤]_ [1][ where][ λ][ (resp.][ µ][) is the] expected value of the arrival rate process (resp. service time). But finite queue systems are always stable since the maximal number of pending items is always finite and are subject to loss instead. To model the drop of incoming item in the queue we use the ratio ρe = _[λ]µ[e]_ [where][ λ][e][ is known as the effective arrival rate and can] be determined thanks to equation (1). _λe = λ(1 −_ _πK) = µ(1 −_ _π0)_ (1) Where π0 (resp. πK) in the above equation (1) refers to the probability to the queue at equilibrium to be empty (resp. full). **2.3** **Related Work** Chua, Ward, Zhang, et al. [10] present an heuristic and an Mixed Integer Programming approach to optimize Service Functions Chain provisioning when using Network Functions Virtualization for a service provider. Their approach relies on minimizing a trade-off between the expected latency and infrastructures resources. Such optimization routing flow in SDN may need additional information to be exchanged between the nodes of a network. This results in an increase of the volume of signalization, by performing some measurements such as in [11]. This is not a consequent problem in unconstrained networks, i.e. static wired networks with near-infinite bandwidth but may decrease performance of wireless network with poor capacity. An interesting solution to save bandwidth would be to predict some of the KPIs from other KPIs and data exchanged globally between nodes. In [12, 13], authors proposed a MANETs application of SDN in the domain of tactical networks. They proposed a multi-level SDN controllers architecture to build both secure and resilient networking. While orchestrating communication efficiently under military constraints such as: high-level of dynamism, frequent network failures, resources-limited devices. The proposed architecture is a trade-off between traditional centralized architecture of SDN and a decentralized architecture to meet dynamic in-network constraints. Jahromi, Hines, and Delanev [14] proposed a Quality of Experience (QoE) management strategy in a SDN to optimize the loading time of all the tile of a mapping application. They have shown the impact of several KPIs on their application using a Generalized Linear Model (GLM). This mechanism make the application aware of the current network state. Promising works rely on estimating KPIs at a graph-level. Note that it is very difficult, if not impossible, to address this analytically since computer networks models a complex structure of chained interfering queues for each flow in the network. Rusek, Suárez-Varela, Mestres, et al. [4] used GNNs for predicting KPIs such as latency, error-rate and jitter. They relied on the Routenet architecture of Figure 2. The idea is to model the problem as a bipartite hypergraph mapping flows to links as depicted on Figure 3. Aggregating messages in such graph may result in predicting KPIs of the network in input. The model needs to know the routing scheme, traffic and links properties. Their result is very promising and has been the subject of two ITU Challenge in 2020 and 2021 [5, 1]. These ITU challenges have very good results since the top-3 teams are around 2% error in delay prediction in the sense of Mean-Absolute Percentage Error (MAPE). In [6], very promising results were obtained with a a near 1% GNN model error (in the sense of MAPE) on the test set. The model mix analytical M/M/1/K queueing theory used to create extra-features to feed GNN model. In order to satisfy the constraint of scalability proposed by the challenge, the first part of model operates at the link level. **2.4** **Dataset** We use public data from the challenge [1] The dataset models static networks that have run for a certain amount of time; the obtained data is a mean of the global working period. The data contains information about 3 **Note: This paper has been accepted for publication at IEEE 13th ICCCNT 2022. ©2022 IEEE. Personal use of this** material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes creating new collective works for resale or ----- Figure 2: Routenet Architecture [4] D _F1_ _F3_ (a) Simple topology (b) Paths-links Hypergraph of (a) Figure 3: Routenet [4] paths-links hypergraph transformation applied on a simple topology graph carrying 3 flows. (a) Black circles represents communication node, double headed arrows between them denotes available symmetric communications links and dotted arrows shows flows path. (b) Circle (resp. dotted) represents links (resp. flows) entities defined in the first graph (Lij is the symmetric link between node i and node j.). Unidirectional arrows encode the relation "<flow> is carried by <link>". the topology of the network, participants and available link characteristics, traffic and routing information. The aim of the GNN ITU Challenge [1] was to build a scalable GNN model in order to predict end-to-end flow latency. Nevertheless, train on one hand and test and validation on the other hand model very different networks. Whereas training dataset models network between 25 and 50 nodes (120,000 samples), test (1,560 samples) and validation (3,120 samples) datasets model networks up to 300 nodes. This results in a very different distribution among these different splits as shown on Figure 4. Figure 4: End-to-end latency distribution on train and test datasets of ITU Challenge 2021, where train and test datasets describe networks of very different sizes.[1] It is important to point out that the proposed data is not in accordance with M/M/1/K queue models since process service time depends on the size of the packet. The size of the packet for each flow follows a Binomial distribution; it can be approximated by a Normal distribution inducing a general service time. Nevertheless, it turns out that the system does not have the behavior of a M/G/1/K queue system globally but that of a complex system with interconnected queues that cannot be easily modeled. 4 **Note: This paper has been accepted for publication at IEEE 13th ICCCNT 2022. ©2022 IEEE. Personal use of this** material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes creating new collective works for resale or ----- Hence, approximating the system locally by a mixed of a simple analytical theory (M/M/1/K) and blackbox optimization (GNNs), as was proposed in [6], is a good approach despite the lack of explicability or interpretability and the high-computational requirements with a lot of parameters to train. We show below that it is possible to obtain comparable performances with other regression approaches. ### 3 Our approaches The main question is to define an estimator ˆy of the occupancy y according to the various available characteristics of the system, with a joint objective of low complexity and performance. In the following, we present regression approaches based on machine learning and then approaches based on curve-fitting. Once an estimate of occupancy is obtained, it is possible to get the latency prediction _d[ˆ]n for a specific link n_ by the simple relation _dˆn = ˆyn_ E(|Pn|) _cn_ where E(|Pn|) is the observed average packet size on link n and cn the capacity of this link. Performances will be evaluated using the MAPE loss-function (2) ���� (ˆy, y) = [100%] _L_ _N_ _N_ � _n=1_ _yˆn −_ _yn_ ���� _yn_ which is preferred to Mean Squared Error (MSE) because of its scale-invariant property. **3.1** **Feature Engineering and Machine Learning** Based on the assumption that the system may be approximated by a model whose essential features come from M/M/1/K and M/G/1/K queue theory, we took essential parameters characterizing queueing systems, such as: ρ, ρe, π0, πK, etc. and built further features by applying interactions and various non-linearities (powers, log, exponential, square root). Then, we selected features in this set by a forward step-wise selection method; i.e. by adding in turn each feature to potential models and keeping the feature with best performance. Finally, we selected the model with best MAPE error. For a linear regression model, this led us to select and keep a set of 4 simple features, which interestingly enough, have simple interpretations:    _π0 =_ 1−1ρ−[K]ρ[+1] _L = ρ + π0_ �k _[kρ][k]_ _ρe =_ _[λ]λ[e]_ _[ρ][ =][ λ]µ[e]_ _Se =_ [�]k _[kρ]e[k]_ (3) where L is the expected number of packets in the queue according to M/M/1/K, π0 the probability that the queue is empty according to M/M/1/K theory, ρe the effective queue utilization, and Se the unnormalized expected value of the effective number of packet in the queue buffer. These features can be thought as a kind of data preprocessing, before applying ML algorithms, and this turns out to be a key to achieving good performances. The 4 previous features have been kept as input for all the machine learning models. Next we considered several machine learning algorithm, fitted on the training split and performances were evaluated by test split of a public dataset [1]. Algorithms that were considered are: Multi-Layer Perceptron model (MLP) with 4 layers and with ReLU activation function, Linear Regression, Gradient Boosting Regression Tree (GBRT) with an ensemble of n = 100 estimators, Random Forest of n = 100 trees and Generalized Linear Model (GLM) with Poisson family and exponential link. All results of these methods are shown in Table 1. 5 **Note: This paper has been accepted for publication at IEEE 13th ICCCNT 2022. ©2022 IEEE. Personal use of this** material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes creating new collective works for resale or ----- **3.2** **Curve Regression for occupancy prediction** There is a high interdependence of the features we selected in Equation 3, since all these features can be expressed in term of ρe. Furthermore, it is confirmed by data exploration that ρe is the prominent feature for occupancy prediction (and in turn latency prediction), as exemplified in Figure 5. It is then tempting to try to further simplify our features space and to try to estimate the occupancy from a non-linear transformation of the single feature ρe, as: _yˆ = g(ρe)_ (4) where ˆy is the estimate of the occupancy y. The concerns are of course to define simple and efficient functions _g, with a low number of parameters, that can model the kind of growth shown in Figure 5, and of course to_ check that the performance remains interesting. We followed three approaches to design the estimator g in order to predict links occupancy and end-to-end flow latency. In all cases, the parameters of g were computed by minimizing the mean squared of the regression error. Figure 5: Data of ITU Challenge 2021 [1], ρe vs queue occupancy. Color-scale is an indicator of points cloud density. **3.2.1** **Exponential of polynomial** The simplest approach is to use a curve-fitting regression of the form _yˆ = g(ρe) = e[p][n][(][ρ][e][)]_ (5) where pn(x) ∈ Rn[x] is a polynomial of degree n with real coefficients. In order to find coefficients of pn one can obviously consider predicting log(y) (where y denotes the queue occupancy). Choosing an arbitrary high polynomial degree results to oscillations and increases largely computation time. However choosing a too small degree does not allow the prediction of high occupancy. **3.2.2** **Generative polynomials** The estimator g is defined as a linear combination of simple functions (fn): � _yˆ = g(ρe) =_ _αn · fn(ρe)_ (6) _n_ 6 **Note: This paper has been accepted for publication at IEEE 13th ICCCNT 2022. ©2022 IEEE. Personal use of this** material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes creating new collective works for resale or ----- **Generative polynomial similar to M/M/1/K theory** The idea here is to use a polynomial fn[K] [that will] match approximately the expression in Equation 3 of the expected number of packets in the queue L. φfγnn[K]n[K] =[(][(][x][x] φ[) =][) =][K]n _[ x][ φ]�_ _[n]nn[K]γ+1nn1[(][x]−(1[)]�x−[K]x[+1])_ _nn ≥ ≥_ 00 ∀x ∈ [0; 1[ (7) where K is the size of the queue[1] The sequence of (fn)[K]n=0 [is finite and defined in interval][ [0; 1[][.] In order to improve regression capabilities, each fn[K] [is defined as][ φ]n[K] [normalized by][ γ][n][, a local maximum of] _φ[K]n_ [in the interval][ [0; 1[][.] **Bernstein Polynomials** The previous method relies on polynomial approximation. Since the expected value L can be expressed theoretically in terms of a polynomial of degree K, we are driven to the Bernstein polynomials that form a basis in the set of polynomial in the interval [0; 1]: �K _fn[K][(][x][) =]_ _n_ � _x[n](1_ _x)[K][−][n]_ (8) _−_ The approximation of any continuous function on [0; 1[ by a Bernstein polynomial converges uniformly. **3.2.3** **Implicit function** The idea here is to define a set of N points θn = (an, bn) and approximate the underlying function by linear interpolation between those points. To obtain a good positioning of these points, we select them as the solution of the following optimization problem: minθ _L(fθ(x), y) +_ _N[α]_ � _n_ _∥⃗un × ⃗un+1∥[2]_ _∥⃗un∥[2]∥⃗un+1∥[2]_ s.t. _⃗un = θn+1_ _θ0n_ _−_ _a0 = 0_ _aN = bN = 1_ _an+1 −_ _an ≥_ 0 _θn = (an, bn)[T]_ _∈_ [0; 1][2] (9) Equation 9 includes a first term for minimizing the interpolation error, and a second term weighted by a parameter α ≥ 0, to force θn sequence to be as aligned and as far as possible. This implies that our sampling will be refined in high curvature zone of our function. The constraint formulated makes θn an increasing sequence along the feature axis in order to get a correct interpolation of the curve, especially when N is high enough. ### 4 Comparison and Discussion In this section, we evaluate our methods on the data from the GNN ITU Challenge 2021, described in subsection 2.4. We compare our results to those of the challenge winners, which establish the state-of-the-art in terms of pure performance. Since the actual labeled test dataset used for the challenge was released after the end of the challenge, all evaluations are performed on this particular dataset. The Table 1 presents the characteristics of the methods, in terms of the number of input features and parameters to be learned; their performance in the sense of MAPE and MSE; and the values of the execution times, both in learning time and inference time. All results were obtained with the same computer configuration: 120 Go RAM, 1 CPU Intel i9-9920X @ 3.50 GHz with 24 cores and 2 GPUs Nvidia TITAN RTX2080 24Go. The methods used for comparison are divided into 3 groups, the first being the set of GNN approaches. 1In results shown Table I., we consider K = 32 in order to match the data contained in the ITU challenge dataset [1]. 7 **Note: This paper has been accepted for publication at IEEE 13th ICCCNT 2022. ©2022 IEEE. Personal use of this** material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes creating new collective works for resale or ----- Sampling Optimization (N = 12, α =1e-5)[*] 24 1.77% 3.18e-5 _≈1min_ 0.306s Table 1: Results Synthesis of various models for flow latency prediction. Test dataset from [1] *only 500,000 samples used for training (2.25% of training dataset); **only 5,000,000 samples used for training (22.5% of training dataset); [+]under-estimation/over-estimation occurs on high queue occupancy prediction In the second group, we used classical machine learning models with only 4 input features obtained by stepwise selection, as presented in Equation 3. In the third group, we group curve regression models using a single well-chosen feature, namely ρe, as presented in subsection 3.2. As we can observe, the proposed approaches achieve a much lower computational time than the GNN approaches, both in terms of learning time and inference time; this at the cost of a marginal performance degradation. Moreover, non-GNN approaches provide a more local solution since predictions are performed at the link level and not at the whole graph level (Models predict queues occupancy, then compute analytically delay for each link and finally aggregate along path). This would allow to use them for simple local predictions, without having to rely on the global knowledge and prediction of the network. The consequent gain in computational time of our low-complexity approaches is that they use far fewer parameters, which reduces the amount of data needed for training. The reduction in the number of parameters and the architecture (number of operations) of the solutions explains the drop in learning and inference times. Nevertheless, when we match the distribution as presented in Figure 5, we notice that most of our data are on a low occupancy level. In practice, some models have a kind of limited behavior when the occupation of the targeted queue is close to 100%: there is a significant over- or under-prediction. However, this behavior does not really affect the overall performance due to the low density of this scenario in our dataset and the predicted values are close enough to the targets. ### 5 Conclusion In this paper, we considered the problem of designing efficient and low-cost algorithms for KPI prediction, implementable at the local level. We have argued and proposed several alternatives to GNNs for predicting the queue occupancy of a complex system using simple ML models with carefully chosen features or general curve-fitting methods. At the cost of a marginal performance loss, our proposals are characterized by low complexity, significantly lower learning and inference times compared to GNNs, and the possibility of local deployment. Thus, this type of solution can be used for continuous performance monitoring. The low complexity and structures of linear regression algorithms or curve-fitting solutions should also be suitable for adaptive formulations. These last two points are current perspectives of this work. Of course, the approaches considered here will have to be considered and adapted for other types of KPI, such as error-rate or jitter. 8 **Note: This paper has been accepted for publication at IEEE 13th ICCCNT 2022. ©2022 IEEE. Personal use of this** material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes creating new collective works for resale or |Approaches|Input Features|Model Parameters|MAPE|MSE|Wall Training Time|Wall Inference Time| |---|---|---|---|---|---|---| |Routenet [4]|Topology Traffic matrix Routing Scheme|-|≫100%|(N/A)|≈12h|-| |Top-1 ITU Challenge Team (PARANA)[6]||654006|1.27%|1.10e-5|≈8h|214s| |MLP|π0 L Se ρe|291|1.91%|3.18e-5|≈45min|8.26s| |Linear Regression*+||4|1.74%|3.20e-5|<1sec|0.296s| |GBRT (n=100)*||4|1.73%|2.90-5|≈1min|0.867s| |Random Forest (n=100)*||4|1.69%|3.00e-5|<1sec|0.994s| |GLM - Poisson+||4|3.68%|5.09e-4|≈1min|0.481s| |Curve-fitting exponential (deg=3)+|ρe|4|3.94%|3.75e-4|≈1sec|0.311s| |Curve-fitting exponential (deg=8)+||9|1.70%|3.53e-5|≈5secs|0.320| |Curve-fitting M/M/1/K**+||33|2.04%|4.42e-5|≈3min|3.55s| |Curve-fitting Bernstein**+||33|1.68%|3.13e-5|≈2min|3.14s| |Sampling Optimization (N = 12, α = 0)*+||24|1.77%|3.18e-5|≈1min|0.281s| |Sampling Optimization (N = 12, α =1e-5)*||24|1.77%|3.18e-5|≈1min|0.306s| ----- Finally, a last point that deserves interest is the fact that these low complexity models can be interpreted/explained either by direct inspection (visualization), or by using tools such as Shapley values [15] which allow to interpret output values by measuring contributions of each input feature on the prediction. ### References [1] J. Suárez-Varela et al., “The graph neural networking challenge: A worldwide competition for education in AI/ML for networks,” ACM SIGCOMM Computer Communication Review, vol. 51, no. 3, pp. 9–16, [2021. DOI: 10.1145/3477482.3477485.](https://doi.org/10.1145/3477482.3477485) [2] S. Singh and R. K. Jha, “A survey on Software Defined Networking: Architecture for next generation network,” Journal of Network and Systems Management, vol. 25, no. 2, pp. 321–374, 2017. [3] R. Amin, M. Reisslein, and N. Shah, “Hybrid SDN networks: A survey of existing approaches,” IEEE _[Communications Surveys & Tutorials, vol. 20, no. 4, pp. 3259–3306, 2018. DOI: 10.1109/COMST.](https://doi.org/10.1109/COMST.2018.2837161)_ ``` 2018.2837161. ``` [4] K. Rusek, J. Suárez-Varela, A. Mestres, P. Barlet-Ros, and A. Cabellos-Aparicio, “Unveiling the potential of graph neural networks for network modeling and optimization in SDN,” in Proceedings of _the 2019 ACM Symposium on SDN Research, 2019, pp. 140–151._ [5] _[The graph neural networking challenge 2020. https://bnn.upc.edu/challenge/gnnet2020.](https://bnn.upc.edu/challenge/gnnet2020)_ [6] [B. K. de Aquino Afonso, GNNet challenge 2021 report (1st place), https://github.com/ITU-AI-](https://github.com/ITU-AI-ML-in-5G-Challenge/ITU-ML5G-PS-001-PARANA) ``` ML-in-5G-Challenge/ITU-ML5G-PS-001-PARANA, 2021. ``` [7] W. L. Hamilton, “Graph representation learning,” Synthesis Lectures on Artifical Intelligence and _Machine Learning, vol. 14, no. 3, pp. 1–159, 2020._ [8] D. Bacciu, F. Errica, A. Micheli, and M. Podda, “A gentle introduction to deep learning for graphs,” _Neural Networks, vol. 129, pp. 203–221, 2020._ [9] R. B. Cooper, “Introduction to queueing theory,” Edward Arnold, London, 1981. [10] F. C. Chua, J. Ward, Y. Zhang, P. Sharma, and B. A. Huberman, “Stringer: Balancing latency and resource usage in service function chain provisioning,” IEEE Internet Computing, vol. 20, no. 6, pp. 22–31, 2016. [11] S. T. V. Pasca, S. S. P. Kodali, and K. Kataoka, “AMPS: Application aware multipath flow routing using machine learning in SDN,” in 2017 Twenty-third National Conference on Communications (NCC), IEEE, 2017, pp. 1–6. [12] K. Poularakis, G. Iosifidis, and L. Tassiulas, “SDN-enabled tactical ad hoc networks: Extending programmable control to the edge,” IEEE Communications Magazine, vol. 56, no. 7, pp. 132–138, 2018. [13] K. Poularakis, Q. Qin, E. M. Nahum, M. Rio, and L. Tassiulas, “Flexible SDN control in tactical ad hoc [networks,” Ad Hoc Networks, vol. 85, pp. 71–80, 2019, ISSN: 1570-8705. DOI: 10.1016/j.adhoc.](https://doi.org/10.1016/j.adhoc.2018.10.012) ``` 2018.10.012. ``` [14] H. Z. Jahromi, A. Hines, and D. T. Delanev, “Towards application-aware networking: Ml-based end-toend application KPI/QoE metrics characterization in SDN,” in 2018 Tenth International Conference on _Ubiquitous and Future Networks (ICUFN), IEEE, 2018, pp. 126–131._ [15] S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” Advances in _neural information processing systems, vol. 30, 2017._ 9 **Note: This paper has been accepted for publication at IEEE 13th ICCCNT 2022. ©2022 IEEE. Personal use of this** material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes creating new collective works for resale or -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2302.00004, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://hal.science/hal-03957811/file/main.pdf" }
2,022
[ "JournalArticle", "Conference" ]
true
2022-10-03T00:00:00
[ { "paperId": "546cdc1f6318fb1ff12e245919424be5cab34594", "title": "The graph neural networking challenge" }, { "paperId": "45f4585f12ed7310c2dbc487a2aa1ba23c568071", "title": "Graph Representation Learning" }, { "paperId": "2e3a86c4b8f6883b371f718eb0a35857a6bf9b95", "title": "A Gentle Introduction to Deep Learning for Graphs" }, { "paperId": "5056d98741060e1e51098a568f170bc545677f0d", "title": "Flexible SDN control in tactical ad hoc networks" }, { "paperId": "53bc1dd1dbfa61bcb5bba6fcaee3f2c2323a0f49", "title": "Unveiling the potential of Graph Neural Networks for network modeling and optimization in SDN" }, { "paperId": "e9764cbb19a8d3895e3c47ac08a77cb42b4ceeb3", "title": "Towards Application-Aware Networking: ML-Based End-to-End Application KPI/QoE Metrics Characterization in SDN" }, { "paperId": "24bf243d9b5f0b8fe6ae840e22b97423d1249234", "title": "Hybrid SDN Networks: A Survey of Existing Approaches" }, { "paperId": "ebffa220a8109feb7f34a15aca3ede4bcd2d2161", "title": "SDN-Enabled Tactical Ad Hoc Networks: Extending Programmable Control to the Edge" }, { "paperId": "442e10a3c6640ded9408622005e3c2a8906ce4c2", "title": "A Unified Approach to Interpreting Model Predictions" }, { "paperId": "ee40e107a2e34a7ee67a913253d12d0dde4943f9", "title": "AMPS: Application aware multipath flow routing using machine learning in SDN" }, { "paperId": "f0c51b5804af7b96bf2d45a5c3ad39987ac46001", "title": "A Survey on Software Defined Networking: Architecture for Next Generation Network" }, { "paperId": "05380d21cd612184e89e82005c9392ac16554352", "title": "Stringer: Balancing Latency and Resource Usage in Service Function Chain Provisioning" }, { "paperId": "7d6e65fde7ce51c281a65a17d8fa86395a3526c2", "title": "Virtual Conference" }, { "paperId": "49bea1156edc540b32cd456a4f7252b5033c8a2f", "title": "Introduction to Queueing Theory" }, { "paperId": null, "title": "GNNet challenge 2021 report (1st place)" }, { "paperId": null, "title": "proposed a Quality of Experience (QoE) management strategy in a SDN to optimize the loading time of all the tile of a mapping application" }, { "paperId": null, "title": "Note : This paper has been accepted for publication at IEEE 13th ICCCNT 2022. ©2022" } ]
8,861
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Business", "source": "external" }, { "category": "Engineering", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Law", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0159c5f17225c6e8aa183f7b39e74aaffa185a1e
[ "Computer Science", "Business", "Engineering" ]
0.896982
Building a Blockchain Application thatComplies with the EU General DataProtection Regulation
0159c5f17225c6e8aa183f7b39e74aaffa185a1e
MIS Q. Executive
[ { "authorId": "4727858", "name": "Alexander Rieger" }, { "authorId": "35440237", "name": "Florian Guggenmos" }, { "authorId": "79347889", "name": "J. Lockl" }, { "authorId": "1711202", "name": "G. Fridgen" }, { "authorId": "2986366", "name": "Nils Urbach" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
# Building a Blockchain Application that Complies with the EU General Data Protection Regulation ### Complying with the EU General Data Protection Regulation (GDPR) poses significant challenges for blockchain projects, including establishing clear responsibilities for compliance, securing lawful bases for processing personal data, and observing rights to rectification and erasure. We describe how Germany’s Federal Office for Migration and Refugees addressed these challenges and created a GDPR-compliant blockchain solution for cross-organizational workflow coordination. Based on the lessons learned, we pro­ vide three recommendations for ensuring blockchain solutions are GDPR-compliant.[1,2] #### Alexander Rieger University of Augsburg (Germany) FIM Research Center Jannik Lockl University of Bayreuth (Germany) Project Group Business & Information Systems Engineering of the Fraunhofer FIT Nils Urbach University of Bayreuth (Germany) Faculty of Law, Business, and Economics #### Florian Guggenmos University of Bayreuth (Germany) Project Group Business & Information Systems Engineering of the Fraunhofer FIT Gilbert Fridgen University of Luxembourg (Luxembourg) SnT - Interdisciplinary Centre for Security, Reliability and Trust ## The EU General Data Protection Regulation Poses Significant Challenges for Blockchain Projects[12] Blockchain technology provides an innovative means of fostering collaboration, especially in cross-organizational workflows. Blockchain solutions allow the organizations involved in the workflow to maintain control over their respective activities but, at the same time, enable them to establish a “shared and persistent truth” on the state of the workflow at any given time. This truth can act as a point of reference if conflicts need to be resolved at a later point. By extension, this allows the organizations to use updates on the blockchain as reliable 1 Carsten Sørensen is the accepting senior editor for this article. 2 We developed this article as part of an applied research project with Germany’s Federal Office for Migration and Refugees. The authors would like to thank everyone involved for their support. We would also like to express our gratitude to Carsten Sørensen, Mary Lacity, Rajiv Sabherwal, and three anonymous reviewers for their guidance and comments, which considerably improved this article. **DOI: 10.17705/2msqe.00020** ----- triggers for subsequent activities. Moreover, the continuous distribution of updates throughout the network means that these triggers are readily available. If required, smart contracts can also allow the automated activation of certain steps of the workflow and its monitoring. In simple terms, blockchain technology offers a promising alternative to centralized workflow management systems where the delegation of workflow governance to a central authority is not possible or desirable.[3] However, when blockchain projects move beyond the proof-of-concept stage, they begin to encounter the limiting effects of regulations and legal barriers. Foremost among these is the European Union (EU) General Data Protection Regulation (GDPR).[4] The GDPR protects a “natural person”[5] from unregulated processing of their personal data and establishes rules governing the free movement of their personal data. It codifies several essential rights of natural persons, such as the right to have inaccurate personal data rectified, or completed if it is incomplete, and to have their personal data erased. Moreover, it establishes clear responsibilities for compliance with the regulation and prohibits the processing of personal data without a lawful basis, such as requiring explicit consent if the action is necessary to fulfill obligations of a law or contract. At first glance, many of the GDPR requirements appear to conflict with the basic properties of blockchain technology. For instance, the technology does not envisage the data being erased at a later point. Moreover, the decentralized nature of blockchain networks seems to prevent the designation of clear responsibilities. Also, the need to obtain a lawful 3 For a detailed discussion on the prospect of using blockchains for the management of business processes and workflows, see Mendling, J. et al. “Blockchains for Business Process Management : Challenges and Opportunities,” ACM Transactions on Management Information _Systems (9:1), February 1, 2018, pp. 1-16._ 4 _Regulation (EU) 2016/679 of the European Parliament and_ _of the Council of 27 April 2016, Council of the European Union,_ European Parliament; the full text of the GDPR is available at https:// publications.europa.eu/en/publication-detail/-/publication/3e485e1511bd-11e6-ba9a-01aa75ed71a1/language-en. While the GDPR is an EU regulation, many global platforms and other cross-border firms observe its requirements. 5 The GDPR regulates the processing of information relating to an identified or identifiable natural person—i.e., an individual human being. It does not regulate the processing of information relating to legal persons. basis for processing personal data at each node appears daunting. As we show in this article, however, these challenges can be resolved. We describe how the Bundesamt für Migration und Flüchtlinge (the BAMF—Germany’s Federal Office for Migration and Refugees) created a GDPR-compliant blockchain solution for processing applications for asylum. (The German asylum procedure is described in Appendix A.) The key learnings from this project give rise to three recommendations for the management of GDPR requirements and the design of GDPR-compliant blockchain solutions. (Appendix B describes the research we conducted in preparing to write this article.) ## A Brief Introduction to the EU General Data Protection Regulation Data privacy has been an important focus of European lawmaking since the 1970s. A key multilateral milestone was the EU’s 1981 signing of the Convention for the Protection of Individuals, which addressed the automatic processing of personal data. The most recent and comprehensive regulatory step was the passing of the General Data Protection Regulation in 2016, which took effect across all member states of the EU in May 2018. The GDPR applies to any act of wholly or partially automated processing[6] of any information relating to an identified or identifiable natural person in the EU, and to any such act by a data controller[7] or a data processor[8] that operates on that person’s behalf, in the European Union. Importantly, it relates not only to data that is obviously personal, such as names 6 As set out in Article 4(2) of the GDPR, the term “processing” en­ compasses a wide variety of conceivable actions, such as recording, storing, and disseminating data. 7 Article 4(7) of the GDPR defines a data controller as a “natural or legal person, public authority, agency or other body which, alone or jointly with others, determines the purposes and means of the processing of personal data ….” 8 Article 4(8) of the GDPR defines a data processor as a “natural or legal person, public authority, agency or other body which processes personal data on behalf of the controller.” ----- but also to data that, in combination with other means, can be used to identify a natural person.[9] The GDPR aims to foster the free movement of personal data within EU member states by standardizing the rules for the processing of personal data by both private and public data controllers. It builds on six principles, including purpose limitation and data minimization, and enshrines privacy by design and by default. Importantly, it outlaws any processing of personal data unless the data controller has a lawful basis. Chapter 3 of the GDPR also establishes the various rights of data subjects[10] (Articles 12 to 23). These rights include, among others, the right to rectification (Article 16)[11] and the right to erasure (“the right to be forgotten”) (Article 17)[12]. This means that data subjects can hold controllers and processors of their data accountable, and violators can incur hefty fines. In particular, Article 83(5) of the GDPR prescribes administrative fines of up to €20 million ($22.29 million)[13] or, in the case of companies, up to 4% of total worldwide annual revenue from the preceding financial year, whichever is higher. ## Reconciling Blockchain Solutions with the GDPR Most guidelines on the management of GDPR requirements presuppose a single identifiable controller and skirt around the particularities of decentralized networks in general and blockchain technology in particular. Blockchain projects therefore face genuine challenges in observing the requirements of the GDPR. Chief among these challenges is the need to establish clear responsibilities for compliance, to secure lawful 9 In particular, the GDPR also applies to data that allows attribu­ tion through the analysis of patterns of use and context. In many instances, this includes public keys. For more details on the resulting challenges, see Lyons, T., Courcelas, L. and Timsit, K. Blockchain _and the GDPR, The European Union Blockchain Observatory and_ Forum, October 16, 2018, available at https://www.eublockchainfo­ rum.eu/sites/default/files/reports/20181016_report_gdpr.pdf. 10 The GDPR uses the term “data subject” as a synonym for any identified or identifiable natural person. 11 Article 16 of the GDPR grants each data subject “the right to obtain from the controller without undue delay the rectification of inaccurate personal data.” 12 Article 17 of the GDPR states that an individual has “the right to obtain from the controller the erasure of personal data concerning him or her without undue delay” when one of the defined reasons applies. 13 Euro/dollar conversion rate as of October 2019. bases for processing personal data, and to comply with the rights to rectification and erasure. **Establishing** **Clear** **Responsibilities** **for Compliance. The GDPR requires that** responsibilities for compliance with its articles are identified and designated, especially when several parties jointly determine the purposes and means of processing (“joint control”).[14] For conventional databases, the establishment of responsibilities is comparatively easy. In blockchain networks, defining responsibility is often difficult. In particular, legal opinions differ as to which participants qualify as standalone controllers and which as joint controllers. The distinction is important because joint controllers are jointly accountable and have to create an arrangement that identifies each joint controller and determines their respective responsibilities, and that is transparent to the affected data subjects.[15] **Securing Lawful Bases for Processing** **Personal data. Article 5 of the GDPR specifies** six lawful bases for processing personal data, including documented authorization by the data subject or processing that is required to fulfill obligations under law or contract;[16] without one of these lawful bases, a data controller cannot legally process personal data. Establishing lawfulness for each data-processing action in a blockchain network can be particularly 14 The primary criterion for qualifying as joint controllers is the joint determination of the purpose of processing (“primacy of the purpose criterion”); simple participation in the determination of the means does not necessarily qualify a participant of a blockchain net­ work as a joint controller. For a detailed discussion of joint control­ lership in the context of blockchains, see Blockchain and the General _Data Protection Regulation: Can distributed ledgers be squared with_ _European data protection law?, European Parliamentary Research_ Service, July 2019, available at http://www.europarl.europa.eu/Reg­ Data/etudes/STUD/2019/634445/EPRS_STU(2019)634445_EN.pdf 15 The national data protection authority of France (CNIL), for instance, considers participants of blockchain networks to be data controllers “when the … participant is a natural person and … the personal data processing operation is related to a professional or com­ mercial activity” or “when the … participant is a legal person and … it registers personal data in a blockchain.” When these controllers do not designate a single controller who determines the purposes and mean of processing, regulators and courts may easily decide to hold them accountable as joint controllers. The CNIL’s detailed opinion is in Blockchain: Solutions for a responsible use of the blockchain in _the context of personal data, 2018, available at https://www.cnil.fr/_ sites/default/files/atoms/files/blockchain.pdf 16 Lawfulness has to be established for three essential processing steps: the submission of new data to the blockchain by a submitting participant; its validation, distribution, and replication by the nodes of the blockchain network; and its reading from the blockchain by another participant. ----- |Table 1: Advantages and Disadvantages of the Central Authority, Shared Responsibility, and Pseudonymization Approaches|Col2|Col3|Col4| |---|---|---|---| ||Description (in terms of controlling and complying with the right to erasure)|Advantages|Disadvantages| |Central Authority|●● The network nominates a central authority that acts as the network’s single controller ●● The right to erasure is waived by way of contracts between the central authority and the network’s partci ipants, and in consultatoi n with afef cted third partei s if necessary|●● Easy identfi ci atoi n of the data controller ●● Requires a less intricate solutoi n architecture|●● Requires centralized control over network rights ●● If any of the erasure contracts become void, the blockchain may have to be modifei d| |Shared Responsibility|●● All partci ipants in the blockchain network act as joint controllers ●● The right to erasure is waived by way of mutual contracts between the network’s partci ipants, and in consultatoi n with afef cted third partei s if necessary|●● Does not require centralized control over network rights ●● Requires a less intricate solutoi n architecture|●● There must be a legal basis for processing personal data for each partci ipant ●● If any of the erasure contracts become void, the blockchain may have to be modifei d| |Pseudonymizatoi n|●● Data on the blockchain is pseudonymized; only those partci ipants who possess the additoi nal informatoi n required for atrt ibutoi n are (joint) controllers ●● The blockchain solutoi n can comply with the right to erasure by eliminatni g the additoi nal informatoi n|●● Does not require centralized control over network rights ●● The right to erasure is upheld by design|●● Requires an intricate solutoi n architecture to ensure that the additoi nal informatoi n required for atrt ibutoi n can be securely shared and reliably eliminated ●● The blockchain may have to be modifei d if there is inadvertent atrt ibutoi n from examining patet rns of use or context (linkability risk) or any other inadvertent reversal of the pseudonymizatoi n (reversal risk)| ----- burdensome. Moreover, any lawful basis may cease to exist or apply in the future (e.g., with the withdrawal of consent or amendment of the law). In these circumstances, storage of the relevant personal data is no longer permitted and the data must be erased. **Complying with the Rights to Erasure** **and Rectification.** The GDPR states that data subjects can request that data controllers rectify their personal data if there are errors and erase the data once a lawful basis ceases to exist. This implies that modifications to data on a blockchain must be made on each copy of the blockchain. #### Three Potential Approaches for Ensuring Blockchain Solutions Are GDPR-Compliant From a data-privacy perspective, addressing the three challenges described above requires a combination of organizational and technical measures. We have identified three potential blockchain solution approaches—“central authority,” “pseudonymization,”[17] and “shared responsibility.”[18] Table 1 lists the advantages and disadvantages of these approaches, and we describe them below. To the best of our knowledge, there is no single best approach for each application and context. Moreover, the approaches are not comprehensively exhaustive, and some blockchain projects may identify other ways of ensuring they comply with the requirements of the GDPR. **Central Authority Approach. The central** authority approach addresses conflicts between GDPR requirements and a blockchain solution through organizational measures and by delegating responsibility to a central authority. This authority may be a single participant in the 17 Article 4(5) of the GDPR defines pseudonymization as “the processing of personal data in such a manner that the personal data can no longer be attributed to a specific data subject without the use of additional information, provided that such additional informa­ tion is kept separately and is subject to technical and organizational measures to ensure that the personal data are not attributed to an identified or identifiable natural person.” Pseudonymization is differ­ ent from anonymization, which renders personal data “anonymous in such a manner that the data subject is not or no longer identifiable.” (Recital 26 of the GDPR). 18 For a comprehensive discussion of the three approaches, see Fridgen, G., Guggenberger, N., Hoeren, T., Prinz, W. and Urbach, N. _Chancen und Herausforderungen von DLT (Blockchain) in Mobilität_ _und Logistik, (the management summary is in English), Bundesmin­_ isteriums für Verkehr und digitale Infrastruktur, May 2019, available at https://www.bmvi.de/SharedDocs/DE/Anlage/DG/blockchaingutachten.pdf?__blob=publicationFile. blockchain network or a group of participants. The central authority assumes the role of the data controller and the responsibility for compliance with the GDPR. Moreover, it establishes the rights of network participants and creates, using a contract or another legal instrument, agreements for processing personal data with the operators of the blockchain nodes. The authority also secures the lawful bases for processing personal data and handles any related matters. When the blockchain network processes the personal data of network participants, the central authority has to create contracts with each network participant. When the network processes the personal data of third parties, the central authority must secure the lawful bases for processing the data of those third parties. The right to erasure of personal data is waived by way of contracts between the central authority and the network’s participants, and, if necessary, in consultation with affected third parties. If any of these contracts become void, the blockchain network must erase the personal data from the blockchain. This can be done in several ways. For instance, each node can remove the data from its block and recalculate all subsequent blocks. This recalculation can be documented in another blockchain. Another option is to use redactable[19] blockchains. The right to rectify data can be achieved through technical means by submitting a rectification transaction to the blockchain. More specifically, the original transaction is invalidated by the rectification transaction, but it remains on the blockchain. The central authority approach is appropriate for blockchain solutions that permit the designation of a single data controller with farreaching competencies. **Shared** **Responsibility** **Approach.** The shared responsibility approach is very similar to the central authority approach but builds on the premise of sharing responsibilities among the participants of the blockchain network. All participants in the network act as joint controllers and establish an arrangement that sets out the respective responsibilities of each participant. The lawful basis for processing 19 For a discussion of redactable blockchains, see Ateniese, G., Magri, B., Venturi, D. and Andrade, E. “Redactable Blockchain – or – Rewriting History in Bitcoin and Friends”, 2017 IEEE Symposium on Security and Privacy, May 11, 2017 pp. 111-126. ----- personal data relating to network participants and/or third parties is ideally ensured through mutual contracts. As with the central authority approach, the right to erasure is waived by way of contracts between the network’s participants and, if necessary, with affected third parties. Again, the right to rectification can be achieved through rectification transactions. The shared responsibility approach is appropriate for blockchain networks where all participants have lawful bases for processing all the personal data exchanged. **Pseudonymization** **Approach.** As its name suggests, this approach is based on pseudonymizing the data on the blockchain so that it only qualifies as personal data when participants possess certain additional off-chain information that allows the data to be attributed to a natural person. Pseudonymization of the data can be achieved using encryption, cryptographic hash functions, or pseudonymous identifiers.[20] Only those participants who possess the additional information required for attribution are controllers. When these controllers jointly determine the purposes and means of processing the pseudonymized data and the data required for attribution, they are joint controllers. As such, they need to establish, through a joint control arrangement, their respective responsibilities for compliance with the GDPR and for establishing lawful bases for processing personal data. Alternatively, they can create data processing agreements to establish clear responsibilities for compliance. Controllers and processors can uphold the right to erasure by eliminating the additional information—that is, by depriving themselves of the ability to attribute data to specific individuals. This technical measure is considerably more reliable than an organizational measure based on waivers but requires a solution that ensures that the additional information needed for attribution can be securely shared and reliably eliminated. The process for rectification mirrors the central authority and shared responsibility approaches. The pseudonymization approach is appropriate for blockchain networks where the 20 In the first case, the additional information required for attribu­ tion is the decryption key. In the second case, the additional informa­ tion is the unhashed information, and in the third case, the additional information required for attribution is the mapping of a pseudony­ mous identifier to a specific identifier. designation of a central authority is not viable or desirable, and where not all participants have lawful bases for the processing of all the personal data exchanged. ## Background of the Choice of Blockchain Technology for the German Asylum Procedure In Germany, the asylum procedure involves close collaboration between various authorities at the municipal, state and federal levels, with the BAMF playing a pivotal central role because it handles and issues decisions regarding asylum applications. State-level migration authorities are responsible for the initial registration of asylum seekers, and for their eventual integration or repatriation. Several security agencies are involved in background checks; municipal governments generally handle housing, and various health authorities provide medical care. #### Lessons Learned from Early Efforts to Introduce Centralized Support Systems for the Asylum Procedure Federal separation of competencies prevents the delegation of workflow governance to a central authority, such as the BAMF. This separation also leads to a significant degree of variation between workflows, and complicates the creation of a common workflow model and the introduction of a conventional workflow management system. One essential step in managing the resulting complexities was to transform the Central Register of Foreign Nationals (Ausländerzentralregister, or AZR for short), a database that contains personal information on about 20 million foreign nationals, into a shared repository for certain master data, such as names and fingerprints. However, this transformation did not include workflow management features. Moreover, the transformation revealed three challenges for creating a centralized solution for the German asylum procedure. First, centralization requires the redistribution of competencies, which, in turn, requires considerable legislative action. In particular, the existence of the AZR requires a specific AZR law. While this law provides a solid legal foundation, it also reduces the AZR’s flexibility, as technical ----- updates first require Germany’s parliament to make a formal legislative update to the AZR law. Second, centralization creates unbalanced data guardianship arrangements. In particular, the BAMF has to assume full responsibility for the lawfulness of the subsequent processing of any data in the AZR. Third, centralization leads to the development of solutions that do not take account of the specifics of individual workflows. In particular, the AZR’s data model includes only a fraction of the data typically exchanged between authorities over the course of the workflow involved in processing asylum applications. #### Identifying Blockchain as a Potential Solution for the Asylum Procedure These shortcomings encouraged the BAMF to explore decentralized alternatives for cross-organizational workflow coordination, which would require neither the delegation of workflow governance to a single authority nor the extension of the AZR. After a preliminary evaluation, the BAMF narrowed down its technological options and decided to consider a blockchain solution. This choice was based on best practices for the identification of blockchain use cases and essentially followed the first seven questions of the ten-step decision path described by Pedersen et al.[21] The solution the BAMF sought was a shared common database for event logs (Question 1 in the ten-step path) that would be used by multiple parties (Question 2). Although trust is not necessarily an issue between the authorities involved in the German asylum procedure, the federal nature of the process means that it incorporates a multitude of interests that are often not fully aligned (Question 3). Concerns about competencies, data guardianship, and flexibility caused the BAMF to seek a decentralized solution (Question 4). Moreover, it argued that a solution for cross-organizational workflow coordination would have to offer tiered rights of access because most authorities involved in the procedure are only entitled to view specific data (Question 5). The rules of the procedure, meanwhile, would remain predominately the 21 Pedersen, A. B., Risius, M., and Beck, R. “A Ten-Step Decision Path to Determine When to Use Blockchain Technologies,” MIS _Quarterly Executive (18:2), June, 2019, pp. 99-115. This article_ provides a comprehensive discussion of what constitutes a genuine blockchain use case. same (Question 6), and the BAMF was interested in creating an immutable log that would facilitate process forensics at a later point (Question 7). #### Choosing the Blockchain Design Access right considerations caused the BAMF to choose a private permissioned blockchain design. Blockchain networks are deemed “private” when reading access is limited to a certain set of participants, such as the authorities involved in the asylum procedure, whereas a public blockchain network allows anyone to read transactions. “Permissioned” means that only preregistered participants can submit new transactions, validate those transactions, and append new blocks; in a permissionless network, any participant can do so.[22] The BAMF chose to make its blockchain solution permissioned because the authorities in the asylum procedure are known and have clearly designated roles and competencies. A private permissioned blockchain solution offered the BAMF several functional and technical benefits over the status quo. Functionally, such a solution would improve integrity and increase the speed of procedures. Lengthy asylum procedures regularly result in undue hardship for applicants, negative press coverage, and protracted revisions in court. The BAMF was particularly interested in blockchain technology’s ability to use event logs to quickly establish a shared truth on the status and course of asylum applications, as illustrated by the manager of the BAMF’s blockchain project: _“Blockchain is a promising technology that_ _can support communication and collaboration_ _among the public authorities involved in asylum_ _procedures. It offers many advantages, especially_ _for sharing status updates quickly and securely:_ _the authorities involved can obtain an overview_ _of the course of an applicant’s asylum procedure_ _via the blockchain and can call up the status_ _almost in real time.” Haris Trtovac, Manager of_ the BAMF’s blockchain project Technically, a blockchain solution could provide the BAMF with flexibility, which would 22 For detailed information on the differences between these block­ chain design choices, see Androulaki, E. et al. “Hyperledger Fabric: A Distributed Operating System for Permissioned Blockchains,” _Proceedings of the Thirteenth EuroSys Conference, April 23-26,_ 2018, ACM Digital Library, available at https://dl.acm.org/citation. cfm?id=3190538. ----- only require agreement on data models and application programming interfaces (APIs). Moreover, it recognized blockchain’s potential to further the once-only principle:[23] _“In the future, we should no longer copy data_ _into large nationwide databases. Rather, we_ _should leave the data where we collect it and_ _use a logging layer to make transparent when_ _and where status changes occurred. With a_ _lightweight blockchain solution, we can more_ _easily implement this logging layer than with an_ _expansion of the existing and already complex IT_ _solutions.” Markus Richter, Vice President of the_ BAMF ## How the BAMF Ensured its Blockchain Solution Is GDPR compliant #### Proof-of-Concept and Pilot Stages The BAMF began its blockchain project in January 2018 with a proof of concept intended to demonstrate that a blockchain solution could offer the functionality required to coordinate the workflow underlying the German asylum procedure. The prototype used a blockchain to log and propagate the completion of essential steps in the procedure. It matched these event logs to asylum applications using AZR identification numbers. Although the prototype was successful in demonstrating blockchain technology’s functional merits, the BAMF was concerned about compliance with the GDPR, which took effect in May 2018. The BAMF therefore commissioned a legal opinion,[24] which raised serious concerns about the prototype’s data model. In particular, 23 The European Commission’s Communication on the eGovern­ _ment Action Plan 2016 – 2020 sets out several principles, including_ the “once only principle,” which states that “public administrations should ensure that citizens and businesses supply the same infor­ mation only once to a public administration. Public administration offices take action if permitted to internally reuse this data, in due respect of data protection rules, so that no additional burden falls on citizens and businesses.”, available at https://ec.europa.eu/digitalsingle-market/en/news/communication-eu-egovernment-action-plan2016-2020-accelerating-digital-transformation 24 For the full opinion (in German only), see Hoeren, T. and Baur, J. Datenschutzrechtliche Zulässigkeit der Übermittlung von Infor­ _mationen über Migranten zwischen öffentlichen Stellen mittels einer_ _Permissioned-Biockchain, 2018, available at https://fragdenstaat.de/_ anfrage/gutachten-blockchainbamf/302470/anhang/ifg_gutachten_ blockchain.pdf. the opinion argued that, while the event logs did not themselves qualify as personal data, the use of the AZR identifiers turned each event log into personal data, which would eventually have to be erased. The opinion urged the BAMF to address three issues: 1. Define the responsibilities for compliance with the requirements of the GDPR 2. Establish the lawful bases for processing personal data 3. Create a design that would allow personal data to be rectified and erased. Ideally, the design would either use a so-called redactable blockchain or pseudonymize the personal data. The BAMF addressed these issues during the subsequent pilot phase. To limit complexity, the BAMF decided to focus on the Saxony Arrival, Decision, and Return (AnkER) facility, which opened in Dresden mid-2018. (The aim is for the initial processing of all asylum seekers to take place in AnkER facilities.) To improve information exchange and expedite procedures, several authorities are involved in the AnkER procedure. The BAMF approached Saxony’s central immigration authority (the LDS), with the aim of jointly creating and testing a blockchain solution for coordinating those parts of the AnkER procedure that required the closest collaboration between the BAMF and the LDS. To mitigate the lack of best practices for managing the requirements of the GDPR and developing a GDPR-compliant solution, the BAMF held several idea-generation workshops and architectural refinement meetings. The BAMF also met with Germany’s Federal Commissioner for Data Protection and Freedom of Information (BfDI). In two workshops, the BAMF and experts from the BfDI discussed the prototype and the BAMF’s propositions for a GDPR-compliant solution. #### Choosing the Blockchain Solution Approach Because it wanted to avoid the creation of a central authority, the BAMF used the pseudonymization approach to ensure that its blockchain solution is GDPR-compliant. It also determined that encryption and hashing were impractical choices for the pseudonymization ----- |Blockchain <H…ash Value <Status Upd <Time-stamp <Authority ID <Pseudonym Blockchain Node Blockchain Node|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|<H…ash Value <Status Upd <Time-stamp <Authority ID <Pseudonym|Col11|> ate > > ou|> s ID>| |---|---|---|---|---|---|---|---|---|---|---|---|---| ||Blockchain Node||||<Pseudon Blockchain Node||||<Pseudon|||| |||||||<Pseudo Blockchain Node|||<Pseudo|||| |||||||||||||| |||||||||||||| |Privacy Services|Privacy Service||||||Privacy Service|||||| ||||Secure co|m|munication|||||||| |||||||||||||| |||||||||||||| |Dashboard Services|Dashboard Service||||Dashboard Service|||||||| |||||||||||||| |||||||||||||| |Back-end Systems|Workflow Management Database System||||Workflow Management Database System|||||||| ||BAMF||||Saxony’s Central Immigration Authority|||||||| of event logs on the blockchain. It rejected encryption because this would limit the network’s ability to validate transactions and because managing and distributing individual encryption keys for each event log would create substantial complexity. Encryption with static keys might eventually lead to all participants being able to decrypt all event logs, which would, in turn, make all participants joint controllers. It chose not to use cryptographic hash functions because this would reduce the blockchain to a simple notarization[25] solution with very limited options for the use of smart contracts. Moreover, such a solution would require the redundant exchange of event logs via another channel. Instead, the BAMF decided to implement a pseudonymous identifier solution with so 25 According to the National Notary Association, “Notarization is the official fraud-deterrent process that assures the parties of a trans­ action that a document is authentic, and can be trusted.” For more information, see https://www.nationalnotary.org/knowledge-center/ about-notaries/what-is-notarization. called privacy services. With this solution, each participant operates an off-chain service that maps pseudonymous identifiers on the blockchain to the IDs used by the participant, and does so in a privacy-compliant, erasable, and rectifiable manner. Without the mapping, the BAMF (and other authorities involved in the blockchain solution) cannot attribute the data on the blockchain to a natural person. In order to enable the sharing of meaningful information, privacy services can exchange mapping information through secure communication channels. #### Creation of a Joint Control Arrangement Through an administrative agreement, the BAMF created a joint control arrangement with the LDS that established the purpose and means of processing and assigned responsibilities for GDPR compliance. In terms of purpose and ----- means, the agreement specified the storage and exchange of event logs required for collaborating, via the blockchain solution throughout the AnkER procedure. In terms of responsibilities, the agreement specified that the BAMF would host and assume responsibility for the data stored on the blockchain and for the privacy services. However, for each event log submitted to the blockchain, such as the BAMF’s ruling on an asylum application, the LDS and the BAMF would have to independently verify whether they have a lawful basis for submission; once the event log is written to the blockchain, the other authority is responsible for establishing its own lawful basis before reading the log. For each piece of mapping information exchanged between the privacy services, the sending authority must verify that it has a lawful basis for sending, and the receiving authority must to establish whether it has a lawful basis for adding the information to its mapping database. To minimize complexity, the BAMF and the LDS consulted the relevant legislation to establish up-front the required lawful bases for each conceivable type of data exchange. #### The BAMF’s Blockchain Solution Architecture In terms of technical measures, the BAMF implemented a blockchain architecture with three layers (see Figure 1). Layer 1 (back-end systems) holds the existing workflow management systems and data repositories of the authorities involved. The other two layers do not need to be integrated with these back-end systems; instead, they can be loosely coupled through a set of APIs. Layer 2 (integration) hosts dashboard services, which create the event logs and can display to users data from both the back-end systems and the blockchain (Layer 3). Layer 2 also hosts privacy services, which map the pseudonymous blockchain IDs with the specific IDs used in the back-end systems. The design of Layers 1 and 2 can vary between the authorities involved in the blockchain solution; only the blockchain layer is standardized across all authorities. **Blockchain Layer. The blockchain layer** propagates pseudonymized event logs, with each entry consisting of four elements—a status update, a time-stamp, the ID of the authority that created the status update and a pseudonymous ID. From a functional perspective, these elements reflect the minimum amount of data required for effective use. From a GDPR perspective, they are sufficiently nonspecific to limit the risk of inadvertent attribution—for example, through the analysis of the trail of event logs.[26] **Integration** **Layer—Privacy** **Services.** In order to attribute the event logs on the blockchain, the BAMF created a network of authority-specific privacy services, with each authority hosting a standalone privacy service. Each service contains databases that map the pseudonymous IDs to the specific IDs— such as application or personal identification numbers—used in the authority’s back-end systems. The privacy services support role-based access procedures for different user groups within authorities, and can exchange mapping information. Such an exchange is important for the handover of an asylum application to another authority.[27] Moreover, the services can exchange requests for the erasure of mappings related to a pseudonymous ID. **Integration** **Layer—Dashboard** **Services.** In order to submit event logs to the blockchain and display data from both the blockchain and the back-end systems, the BAMF implemented dashboard services. Event logs can be submitted manually to the dashboard services, or by drawing (pull-based mechanism) or receiving (push-based mechanism) the data from the back-end systems. The dashboard services then convert the event log data to comply with the blockchain’s data model. To display the data, users access the dashboard services through a web browser and enter various commands, such as “display the history of a certain procedure” or “display all procedures that meet certain conditions. The dashboard will then, in accordance with the access rights of the user and the mapping information in the privacy service, collect and display attributed event logs from the 26 The risk of inadvertent attribution from spatiotemporal data— i.e., data points with both location and time attributes—is high because the four data points can be sufficient to uniquely identify a person (linkability risk). For a detailed discussion of the linkability risk of anonymized mobility data, see de Montjoye, Y.-A., Hidalgo, C. A., Verleysen, M. and Blondel, V. D. “Unique in the Crowd: The privacy bounds of human mobility,” Scientific Reports (3), March 2013, Article 1376. 27 From a legal perspective, every such exchange equates to the processing of personal data, and requires both the sending and receiv­ ing authority to establish a lawful basis. ----- blockchain layer and further information from the back-end systems of the authority. Importantly, a user can only view information for which the authority and the user have clearance and a lawful basis. #### Ensuring Privacy by Design **Erasure by Design.** Erasure of personal data from a blockchain may become necessary for several reasons, such as simple errors in entering data, or the expiration of a lawful basis. Explicit time limits in the German Asylum Act, for instance, ensure that authorities do not store personal data for more than a maximum of ten years after the completion of a procedure. The erasure procedure implemented in the BAMF’s blockchain solution is triggered by an authority issuing a command to its privacy service, which deletes the respective mapping and submits a so-called “erasure event log” to the blockchain. An erasure event log on the blockchain invalidates the pseudonymous blockchain ID and prevents further use of this ID by all authorities in the blockchain network. Moreover, the log informs other authorities of the erasure. Each joint controller who receives this information can then use the erasure event log as a trigger to re examine all the lawful bases. For those events for which the joint controllers still have a lawful basis, they can create and submit copies to the blockchain under a new pseudonymous ID. Conceptually, the erasure procedure could also be useful for off-chain information exchange related to an event log. Currently, authorities check whether data requests from other authorities are legitimate, but often do not keep a record of these requests or the data they forward. This means that authorities are unable to direct requests for erasure and rectification to specific authorities. The blockchain solution, however, would ensure that such requests for erasure reach all authorities in the blockchain network. **Rectification** **by** **Design.** In addition to the erasure procedure, the BAMF also implemented a rectification procedure. Rectification may become necessary if, for example, false event logs are submitted to the blockchain or duplicate blockchain IDs need to be reconciled. The rectification procedure mirrors the erasure procedure and is triggered by specific rectification actions in the back-end systems. Rectification actions are submitted to the blockchain as “rectification events.” Other authorities can respond to rectification events by, for instance, approaching the issuing authority for further information, and/or stopping or reversing subsequent steps in the asylum procedure. If there are duplicates, the privacy service adjusts its mapping and retires one of the duplicate blockchain IDs. ## Recommendations for Ensuring Blockchain Solutions Are GDPR-Compliant We distilled three key learnings from the BAMF project and have translated these into three recommendations. These recommendations should be interpreted as high-level guidelines rather than as a reference architecture or legal advice. In line with the European parliament,[28] we advise each blockchain project to seek its own legal assessment and to design its own portfolio of organizational and technical measures. #### 1. Avoid Storing Personal Data on a Blockchain Blockchain solutions should be designed so that it is not necessary to store personal data on the blockchain. Instead, personal data should remain in systems that permit rectification and erasure. This advice also applies to any attribute on a blockchain that allows identification of an individual by analyzing patterns of use or context. #### 2. A Blockchain Solution that Needs to Process Personal Data Should Use a Private and Permissioned Pseudonymization Approach If a blockchain solution will process personal data, we recommend using the pseudonymization approach, because a central authority or shared responsibility approach will be impractical in most instances. Moreover, a pseudonymization approach simplifies the identification of controllers. All those who hold the additional information required for attribution qualify as 28 “Compatibility between distributed ledgers and the GDPR can only be assessed on the basis of a detailed case-by-case analysis that accounts for the specific technical design and governance set-up of the relevant blockchain use case.” European Parliament, 2019. ----- (joint) controllers unless otherwise specified in an agreement for processing personal data. When the solution requires two or more participants to share additional information for attribution, we strongly recommend establishing a private and permissioned blockchain network. This will simplify the establishment and management of arrangements required for joint control or agreements for processing personal data. In particular, a private network enables the establishment of a controlled introduction process during which new participants can be added to the arrangements or agreements. A permissioned network facilitates the creation of a flexible and role-based model for the allocation of responsibilities. To avoid inadvertent attribution, however, even pseudonymized data should be limited to an absolute minimum. Moreover, the solution should store information required for attribution in a highly secure manner, as any uncontrolled disclosure may require the blockchain to be modified. #### 3. A Blockchain Solution that Needs to Coordinate Cross-Organizational Workflows Should Use a Private and Permissioned Pseudonymization Approach with Identifier Mapping For cross-organizational workflows, the pseudonymization approach with identifier mapping—i.e., separate mapping databases for each participant—provides the best trade-off between value and security. Although storing only hashed event logs on the blockchain would be more secure, this approach would require the redundant exchange of the unhashed data and would limit the use of the blockchain solution to simple notarization. Storing encrypted event logs on the blockchain would be just as useful as identifier mapping but would require each event log to be encrypted with a separate encryption key, which would significantly increase the complexity and vulnerability of the overall blockchain solution. ## Concluding Remarks _“GDPR_ _compliance_ _is_ _not_ _about_ _the_ _technology, it is about how the technology_ _is used. Just like there is no GDPR-compliant_ _Internet or GDPR-compliant artificial intelligence_ _algorithm, there is no such thing as a GDPR-_ _compliant blockchain technology. There are only_ _GDPR-compliant use cases and applications.”[29]_ The BAMF has created a GDPR-compliant blockchain application through a combination of organizational and technical measures. The BAMF application for processing asylum applications thus demonstrates that blockchain technology and the GDPR are not incompatible and suggests that organizations should continue to explore and develop blockchain solutions that will involve the processing of personal data. Because blockchain solutions emphasize decentralized governance, they could be a particularly promising alternative in cross-organizational settings that prevent the delegation of workflow governance to a central authority. A next essential step for the widespread deployment of GDPR-compliant blockchain applications will be to establish standards and reference architectures that ensure the interoperability of various blockchain technologies and solutions. ## Appendix A: The German Asylum Procedure The German Constitution grants anyone persecuted on political grounds the right to asylum. This right also extends to those fleeing from violence, war, or terrorism. 29 Lyons, T., Courcelas, L. and Timsit, K., op. cit., October 16, 2018. ----- |Table 2: BAMF Blockchain Team Members Interviewed|Col2| |---|---| |Role in the Blockchain Project|Focus| |Director of the AnkER and functoi nal project lead with more than 15 years’ experience|Functoi nal benefti s, design principles, and data privacy| |Business process manager with more than 15 years’ experience|Functoi nal benefti s, design principles, and data privacy| |Lawyer, GDPR compliance-responsible team member with more than 15 years’ experience|Data privacy| |Lawyer, GDPR compliance-responsible team member with more than 15 years’ experience|Functoi nal benefti s, design principles, and data privacy| |Project manager with more than 20 years’ experience, responsible for communicatoi n with the c-suite|Functoi nal benefti s, design principle, and data privacy| Figure 2 shows a simplified version of the German asylum procedure. On arriving in Germany, federal law requires asylum seekers to immediately report to federal or state police and make a request for asylum. The police will then take them to the closest registration agency, where they will have access to medical care, and the registration agency provides them with a proof-of-arrival document that grants a temporary right to stay. While at the registration agency, asylum seekers can also register their application with the BAMF. The BAMF checks if another member state of the European Union has previously registered the applicant. If that check is positive, the Dublin Regulation stipulates that the refugee must be returned to the member |Table 3: External Blockchain Experts Interviewed|Col2|Col3| |---|---|---| |Interviewee|Experience|Focus| |Serial blockchain entrepreneur|Founder and CEO of a blockchain startup that has implemented a blockchain-based payment system in the refugee context|Functoi nal beneftis, design principles, and principles for blockchain decision paths| |Blockchain consultant|Blockchain consultant who has worked since 2015 for T-Systems MMS, and has been involved with multpi le blockchain proofs of concept and pilots|Functoi nal beneftis, design principles, and principles for blockchain decision paths| |Blockchain researcher and consultant|Blockchain researcher and solutoi n architect who has worked since 2018 for Centrifuge, which provides an open, decentralized operatni g system that aims to connect the global fni ancial supply chain|Functoi nal beneftis, design principles, and impact of blockchain on IT strategies| |Blockchain researcher and consultant|Associate partner who has worked since 2008 for a Fortune 500 technology company closely involved with Hyperledger Fabric|Functoi nal beneftis, design principles, and data privacy| |Blockchain developer|Blockchain developer and solutoi n architect who has worked since 2016 for the NEM Foundatoi n, which provides technical support for the NEM ecosystems|Functoi nal beneftis, design principles, and data privacy| |Blockchain researcher and consultant|Founder and CEO of a blockchain startup founded in 2016 to provide secure and GDPR-compliant data exchange|Functoi nal beneftis, design principles, and data privacy| |Blockchain researcher and consultant|Junior IT manager who has worked since 2017 for a globally actvi e automotvi e supplier on technology research and implementatoi n|Functoi nal beneftis, design principles, and data privacy| |Blockchain developer|Blockchain developer and solutoi n architect who has worked since 2016 for a globally actvi e automotvi e supplier|Functoi nal beneftis, design principles, and data privacy| |Blockchain entrepreneur|Co-founder of a blockchain startup that ofef rs digital infrastructure services for innovatvi e electricity tarifsf|Functoi nal beneftis, design principles, and data privacy| |Blockchain researcher and consultant|Blockchain Ph.D. student and consultant who has worked since 2014 for one of the largest research insttiutoi ns in Europe|Functoi nal beneftis, design principles, and data privacy| ----- state in which he or she was first registered. This check, however, can take up to several days. Meanwhile, refugees may have to relocate to a different registration agency based on their nationality and Germany’s federal quota system. If the check is negative, the BAMF will hold a personal interview at the closest appropriate registration agency or a regional office. A BAMF caseworker will then decide whether to approve or reject the application for asylum. The caseworker justifies the decision in a written document that is given to the applicant. If the caseworker rejects the application, the applicant can appeal the decision in court. Favorable decisions result in the applicant being granted a residence permit. If the application is rejected, the relevant immigration authority repatriates the applicant. More details on the German asylum procedure are available in the BAMF’s overview document.[30] ## Appendix B: Research Method There is a dearth of detailed accounts of and knowledge about developing GDPR-compliant blockchain applications. In the public sector, in particular, most governmental agencies remain unfamiliar with blockchain technology. Our research thus required us to provide substantial guidance to the BAMF on developing its blockchain solution, as well as to other agencies, such as the Federal Commissioner for Data Protection and Freedom of Information, to help them assess the solution’s GDPR-compliance. As a consequence, we chose an action research[31] approach, with three of our co-authors providing advisory services to the BAMF’s blockchain project from January 2018 onward. These three co-authors familiarized the BAMF team with blockchain technology and organized an ongoing cycle of cross-team reflections, 30 _The stages of the German Asylum Procedure: An Overview of_ _the Individual Procedural Steps and the Legal Basis, 2016, Federal_ Office for Migration and Refugees, available at http://www.bamf.de/ SharedDocs/Anlagen/EN/Publikationen/Broschueren/das-deutscheasylverfahren.pdf?__blob=publicationFile. 31 Action research emphasizes (participatory) observation in the field to address a specific problem (in this case, enabling digital federalism through a GDPR-compliant blockchain architecture). For more information on action research, see Baskerville, R. and Myers, M. D. “Special Issue: Action Research in Information Systems,” MIS _Quarterly (28:3), September 2004, pp. 329-335._ which continued throughout the project.[32] One co-author, for instance, worked closely with the IT vendor hired by the BAMF to implement the blockchain solution and guided the BAMF’s architectural board. Two other co-authors were not involved with the project team’s operations but acted as external observers. The combination of three collaborating and two observing researchers allowed us to maintain high standards of evidence gathering and academic rigor. In the course of the project, we gathered evidence from four different sources: 1. We held various workshops on functional, technical, and data privacy issues 2. We regularly participated in and contributed to developer meetings and architectural reviews 3. We analyzed public blockchain interviews of BAMF employees and conducted 15 additional semistructured interviews with blockchain project team members and blockchain experts. These interviews lasted between 40 minutes and two hours and each was recorded 4. We reviewed and analyzed various internal and external documents on the blockchain project. #### Blockchain Workshops and Contribution to Technical Meetings During the project, the three collaborating co-authors held nearly 30 blockchain workshops. The range of attendees included BAMF employees, employees of Saxony’s central immigration authority (the LDS), employees of the Federal and the Saxony Ministries of the Interior, a delegation from the Federal Commissioner for Data Protection and Freedom of Information, employees of the Dutch Immigration and Naturalization Service, and several other organizations. In these workshops, we focused on various functional, technological and data privacy issues. To deliver the educational segments of these workshops, 32 Avison, D., Baskerville, R., Myers, M. and Wood-Harper, T. “IS action research: can we serve two masters? (panel session),” Kock, N., panel chairman, Proceedings of the 20th International Conference _on Information Systems, December 1999, pp. 582-585._ ----- |Table 4: BAMF Employees Public Blockchain Interviews Analyzed|Col2|Col3| |---|---|---| |Public Interview Reported in:|Interviewee and Position|Focus| |Behörden Spiegel|Dr. Markus Richter (BAMF CIO from Jan 2018 – July 2018 and BAMF vice president since July 2018)|Functoi nal and technical beneftis and data privacy| |Der Spiegel|Dr. Markus Richter (BAMF CIO from Jan 2018 – July 2018 and BAMF vice president since July 2018)|Functoi nal and technical beneftis and data privacy| |Bundesamt für Migratoi n und Flüchtlinge – Digitalisierungsagenda 2020|Haris Trtovac (BAMF blockchain project manager since April 2018)|Functoi nal and technical beneftis and data privacy| |Bundesamt fürMigratoi n und Flüchtlinge – Digitalisierungsagenda 2020|Kausik Munsi (BAMF CTO)|Functoi nal and technical beneftis, data privacy and impact of blockchain on the BAMF’s IT strategy| we adapted the method of Fridgen et al.[33] In the conceptual segments, we used creative elements to access the attendees’ prior experiences and knowledge and to further their involvement. In addition to these workshops, we collaborated with the BAMF team members on a daily basis in stand-up meetings, development meetings, and management calls. We were routinely involved in architectural as well as sprint review and planning meetings. In particular, we suggested multiple refinements to the blockchain solution and helped resolve technical and data-privacy issues. For instance, we developed the erasure and rectification concepts and contributed essential elements to the privacy service concept. #### Interviews with BAMF Stakeholders, Team Members, and Experts Given the novelty of blockchain technology and the related challenges, we complemented our action research approach by conducting interviews, which are a preferred method for extracting explorative knowledge. In total, we conducted 15 interviews, five with project team members and 10 with various blockchain experts. We used an interview guide for these semistructured interviews, which allowed the interviews to flow naturally but also ensured 33 Fridgen, G., Lockl, J., Radszuwill, S., Rieger, A., Schweizer, A. and Urbach, N. “A Solution in Search of a Problem: A Method for the Development of Blockchain Use Cases,” Proceedings of the Ameri­ _cas Conference on Information Systems, August 2018, pp. 1-10._ comparability between the interviews. An open dialogue, rather than the rigorous use of predefined questions, helped to maximize the depth of insights provided by interviewees, who thus delivered valuable knowledge that supported the subsequent development of the recommendations.[34] Because all blockchain team members preferred to remain anonymous, the table below provides only anonymized information on their roles in the blockchain project and their prior experience. We also conducted ten semistructured interviews with external blockchain experts (as listed in the next table), some of whom preferred to remain anonymous. In addition, we analyzed public blockchain interviews given by four BAMF employees, listed in Table 4. #### Analysis of Internal and Public Documents We also analyzed several hundred pages of BAMF internal memos, reports, analyses, and meeting minutes. Importantly, these internal documents included highly relevant strategy papers, data privacy analyses, and architectural specifications. We also reviewed the BAMF’s public documents, such as its digitalization 34 Urquhart, C., Lehmann, H. and Myers, M. D. “Putting the ‘theory’ back into grounded theory: guidelines for grounded theory studies in information systems,” Information Systems Journal (20:4), July 2010, pp. 357-381. ----- agenda and blockchain webpage. Lastly, but importantly, we reviewed legal analyses and the data privacy advice issued by lawyers and renowned German scholars concerning blockchain, legal decisions in comparable scenarios, and governmental papers on comparable blockchain use cases. #### Analyzing the Evidence from the Sources To analyze the evidence, we first consolidated our sources of data and the data itself to eliminate redundancies. Next, we clarified imprecise statements and added—where needed—explanatory comments to data points. Third, we assigned codes to the data points and developed tentative principles through open and, later, axial coding. Where the data related to new phenomena, we marked the passages and discussed them within the research team, building new principles when necessary.[35] We iteratively adapted the codes until they were collectively exhaustive and mutually exclusive. Subsequently, we discussed the resultant principles with the practitioners in order to gain other perspectives. ## About the Authors **Alexander Rieger** Alexander Rieger (alexander.rieger@fim rc.de) is a doctoral candidate at the Finance & Information Management (FIM) Research Center and the Project Group Business & Information Systems Engineering of the Fraunhofer FIT, University of Augsburg. His professional interests include innovative digital technologies such as blockchain and artificial intelligence, and, more specifically, their strategic implications and adoption. Prior to joining the BAMF’s blockchain project in February 2018, Alex spent several years working in industry and consulting. **Florian Guggenmos** Florian Guggenmos (florian.guggenmos@fim rc.de) is a doctoral candidate at the Finance & Information Management (FIM) Research Center and the Project Group Business & Information Systems Engineering of the Fraunhofer FIT, University of Bayreuth. His current research 35 Strauss, A. and Corbin, J. M. Basics of Qualitative Research: _Grounded Theory Procedures and Techniques, 1990, Sage Publica­_ tions. focuses on systemic risk management as well as data privacy and information security, particularly in the context of digitalization projects. Florian has also worked on a range of applied research projects. He joined the BAMF’s blockchain project in February 2018. **Jannik Lockl** Jannik Lockl (jannik.lockl@fim-rc.de) is a doctoral candidate at the Finance & Information Management (FIM) Research Center and the Project Group Business & Information Systems Engineering of the Fraunhofer FIT, University of Bayreuth. His main focus is on the Internet of Things (IoT) as well as the wider adoption of digital technologies and the socioeconomic embedding of blockchain applications. Jannik worked as a consultant on a variety of industry projects before joining the BAMF’s blockchain project in February 2018. **Gilbert Fridgen** Gilbert Fridgen (gilbert.fridgen@uni.lu) is PayPal-FNR PEARL Chair in Digital Financial Services in the Interdisciplinary Center for Security, Reliability and Trust (SnT) at the University of Luxembourg. His work focuses on smart grids, the machine economy, and blockchain technology in both the public and private sectors. Gilbert’s work has been published in several prominent IS, management, computer science and engineering journals. He has also managed various industry research projects and received multiple research grants. Gilbert has served as expert counsel to many German government bodies, including the Bundestag and six German federal ministries, and also to the European Commission through its European Blockchain Partnership. **Nils Urbach** Nils Urbach (nils.urbach@fim-rc.de) is a professor of information systems and strategic IT Management at the University of Bayreuth. He is deputy director of the Finance & Information Management (FIM) Research Center and of the Project Group Business & Information Systems Engineering of Fraunhofer FIT. He is also a co-founder and director of the Fraunhofer BlockchainLab. Nils’ research focuses on digital transformation, blockchain, and the management of artificial intelligence. His work has been published in leading journals including _MIS_ ----- _Quarterly_ _Executive,_ _Journal_ _of_ _Information_ _Technology,_ and _The_ _Journal_ _of_ _Strategic_ _Information Systems. Before his academic career,_ Nils worked for several years as a management consultant. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.17705/2msqe.00020?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.17705/2msqe.00020, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://www.fim-rc.de/Paperbibliothek/Veroeffentlicht/941/wi-941.pdf" }
2,019
[ "JournalArticle" ]
true
2019-12-01T00:00:00
[]
15,269
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/0159df66618cb667a54a529247fedac5640af2da
[ "Computer Science" ]
0.921355
Investigating Service Discovery, Management and Network Support for Next Generation Object Oriented Services
0159df66618cb667a54a529247fedac5640af2da
Intelligence in Networks
[ { "authorId": "2133614", "name": "Bilhanan Silverajan" }, { "authorId": "39982166", "name": "J. Hartman" }, { "authorId": "153831770", "name": "Jani Laaksonen" } ]
{ "alternate_issns": null, "alternate_names": [ "IFIP Int Conf Intell Netw Telecommun Netw Intell", "IFIP International Conference on Intelligence in Networks: Telecommunication Network Intelligence", "SMARTNET", "Intell Netw" ], "alternate_urls": null, "id": "bf640a72-6f1c-4143-aaec-ae21f1bb6eff", "issn": null, "name": "Intelligence in Networks", "type": "conference", "url": null }
null
###### Investigating Service Discovery, Management and Network Support for Next Generation Object Oriented Services Bilhanan Silverajan, Joona Hartman, and Jani Laaksonen _Dept. oflnformation Technology, Tampere University ofTechnology, P.O. Box 553,_ _FIN-3310J_ _Tampere, Finland_ _{bilhanan_ I hartmanj IJjlaakso}@cs.tutfi Abstract The network computing industry has eagerly embraced technologies, welcoming an ever-increasing variety ofnew service discovery protocols and object architectures. With this abundance now offered across a wide collection of environments, technologies that offer standardized interfaces for the discovery process, while supporting communication for several different types of service access technologies, will provide the greatest achievable interoperability and resilience in the long-term. In this paper, we introduce a distributed architecture based on using directory services to significantly reduce the complexity ofmanaging the information and services required to support next-generation networked applications, by providing automatic service discovery and a single coherent model for representing the data managed by supporting services. Standards-based solutions are used, and a prototype implementation ofthe CORBA Naming Service that has been designed to iIIustrate how the architecture incorporates distributed object models, directory services and multicast-based dynamic service discovery is presented. Keywords: Distributed Object Architectures, Next-Generation Applications, Directory Services, Service Discovery The original version of this chapter was revised: The copyright line was incorrect. This has been ----- ###### 1. INTRO DUC TI ON Recent years have witnessed the network and distributed computing industry embracing various technologies, media and content as a direct consequence of converging technical and business interests of telecommunications networks and Internet service providers. In the long run, this will likely lead to integration of all types of communications: a bewildering array of devices such as mobiles, fixed phones, PDAs, embedded devices, workstations and pes will all be used to provide seamless communications between users and services. This has led to significant investment by many service and content providers into developing and furthering network and protocol-oriented technologies, for it is widely accepted that such a convergence willlikely result in a structurally stable core packet-switched public network such as the Internet, having many edges acting as private gateways to proprietary and subscriber-based intranets offering a multitude of services using several different kinds of technology. These private networks may be both circuit and packet switched networks and will comprise subscriber-based IP connections, enterprise nodes and services, fixed and mobile telephone connections, intelligent network services and so on. In this paper, we introduce a distributed architecture based on using directory services to significantly reduce the complexity of managing the information and services required to support next-generation networked applications, by providing automatic service discovery and a single coherent model for representing the data managed by supporting services. The architecture promises to be scalable and flexible enough to address basic issues that will arise in distributed computing as weIl as user, device and application mobility in the future. Section 2 discusses the type of network connectivity and support necessary to cope with the rising demand of mobile networked users and applications. Section 3 then sets out the aims and objectives ofthe proposed architecture while Section 4 provides an overview and a discussion of the proposed architecture, as well as describing a prototype that was implemented. Section 5 presents some performance measurements that were made, with the conclusions in Section 6. ###### 2. REQUIREMENTS FOR NETWORK SUPPORT Regardless of how services are utilized, two dominant scenarios are nevertheless heavily anticipated to influence and drive the development of next generation networks: User (and device) mobility and application mobility. ----- ###### 2.1 User mobility Network connectivity and support for user mobility is a necessary measure which is already being provisioned for in many organizations. The overall demand for mobility has indeed not shown any significant decrease, as evident by the tremendous popularity and ever-increasing adoption of low cost portable devices such as mobile computing devices, PDAs and phones for computing and networking needs. This trend will very likely be even more widespread, given that roaming agreements signed amongst mobile operators may not only inc1ude current demands for seamless GSM-based voice or data calls, but might encompass future GPRS and UMTS enabled service. In supporting user mobility, user identities, profiles, and sessions will need to be preserved across network boundaries. Very often user mobility is associated with device mobility, in which the mobile device retains its identity when roaming. The user may thus have an associated device which moves with hirn but the user may remain unaware of the underlying networking issues such as whether the device assuming temporary network addresses on roaming networks or retaining its address during base station handovers. However, some applications resident on the device may need an awareness of the current location to access local services such as printers and file servers and might thus have to perform service discovery to a certain extent. Furthermore, visiting devices mayaiso enrich the network being visited by offering their own services to the various applications and users that are resident on that particular network. Access control mechanisms would thus need to be exerted by the network hosting the user to ascertain exactly which local services are visible and what levels of service utilization are permitted. There are several technologies which exist today that inherently support user and device mobility. These include wireless LAN technologies such as 802.11 b, Bluetooth technology for piconet and scatternet oriented ad-hoc networking, as well as movement across networks as specified by mobile IP. ###### 2.2 Application mobility Application mobility can occur independently of user and device mobility, in a sense that all or certain parts of an application may autonomously or semi-autonomously mi grate across heterogenous network spaces. An application mayaiso consist of numerous non-mobile components or objects which transparently reside in several different parts of a network, with each component offering weH known services to the others. ----- Network connectivity and support for application mobility on the other hand is foreseen in the medium to long-term future. It can be regarded as the next major wave in mobility after user and device mobility, which will have an important bearing on the underlying infrastructure. Such an approach for application development is being undertaken in mobile code and agent technologies and frameworks, as weIl as distributed object-oriented architeetures. Current examples include Java programs, automatie downloads of codecs by multimedia applications, Jini[1]-enabled programs, CORBA [2] applications and the forthcoming distributed applications of the Microsoft .NET framework [3]. Location awareness will be of lesser importance for migration when compared to computing resource availability ofthe present host and network. Owing to the transparent, distributed and seemingly autonomous nature of such applications, it is rather unlikely they would enjoy as rapid and widespread an adoption across untrusted commercial networks as user and device mobility would. Rather, it is more probable that application mobility would gamer more support and fill a demand for important niche services in enterprise level computing and service provisioning, within corporate intranets as weIl as private research networks of a more academic nature. ###### 2.3 Supporting requirements Such an environment of connectivity and mobility will have a significant impact on the way next-generation applications, both autonomous and interactive are implemented, discovered, managed and accessed. One can easily perceive that both user mobility and application mobility will become killer scenarios for existing networks and services. These will create an overwhelming strain in the way current networks and legacy applications manage, interact and serve the necessary information that needs to be delivered to the end-users and applications. In both cases, a more robust supporting infrastructure for handling next-generation applications needs to be developed in a standardized manner that must also prove scalable and adapt to accommodate the research and development challenges that will manifest themselves. The overall design must possess some flexibility, as applications being built today are likely to function as part of bigger, integrated systems tomorrow. The architect of tomorrow's distributed systems will face large- scale distributions both in terms of the number of users and connected systems running atop heterogenous networks. There may not be a revolutionary change in the way networks and services will be designed for supporting next-generation applications. In fact, all indications point to the development being evolutionary, with network architects and scientists ----- examining, using and extending existing solutions and standards to support new applications and services as the need arises. The huge potential of being able to harness an increased subscriber base of a converged market has led to several sets of standards continuously being introduced for communication protocols and architectures. In many ways, the objectives and target applications to be supported by these technologies can be remarkably similar. Even the technologies may mirror each other. However, in reality, getting applications and services supported by one set of standards to communicate with another pose interoperability challenges. This also leads to an unfortunate effect of leading to fragmented or duplicate solutions. As an example, IETF's Service Location Protocol [4], Sun Microsystems' Jini, Microsoft's UPnP [5] and wireless technologies promoted by IrDA and Bluetooth all implement their own service discovery mechanisms for potentially common types of services such as printing. Another notable example would include an object location service such as CORBA and RMI naming services. Thus, a service designer targeting a product for multiple networking environments, object services or service discovery methods may be faced with the dilemma having to develop and managing several different kinds of applications, each supporting a specific protocol or object model but in all probability having duplicate information models. Instead, trying to achieve the same goal by supporting protocol bridging and conversion methods for several service discovery methods and location services, with different front-ends using a common back-end application- level information repositorystoring service records and attributes, would be a far less harrowing experience. The challenge then becomes one of developing a unified yet scalable service information management system that stores and manages service data in a standardized and generic manner, but still provides the appropriate data to requesting application via a specific discovery method, using gateways and information mapping and translation. ###### 3. AIMS AND OBJECTIVES A prototype distributed environment that would enable devices, object- oriented client applications, and services to automatically search for a particular capability, and to request and establish interoperable sessions with other clients or servers was investigated to meet the following objectives: - To provide applications, services and devices a standard method for describing and advertising·their capabilities ----- - To allow mobile users and applications the possibility to discover local services upon application start-up or movement within network space or administrative zones automatically or with as little static service discovery configuration as possible - Interworking non-invasively or transparently with existing code and applications - Ease of administration and maintenance by providing a single consistent model of available services that can scale, at the very minimum, to meet enterprise computing needs, thus reducing or eliminating the need for duplicity of information needed to support various access methods - Integration into existing or commonly prevalent network resources, supporting components and infrastructure that may already be in place, but remaining extensible enough to meet future needs - Well-known user and application access control and security mechanisms can be enforced, and existing authentication practices and encryption methods within the organization can be used. ###### 4. ARCHITECTURAL OVERVIEW Figure 1 illustrates a prototype implementation of a CORBA Naming Service [6] that has been designed to fulfill the objectives mentioned in section 3. CORBA has become increasingly important in distributed computing for the Internet aswell as telecommunications owing to its language independence and separation of the object interface away from its implementation, and the increased use of distributed objects needing fundamental CORBA based object services such as the Naming Service cannot be ignored. In this prototype, the back end service repository was implemented as a directory service using the Lightweight Directory Access Protocol (LDAP) [7]. This allowed the possibility to leverage the existing prevalence of using LDAP-based services and servers based on a well-known IETF-standard protocol already commonly deployed at the organizational level. Moreover much research is being done to make LDAP services secure, scalable, distributed and fault-tolerant. Clients would be able to retrieve information about services either natively through LDAP, or through a lightweight gateway/proxy that will implement the specific mappings necessary to support the particular type of service technology (in this case the CORBA Naming Service) the client would be using. As the figure shows, the model would be capable of supporting other kinds of front-ends as weIl. ----- ##### ,_._.-.-., CORBA , Java , _Multicast and_ Name _SLP Aware Service Agenls_ Server i JNDI/RMI i ##### ,_.-.-._., Front-end i Front-end i Microsoft i ###### ! Active , D' , . lrectory. ###### ! Front-end ! _._-- -_ ... Back-End _All TCPIIP Network_ _componeniS and_ _applications running either_ _over IPv4 01" IPv6_ _Figure I._ Prototype directory-based CORBA Naming Service The discovery of the CORBA Naming Service by CORBA client and server objects was made possible using the Service Location Protocol (SLP) as a bootstrapping protocol for multicast-based for service discovery. ###### Abrief description of SLP, LDAP and the CORBA Naming Service together with how they have been used or implemented are discussed in the following subsections. ###### 4.1 Automatic service dis'covery using SLP The Service Location Protocol (SLP) provides a flexible and scalable framework for providing hosts with access to information about the existence, location, and configuration of networked services. Traditionally, users have had to find services by knowing the name of a network host or its network address [4]. SLP eliminates the need for a user to know the name of a network host supporting a service. Rather, the user supplies the desired type of service and a set of attributes which describe the service. Based on that description, SLP resolves the network address of the service for the user. SLP models client applications as User Agents (UAs), while services are advertised by Service Agents (SAs). Applications which would like to ----- contact the SA running on the same host to register a Uniform Resource Location (URL) to be advertised, such as "service:ftp://ftp.company.coml". A service can moreover have attributes that describe the service (or the URL) in more detail. As an example, a web server would have its URL as .. service:http://www.company.com .. with attributes possibly listing the name, email address or a telephone number of the webmaster. SLP also allows services to be administratively grouped into scopes that could then be controlled for service provisioning to UAs. Client programs (UAs) that would like to find services in a network also use standardized IETF function calls to find available services in a network and are capable of querying available service types, specific service type URLs as weIl as attributes associated with a specific URL. Directory Agents (DAs) are an optional third category ofagents that SLP defines which allows for scalability by caching SA advertisements for direct interaction with UAs. Service discovery can either be statically configured or allowed to take advantage of multi casting or broadcasting features in a network for dynamic configuration. The choice of using SLP as a primary service discovery protocol was influenced by the fact that it has been specified by the IETF, and hence did not need to be adapted to be used over TCP/IP. Moreover, it is generic and independent of any particular language and object architecture. SLP's simplicity combined with its feature set also enables it to interwork with various other service discovery methods and environments such as Jini using bridging [8] and the Apple Computer's next-generation AppleTalk [9]. Much standardization work is also in progress, such as efforts in combining SLP with LDAP [10], and extensions for SLP for use over IPv6 [11]. An open-source implementation of SLP, named OpenSLP [12] that implements SLP version 2, was primarily used in developing the SLP UA and SA components of the architecture. In the later stages of development and testing, the SLP system libraries of Sun's Solaris 8 were also used successfuIly. As only a minimum configuration was involved, no Directory Agents were used, and the architecture was developed to use dynamic discovery and communication between UAs and SAs using multicasting instead. ###### 4.2 Information storage and retrieval using LDAP LDAP was originally developed as a front end to X.SOO, the OSI directory service. X.SOO defines the Directory Access Protocol (DAP) for clients to use when contacting directory servers. Currently at version 3, LDAP however has evolved to eventually provide most of the functionality of DAP at a much lower cost and defines a reasonably simple mechanism for ----- Internet clients to query and manage an arbitrary database of hierarchical attribute/value pairs over a _TCP/IP_ connection. LDAP has rapidly gaining significant Internet support, including the support of many companies, such as NovelI, Sun, HP, IBM, SOl, AT&T and Banyan, and is the focus ofmuch standardization activity in the IETF. The LDAP directory service model is based on entries. An entry is a collection of attributes that has a name, called a distinguished name (DN). The DN is used to refer to the entry unambiguously. Each of the entry's attributes has a type and one or more values. The types are typically mnemonic strings, like "cn" for common name, or "mail" for e-mail address. The values depend on what type of attribute it iso Directory entries are arranged in a hierarchical tree-like structure that reflects political, geographie, and/or organizational boundaries, representing people, organizational units, printers, documents, or just about anything else one can think of. The information model stored in the LDAP directory backend of our prototype was loosely modeled upon the campus computer science department building, in both the physical and organizational sense. The top level root entry (o=CS,c=FI) represents the computer science building which was then subdivided into four floors representing the structure, personneI, laboratories, services and IP subnets resident on these floors. The basic idea in storing service and object information in the LDAP directory in this manner was not only to store the access method ·of the object or the URL of the service, but also to append the geological or organizational location information ofthe respective services and objects. Any LDAP browser could then be used in obtaining detailed information about services offered within particular subnet or subnets (such as printers), by inspecting the relevant attributes of the LDAP entries, such as their network addresses, physical location as weIl as different access methods (network printing, BluetoothlIrDA connectivity, Appletalk, and so on). The OpenLDAP [13] project provided the necessary tools and libraries needed for this part of the architecture. ###### 4.3 Object location discovery using CORBA Naming Service The Naming Service was implemented in C++ with MICO [14], an open- source CORBA implementation which conforms to CORBA 2.3, using the DSI (Dynamic Skeleton Interface) to dynamically handle the object invocations. Apart from being designed specifically to aid in the development of gateways, the CORBA DSI functionality will allow the possibility to expand the interface of the CORBA naming service to possibly ----- serve other kinds of naming services, since type-checking is done at runtime. Clients are free to use either the DII (Dynamic Invocation Interface) or the SII (Static Invocation Interface) to invoke the methods on the server object. The CORBA objects are stored in the directory as stringified object references. LDAP offers a schema for representing CORBA Object References in an LDAP Directory [15], with each entry storing a textual description of the CORBA object as weIl as its IOR (Interoperable Object Reference). The IOR is the attribute with the single-most relevance in this case, as it is the primary means used by CORBA clients to locate and communicate with CORBA server objects in a network. Figure 2 displays the LDAP directory as seen by a freeware LDAP browser, Softerra LDAP Browser [16] running on Windows2000, after the Naming Service was launched and was used to register a simple CORBA Messaging server application, showing its various attributes, location as weIl as its DN in the window titlebar. s-!iI Browser root 8-EI o=CS,c=Fl cn=Manager ou=fOu"ttfIoor 14 20 "I ou=SqlaProcessng 29 " ou= Telecom mlTiCaUons OU=l-B420 #### I ou=HC414 ### I L cn=CorbaMessageBroker ou:people #### I - ou..prhters ##### I ou=rooms ## te:=S ##### :k: ou=users _Figure 2._ LDAP Browser showing an entry representing a CORBA object Because the CORBA Naming Service that has been specified by the Object Management Group directly supports the notion of name-to-object associations using name bindings relative to a naming context [6], an entire Naming Service is conceptually a naming graph having hierarchical relationships among parent and child nodes representing contexts and directed edges forming names. This relationship is directly supported by the hierarchical nature of an LDAP directory structure. Also, a name is comprised of a concatenation of components, where components each ----- denoting the bound object. The Naming Service specifications do not restrict the way in which the naming system should interpret, assign or manage these attributes. Thus our prototype maps these structures to be directly relevant to the entries in the LDAP tree, in such a way that the _kind_ attribute is mapped to the corresponding LDAP entry's attribute type, such as "cn" or "ou", and the _id attribute is mapped to its value. Therefore,_ if the CORBA Naming Service binds itselfto the LDAP Tree with the DN of"o=CS,c=FI", then "ou=FourthFloor,ou=Telecommunications,ou=HC414,cn=CorbaMessageBro ker" would actually be a valid name string for binding or resolving an object using our prototype Naming Service, as it would convert the given name sequence to a string representing the DN ofthe LDAP entry. At the moment, the Naming Service implements four methods: _bind,_ _unbind and rebind are used by CORBA server objects to register themselves,_ and _resolve_ is used by CORBA client objects to find the IORs of server objects corresponding to their names from the Naming Service. ###### 4.4 Example usage scenario For starting the Naming Service, the following steps are undertaken: I. OpenLDAP server daemon slapd, must first be running. 2. The OpenSLP server daemon, slpd is launched on all machines which desire to host SAs. 3. The CORBA Naming Service application is launched, and binds to a weIl known location to the LDAP back-end. This will be known as the root context of the Naming Service. At the moment, this is configured statically. However this can be easily extended to be dynamic as weIl. 4. The Naming Service registers itselfto its local slpd, passing its VRL (e.g. "service:namingservice://myhost.company.com")and its stringified IOR as an attribute. 5. The Naming Service is now ready to serve CORBA objects. Registering a CORBA server application with the Naming Service would proceed in the foIlowing manner: I. CORBA Server starts up, and behaves as a VA which multicasts a Service Request packet to the Administratively Scoped SLP Multicast Address [17], 239.255.255.253, requesting for a service type "namingservice". Currently, VAs use default scopes with a multicast packet TTL value of 8. 2. The SA responsible for the Naming Service unicasts a Service Reply back to the requesting VA, retuming its URL and thus specifying the location of the Naming Service. ----- 3. Armed with the VRL, the VA once again multicasts an Attribute Request packet requesting for the IOR attribute for that VRL. 4. The SA once again unicasts directly to the VA with an Attribute Reply, furnishing it with the IOR attribute registered with that VRL. 5. Armed with enough knowledge now, the CORBA Server application contacts the Naming Service directly with the IOR it possesses, and registers itself with the Naming Service using the bind call with 2 parameters: its name and its IOR. 6. The Naming Service maps the information into an LDAP call and stores the information in its LDAP database backend. In order for a CORBA client finding a CORBA server, it does the following: 1. The CORBA client first discovers the Naming Service in a similar way to the CORBA Server. 2. The Client calls the resolve call of the Naming Service, passing the name of the CORBA server object as a parameter and obtaining an IOR from the Naming Service with which the c1ient then invokes methods on the CORBA server object. ###### 5. RESULTS The basic network topology used for developing and testing the prototype implementation is depicted in Figure 3. All 3 subnets were multicast aware, with the FreeBSD machine hosting the LDAP, Naming Service, CORBA server objects and SLP Service Agents. As an initial indication of the performance of the prototype, the timing measurements that were recorded are tabulated in Table 1. All SLP and CORBA c1ient calls in Table 1 were made from the SunBlade machine to the FreeBSD machine. For the Naming Service, the time taken to execute a single resolve call was measured. The time taken to execute the same call was also measured against MICO's native naming service, which runs as a standalone daemon in the network. For SLP measurements we timed the discovery of the Naming Service. No DAs were employed in this topology, hence communication occurred directly between the _VA_ in the SunBlade machine and the SA in the FreeBSD machine. The measurements were taken over aperiod of approximately 3 days. ----- Cisco 6509 Integrated Pentlum 133MHz/64MB Switch/Router FreeBSD 4.2-STABLE Cisco 7206 VXR Subnet B Router IOOMbps IOOMbps Fast-Ethernet Fast-Ethernet IOOMbps SunBlade 100/512MB UltraSPARC lIe. Solarls 8 Fast-Ethernet Subnet C IOOMbps IOOMbps Fast-Ethernet Fast-Ethernet Cisco3548 Switch _Figure 3._ Physical Network Topology Approximately two-thirds of the execution time of the resolve call was consumed by the LDAP operation, while a small fraction was consumed in the Naming Service front-end that serves as a gateway by mapping parameters and return results to and from the CORBA Naming Service API and the OpenLDAP API. On the other hand, to use MICO's own Naming Service, it is necessary to initially ron an object adapter daemon called micod, create an entry for the naming service in an implementation repository and pass the c1ients the address of the naming service. A fair bit of manual intervention, static configuration and using commandline options are hence necessary. Also, the entries of the Naming Service are stored in memory. _Table J._ Timing measurements in milliseconds Prototype MICO's Naming Service SLP Average duration 36.5843 3.60773 4.24144 Maximum duration 148.506 34.252 10.203 Minimum duration 34.289 3.213 3.960 Thus, bearing in mi nd the objectives laid out in Section 3, we firmly believe that the offset in performance of our prototype implementation is justified by virtue of the scalable replicating and fault-tolerant properties ----- inherent from the LDAP component of the architecture. Automatic service discovery is also ideal for object services, owing to their more volatile and migratory nature as compared to fixed network services. The design of the architecture with the components described also allows for the flexible interaction amongst them, depending on the needs of the organization. As an example, the OpenLDAP server could be modified to become SLP-aware so that even the Naming Service front-end could perform service discovery to dynamically discover its backend, as opposed to the configuration that our prototype used. Interworking between SLP and LDAP is being studied, so that if Directory Agents are used in the architecture, SLP service URLs and attributes can be stored in an LDAP Directory. [10] ###### 6. CONCLUSIONS Almost all the software components used in designing, buildingand testing the prototype implementation have been based on code obtained from highly active open-source projects. This has made troubleshooting and bug- solving relatively easy through open mail-list forums and by simply browsing through the code. Minor problems, such as UAs in OpenSLP insisting on multicasting to search for DAs even when configured not to, were quickly solved, and more optimal service discovery times were achieved (from a very rough estimate, UAs using Solaris's SLP library for service discovery took far longer, about 12 seconds). At the moment, SLP over IPv6 within our prototype has yet to be tested. However, the following are changes required to have the Service Location Protocol work over IPv6. These changes include [11]: - Eliminating support for broadcast SLP requests - Address Specification for IPv6 Addresses in URLs - Use of IPv6 multicast addresses and IPv6 addressscopes - Restricted Propagation of Service Advertisements The architecture proposed in this paper also does not preclude the use of any security models and can remain fully conformant to any and all security mechanisms standardized by the many RFCs and other specification documents for its various components. As an example, SLP authentication and LDAP security and access control mechanisms can be used to enhance the architecture to allow employees and regular users to use all standard services of the organization's internal network, but restricting visitors and anonymous users to a smaller sub set, perhaps by presenting a different location service front-end during service discovery which has a more limited level of access and visibility of the LDAP directory. ----- Being able to store, access and manipulate data stored in a common directory format with LDAP also has a huge advantage in being able to use common LDAP browsers (with varying levels of quality) supported in many platforms. This allows a far greater ease of management and maintenance than having to maintain several different service models, each having their own customized administration tools. This also significantly reduces the leaming curve for efficient tool use. The network computing industry has eagerly embraced technologies, welcoming an ever-increasing variety of new service discovery protocols and object architectures. With this abundance now offered across a wide collection of environments, technologies that offer standardized interfaces for the discovery process, while supporting communication for several different types of service access technologies, will provide the greatest achievable interoperability and resilience in the long-term. In this respect, the proposed architecture in this paper holds good promise for supporting enterprise-Ievel next-generation computing needs. ###### REFERENCES [1] Sun Microsystems: Jini Network Techno1ogy, http://www.sun.comljini/ [2] OMG: The Common Object Request Broker: Architecture and Specification. CORBA V2.3, June 1999. [3] Microsoft Corporation: .NET http://msdn.microsoft.comlnet [4] IETF: RFC 2608, "Service Location Protoco1, Version 2", June 1999. [5] Microsoft Corporation: Universal P1ug and Play forum, http://www.upnp.org [6] OMG: Naming Service Specification, February 2001. [7] IETF: RFC 2251, "Lightweight Directory Access Protocol (v3)", Dec 1997. [8] Erik Guttman, James Kempf: Automatic Discovery of Thin Servers: SLP, Jini and the SLP-Jini Bridge, IECON, San Jose, 1999. [9] Apple Computer Inc.: Mac OS X Server - Network & Security http://www.app1e.comlmacosxl server/networksecurity .html [lO]IETF: RFC 2609, "Service Temp1ates and Service: Schemes", June 1999. [11]IETF: RFC 3111, "Service Location Protocol Modifications for IPv6", May 2001. [12] The OpenSLP Project, http://www.opens1p.org [l3]The OpenLDAP Project, http://www.openldap.org [14]MICO - Mico Is COrba http://www mico org ----- [15]IETF: RFC 2714, 11 Schema for Representing CORBA Objeet Referenees in an LDAP Direetory", Oetober 1999. [16] Softerra LDAP Browser, http://www.ldapadministrator.com [17] IETF: RFC 2365, "Administratively Scoped IP Multicast", July 1998. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-0-387-35584-9_5?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-0-387-35584-9_5, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://link.springer.com/content/pdf/10.1007/978-0-387-35584-9_5.pdf" }
2,002
[ "JournalArticle" ]
true
2002-04-08T00:00:00
[]
7,739
en
[ { "category": "Law", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/015d50faf4b17a5b7a910e87a7327d9a7896e62d
[]
0.90359
It's Time to Regulate Stablecoins as Deposits and Require Their Issuers to Be FDIC-Insured Banks
015d50faf4b17a5b7a910e87a7327d9a7896e62d
Social Science Research Network
[ { "authorId": "119536238", "name": "Arthur E. Wilmarth" } ]
{ "alternate_issns": null, "alternate_names": [ "SSRN, Social Science Research Network (SSRN) home page", "SSRN Electronic Journal", "Soc Sci Res Netw", "SSRN", "SSRN Home Page", "SSRN Electron J", "Social Science Electronic Publishing presents Social Science Research Network" ], "alternate_urls": [ "www.ssrn.com/", "https://fatcat.wiki/container/tol7woxlqjeg5bmzadeg6qrg3e", "https://www.wikidata.org/wiki/Q53949192", "www.ssrn.com/en", "http://www.ssrn.com/en/", "http://umlib.nl/ssrn", "umlib.nl/ssrn" ], "id": "75d7a8c1-d871-42db-a8e4-7cf5146fdb62", "issn": "1556-5068", "name": "Social Science Research Network", "type": "journal", "url": "http://www.ssrn.com/" }
null
[GW Law Faculty Publications & Other Works](https://scholarship.law.gwu.edu/faculty_publications) [Faculty Scholarship](https://scholarship.law.gwu.edu/faculty_scholarship) 2021 # It's Time to Regulate Stablecoins as Deposits and Require Their It's Time to Regulate Stablecoins as Deposits and Require Their Issuers to Be FDIC-Insured Banks Issuers to Be FDIC-Insured Banks Arthur E. Wilmarth Jr. George Washington University Law School, awilmarth@law.gwu.edu [Follow this and additional works at: https://scholarship.law.gwu.edu/faculty_publications](https://scholarship.law.gwu.edu/faculty_publications?utm_source=scholarship.law.gwu.edu%2Ffaculty_publications%2F1576&utm_medium=PDF&utm_campaign=PDFCoverPages) [Part of the Law Commons](http://network.bepress.com/hgg/discipline/578?utm_source=scholarship.law.gwu.edu%2Ffaculty_publications%2F1576&utm_medium=PDF&utm_campaign=PDFCoverPages) Recommended Citation Recommended Citation 41 Banking & Financial Services Policy Report No. 2 (Feb. 2022), at 1-20. This Article is brought to you for free and open access by the Faculty Scholarship at Scholarly Commons. It has been accepted for inclusion in GW Law Faculty Publications & Other Works by an authorized administrator of [Scholarly Commons. For more information, please contact spagel@law.gwu.edu.](mailto:spagel@law.gwu.edu) ----- **It’s Time to Regulate Stablecoins as Deposits and** **Require Their Issuers to Be FDIC-Insured Banks** Arthur E. Wilmarth, Jr.* December 16, 2021 **Introduction** In November 2021, the President’s Working Group on Financial Markets (PWG) issued a report analyzing the rapid expansion and growing risks of the stablecoin market. As explained in PWG’s report, “[s]tablecoins are digital assets that are designed to maintain a stable value relative to a national currency or other reference assets.”[1] PWG’s report determined that stablecoins pose a wide range of potential hazards, including the risks of inflicting large losses on investors, destabilizing financial markets and the payments system, supporting money laundering, tax evasion, and other forms of illicit finance, and promoting dangerous concentrations of economic and financial power. PWG’s report called on Congress to pass legislation that would (i) require all issuers of stablecoins to be banks that are insured by the Federal Deposit Insurance Corporation (FDIC), and (ii) “ensure that payment stablecoins are subject to appropriate federal prudential oversight on a consistent and comprehensive basis.” PWG also recommended that federal agencies and the Financial Stability Oversight Council (FSOC) should use their “existing authorities” to “address risks associated with payment stablecoin arrangements . . . to the extent possible.”[2] *Professor Emeritus of Law, George Washington University Law School. 1 President’s Working Group on Financial Markets et al., Report on Stablecoins (Nov. 2021) (quote at 1) [[hereinafter PWG Stablecoin Report], https://home.treasury.gov/system/files/136/StableCoinReport_Nov1_508.pdf;](https://home.treasury.gov/system/files/136/StableCoinReport_Nov1_508.pdf) _see also Alexis Goldstein, Written Testimony before the Senate Comm. on Banking, Housing, and Urban Affairs 1_ (Dec. 14, 2021) (“Stablecoins are crypto assets that attempt to maintain a stable value, either through a basket of reserve assets acting as collateral (asset-backed stablecoins), or through algorithms (algorithmic stablecoins).”) [[hereinafter Goldstein Testimony], https://www.banking.senate.gov/imo/media/doc/Goldstein%20Testimony%2012-](https://www.banking.senate.gov/imo/media/doc/Goldstein%20Testimony%2012-14-21.pdf) [14-21.pdf. This paper focuses on asset-backed stablecoins, which account for most of the stablecoin market.](https://www.banking.senate.gov/imo/media/doc/Goldstein%20Testimony%2012-14-21.pdf) 2 PWG Stablecoin Report, supra note 1, at 16, 18. ----- At present, stablecoins are mainly used to make payments for trades in cryptocurrency markets and to provide collateral for derivatives and lending transactions involving cryptocurrencies.[3] However, technology companies are exploring a much broader range of potential uses for stablecoins, including their use as digital currencies for making purchases and sales of goods and services as well as person-to-person payments. In October, Facebook launched a “pilot” of its Novi “digital currency wallet,” which uses the Pax Dollar stablecoin as its first digital currency.[4] Novi enables its customers to make person-to-person payments within and across national borders and is part of Facebook’s larger plan to establish itself as “a challenger in the payments system.”[5] Facebook intends to “migrate” Novi’s digital wallet to its proposed Diem stablecoin as soon as Facebook receives regulatory approvals for Diem.[6] Facebook’s launch of Novi indicates that stablecoins could potentially become a form of “private money” that is widely used in consumer and commercial transactions. Federal agencies have not yet issued rules governing the issuance and distribution of stablecoins. Federal and state officials have only rarely enforced consumer and investor protection laws against issuers and distributors of stablecoins. PWG’s report calls on federal agencies and Congress to take 3 _Id. at 7-10; see also Andrew Ackerman, “Stablecoins in Spotlight as U.S. Begins to Lay Ground for Rules on_ Cryptocurrencies,” Wall Street Journal (Sept. 25, 2021) (“For now, stablecoins are used mainly by investors to buy and sell crypto assets on exchanges . . . [and] as collateral for derivatives”), [https://www.wsj.com/articles/stablecoins-in-spotlight-as-u-s-begins-to-lay-ground-for-rules-on-cryptocurrencies-](https://www.wsj.com/articles/stablecoins-in-spotlight-as-u-s-begins-to-lay-ground-for-rules-on-cryptocurrencies-11632562202?mod=article_inline) [11632562202?mod=article_inline; Statement by SEC Chair Gary Gensler, “President’s Working Group Report on](https://www.wsj.com/articles/stablecoins-in-spotlight-as-u-s-begins-to-lay-ground-for-rules-on-cryptocurrencies-11632562202?mod=article_inline) [Stablecoins” (Nov. 1, 2021) [hereinafter Gensler Statement], https://www.sec.gov/news/statement/gensler-](https://www.sec.gov/news/statement/gensler-statement-presidents-working-group-report-stablecoins-110121) [statement-presidents-working-group-report-stablecoins-110121 (stating that “more than 75 percent of trading on all](https://www.sec.gov/news/statement/gensler-statement-presidents-working-group-report-stablecoins-110121) crypto trading platforms occurred between a stablecoin and some other token” in October 2021). 4 See infra notes 17-21, 45-47 and accompanying text. 5 Hannah Murphy & Siddarth Venkataramakrishnan, “Facebook says ready to launch digital wallet,” Financial _Times (Aug. 18, 2021) (quoting David Marcus, who was then the leader of Facebook’s Novi project),_ [https://www.ft.com/content/a8512417-1fde-481a-b282-2f892e3c3b51; Siddarth Venkataramakrishnan & Hannah](https://www.ft.com/content/a8512417-1fde-481a-b282-2f892e3c3b51) Murphy, “Facebook launches digital wallet Novi,” Financial Times (Oct. 19, 2021), [https://www.ft.com/content/b9a61950-a32c-4c77-95fe-fc0d00021a0f.](https://www.ft.com/content/b9a61950-a32c-4c77-95fe-fc0d00021a0f) 6 Venkataramakrishnan & Murphy, supra note 5 (quoting David Marcus). 2 ----- immediate steps to establish a federal oversight regime that could respond effectively to the dangers created by stablecoins.[7] This paper strongly supports three regulatory approaches recommended in PWG’s report. First, the Securities and Exchange Commission (SEC) should use its available powers to regulate stablecoins as “securities” and protect investors and securities markets. However, the scope of the SEC’s authority to regulate stablecoins is not clear, and federal securities laws do not provide adequate safeguards to control the systemic threats that stablecoins pose to financial stability and the payments system. Second, the Department of Justice (DOJ) should designate stablecoins as “deposits” and should bring enforcement actions to prevent issuers and distributors of stablecoins from unlawfully receiving “deposits” in violation of Section 21(a) of the Glass-Steagall Act. Section 21(a) offers a promising avenue for regulatory action, but its provisions contain uncertainties and gaps and do not provide a complete remedy for the hazards created by stablecoins. The most significant gap in Section 21(a) allows state (and possibly federal) banking authorities to charter special-purpose depository institutions that could issue and distribute stablecoins without obtaining deposit insurance from the FDIC. Third, Congress should adopt legislation mandating that all issuers and distributors of stablecoins must be FDIC-insured banks. That requirement would compel all stablecoin issuers 7 PWG Stablecoin Report, supra note 1, at 2-3, 10-18. For additional discussions of the dangers created by stablecoins and possible regulatory responses to those perils, see Ackerman, supra note 3; Testimony by Hilary J. Allen before the Senate Comm. on Banking, Housing, and Urban Affairs (Dec. 14, 2021) [hereinafter Allen [Testimony], https://www.banking.senate.gov/imo/media/doc/Allen%20Testimony%2012-14-211.pdf; Dan Awrey,](https://www.banking.senate.gov/imo/media/doc/Allen%20Testimony%2012-14-211.pdf) _Bad Money, 106 Cornell Law Review 1, 6-8, 39-45 (2020); Nate DiCamillo, “The US is dragging its heels on critical_ [stablecoin regulations,” Quartz (Nov. 8, 2021), https://qz.com/2083636/what-are-stablecoins-and-how-will-they-be-](https://qz.com/2083636/what-are-stablecoins-and-how-will-they-be-regulated/) [regulated/; Gary Gorton & Jeffrey Y. Zhang, “Taming Wildcat Stablecoins” (Sept. 30, 2021), available at](https://qz.com/2083636/what-are-stablecoins-and-how-will-they-be-regulated/) [http://ssrn.com/abstract=3888752; Jeanna Smialek, “Why Washington Worries About Stablecoins,” New York Times](http://ssrn.com/abstract=3888752) [(Sept. 23, 2021), https://www.nytimes.com/2021/09/17/business/economy/federal-reserve-virtual-currency-](https://www.nytimes.com/2021/09/17/business/economy/federal-reserve-virtual-currency-stablecoin.html) [stablecoin.html.](https://www.nytimes.com/2021/09/17/business/economy/federal-reserve-virtual-currency-stablecoin.html) 3 ----- and distributors and their parent companies to comply with federal laws that protect the safety, soundness, and stability of our banking system and obligate banks to operate in a manner consistent with the public interest. Requiring stablecoin issuers and distributors to be FDIC insured banks would also maintain the longstanding U.S. policy of separating banking and commerce. It would prevent Facebook and other Big Tech firms from using stablecoin ventures as building blocks for “shadow banking” empires that would erode consumer protections, impair competition, subvert the effectiveness of financial regulation, and potentially unleash systemic crises across our financial and commercial sectors during severe economic downturns and financial disruptions. **Analysis** **1.** **The Rapid Expansion and Escalating Risks of Stablecoins** The volume of outstanding stablecoins has mushroomed during the past two years, growing from less than $6 billion in January 2020 to $150 billion in December 2021.[8] The rapid expansion of the stablecoin market has mirrored the explosive growth of all cryptocurrency markets. The total market capitalization of cryptocurrencies increased almost nine-fold – from $350 billion to $3 trillion – between September 2020 and November 2021.[9] At present, stablecoins are mainly used to speculate in cryptocurrencies and other digital assets. Stablecoins are the leading form of payment for trades executed on cryptocurrency exchanges, and stablecoins are used as collateral for derivatives and lending transactions involving digital assets. In addition, stablecoins play “a central role in facilitating trading, 8 The Block, “Stablecoins: Total Stablecoin Supply” (visited on Dec. 15, 2021), [https://www.theblockcrypto.com/data/decentralized-finance/stablecoins.](https://www.theblockcrypto.com/data/decentralized-finance/stablecoins) 9 Office of Financial Research, Annual Report to Congress 2021, at 49 [hereinafter OFR 2021 Annual Report], [https://www.financialresearch.gov/annual-reports/files/OFR-Annual-Report-2021.pdf; Yvonne Lau,](https://www.financialresearch.gov/annual-reports/files/OFR-Annual-Report-2021.pdf) “Cryptocurrencies hit market cap of $3 trillion for the first time as Bitcoin and Ether reach record highs,” Fortune [(Nov. 9, 2021), https://fortune.com/2021/11/09/cryptocurrency-market-cap-3-trillion-bitcion-ether-shiba-inu/.](https://fortune.com/2021/11/09/cryptocurrency-market-cap-3-trillion-bitcion-ether-shiba-inu/) 4 ----- lending, and borrowing activity” in decentralized finance (DeFi) transactions. DeFi transactions are completed by using smart contracts and “autonomous” distributed ledgers instead of organized exchanges.[10] Stablecoins enable participants to trade in cryptocurrencies and engage in other digital asset transactions while avoiding the use of fiat currencies and traditional financial institutions. Stablecoins offer a much higher degree of anonymity in conducting such transactions, and many participants use stablecoins to avoid complying with “Know Your Customer” (KYC) requirements, anti-money laundering (AML) laws, tax laws, and sanctions against terrorist financing.[11] “[V]irtually no KYC/AML checks” are conducted for DeFi transactions, and criminals can “launder proceeds of crime” by exchanging other assets for stablecoins (or vice versa) while “hiding the blockchain money trail.”[12] Tether, the largest issuer of stablecoins, has issued over $80 billion of stablecoins and controls a majority of the stablecoin market.[13] Tether and the issuers of most other leading stablecoins represent to the public that they hold sufficient “reserves” to maintain a 1-for-1 parity 10 PWG Report, supra note 1, at 5-10 (quote at 9); see also Allen Testimony, supra note 7, at 2-3, 6-14; Goldstein Testimony, supra note 1, at 1-5, 10-13; Gary Silverman, “Cryptocurrency: rise of decentralized finance sparks ‘dirty [money’ fears,” Financial Times (Sept. 15, 2021), https://www.ft.com/content/beeb2f8c-99ec-494b-aa76-](https://www.ft.com/content/beeb2f8c-99ec-494b-aa76-a7be0bf9dae6) [a7be0bf9dae6.](https://www.ft.com/content/beeb2f8c-99ec-494b-aa76-a7be0bf9dae6) 11 PWG Report, supra note 1, at 1-2, 10-11, 19-21; Gensler Statement, supra note 3; Goldstein Testimony, supra note 1, at 1-2, 5, 13-15; Smialek, supra note 7; see also Zeke Faux, “Anyone Seen Tether’s Billions?”, Bloomberg _BusinessWeek (Oct. 7, 2021) (“Tether Holdings checks the identity of people who buy coins directly from the_ company, but once the currency is out in the world, it can be transferred anonymously, just by sending a code. A drug lord can hold millions of Tethers in a digital wallet and send it to a terrorist without anyone knowing.”), [https://www.bloomberg.com/news/features/2021-10-07/crypto-mystery-where-s-the-69-billion-backing-the-](https://www.bloomberg.com/news/features/2021-10-07/crypto-mystery-where-s-the-69-billion-backing-the-stablecoin-tether?sref=f7rH2jWS) [stablecoin-tether?sref=f7rH2jWS; JP Koning, “What Happens If All Stablecoin Users Have to Be Identified?”,](https://www.bloomberg.com/news/features/2021-10-07/crypto-mystery-where-s-the-69-billion-backing-the-stablecoin-tether?sref=f7rH2jWS) _CoinDesk (Sept. 14, 2021) (“Right now, a large chunk of stablecoin usage is pseudonymous. That is, you or I can_ hold $20,000 worth of tether or USD coin stablecoins in an unhosted wallet (i.e., not on an exchange), without [having to provide our identities to either Tether or Circle.”), https://www.coindesk.com/policy/2021/02/18/what-](https://www.coindesk.com/policy/2021/02/18/what-happens-if-all-stablecoin-users-have-to-be-identified/) [happens-if-all-stablecoin-users-have-to-be-identified/; Silverman, supra note 10 (reporting on the belief of some](https://www.coindesk.com/policy/2021/02/18/what-happens-if-all-stablecoin-users-have-to-be-identified/) cryptocurrency entrepreneurs that “DeFi innovations . . . will enable them to break free of [KYC] obligations”). 12 Goldstein Testimony, supra note 1, at 5, 14 (including quotes from a report, dated Oct. 18, 2021, by Elliptic, a cryptocurrency compliance firm); see also Silverman, supra note 10 (reporting that DeFi “allows a wave of innovation by people trying to launder money through the system”) (quoting David Jevans, CEO of CipherTrace, a cryptocurrency intelligence company). 13 The Block, supra note 8. 5 ----- between their stablecoins and the U.S. dollar. However, there are substantial doubts about the adequacy of reserves held by Tether and other issuers. Tether and its affiliates paid more than $60 million to settle charges filed by the Office of the New York Attorney General and the Commodity Futures Trading Commission, alleging that Tether’s representations about its reserves were false and materially misleading.[14] In 2021, Tether disclosed that a majority of its reserves consisted of commercial paper and other corporate obligations (reportedly including debts of Chinese companies). At best, the reserves of Tether and other leading stablecoin issuers resemble the assets held by prime money market funds, which experienced systemic investor runs and were forced to accept bailouts from the federal government in 2008 and 2020.[15] As discussed in PWG’s report, some technology companies have “the stated ambition” to create stablecoin programs that can be “used widely by retail users to pay for goods and services, by corporations in the context of supply chain payments, and in the context of international remittances.”[16] Facebook’s launch of Novi in October 2021 is a “pilot” for Facebook’s planned creation of a global digital payments network that will ultimately use Facebook’s proposed Diem stablecoin. Novi is available initially to customers in the U.S. and Guatemala, and its first digital currency is the stablecoin USDP (Pax Dollar), issued by Paxos.[17] 14 Commodity Futures Trading Commission, Press Release No. 8450-21, “CFTC Orders Tether and Bitfinex to Pay [Fines Totaling $42.5 Million” (Oct. 15, 2021), https://www.cftc.gov/PressRoom/PressReleases/8450-21; Office of](https://www.cftc.gov/PressRoom/PressReleases/8450-21) the N.Y. Attorney General, Press Release, “Attorney General James Ends Virtual Currency Trading Platform Bitfinex’s Illegal Activities in New York” (Feb. 23, 2021) (imposing an $18.5 million fine on Tether and its [affiliates), https://ag.ny.gov/press-release/2021/attorney-general-james-ends-virtual-currency-trading-platform-](https://ag.ny.gov/press-release/2021/attorney-general-james-ends-virtual-currency-trading-platform-bitfinexs-illegal) [bitfinexs-illegal.](https://ag.ny.gov/press-release/2021/attorney-general-james-ends-virtual-currency-trading-platform-bitfinexs-illegal) 15 Faux, supra note 11; Goldstein Testimony, supra note 1, at 2-5; OFR 2021 Annual Report, supra note 9, at 51-52; Gorton & Zhang, supra note 7, at 6-16, 21-24; Bill Nelson & Paige Pidano Paridon, “Stablecoins are backed by ‘reserves’? Give us a break,” American Banker (Dec. 10, 2021), available at 2021 WLNR 40403852; Arthur E. Wilmarth, Jr., “The Pandemic Crisis Shows That the World Remains Trapped in a ‘Global Doom Loop’ of Financial Instability, Rising Debt Levels, and Escalating Bailouts,” 40 Banking & Financial Services Policy Report No. 8 [(Aug. 2020), at 1, 9-10, available at https://ssrn.com/abstract=3901967 [hereinafter Wilmarth, “Pandemic Crisis”];](https://ssrn.com/abstract=3901967) Yueqi Yang, “Tether Fails to Dispel Mystery on Stablecoin’s Crucial Reserves,” Bloomberg Law (Dec. 3, 2021). 16 PWG Report, supra note 1, at 8. 17 Venkataramakrishnan & Murphy, supra note 5; see also Novi Financial, Inc., “Meet Novi,” [https://www.novi.com/.](https://www.novi.com/) 6 ----- According to Facebook, “Novi is a digital wallet that helps you send and receive money instantly and securely.” Novi’s customers can send and receive payments by using “digital currencies, starting with USDP (Pax Dollar). When you add money to your Novi account, we’ll convert it to USDP. On Novi, 1 USDP is equal to 1 US dollar.”[18] Novi’s terms of service allow customers to redeem their stablecoins from Novi based on the same 1-for-1 parity between the Pax Dollar and the U.S. dollar.[19] Facebook’s launch of Novi indicates that stablecoins could potentially expand from their current roles in cryptocurrency trading and other digital asset transactions to a much broader range of uses in consumer and commercial transactions. On October 19, 2021, David Marcus, who was then head of Novi, described Facebook’s ambitions to create a general-use digital payments network: Beyond the pilot, our business model is clear. We’re a challenger in payments. We’ll offer free person-to-person payments using Novi. Once we have a solid customer base, we’ll offer cheaper merchant payments and make a profit on merchant services.[20] Marcus explained that “our support for Diem hasn’t changed and we intend to launch Novi with Diem once it receives regulatory approval.” Marcus also confirmed that Novi planned to offer “interoperability” in payments between Facebook’s Diem, Pax Dollar, and other stablecoins.[21] [18 Novi Financial, Inc., “Novi: How It Works,” https://www.novi.com/how-it-works.](https://www.novi.com/how-it-works) 19 Novi Financial, Inc., “Terms of Service (Last Modified: October 19, 2021)” [hereinafter Novi Terms of Service], - 3 (“User Redemption Right”) (“You are entitled to redeem each Digital Currency for one U.S. dollar (USD) with [Novi.”), https://www.novi.com/legal/app/us/terms-of-service?temp_locale=en_US](https://www.novi.com/legal/app/us/terms-of-service?temp_locale=en_US) [20 Tweet by David Marcus (Oct. 19, 2021), https://twitter.com/davidmarcus/status/1450447444379013122 (visited](https://twitter.com/davidmarcus/status/1450447444379013122) on Dec. 15, 2021). 21 _Id.; see also Venkataramakrishnan & Murphy, supra note 5 (reporting on Marcus’ statements about Facebook’s_ ambitions for Novi). 7 ----- As discussed in PWG’s report, stablecoins present a wide array of potential hazards, including deceptive marketing, fraudulent and manipulative trading, abusive and predatory terms, and facilitating evasion of KYC/AML requirements, tax laws, and sanctions against terrorist financing.[22] This paper focuses on four systemic dangers posed by stablecoins, which are also analyzed in PWG’s report. First, investors in stablecoins could suffer large losses from investor runs triggered by concerns about the adequacy of stablecoin reserves. Investor runs on stablecoins would likely resemble the investor runs that occurred in 2008 and 2020 in prime money market funds, which invest (like stablecoins) in securities that are not issued or guaranteed by the federal government. Stablecoins are also similar to the private banknotes that state-chartered banks issued before the Civil War. Many state-chartered banks experienced runs by holders of their banknotes during that period because they did not hold adequate reserves and their notes were not guaranteed by the federal government.[23] Second, the collapse of a major stablecoin could destabilize financial markets. For example, a default by Tether – the leading form of payment used in cryptocurrency transactions – would probably cause widespread trading failures as well as fire sales in cryptocurrency markets. If stablecoins became a widely-accepted medium of payment for purchases and sales of goods and services in the general economy, the failure of a leading stablecoin could trigger a 22 PWG Stablecoin Report, supra note 1, at 1-2, 10-11, 19-21; Goldstein Testimony, supra note 1, at 5-15; see also Letter from Open Markets Institute to federal regulatory agencies, dated Nov. 23, 2021, expressing concerns about “Facebook’s Digital Asset Wallet Pilot” [hereinafter Open Markets Facebook Letter], at 1-7, [https://www.openmarketsinstitute.org/publications/letter-to-regulators-grave-risks-of-facebook-digital-wallet-pilot.](https://www.openmarketsinstitute.org/publications/letter-to-regulators-grave-risks-of-facebook-digital-wallet-pilot) 23 PWG Stablecoin Report, supra note 1, at 1-2, 10-12; see also Ackerman, supra note 3; Awrey, supra note 7, at 36, 11-18, 33-39; Gorton & Zhang, supra note 7, at 21-31; James Mackintosh, “Bitcoin’s Reliance on Stablecoins Harks Back to the Wild West of Finance,” Wall Street Journal (May 27, 2021), [https://www.wsj.com/articles/bitcoins-reliance-on-stablecoins-harks-back-to-the-wild-west-of-finance-](https://www.wsj.com/articles/bitcoins-reliance-on-stablecoins-harks-back-to-the-wild-west-of-finance-11622115246) [11622115246.; Arthur J. Rolnick & Warren E. Weber, “Free Banking, Wildcat Banking, and Shinplasters,” 6](https://www.wsj.com/articles/bitcoins-reliance-on-stablecoins-harks-back-to-the-wild-west-of-finance-11622115246) _Quarterly Review No. 3, at 10-19 (Fed. Res. Bank of Minneapolis, Fall 1982)._ 8 ----- generalized run on stablecoins that might shut down the payments system and inflict widespread losses on consumers, business firms, and financial institutions.[24] Third, issuers and distributors of stablecoins are rapidly becoming a new category of systemically important “shadow banks.” Shadow banks provide functional substitutes for deposits (“shadow deposits”) and offer other financial services that mimic the activities of banks while avoiding compliance with federal laws that establish essential safeguards for the safety, soundness, and stability of our banking system. The systemic significance of stablecoin issuers would increase exponentially if stablecoins are widely accepted as a medium of payment in consumer and commercial transactions. Under those circumstances, stablecoins would likely become a systemically important form of “private money” comparable to money market funds, which do not have explicit government backing but rely on general expectations of government support during severe economic downturns or financial crises.[25] Fourth, issuers and distributors of stablecoins are permitted to combine their financial activities with commercial ventures because they are not defined as “banks” for purposes of the Bank Holding Company Act (BHC Act). Like other shadow banks, issuers and distributors of stablecoins are not subject to the BHC Act’s longstanding policy of separating banking and commerce.[26] As the PWG’s report correctly pointed out, 24 PWG Stablecoin Report, supra note 1, at 1-3, 12-14; see also OFR 2021 Annual Report, supra note 9, at 49-54; Sam Knight, “Biden Administration Is Playing With Fire by Failing to Regulate Cryptocurrency,” Truthout (Nov. [16, 2021), https://truthout.org/articles/biden-administration-is-playing-with-fire-by-failing-to-regulate-](https://truthout.org/articles/biden-administration-is-playing-with-fire-by-failing-to-regulate-cryptocurrency/) [cryptocurrency/.](https://truthout.org/articles/biden-administration-is-playing-with-fire-by-failing-to-regulate-cryptocurrency/) 25 PWG Stablecoin Report, supra note 1, at 1-3, 7-14; see also Gorton & Zhang, supra note 7, at 3-6, 21-24, 33, 38; Wilmarth, “Pandemic Crisis,” supra note 15, at 6-13, 16-17; Arthur E. Wilmarth, Jr., Taming the Megabanks: Why _We Need a New Glass-Steagall Act 150-57, 279-88, 341-44, 353-56 (Oxford Univ. Press, 2020) [hereinafter_ Wilmarth, Taming the Megabanks]. 26 12 U.S.C. §§ 1841(c), 1843; Arthur E. Wilmarth, Jr., “The OCC’s and FDIC’s Attempts to Confer Banking Privileges on Nonbanks and Commercial Firms Violate Federal Laws and Are Contrary to Public Policy,” 39 _Banking & Financial Services Policy Report No. 10 (Oct. 2020), at 1, 6-11, available at_ [https://ssrn.com/abstract=3750964 [hereinafter Wilmarth, “Banking Privileges”]; see also Gorton & Zhang, supra](https://ssrn.com/abstract=3750964) note 7, at 17-19. 9 ----- [T]he combination of a stablecoin issuer or wallet provider and a commercial firm could lead to an excessive concentration of economic power. These policy concerns are analogous to those traditionally associated with the mixing of banking and commerce, such as advantages in accessing credit or using data to market or restrict access to products. This combination could have detrimental effects on competition and lead to market concentration in sectors of the real economy.[27] As explained below in Part 2(c), permitting issuers and distributors of stablecoins to operate without being chartered and regulated as FDIC-insured banks would enable Facebook and other Big Tech firms to enter the banking business and undermine the BHC Act’s policy of separating banks from commercial enterprises. Allowing Big Tech firms to subvert that policy would inflict great harm on our financial system, economy, and society. **2.** **Regulatory Strategies for Controlling the Dangers of Stablecoins** This section strongly endorses three regulatory approaches discussed in PWG’s report for addressing the perils created by stablecoins. First, the SEC should use its existing powers to regulate stablecoins as “securities” and protect investors and securities markets. Second, DOJ should designate stablecoins as “deposits” and bring enforcement actions to prevent issuers and distributors of stablecoins from violating Section 21(a) of the Glass-Steagall Act. Third, to overcome uncertainties and gaps that limit the effectiveness of SEC and DOJ remedies, Congress should pass legislation requiring all issuers and distributors of stablecoins to be FDIC-insured banks.[28] 27 PWG Stablecoin Report, supra note 1, at 14. 28 _Id. at 2-3, 15-18._ 10 ----- **a.** **The SEC should use its existing powers to regulate stablecoins as** **“securities.”** The SEC should exercise its existing authority to regulate stablecoins as “securities,” thereby requiring issuers and distributors of stablecoins to comply with federal securities laws that protect investors and securities markets by (1) prohibiting fraud and manipulation in purchases and sales of securities and (2) imposing registration and disclosure duties on those who sell securities to the public. As discussed below, the SEC would face difficult legal challenges in regulating stablecoins as “securities.” In addition, the SEC does not possess the broad prudential oversight powers that federal bank regulators can wield to address systemic risks and promote financial stability. Consequently, vigorous efforts by the SEC to regulate stablecoins as “securities” would be a very helpful step, but it would not provide an adequate remedy for the systemic perils created by stablecoins. The SEC would confront potentially significant obstacles in showing that stablecoins are “securities,” especially with regard to stablecoins that do not pay interest and are used solely for the purpose of buying and selling goods and services for consumption. To establish legal grounds for regulating stablecoins as “securities,” the SEC must show that stablecoins are “investment contracts” or “notes” (debt obligations) as defined in federal securities laws.[29] Under the Supreme Court’s Howey decision, an “investment contract” is a “scheme [that] involves an investment of money in a common enterprise with profits to come solely from the efforts of others.”[30] In SEC v. Edwards, the Supreme Court explained that the “profits” referred to in Howey are “the profits that investors seek on their investment . . . in the sense of income or 29 _See 15 U.S.C. §§ 77b(a)(1), 78c(a)(10), 80a-2(a)(36); Todd Phillips, The SEC’s Regulatory Role in the Digital_ _[Assets Markets 5-7 (Center for American Progress, Oct. 2020), available at http://ssrn.com/abstract=3964632.](http://ssrn.com/abstract=3964632)_ 30 _SEC v. W.J. Howey & Co., 328 U.S. 293, 301 (1946); see also Phillips, supra note 29, at 5-6._ 11 ----- return, to include, for example, dividends, other periodic payments, or the increased value of the investment.” The Court also held in Edwards that “fixed returns” on “investments pitched as low-risk” would satisfy the Howey test, and the ability of investors to redeem their investments would not affect that outcome.[31] In Reves v. Ernst & Young, the Supreme Court held that every promissory “note” is presumptively a “security.” However, that presumption can be rebutted based on several factors, including whether “the buyer is interested primarily in the profit the note is expected to generate,” or, in contrast, whether “the note is exchanged to facilitate the purchase and sale of a minor asset or consumer good.” Reves held that courts should also consider (1) whether the note is an instrument in which there is “common trading for speculation or investment,” and (2) whether “the existence of another regulatory scheme significantly reduces the risk of the instrument, thereby rendering application of the Securities Acts unnecessary.”[32] Federal district courts concluded in several cases that cryptocurrencies created by sellers with fluctuating values were “investment contracts” and “securities” under federal securities laws. In those cases, the sellers represented that their cryptocurrencies could increase in value and provide investment gains to the buyers.[33] None of those cases involved stablecoins having a fixed value with reference to widely-used fiat currencies or other ostensibly “safe” assets, and there do not appear to be any reported court decisions addressing the issue of whether such stablecoins are “securities.” 31 _SEC v. Edwards, 540 U.S. 389, 394-97 (2004)._ 32 _Reves v. Ernst & Young, 494 U.S. 56, 64-67 (1990); see also Phillips, supra note 29, at 6._ 33 _See, e.g.,_ _SEC v. NAC Foundation, LLC, 512 F. Supp. 2d 988, 994-97 (N.D. Cal. 2021); SEC v. Kik Interactive_ _Inc., 492 F. Supp. 2d 169, 177-80 (S.D.N.Y. 2020); SEC v. Telegram Group Inc., 448 F. Supp. 2d 352, 364-79_ (S.D.N.Y. 2020); Balestra v. ATBCoin LLC, 380 F. Supp. 2d 340, 352-57 (S.D.N.Y. 2019); SEC v. Shavers, No. 4:13-CV-416, 2014 WL 12622292, at *4-*8 (E.D. Tex. Aug. 26, 2014). 12 ----- Issuers of the most widely-used stablecoins (including Tether, USD Coin, and Pax Dollar) represent that their stablecoins will maintain a 1-to-1 parity with the U.S. dollar by holding reserves that include cash, government securities, and (in most cases) corporate debt obligations. Most leading stablecoins do not pay interest to their holders. Thus, instead of promising potential gains, issuers of most prominent stablecoins assure investors that they will not suffer losses from buying and holding stablecoins. Those stablecoins are different from cryptocurrencies that have fluctuating values and offer buyers the possibility of making profits from trading.[34] The SEC could argue that stablecoins should be treated as “investment contracts” or “notes” because (1) issuers and distributors offer and sell stablecoins to investors with the shared understanding that stablecoins are the most widely-used form of payment for speculating in cryptocurrencies and other digital assets;[35] and (2) issuers and distributors expect that most buyers of stablecoins will use their coins to pursue speculative profits by trading in digital assets or by lending their coins to other traders.[36] Thus, a purchase of stablecoins could reasonably be viewed as the payment of an “entry fee” enabling the buyer to speculate in cryptocurrency markets, just as the purchase of poker chips permits a gambler to participate and place bets in 34 Ackerman, supra note 3; Awrey, supra note 7, at 60 n.221; Nikhilesh De, “SEC Chair Hints Some Stablecoins [Are Securities,” CoinDesk (Sept. 14, 2021), https://www.coindesk.com/markets/2021/07/21/sec-chair-hints-some-](https://www.coindesk.com/markets/2021/07/21/sec-chair-hints-some-stablecoins-are-securities/) [stablecoins-are-securities/; DiCamillo, supra note 7; Gorton & Zhang, supra note 7, at 3, 6-8, 12-16; Smialek, supra](https://www.coindesk.com/markets/2021/07/21/sec-chair-hints-some-stablecoins-are-securities/) note 7; Wilmarth, “Pandemic Crisis,” supra note 15, at 9-10. 35 _See supra notes 3 & 10 and accompanying text._ 36 For court decisions that involved interest-bearing debt instruments but also indicated that financial instruments sold for the purpose of encouraging speculation could be treated as “securities,” see, e.g., _Gary Plastic Packaging_ _Corp. v. Merrill Lynch, Pierce, Fenner & Smith, Inc., 756 F.2d 230, 240-42 (2d Cir. 1985) (holding that the_ defendant broker-dealer sold “investment contracts” that were subject to regulation as “securities” because the defendant sold negotiable bank certificates of deposits (CDs) accompanied by promises that the defendant would monitor the quality of the issuing banks, repurchase the CDs on demand, and maintain a “secondary market” in the CDs, thereby enabling customers to resell their CDs for potential gains without risking any loss of their principal or accrued interest); Stoiber v. SEC, 161 F.3d 745, 747-52 (D.C. Cir. 1998) (holding that the defendant broker sold “notes” that were subject to regulation as “securities” because the defendant sold interest-bearing promissory notes to customers with the understanding that the defendant would use most of the sale proceeds to trade in commodities and generate profits to pay off the notes). 13 ----- poker games and tournaments.[37] SEC Chair Gary Gensler recently observed that stablecoins are primarily bought and used for speculative purposes, and he described stablecoins as “acting almost like poker chips at the casino.”[38] The SEC could argue that it would be proper to classify stablecoins as “notes” because buyers of stablecoins are “primarily motivated by the opportunity to earn a profit on their money” by using their stablecoins to pay for subsequent speculative transactions.[39] In contrast, if issuers created stablecoins that could be used only to buy and sell goods and services for consumption, and that could not be used for speculation, it would be much more difficult for the SEC to characterize those stablecoins as “securities.” As explained above, court decisions defining “investment contracts” and “notes” have excluded financial instruments that are purchased solely for the purpose of buying and selling goods, other property, or services for consumption, and not for potential investment gains.[40] Special-purpose, consumption-only stablecoins do not appear to be part of the present digital asset landscape. However, issuers 37 _See Tschetschot v. Commissioner, T.C. Memo. 2007-38, 2007 WL 518989, at *3 (U.S.T.C., Feb. 20, 2007)_ (stating that participants in poker tournaments bought poker chips as part of their “entry fees” for the purpose of “placing bets, hoping to win” prizes). 38 Gensler Statement, supra note 3; Tory Newmyer, “SEC’s Gensler likens stablecoins to ‘poker chips’ amid calls for tougher crypto regulation,” Washington Post (Sept. 21, 2021) (quoting from interview with Mr. Gensler), [https://www.washingtonpost.com/business/2021/09/21/sec-gensler-crypto-stablecoins/; see also Opening Statement](https://www.washingtonpost.com/business/2021/09/21/sec-gensler-crypto-stablecoins/) of Sen. Sherrod Brown at a hearing before the Senate Comm. on Banking, Housing, and Urban Affairs (Dec. 14, 2021) (“Stablecoins make it easier than ever to risk real dollars on cryptocurrencies.”), [https://www.banking.senate.gov/imo/media/doc/Brown%20Statement%2012-14-21.pdf.](https://www.banking.senate.gov/imo/media/doc/Brown%20Statement%2012-14-21.pdf) 39 _Stoiber v. SEC, 161 F.3d at 750._ 40 _See, e.g., SEC v. Kik Interactive, 492 F. Supp. 2d at 179-80 (distinguishing between digital assets purchased for a_ “consumptive use” and those bought primarily for their “profit-making potential”); Solis v. Latium Network, Inc., No. 18-10255 (SDW) (SCM), 2018 WL 6445543, at *1-*3 (D.N.J., Dec. 10, 2018) (holding that digital tokens sold by defendants were “securities” because defendants encouraged plaintiffs to “expect a profit” from investing in the tokens, even though the tokens could also potentially be used to purchase services). For additional analysis of the distinction between financial instruments used solely for consumption and those purchased for investment gains, see SEC “Finhub” Staff, “Framework for ‘Investment Contract’ Analysis of Digital Assets” (April 3, 2019), Part II.C.3, [https://www.sec.gov/corpfin/framework-investment-contract-analysis-digital-assets; Jay B. Sykes, “Securities](https://www.sec.gov/corpfin/framework-investment-contract-analysis-digital-assets) Regulation and Initial Coin Offerings: A Legal Primer,” at 14-19, 26-32 (Congressional Res. Serv. Rep. R45301, [Aug. 31, 2018), https://sgp.fas.org/crs/misc/R45301.pdf.](https://sgp.fas.org/crs/misc/R45301.pdf) 14 ----- might decide to create such instruments if the SEC succeeded in classifying stablecoins used for speculation as “securities.” The SEC could also seek to regulate issuers of stablecoins as investment companies under the Investment Company Act of 1940 (1940 Act). Issuers of stablecoins that engage primarily in the business of investing and trading in securities, or engage in that business and hold more than 40% of their assets in non-government securities, could potentially be treated as investment companies. There are numerous exemptions in the 1940 Act that might allow some issuers of stablecoins to avoid being treated as investment companies, and an analysis of those exemptions is beyond the scope of this paper.[41] The SEC’s track record with money market funds – financial instruments that closely resemble stablecoins – does not inspire confidence that the SEC could effectively control the systemic dangers of stablecoins by regulating them as investment companies. The SEC’s regulation of money market funds under the 1940 Act failed to ensure the resilience of those funds after Lehman Brothers collapsed in September 2008. Lehman’s bankruptcy and default on its commercial paper triggered systemic runs by investors on money market funds, and the Treasury Department and Federal Reserve (Fed) were forced to arrange a comprehensive bailout of those funds. Despite that calamity, the SEC rejected numerous recommendations after 2008 – including one from FSOC – calling on money market funds to stop offering fixed net asset values (NAVs) of $1 per share and to use floating NAVs like other mutual funds. The SEC required institutional prime and tax-exempt money market funds to adopt floating NAVs, and it permitted non-government money market funds to impose restrictions on redemption. However, the SEC allowed retail prime money market funds and institutional and retail government money 41 15 U.S.C. § 80a-3; see Phillips, supra note 29, at 6-7; SEC, “Investment Company Registration and Regulation [Package,” https://www.sec.gov/investment/fast-answers/divisionsinvestmentinvcoreg121504htm.html.](https://www.sec.gov/investment/fast-answers/divisionsinvestmentinvcoreg121504htm.html) 15 ----- market funds to continue offering deposit-like treatment with fixed NAVs of $1 per share. Money market funds experienced another series of systemic runs by investors in March 2020 and were rescued for a second time by the Treasury and the Fed.[42] In December 2021, the SEC issued a proposal to amend its money market fund rules to address the problems revealed by the investor runs of 2020. The SEC’s proposal would increase liquidity requirements and modify redemption terms for money market funds in order to reduce incentives for investor runs during periods of financial stress. At the same time, the proposal conceded that the SEC’s changes to its money market fund rules in 2010 and 2014 did not achieve their intended purpose and failed to prevent the investor runs of 2020.[43] The SEC’s proposal considered – and rejected – the alternative possibility of requiring all money market funds to use floating NAVs. The proposal acknowledged that a new rule requiring floating NAVs for all funds would increase transparency about the risk of money market fund investments. . . . To the degree that investors in stable NAV funds are currently treating them as if they were holding U.S. dollars due to a lack of transparency about risks of such funds, expanding the scope of the floating NAV requirements may enhance 42 Michael S. Barr, Howell E. Jackson & Margaret E. Tahyar, Financial Regulation: Law & Policy 1302-24 (2d ed. 2018); Gorton & Zhang, supra note 7, at 21-24; Wilmarth, “Pandemic Crisis,” supra note 15, at 4-8, 11-12; Wilmarth, Taming the Megabanks, supra note 25, at 153-57, 279-88, 341-44; see also Marco Cypriani & Gabrielle La Spada, “Sophisticated and Unsophisticated Runs” (Fed. Res. Bank of N.Y. Staff Rep. No. 956, Dec. 2020), [https://www.newyorkfed.org/medialibrary/media/research/staff_reports/sr956.pdf; Lei Li, Yi Li, Marco](https://www.newyorkfed.org/medialibrary/media/research/staff_reports/sr956.pdf) Macchiavelli & Xing (Alex) Zhou, “Liquidity Restrictions, Runs, and Central Bank Interventions: Evidence from [Money Market Funds” (May 24, 2021), available at https://ssrn.com/abstract=3607593.](https://ssrn.com/abstract=3607593) 43 Securities and Exchange Commission, “Money Market Fund Reforms: Proposed rule” (Dec. 15, 2021) [[hereinafter SEC Money Market Fund Proposal], https://www.sec.gov/rules/proposed/2021/ic-34441.pdf. For the](https://www.sec.gov/rules/proposed/2021/ic-34441.pdf) SEC’s explanation of why its changes to money market fund rules in 2010 and 2014 were inadequate and did not prevent the investor runs of March 2020, see id. at 10-31, 81-97. 16 ----- investor protections and enable investors to make more informed investment decisions.[44] The SEC’s proposal also recognized that requiring floating NAVs for all funds “would reduce the distortions arising out of implicit government guarantees of money market funds” and likely cause “investors of stable NAV funds to reallocate capital into cash accounts subject to deposit insurance.”[45] The resulting shrinkage of the money market fund industry would reduce the demand for short-term wholesale funding instruments, such as securities repurchase agreements (repos) and commercial paper. Money market funds are the most prominent investors in repos and commercial paper. Those short-term debt instruments also function as “shadow deposits” (functional substitutes for bank deposits), and they experienced their own systemic breakdowns in 2008 and 2020.[46] The SEC’s proposal admitted that the support provided by money market funds for short-term wholesale funding markets “may be sustainable, in part, due to perceived government backstops of money market funds and lack of transparency to investors about the risks inherent in money market fund investments.”[47] Thus, the SEC’s proposal recognized that money market funds with fixed NAVs produce market distortions, depend on implicit government guarantees, and do not provide full transparency to investors. Nevertheless, the SEC’s proposal rejected the option of requiring all money market funds to adopt floating NAVs. That rejection suggests that the SEC would be equally unprepared to force stablecoins to abandon their promised 1:1 parity with the U.S. dollar – a promise that conveys to investors the same illusion of deposit-like status. 44 _Id. at 234._ 45 _Id. at 236-38._ 46 _Id. at 237-38; see also Wilmarth, “Pandemic Crisis,” supra note 15, at 2-8, 12; Wilmarth, Taming the Megabanks,_ _supra note 15, at 153-57, 278-87, 341-44, 353-56._ 47 SEC Money Market Fund Proposal, supra note 43, at 238. 17 ----- Like money market funds, stablecoins are “shadow deposits” – a type of “private money” that is designed to serve as a functional substitute for federally-insured bank deposits. The bailouts of money market funds in 2008 and 2020 and the close similarities between those funds and stablecoins strongly support the conclusion that both types of financial instruments should be regulated in the same way as bank deposits to control their systemic dangers.[48] As shown by the vicissitudes of money market funds, regulating stablecoins as investment companies under the 1940 Act would not provide an adequate substitute for requiring stablecoins to comply with the regulatory regime governing FDIC-insured bank deposits. An additional factor supporting that conclusion is that the SEC lacks broad financial stability powers comparable to the extensive prudential regulatory and supervisory authorities of federal banking agencies. The SEC’s core mission is to protect investors and preserve the integrity of securities markets. The SEC generally has not attempted to act as a financial stability regulator.[49] As shown in the next two sections, regulating stablecoins as deposits and requiring their issuers and distributors to become FDIC-insured banks would provide the most promising approach for controlling their systemic perils. **b.** **The Department of Justice should enforce Section 21(a) of the Glass-** **Steagall Act against issuers and distributors of stablecoins.** Novi provides deposit-like treatment for stablecoins that its customers buy and hold in their accounts. Novi sells stablecoins (currently Pax Dollars) to its customers at a fixed price of $1 per coin, and Novi agrees to redeem stablecoins from its customers at the same price of $1 per 48 Gorton & Zhang, supra note 7, at 21-24, 33-35; Wilmarth, “Pandemic Crisis,” supra note 15, at 6-10. 49 _See Congressional Research Service, “Who Regulates Whom? An Overview of the U.S. Financial Regulatory_ [Framework” (CRS Report No. R44918, updated Mar. 10, 2020), https://sgp.fas.org/crs/misc/R44918.pdf; Daniel K.](https://sgp.fas.org/crs/misc/R44918.pdf) Tarullo, “The SEC should – and can – pay more attention to financial stability” (May 13, 2021), [https://www.brookings.edu/blog/up-front/2021/05/13/the-sec-should-and-can-pay-more-attention-to-financial-](https://www.brookings.edu/blog/up-front/2021/05/13/the-sec-should-and-can-pay-more-attention-to-financial-stability/) [stability/; see generally Barr, Tahyar & Jackson, supra note 42, at 444-502, 535-64.](https://www.brookings.edu/blog/up-front/2021/05/13/the-sec-should-and-can-pay-more-attention-to-financial-stability/) 18 ----- coin.[50] Novi’s customers own the stablecoins held in their accounts, and Novi agrees to maintain custody of its customer’s stablecoins until they are redeemed, withdrawn, or transferred to other persons.[51] Novi enables its customers to (i) purchase and hold Digital Currency in your Account, (ii) conduct person-to person transfers of Digital Currency, (iii) set up recurring Digital Currency transactions, (iv) convert your Digital Currency to local currency and pick up cash; (iv) convert your Digital Currency to local currency and transfer to your linked bank account via an automated clearing house (“ACH”) transaction; and (vi) use any additional features we may provide through your use of the Services.[52] The services that Novi provides to its customers through their stablecoin accounts satisfy both of the key functional characteristics of “deposits”: (1) the placing of funds with another person for custody and safekeeping, and (2) the ability of the depositor to withdraw or transfer those funds on demand or at a definite time. In a 1991 decision, the Second Circuit Court of Appeals held that, “[a]s commonly understood, the term ‘deposit’ means a sum of money placed in the custody of a bank, to be withdrawn at the will of the depositor.”[53] Similarly, in a 2016 decision, the Fifth Circuit Court of Appeals explained: 50 Novi Financial, Inc., “How It Works: Add money” (“Simply add a debit card to put money in your account, and [it’ll be converted to USDP. On Novi, 1 USDP is equal to 1 US dollar.”), https://www.novi.com/how-it-works; Novi](https://www.novi.com/how-it-works) Terms of Service, supra note 19, ¶ 3 (“User Redemption Right”) (“You are entitled to redeem each Digital Currency for one U.S. dollar (USD) with Novi.”). 51 Novi Terms of Service, supra note 19, ¶ 3 (“Title and Ownership”) (“Your Account will give you access to buy, sell, transfer, and manage your Digital Currency. The Digital Currency is held by Novi on a blockchain in one or more blockchain addresses (each, a ‘Wallet’). . . . Novi controls the Wallet that holds your Digital Currency. . . . [Y]ou own all beneficial interest in the Digital Currency in your Account.”); id. ¶ 3 (“Custody”) (“To more securely custody [sic] Digital Currency, we may use one or more shared, commingled Wallets to hold Digital Currency on your behalf and on our own behalf.”). 52 Novi Terms of Service, supra note 19, ¶ 5 (“Description of the Services”). 53 _United States v. Jenkins, 943 F.2d 167, 174 (2d Cir.) (citations omitted), cert. denied, 502 U.S. 1014 (1991)._ 19 ----- The relevant authorities demonstrate that the essential elements of a “deposit” include the following. First, a deposit must involve the placement of funds with another for “safekeeping.” . . . Second, those funds must be subject to the control of the depositor such that they are repayable on demand or at a fixed time.[54] As shown above, Novi clearly accepts “deposits” by agreeing to (i) receive funds from its customers, (ii) convert those funds into stablecoins at a fixed 1:1 parity with U.S. dollars, (iii) maintain custody of the stablecoins owned by its customers, (iv) repay its customers’ funds based on the same fixed parity when its customers decide to redeem their stablecoins, and (v) allow its customers to transfer their stablecoins to other persons. Courts have repeatedly held that nonbanks are deemed to receive “deposits” if they accept funds from other persons while agreeing to hold those funds and repay them on demand or at a specified time. The ability of customers to transfer their funds to third parties is not a prerequisite for status as “deposits,” but their right to transfer their funds to third parties provides additional evidence that a deposit relationship has been formed.[55] Section 21(a) of the Glass-Steagall Act establishes two overlapping prohibitions against the receipt of deposits by nonbanks. Section 21(a)(1) focuses on persons who are engaged in securities activities. Section 21(a)(1) bars issuers, underwriters, distributors, and sellers of 54 _MoneyGram Int’l, Inc. v. Commissioner, No. 15-60527, 664 Fed. Appx. 386, 392 (5th Cir., Nov. 15, 2016)_ (citations omitted); see also MoneyGram Int’l, Inc. v. Commissioner, 999 F.3d 269, 274-76 (5th Cir. 2021). 55 _United States v. Jenkins, 943 F.2d_ at 174 (holding that the defendant (an individual) accepted a “deposit” when he “took custody” of $150,000 on behalf of a purported foreign bank and “agreed to return it at the will” of the depositor, stating “[y]our money will be here for your use”); In re Thaxton Group, Inc., Securities Litigation, C/A No. 8:02-2612-GRA, 2006 WL 8462530, at *1–*3, *9–*14 (D.S.C., Mar. 20, 2006) (holding that the defendant (a nonbank finance company) accepted “deposits” by selling $121 million of demand notes to 5,000 investors, thereby “taking money from investors in return for a promise to return the funds on demand,” and explaining that the “notes were designed to imitate bank certificates of deposit and money market accounts in order to attract bank depositors to the note program”); S & N Equipment Co. v. Casa Grande Cotton Finance Co., 97 F.3d 337, 340–45 (9th Cir. 1996) (determining that the defendant (a nonbank finance company) accepted “demand deposits” for purposes of the BHC Act because the defendant “accepted funds from its customers,” placed those funds in “credit accounts,” and allowed customers to “withdraw funds as needed” and transfer funds to third parties). 20 ----- “stocks, bonds, debentures, notes, or other securities” from also “engag[ing] at the same time to any extent whatever in the business of receiving deposits subject to check or repayment upon presentation of a passbook, certificate of deposit, or other evidence of debt, or upon request of the depositor.”[56] If Novi’s stablecoins are determined to be “securities,” Novi would be “engag[ing] at the same time” in both (1) issuing, underwriting, distributing, or selling “securities” and (2) receiving “deposits” that are (A) withdrawn or transferred by customers using functional equivalents of “checks” or (B) repaid to customers upon their request. Section 21(a)(1) clearly forbids that combination of activities.[57] Court decisions have treated “deposits” as also being “securities” for purposes of the federal securities laws unless those deposits are accepted either by FDIC-insured U.S. banks or by foreign banks subject to a regulatory regime that provides comparable protections to depositors.[58] Based on those decisions, Novi’s stablecoins would be subject to regulation as both “deposits” and “securities” because Novi is not chartered or regulated as a bank and does not have FDIC insurance. Accordingly, DOJ should determine that Novi’s stablecoins violate Section 21(a)(1) if those stablecoins are found to be “securities.” Section 21(a)(2) of the Glass-Steagall Act is a broader and more sweeping provision. Section 21(a)(2) prohibits all persons (regardless of whether they are also involved in “securities” activities) from “engag[ing], to any extent whatever . . . in the business of receiving deposits” – described with the functional characteristics included in Section 21(a)(1) – unless 56 12 U.S.C. § 378(a)(1). 57 For court decisions applying Section 21(a)(1) and holding that the relevant terms – including “securities,” “underwriting,” and “dealing” – are generally given the same meaning under federal securities laws and the GlassSteagall Act, see Securities Indus. Ass’n v. Board of Governors, 468 U.S. 137, 148-52 (1984); Investment Co. _Institute v. Conover, 790 F.2d 925, 927-28, 933-34 (D.C. Cir.), cert. denied sub nom. Investment Co. Institute v._ _Clarke, 479 U.S. 939 (1986)._ 58 _Marine Bank v. Weaver, 455 U.S. 551, 555-59 (1982); SEC v. McDuffie, Civil Action No. 12-cv-02939, 2014 WL_ 4548723, at *3-*7 (D. Colo., Sept. 15, 2014); SEC v. Stanford Int’l Bank, Ltd., Civil Action No. 3:09-CV-0298-N, 2011 WL 13160374, at *3-*5 (N.D. Tex., Nov. 30, 2011). 21 ----- those persons satisfy one of three alternative sets of regulatory criteria. Under Section 21(a)(2), a person who engages in the business of receiving deposits must (A) be chartered and authorized to “engage in such business” by, and subject to examination and regulation under, federal laws or the laws of a state, U.S. territory, or the District of Columbia, or (B) be “permitted by” federal laws or the laws of a state, U.S. territory, or the District of Columbia to “engage in such business” and also be subject under the laws of that jurisdiction to “examination and regulation,” or (C) submit to “periodic examination by the banking authority” of the state, territory, or District of Columbia where “such business is carried on,” and publish “periodic reports of its condition,” in “the same manner and under the same conditions” as are required by the laws of such state, territory, or District for chartered banks “engaged in such business in the same locality.”[59] It bears repeating that Section 21(a)(2) – unlike Section 21(a)(1) – applies to all _persons who engage in the business of receiving deposits, regardless of whether they are also_ issuing, underwriting, distributing, or selling “securities.” Paragraphs (A) and (B) of Section 21(a)(2) cover much of the same ground – as both paragraphs refer to institutions that are legally authorized to engage in “the business of receiving deposits” – except that paragraph (A) refers to chartering, examination, and regulation while paragraph (B) refers to examination and regulation but not chartering. Paragraph (C) describes persons who are subject to “periodic examination” by a state, District, or territorial “banking authority” and who also submit “periodic reports,” with such examination and reports to be made “in the same manner and under the same conditions” as are required for chartered banks engaged in the business of receiving deposits in the same state, District, or territory. The crucial point is that all three paragraphs in Section 21(a)(2) refer to institutions that are either chartered as, 59 12 U.S.C. § 378(a)(2). 22 ----- regulated as, or subject to the same examination and reporting requirements as, deposit-taking banks. Persons who do not satisfy the criteria set forth in any of the three paragraphs would violate Section 21(a)(2) if they engage “to any extent whatever . . . in the business of receiving deposits.”[60] As explained above, Novi’s activities satisfy Section 21(a)’s functional description of engaging in the “business of receiving deposits subject to check or repayment . . . upon request of the depositor.” Novi’s deposit-taking business violates Section 21(a)(2) – regardless of whether its stablecoins are treated as “securities” – because Novi does not satisfy any of the criteria set forth in paragraphs (A), (B), or (C). Novi is not chartered as a deposit-taking bank under federal or state laws. Novi is not permitted by federal or state laws to engage in the business of receiving deposits while being examined and regulated in connection with that business. Novi also is not complying with the same examination and reporting requirements as are applied by the “banking authority” of the relevant state, District, or territory to chartered, deposit-taking banks. Section 21(b) imposes criminal sanctions on persons who violate Section 21(a), and DOJ is therefore responsible for enforcing the statute. In October 1979, a New York savings bank sent a letter to DOJ and the SEC alleging that Merrill Lynch was violating Section 21(a) by offering “cash management” money market funds that were unlawful “deposits.” DOJ’s Criminal Division issued an opinion in December 1979, which rejected the savings bank’s allegations based on a highly formalistic analysis. DOJ classified money market funds as equity investments rather than debt claims, and DOJ concluded that only debt claims could be treated as “deposits” under Section 21(a). DOJ ignored the fact that Merrill Lynch provided its customers 60 _Id.; see United States v. Jenkins, 943 F.2d at 173-74; In re Thaxton Group, Inc., Securities Litigation, supra note_ 55, at *1-*3, *9-*14. 23 ----- with the functional equivalent of deposits by (i) maintaining a stable value – a fixed NAV of $1 per share – for the money market funds held in customer accounts, (iii) allowing customers to withdraw their funds with the same stable value by making redemption requests or writing checks, and (iii) enabling customers to transfer their funds with the same stable value to third parties by writing checks.[61] DOJ’s 1979 opinion should not be considered a binding precedent. That opinion’s formalistic reasoning is not consistent with either Section 21(a)’s functional description of “deposits” or the statute’s purpose to “prohibit[] . . . unregulated private banking so far as practicable.” The 1979 opinion’s reasoning also is not compatible with the functional, pragmatic approach of at least two courts that interpreted Section 21(a) more recently.[62] DOJ should undertake a fresh review of Section 21(a) and should determine that the statute’s functional description of “deposits” includes funds that are received from and held on behalf of customers with the understanding that customers may withdraw or transfer those funds by using “checks” (or functionally equivalent methods of payment) or by making “requests” for repayment. Based on the foregoing determination, DOJ should issue a rule declaring that issuers and distributors of stablecoins providing a fixed 1:1 parity with the U.S. dollar or another widely used fiat currency are “engag[ing] . . . in the business of receiving deposits” if they receive funds from customers, hold stablecoins on behalf of customers, and allow customers to redeem, 61 Gorton & Zhang, supra note 7, at 10-12, 33-35; Howell E. Jackson & Morgan Ricks, “Locating Stablecoins Within the Regulatory Perimeter,” Harvard Law School Forum on Corporate Governance (Aug. 5, 2021), [https://corpgov.law.harvard.edu/2021/08/05/locating-stablecoins-within-the-regulatory-perimeter/.; Wilmarth,](https://corpgov.law.harvard.edu/2021/08/05/locating-stablecoins-within-the-regulatory-perimeter/) _Taming the Megabanks, supra note 25, at 153-54._ 62 _See United States v. Jenkins, 943 F.2d at 173-74; In re Thaxton Group, Inc., Securities Litigation, supra note 55,_ at *1-*3, *9-*14; Jackson & Ricks, supra note 61 (“The legislative history of section 21(a)(2) confirms that the provision was intended to ‘prohibit[]. . . unregulated private banking so far as practicable.’”) (quoting Senate Report No. 1007, 74th Cong., 1st Sess. 15 (1935)). See also Gorton & Zhang, supra note 7, at 10-12, 33-35; Wilmarth, “Pandemic Crisis,” supra note 15, at 8, 20-21 n.45; Wilmarth, Taming the Megabanks, supra note 25, at 137-39, 153-54. 24 ----- withdraw, or transfer their stablecoins. Pursuant to Section 21(a)(1), DOJ’s rule should prohibit issuers and distributors of stablecoins that are determined to be “securities” from also receiving, holding, and allowing redemptions, withdrawals, or transfers of customers’ stablecoins. For stablecoins that are determined not to be “securities,” DOJ’s rule should describe the criteria that issuers and distributors of those stablecoins must satisfy under Section 21(a)(2). DOJ’s rule should make clear that issuers and distributors of stablecoins may not receive, hold, and allow redemptions, withdrawals, or transfers of customers’ stablecoins unless those issuers and distributors are either (A) chartered, examined, and regulated as deposit-taking banks, or (B) legally authorized to engage in the business of receiving deposits while also being examined and regulated in their conduct of that business, or (C) complying with the same examination and reporting requirements as are applied to deposit-taking banks by the “banking authority” of the relevant state, District, or U.S. territory.[63] Some might argue that DOJ should not bring enforcement proceedings against issuers and distributors of stablecoins under Section 21(a) unless DOJ takes similar measures against other nonbanks providing financial services that are functionally equivalent to deposits. Such nonbanks would include money market funds as well as payment service providers such as PayPal and its subsidiary Venmo, which hold customer balances and allow customers to withdraw or transfer those balances to others. I would personally welcome a decision by DOJ to take an across-the-board approach, and I believe DOJ would have authority under Section 21(a) 63 12 U.S.C. § 378(a); see Gorton & Zhang, supra note 7, at 33-35; Jackson & Ricks, supra note 61; Wilmarth, “Pandemic Crisis,” supra note 15, at 8-10, 20-21 n.45 (contending that an issuer or distributor of stablecoins would not comply with Section 21(a)(2) if it merely obtained a state money transmitter license and complied with FinCEN’s AML requirements, as those limited forms of regulation would not be “equivalent to bank regulation and supervision in any meaningful sense”); see also _MoneyGram_ _Int’l Inc. v. Commissioner, 999 F.3d 269 (5th Cir._ 2021) (holding that a state-licensed money transmitter was not a “bank” because it did not accept “deposits”); Awrey, supra note 7, at 7-8, 40-56 (describing the “alarming . . . permissiveness” of state laws regulating money transmitters, a situation that “undermines the credibility of [money transmitters’] monetary commitments”). 25 ----- to institute enforcement proceedings against money market funds, PayPal, and Venmo.[64] However, DOJ is not required to act against all violators of Section 21(a) at the same time. DOJ could reasonably decide to focus on stablecoins as a particularly dangerous form of unauthorized deposit-taking that should be stopped before DOJ determines how to deal with similar problems created by money market funds, PayPal, and Venmo.[65] **c.** **Congress should pass legislation mandating that all issuers and** **distributors of stablecoins must be FDIC-insured banks.** **i.** **Requiring all issuers and distributors of stablecoins to be** **FDIC-insured banks would remove uncertainties and gaps in** **Section 21(a) of the Glass-Steagall Act.** I strongly support PWG’s recommendation that Congress should “promptly” pass legislation requiring all issuers of stablecoins to be FDIC-insured banks.[66] Such legislation is urgently needed to overcome uncertainties and gaps that currently exist in Section 21(a) of the Glass-Steagall Act. As explained in the preceding section, an issuer or distributor of stablecoins could offer services that are functionally equivalent to deposits and avoid violating Section 21(a) if it could show that (1) its stablecoins are not “securities,” and (2) it is either (A) chartered, regulated, and examined as a deposit-taking bank, or (B) legally authorized to engage in the business of receiving deposits and subject to examination and regulation in conducting that 64 Wilmarth, “Pandemic Crisis,” supra note 15, at 7-10; see also Gorton & Zhang, supra note 7, at 33-35; Jackson & Ricks, supra note 61. 65 _See United States v. Central Adjustment Bureau, Inc., 823 F.2d 880 (5th_ Cir. 1987) (holding that “Congress can attack particular evils on a step by step basis,” and Congress had a rational basis for passing the Fair Debt Collection Act, which prohibited abusive practices by independent debt collectors without addressing similar abuses by other types of debt collectors); see also Minnesota v. Clover Leaf Creamery Co., 449 U.S. 456, 461-70 (1981) (quote at 466) (holding that state legislatures “need not ‘strike at all evils at the same time or in the same way,’” and the Minnesota legislature had a rational basis for banning nonreturnable plastic milk jugs to protect the environment without also prohibiting nonreturnable paperboard milk containers) (quoting Semler v. Oregon State Bd. of Dental _Examiners, 294 U.S. 608, 610 (1935))._ 66 PWG Report, supra note 1, at 2, 16. 26 ----- business, or (C) subject to the same examination and reporting requirements as are applied to chartered, deposit-taking banks by the “banking authority” of the relevant state, District, or U.S. territory. The terms of Section 21(a)(2) contain potential ambiguities that would need to be resolved by DOJ and the courts. For example, what precise levels of examination and regulation are needed to satisfy paragraph (B), and what exact types of examinations and reports are required to comply with paragraph (C)? States that wanted to attract entry by stablecoin issuers and distributors could enact laws designed to exploit those ambiguities by granting the most lenient possible treatment to stablecoin providers.[67] Even more troubling, an issuer or distributor of stablecoins would qualify under paragraph (A) of Section 21(a)(2) if it could obtain a charter for an uninsured depository institution from a federal or state banking authority. Until recently, a deposit-taking bank could not receive either a federal or state charter unless it also obtained deposit insurance from the FDIC and became subject to the full panoply of laws governing FDIC-insured banks. Federal law currently requires all national banks that accept deposits to obtain FDIC insurance. Prior to 2019, every state required state-chartered banks to obtain FDIC insurance as a precondition for accepting deposits.[68] A number of the state laws mandating FDIC insurance for state-chartered banks were enacted in response to widespread failures of non-federally-insured depository institutions during the 1980s and early 1990s. During that period, state-sponsored, privately-funded insurance 67 _See Jackson & Ricks, supra note 61._ 68 _See Barr, Jackson & Tahyar, supra note 42, at 173-74 (As of 2018, “Federal law requires that all national banks be_ FDIC insured, and all state laws require that a state-chartered commercial bank obtain FDIC insurance.”); 12 U.S.C. § 222 (“Every national bank in any State shall . . . become a member bank of the Federal Reserve System . . . and shall thereupon be an insured bank under the Federal Deposit Insurance Act”); Wilmarth, “Banking Privileges,” _supra note 26, at 2-7._ 27 ----- systems for state-chartered depository institutions collapsed in several states, with the worst disasters occurring in Ohio, Maryland, and Rhone Island. The injuries suffered by depositors and local economies were particularly severe in Rhode Island, where non-federally-insured depositors lost access to at least some of their deposits for nearly three years.[69] Despite the dismal record of non-federally-insured depository institutions, Wyoming and Nebraska have passed laws during the past two years that authorize charters for uninsured “special purpose depository institutions” (SPDIs) in Wyoming and uninsured “digital asset depositories” (DADs) in Nebraska. Wyoming and Nebraska allow SPDIs and DADs to accept deposits (including deposits of digital assets) and engage in other cryptocurrency-related activities without obtaining FDIC insurance. Wyoming has already approved four SPDI charters, including one awarded to Kraken, a major cryptocurrency venture.[70] In December 2020, Figure Technologies (Figure) applied to the OCC to obtain a charter for a national bank that would accept only “jumbo” deposits larger than $250,000 (the current limit for federal deposit insurance). Figure asserted that it could avoid any obligation to obtain FDIC insurance by accepting only jumbo deposits. The OCC has not acted on Figure’s application, and state regulators have challenged the OCC’s legal authority to approve any 69 Quian Chen et al., “The Macroeconomic Fallout of Shutting Down the Banking System,” 105 Economic Review [No. 2, at 31 (Fed. Res. Bank of K.C., 2020), https://www.kansascityfed.org/documents/8185/v105n2sharma.pdf;](https://www.kansascityfed.org/documents/8185/v105n2sharma.pdf) Walker F. Todd, “Lessons from the Collapse of Three State-Chartered Private Insurance Funds,” Economic _[Commentary (Fed. Res. Bank of Cleve., May 1, 1994), https://www.clevelandfed.org/en/newsroom-and-](https://www.clevelandfed.org/en/newsroom-and-events/publications/economic-commentary/economic-commentary-archives/1994-economic-commentaries/ec-19940501-lessons-from-the-collapse-of-three-state-chartered-private-deposit-insurance-funds.aspx)_ [events/publications/economic-commentary/economic-commentary-archives/1994-economic-commentaries/ec-](https://www.clevelandfed.org/en/newsroom-and-events/publications/economic-commentary/economic-commentary-archives/1994-economic-commentaries/ec-19940501-lessons-from-the-collapse-of-three-state-chartered-private-deposit-insurance-funds.aspx) [19940501-lessons-from-the-collapse-of-three-state-chartered-private-deposit-insurance-funds.aspx; see also](https://www.clevelandfed.org/en/newsroom-and-events/publications/economic-commentary/economic-commentary-archives/1994-economic-commentaries/ec-19940501-lessons-from-the-collapse-of-three-state-chartered-private-deposit-insurance-funds.aspx) Christine Bradley & Valentine V. Craig, “Privatizing Deposit Insurance: Results of the 2006 FDIC Study,” 1 FDIC _Quarterly No. 2, at 23, 28-30 (2007) (discussing the collapses of numerous state-sponsored private insurance_ [systems for depository institutions between the 1830s and the 1990s), https://www.fdic.gov/analysis/quarterly-](https://www.fdic.gov/analysis/quarterly-banking-profile/fdic-quarterly/2007-vol1-2/privatizing-deposit-insurance.pdf) [banking-profile/fdic-quarterly/2007-vol1-2/privatizing-deposit-insurance.pdf.](https://www.fdic.gov/analysis/quarterly-banking-profile/fdic-quarterly/2007-vol1-2/privatizing-deposit-insurance.pdf) 70 _See Gorton & Zhang, supra note 7, at 20; Wyoming Division of Banking, “Special Purpose Depository_ [Institutions,” https://wyomingbankingdivision.wyo.gov/banks-and-trust-companies/special-purpose-depository-](https://wyomingbankingdivision.wyo.gov/banks-and-trust-companies/special-purpose-depository-institutions) [institutions; “Nebraska Financial Innovation Act,” Neb. Rev. Stat. Ch. 8, Art. 30, available at](https://wyomingbankingdivision.wyo.gov/banks-and-trust-companies/special-purpose-depository-institutions) [https://ndbf.nebraska.gov/about/legal/financial-innovation-act.](https://ndbf.nebraska.gov/about/legal/financial-innovation-act) 28 ----- charters for uninsured, deposit-taking national banks.[71] Nevertheless, Figure and other cryptocurrency companies recently met with the Fed, FDIC, and OCC to discuss “how to issue a stablecoin that satisfies” regulators.[72] Wyoming’s and Nebraska’s new laws and Figure’s charter application are designed to produce charters for uninsured deposit-taking banks that can issue and distribute stablecoins and engage in other cryptocurrency activities. As indicated above, uninsured banks – if lawfully chartered and legally authorized to receive deposits – could issue and distribute stablecoins that are not “securities” without violating Section 21(a)(2)(A) of the Glass-Steagall Act. They also would not be required to comply with the Federal Deposit Insurance Act (FDI Act) or the BHC Act. As explained in the next section, the FDI Act and the BHC Act establish crucial public interest safeguards that govern FDIC-insured banks and their parent companies.[73] Congress should make certain that those safeguards apply to all institutions that engage in the business of receiving either bank deposits or “shadow deposits.” Ongoing efforts by cryptocurrency ventures to obtain charters for uninsured, deposit-taking banks have cast a very bright light on a dangerous gap in existing laws. Congress must quickly pass legislation that will close that gap by requiring all issuers and distributors of stablecoins to be FDIC-insured banks. **ii.** **Congress should require all issuers and distributors of** **stablecoins to be FDIC-insured banks, thereby bringing those** **entities within the scope of the FDI Act and the BHC Act.** 71 Lydia Beyoud, “Fintech Charter Suit on Hold as Bank Regulator Reviews Policies,” Bloomberg Law (June 17, 2021); Lydia Beyoud, “Fintech Lender’s Bank Bid Says ‘No Thanks’ on Deposit Insurance,” Bloomberg Law (Dec. 3, 2020). 72 Jesse Hamilton, “Stablecoin Advocates Make Their Case to U.S. Banking Regulators,” Bloomberg Law (Nov. 22, 2021). 73 _See Gorton & Zhang, supra note 7, at 3-6, 17-21, 33-35, 38-39; Wilmarth, “Banking Privileges,” supra note 26, at_ 1, 6-11. 29 ----- PWG’s report urged Congress to act “promptly” in passing legislation that would require all issuers of stablecoins to be FDIC-insured banks.[74] The same requirement should apply to entities, such as Facebook’s Novi, that distribute stablecoins issued by other companies. Firms that distribute stablecoins to the public should not be allowed to avoid compliance with the FDI Act and the BHC Act simply because their deposit-taking and payment networks employ stablecoins issued by other companies. Legislation requiring all issuers and distributors of stablecoins to be FDIC-insured banks would guarantee that those firms must comply with crucial public interest safeguards in the FDI Act. Those safeguards include: (a) deposit insurance coverage, payment of risk-based deposit insurance premiums, and reporting and examination requirements under 12 U.S.C. §§ 1817, 1820 & 1821; (b) supervisory and enforcement powers granted to federal bank regulators under 12 U.S.C. § 1818; (c) procedures for resolving failed and failing banks under 12 U.S.C. §§ 1821(c), 1822 & 1823; (d) risk-based capital requirements and other safety and soundness standards under 12 U.S.C. §§ 1831p-1 & 3901-07; (e) prompt corrective action remedies under 12 U.S.C. § 1831o; (f) safety and soundness requirements and protections for competition that govern proposed changes in control of banks and bank mergers under 12 U.S.C. §§ 1817(j) & 1828(c); (f) prohibitions on abusive tying practices under 12 U.S.C. §§ 1971-77; (f) “source of strength” obligations and capital requirements for parent companies of FDIC-insured banks under 12 U.S.C. §§ 1831o-1 & 5371(b); (g) community reinvestment duties under 12 U.S.C. §§ 3901-08; and (h) expedited funds availability requirements under 12 U.S.C. §§ 4001-10. Mandating status as FDIC-insured banks for all issuers and distributors of stablecoins would also make certain that those entities are treated as “banks” for purposes of the BHC Act.[75] 74 PWG Stablecoin Report, supra note 1, at 2, 16. 75 12 U.S.C. § 1841(c)(1)(A) (defining “bank” for purposes of the BHC Act to include all FDIC-insured banks). 30 ----- The BHC Act requires all companies that own or control FDIC-insured banks to comply with additional public interest safeguards, including (a) safety and soundness standards and protections for competition governing proposed acquisitions of banks under 12 U.S.C. § 1842; (b) limitations on permissible nonbanking activities under 12 U.S.C. § 1843; (c) the Fed’s authority to conduct examinations, require reports, bring enforcement actions, adopt supervisory rules, and impose risk-based capital requirements under 12 U.S.C. §§ 1818, 1844, 1847, & 5371(b); and (d) privacy protections that (i) prohibit financial holding companies from disclosing nonpublic customer information to unaffiliated third parties against their customers’ wishes, and (ii) bar third parties from using false or deceptive practices to obtain such information (15 U.S.C. §§ 6801-09, 6821-27). One of the BHC Act’s most important provisions is 12 U.S.C. § 1843, which prohibits companies that own or control banks from engaging in commercial activities or owning commercial enterprises. Section 1843 prevents the formation of banking-and-commercial conglomerates that would pose grave dangers to our society, financial system, and economy, including (1) hazardous concentrations of economic and financial power and political influence, (2) toxic conflicts of interest that would destroy the ability of banks to act objectively in providing credit and other services, and (3) risks of systemic contagion between financial and commercial sectors that could inflict enormous losses on the federal “safety net” for banks (including the FDIC’s deposit insurance fund, the Fed’s discount window, and the Fed’s guarantee for interbank payments made on Fedwire, as well as the federal government’s explicit and implicit protections for “too big to fail” banking organizations). Requiring all issuers and distributors of stablecoins to be FDIC-insured banks would prevent stablecoin ventures from 31 ----- being owned or controlled by commercial enterprises, including Big Tech firms like Apple, Amazon, Facebook, Google, and Microsoft.[76] Stopping Big Tech firms from acquiring ownership or control of issuers and distributors of stablecoins should be a top priority for financial regulators and Congress. Big Tech firms already enjoy significant advantages over traditional providers of financial services in areas such as automation, artificial intelligence, data management, and mobile payments. The rapid expansion of Ant Financial (Alipay) and Tencent (WePay) in China – prior to the crackdown on both companies by Chinese authorities in 2020 – indicates that Big Tech firms could potentially dominate major segments of our financial industry if they were allowed to offer deposit and payment services. The entry of Big Tech firms into the banking business would create a wide range of potential threats, including unfair competition, market dominance, predatory lending, abusive sharing of customer data and other violations of customer privacy rights, as well as much greater risks of systemic contagion across financial and nonfinancial sectors of our economy during financial crises and severe economic downturns.[77] 76 Wilmarth, “Banking Privileges,” supra note 26, at 6-11; Arthur E. Wilmarth, Jr., “The FDIC Should Not Allow Commercial Firms to Acquire Industrial Banks,” 39 Banking & Financial Services Policy Report No. 5 (May 2020), [at 1, 2-10 [hereinafter Wilmarth, “Industrial Banks”], available at https://ssrn.com/abstract=3613022; Arthur E.](https://ssrn.com/abstract=3613022) Wilmarth, Jr., “Wirecard and Greensill Scandals Confirm Dangers of Mixing Banking and Commerce,” 40 Banking _& Financial Services Policy Report No. 5 (May 2021), at 1, 11-12 [hereinafter Wilmarth, “Wirecard and_ [Greensill”], available at https://ssrn.com/abstract=3849567.](https://ssrn.com/abstract=3849567) 77 _See Raúl Carrillo, Testimony before the Subcomm. on Consumer Protection and Financial Institutions of the_ House Comm. on Financial Services 7-9 (April 15, 2021) [hereinafter Carrillo Testimony], [https://docs.house.gov/meetings/BA/BA15/20210415/111447/HHRG-117-BA15-Wstate-CarrilloR-20210415.pdf;](https://docs.house.gov/meetings/BA/BA15/20210415/111447/HHRG-117-BA15-Wstate-CarrilloR-20210415.pdf) Kathryn Petralia, Thomas Philippon, Tara Rice & Nicholas Véron, Banking Disrupted? Financial Intermediation in _an Era of Transformational Technology 25–38, 44–82 (Geneva Reports on the World Economy 22, 2019),_ [https://www.cimb.ch/uploads/1/1/5/4/115414161/banking_disrupted_geneva22-1.pdf; Investigation of Competition](https://www.cimb.ch/uploads/1/1/5/4/115414161/banking_disrupted_geneva22-1.pdf) _in Digital Markets: Majority Staff Report and Recommendations (Subcomm. on Antitrust, Commercial and_ Administrative Law of the House Comm. on the Judiciary, 2020) [hereinafter 2020 House Staff Report on [Competition in Digital Markets], https://judiciary.house.gov/uploadedfiles/competition_in_digital_markets.pdf;](https://judiciary.house.gov/uploadedfiles/competition_in_digital_markets.pdf?utm_campaign=4493-519) Wilmarth, “Banking Privileges,” supra note 26, at 6-11; Wilmarth, “Industrial Banks,” supra note 76, at 4-10; Wilmarth, “Wirecard and Greensill,” supra note 76, at 11-13. 32 ----- Facebook’s plan to offer deposit and payment services through Novi poses unacceptable threats to consumer privacy and welfare. Facebook has repeatedly abused its customers’ privacy rights and has reportedly marketed products that it knew were harmful to its customers. In 2012, Facebook entered into a consent decree with the Federal Trade Commission (FTC) to settle charges that it deceived customers and violated its promises to allow customers to control the privacy of information they posted on Facebook. In 2019, Facebook paid a $5 billion fine to resolve the FTC’s charges that Facebook violated the privacy commitments included in the 2012 consent decree.[78] The FTC recently launched a new investigation of Facebook after a whistleblower informed Congress that Facebook knew from its internal research that some of its products caused mental health problems in minors as well as other harms to customers.[79] Facebook has long sought to enter the financial services industry to extend its dominance over social networks and expand its access to customer information. In 2012, Facebook founder and CEO Mark Zuckerberg said that the launch of a successful payment service would give Facebook “a pretty awesome combo and a good reason for people to use [Facebook’s] platform,” as well as making it “more acceptable for us to charge them quite a bit more for using [our] platform.”[80] Offering deposit and payment services would greatly increase Facebook’s ability to access, leverage, and monetize its customers’ private information. As Open Markets Institute recently pointed out, 78 “FTC Imposes $5 Billion Penalty and Sweeping New Privacy Restrictions on Facebook,” (Fed. Trade Comm’n, [July 24, 2019), https://www.ftc.gov/news-events/press-releases/2019/07/ftc-imposes-5-billion-penalty-sweeping-](https://www.ftc.gov/news-events/press-releases/2019/07/ftc-imposes-5-billion-penalty-sweeping-new-privacy-restrictions) [new-privacy-restrictions.](https://www.ftc.gov/news-events/press-releases/2019/07/ftc-imposes-5-billion-penalty-sweeping-new-privacy-restrictions) 79 John D. McKinnon & Brent Kendall, “Federal Trade Commission Scrutinizing Facebook Disclosures,” Wall _[Street Journal (Oct. 27, 2021), https://www.wsj.com/articles/facebook-ftc-privacy-kids-11635289993; John D.](https://www.wsj.com/articles/facebook-ftc-privacy-kids-11635289993)_ McKinnon & Ryan Tracy, “Facebook Whistleblower’s Testimony Builds Momentum for Tougher Tech Laws,” _[Wall Street Journal (Oct. 5, 2021), https://www.wsj.com/articles/facebook-whistleblower-frances-haugen-set-to-](https://www.wsj.com/articles/facebook-whistleblower-frances-haugen-set-to-appear-before-senate-panel-11633426201)_ [appear-before-senate-panel-11633426201.](https://www.wsj.com/articles/facebook-whistleblower-frances-haugen-set-to-appear-before-senate-panel-11633426201) 80 Open Markets Facebook Letter, supra note 22, at 12 (quoting Zuckerberg’s statement). 33 ----- Facebook occupies a dominant role in American life and indeed the lives of people around the world, with over 1 billion users for four of its services, including Facebook, Instagram, Messenger, and WhatsApp. Facebook is also a giant in the advertising space, with their 2020 advertising revenue close to $84.2 billion dollars — nearly $1.6 billion each week.[81] A 2020 House subcommittee staff study found that “Facebook has monopoly power in the market for social networking” and has exploited that market power by becoming a “gatekeeper” with “outsized power to control the fates of other companies.”[82] Facebook generates most of its revenues by selling digital advertising. Facebook’s access to the private information of hundreds of millions of customers enables it to command much higher prices for its sales of advertising, compared with its competitors.[83] The House staff study concluded that Facebook’s dominance of the social networking market – like Google’s dominance of the Internet search market – allows Facebook to “abuse consumers’ privacy without losing customers.”[84] As one expert advised the House subcommittee: Facebook and Google have built comprehensive dossiers on almost everyone, and they can sell incredibly targeted advertisement on that basis. . . . But doing so represents an inherent violation of the receiver’s privacy. Every ad targeted using personal information gathered without explicit, informed consent is at some level 81 _Id._ 82 2020 House Staff Report on Competition in Digital Markets, supra note 77, at 12-14, 39-40; see also id. at 132-73 (providing a detailed analysis of Facebook’s exploitation of its market power, including its acquisition of numerous competitors). 83 _Id. at 170-72._ 84 _Id. at 18, 51-53._ 34 ----- a violation of privacy. And Facebook and Google are profiting immensely by selling these violations to advertisers.[85] Novi’s deposit and payment services could enable Facebook to collect and monetize a vast array of data about its customers’ financial assets and transactions. The treasure trove of nonpublic customer information that Big Tech firms could capture by offering deposit and payment services is indicated by the huge data set compiled by JPMorgan Chase Institute (JPMCI). JPMCI has collected and analyzed a massive pool of information, drawn from the records of JPMorgan Chase (JPMC), the largest U.S. bank, containing the “saving, spending, and borrowing habits of the bank’s customers.” The Fed “has used the Institute’s research when weighing interest-rate decisions,” thereby confirming the enormous value of JPMC’s comprehensive information about its customers’ financial dealings.[86] Allowing Facebook and other Big Tech firms to build similar data sets by offering deposit and payment services would increase exponentially their ability to monetize customer information and degrade customer privacy by secretly transferring that information to third-party sellers of goods and services. Requiring all issuers and distributors of stablecoins to be FDIC-insured banks would guarantee that all companies that own or control those entities must comply with the privacy protections governing financial holding companies (15 U.S.C. §§ 6801-09, 6821-27). That requirement would also prevent Facebook and other Big Tech firms from offering deposit and payment services built around stablecoins. The PWG Report correctly determined that “the combination of a stablecoin issuer or [digital] wallet provider and a commercial firm could lead 85 _Id. at 54-55 (quoting testimony of David Heinemeier Hansson, co-founder and chief technology officer of_ Basecamp). 86 David Benoit, “How JPMorgan Is Helping Washington Make Sense of the Pandemic Recovery,” Wall Street _[Journal (Nov. 9, 2021), https://www.wsj.com/articles/how-jpmorgan-is-helping-washington-make-sense-of-the-](https://www.wsj.com/articles/how-jpmorgan-is-helping-washington-make-sense-of-the-pandemic-economy-11636462980)_ [pandemic-economy-11636462980.](https://www.wsj.com/articles/how-jpmorgan-is-helping-washington-make-sense-of-the-pandemic-economy-11636462980) 35 ----- to an excessive concentration of economic power,” which could “restrict access” to credit and other financial services and have “detrimental effects on competition.”[87] Our nation stands at a crossroads. We can maintain the BHC Act’s longstanding policy of separating banking and commerce, thereby preserving a financial sector, a national economy, and a society that are not compromised by toxic conflicts of interest, exploited by unfair competitive advantages, or dominated by the overwhelming economic power and political influence of giant banking-and-commercial conglomerates. Or we can allow Facebook and other Big Tech firms to enter the banking business and leverage their stablecoin ventures to create massive “shadow banking” empires, thereby subverting the BHC Act’s separation of banking and commerce. In that event, Big Tech firms might well gain dominance over our banking industry – either by building their own financial kingdoms or by combining with our largest banks – thereby creating the very evils that the BHC Act was designed to prevent.[88] **iii.** **The FDIC should not approve pass-through deposit insurance** **coverage for stablecoins.** The FDIC is reportedly considering the possibility of allowing FDIC-insured banks to hold reserves deposited by stablecoin issuers and provide “pass-through” deposit insurance coverage to customers of those issuers.[89] The FDIC currently grants pass-through deposit insurance coverage to holders of stored-value cards if the issuers of those cards satisfy the following conditions: (i) each issuer must establish a custodial deposit account at an FDIC insured bank to hold the funds owned by card holders, (ii) the issuer must allow card holders to 87 PWG Stablecoin Report, supra note 1, at 14. 88 Carrillo Testimony, supra note 77, at 7-9; Wilmarth, “Banking Privileges,” supra note 26, at 6-11; Wilmarth, “Industrial Banks,” supra note 76, at 4-12; Wilmarth “Wirecard and Greensill,” supra note 76, at 11-14. 89 Nate DiCamillo, “US FDIC Said to Be Studying Deposit Insurance for Stablecoins,” CoinDesk (Oct. 6, 2021), [https://www.coindesk.com/policy/2021/10/06/us-fdic-said-to-be-studying-deposit-insurance-for-stablecoins/.](https://www.coindesk.com/policy/2021/10/06/us-fdic-said-to-be-studying-deposit-insurance-for-stablecoins/) 36 ----- access their funds at the bank through withdrawals or transfers to third parties, (iii) the bank’s records must confirm that the issuer has established a custodial deposit account holding funds owned by card holders, (iv) either the bank’s records or the issuer’s records must show, on an accurate and current basis, the identity of each card holder and the amount of funds owned by that holder, and (v) the issuer must inform card holders that their funds are held in a custodial account at the bank.[90] Approving pass-through deposit insurance for stablecoins would involve a number of operational difficulties. One of the most significant challenges would be the requirement that either the custodial bank or the stablecoin issuer must maintain current and accurate records showing the identity of each holder of stablecoins and the amount of stablecoins owned by that holder. As the PWG’s report pointed out, The majority of the stablecoin market currently operates on public blockchains where transactions may be pseudonymous, meaning the identity of the sender or the receiver of a transaction is unknown, but other transactional information is available (e.g., the amount, the time, the value, etc.).”[91] Indeed, the relative anonymity of transactions conducted with stablecoins – compared with traditional payment methods other than cash – is a major reason for the popularity of stablecoins.[92] It is difficult to envision how stablecoin issuers and custodial banks could maintain accurate and current records showing the identities of holders of stablecoins and the amounts of 90 FDIC, New General Counsel Opinion No. 8 (Oct. 31, 2008), available at 73 Fed. Reg. 67155-57 (Nov. 13, 2008). 91 PWG Stablecoin Report, supra note 1, at 19 n.39; see also DiCamillo, supra note 84 (explaining that stablecoins are transferred on “public blockchain networks . . . and theoretically anyone with a crypto wallet that hasn’t been blacklisted can receive stablecoins from, and send them to, other wallets.”). 92 _See supra notes 11-12 and accompanying text._ 37 ----- coins they own without changing the fundamental nature of stablecoin transactions that are conducted on distributed ledgers (particularly in DeFi transactions). A financial journalist recently described the following problems that stablecoin issuers would confront if “financial regulators declare that all stablecoin owners must be verified”: Building infrastructure to collect and verify the identity of all users [of stablecoins], and not just the few who redeem or deposit, is expensive. To recoup their costs, issuers . . . may consider introducing fees. All of this could render stablecoins less accessible for people who only want to use them for casual remittances. . . . In DeFI, stablecoins are often deposited into accounts controlled by bits of autonomous code, or smart contracts, which don't have any underlying owner. It’s not evident how a stablecoin issuer can conduct KYC [compliance] on a smart contract.[93] Thus, it would be extremely difficult, if not impossible, for a stablecoin issuer or a custodial bank to maintain an accurate and current record of the identities of stablecoin holders or the amounts of coins they currently own in order to qualify for pass-through deposit insurance coverage from the FDIC. Moreover, granting pass-through deposit insurance coverage to holders of stablecoins would not remove the systemic perils created by the issuers and distributors of those coins. Pass through coverage would allow issuers and distributors of stablecoins and their customers to benefit from access to the FDIC’s deposit insurance fund. Issuers, distributors, and customers would also benefit indirectly from the custodial bank’s access to the Fed’s discount window 93 Koning, supra note 11. 38 ----- advances and payments system guarantees, as well as other components of the federal safety net for banks. In contrast, pass-through deposit insurance coverage would not require issuers and distributors of stablecoins to comply with the FDI Act’s provisions that protect customers, communities, businesses, and the stability of the banking system. In addition, pass-through coverage would allow companies that own or control issuers and distributors of stablecoins to avoid complying with the BHC Act’s safeguards, including the Fed’s regime of consolidated regulation and supervision, the privacy rules governing financial holding companies, and the separation of banking and commerce. In sum, pass-through deposit insurance coverage would enable Facebook and other Big Tech firms to offer deposit and payment services and receive extensive benefits from the federal safety net for FDIC-insured banks without complying with the public interest mandates governing those banks and their parent companies. Pass-through coverage would effectively create a “back door” that allows Big Tech firms to compete directly with traditional banks, undermine the BHC Act’s separation of banking and commerce, and circumvent other important public interest safeguards. Pass-through deposit insurance coverage for stablecoins would produce many of the harmful effects of “rent-a-bank” arrangements, which the OCC approved when it adopted its so called “true lender” rule in October 2020. The OCC’s rule declared that a national bank would be treated as the “true lender” for a loan as long as the bank funded the loan at closing or was named as the lender in the loan agreement, even if the bank transferred its entire interest and entire risk in the loan to a nonbank “partner” the following day. Under the OCC’s rule, nonbanks could have reaped the benefits that their national bank partners enjoyed under federal statutes preempting the application of state usury laws and other state consumer protection laws 39 ----- to national banks.[94] In June 2021, Congress passed a joint resolution that repealed the OCC’s rule under the Congressional Review Act.[95] Members of Congress who supported the joint resolution condemned the OCC’s “true lender” rule for allowing predatory nonbank lenders “to use superficial and deceptive partnerships with [national] banks to skirt state laws and charge outrageous annual percentage rates” on loans they acquired from their national bank partners.[96] The FDIC should reject pass-through deposit insurance coverage for stablecoins for the same reasons that Congress repealed the OCC’s “true lender” rule. Issuers and distributors of stablecoins should not be allowed to obtain the benefits provided by FDIC insurance and other components of the federal safety net for banks unless those entities become FDIC-insured banks. Issuers and distributors of stablecoins and their parent companies should not be allowed to use “rent-a-bank” arrangements to engage in destructive regulatory arbitrage. Instead, they should be required to comply fully with the essential public interest safeguards contained in the FDI Act and the BHC Act.[97] **Conclusion** The PWG’s report provides a welcome blueprint for top-priority actions by regulatory agencies and Congress. The SEC should use its existing powers to regulate stablecoins as 94 Wilmarth, “Banking Privileges,” supra note 26, at 2, 14-17; see also Carrillo Testimony, supra note 77, at 6-7; Adam J. Levitin, “Rent-a-Bank: Bank Partnerships and the Evasion of Usury Laws,” 71 Duke Law Journal 329 (2021). 95 Davis Polk, “Client Update: The OCC’s true lender rule has been repealed” (July 1, 2021), [https://www.davispolk.com/insights/client-update/occs-true-lender-rule-has-been-repealed.](https://www.davispolk.com/insights/client-update/occs-true-lender-rule-has-been-repealed) 96 Senate Comm. on Banking, Housing, and Urban Affairs, “Majority Press Release: President Signs Van Hollen, Brown Legislation to Strike Down Trump-era ‘Rent-a-Bank’ Rule” (June 30, 2021) (quote), [https://www.banking.senate.gov/newsroom/majority/president-signs-van-hollen-brown-legislation-to-strike-down-](https://www.banking.senate.gov/newsroom/majority/president-signs-van-hollen-brown-legislation-to-strike-down-trump-era-rent-a-bank-rule) [trump-era-rent-a-bank-rule; see also House Financial Services Comm., “Press Release: Waters Floor Statement on](https://www.banking.senate.gov/newsroom/majority/president-signs-van-hollen-brown-legislation-to-strike-down-trump-era-rent-a-bank-rule) House Passage of Resolution to Eliminate Trump’s Predatory ‘True Lender’ Rule” (June 25, 2021), [https://financialservices.house.gov/news/documentsingle.aspx?DocumentID=408055.](https://financialservices.house.gov/news/documentsingle.aspx?DocumentID=408055) 97 _See Carrillo Testimony, supra note 77, at 7-9 (supporting the proposed “STABLE Act,” introduced in December_ 2020 by three House members); see also Press Release, “Tlaib, García, and Lynch Introduce Legislation Protecting Consumers Against Cryptocurrency-Related Financial Threats” (Dec. 2, 2020) (describing the proposed STABLE [Act), https://tlaib.house.gov/media/press-releases/tlaib-garcia-and-lynch-stableact.](https://tlaib.house.gov/media/press-releases/tlaib-garcia-and-lynch-stableact) 40 ----- “securities” and protect investors and securities markets. DOJ should designate stablecoins as “deposits” and bring enforcement actions to prevent issuers and distributors of stablecoins from violating Section 21(a) of the Glass-Steagall Act. To overcome uncertainties and gaps in the remedies available to the SEC and DOJ, Congress should pass legislation requiring all issuers and distributors of stablecoins to be FDIC-insured banks. The foregoing measures are urgently needed to counteract the grave dangers that stablecoins pose to our society, financial system, and economy. 41 -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.2139/ssrn.4000795?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.2139/ssrn.4000795, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "CLOSED", "url": "" }
2,021
[]
false
null
[]
28,283
en
[ { "category": "Business", "source": "external" }, { "category": "Law", "source": "s2-fos-model" }, { "category": "Political Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/015ece3905050fcb19ea459d7aebfa88a6d015a3
[ "Business" ]
0.934849
The case for a technically safe environment to protect the identities of anonymous whistle-blowers : a conceptual paper
015ece3905050fcb19ea459d7aebfa88a6d015a3
[ { "authorId": "2799211", "name": "J. Loggerenberg" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
Whistle-blowing is one of the most important aspects in the fight against corruption. In most cases, it is impossible to commit a corrupt deed without at least one other person being involved in, or, knowing about it. Many businesses and public entities have a ‘crime lines’ in place to facilitate whistle-blowing but these facilities are mostly limited to a particular telephone number. Using a telephone to report corruption has inherent problems when anonymity is required. Organisations can easily trace calls made from their facilities. Even calls made from cellular phones can be traced. Eavesdropping, although illegal, is not technically difficult to achieve. The fact is: Telephonic whistle-blowing provides little protection for the whistle-blower who wants to remain anonymous. This paper focuses on the problem of current ways for whistle-blowing and suggests an improvement conceptually. It aims to open up debate and discussion on this topic with the intention to attract further contributions and stimulate research on this topic. Although the paper focuses strongly on the situation in South Africa, it is probably equally applicable anywhere else in the world.   Key words: Whistle-blowing, Information Technology, Anonymity, Onion Routing.
African Journal of Business Management Vol.6 (44), pp. 10799-10806, 7 November 2012 Available online at http://www.academicjournals.org/AJBM DOI: 10.5897/AJBM12.992 ISSN 1993-8233 ©2012 Academic Journals ## Review # The case for a technically safe environment to protect the identities of anonymous whistle-blowers: A conceptual paper ### Johan van Loggerenberg [Department of Informatics, University of Pretoria, Pretoria, 0001 South Africa. E-mail: johan.vl@up.ac.za.](mailto:johan.vl@up.ac.za) Accepted 25 September, 2012 **Whistle-blowing is one of the most important aspects in the fight against corruption. In most cases, it is** **impossible to commit a corrupt deed without at least one other person being involved in, or, knowing** **about it. Many businesses and public entities have a ‘crime lines’ in place to facilitate whistle-blowing** **but these facilities are mostly limited to a particular telephone number. Using a telephone to report** **corruption has inherent problems when anonymity is required. Organisations can easily trace calls** **made from their facilities. Even calls made from cellular phones can be traced. Eavesdropping,** **although illegal, is not technically difficult to achieve. The fact is: Telephonic whistle-blowing provides** **little protection for the whistle-blower who wants to remain anonymous. This paper focuses on the** **problem of current ways for whistle-blowing and suggests an improvement conceptually. It aims to** **open up debate and discussion on this topic with the intention to attract further contributions and** **stimulate research on this topic. Although the paper focuses strongly on the situation in South Africa, it** **is probably equally applicable anywhere else in the world.** **Key words: Whistle-blowing, Information Technology, Anonymity, Onion Routing.** **INTRODUCTION** Reporting suspicious incidents is one of the prime components in the detection of criminal activities. Without such reporting, law enforcement becomes much less effective with the result that many crimes take place without being detected. Suspicious incidents are typically facilitated through the use of “Crime Lines”. These are typically telephone numbers that can be dialled by whistle-blowers. In most cases protection of the identity of the whistle blower is of paramount importance. To report an incident which exposes another individual can be seriously dangerous to the person blowing the whistle and cases are known where such whistle-blowers have lost their lives. Many cases have been documented where whistle blowers chose to reveal their identity. When the identity of the whistle-blower is known, investigation by law enforcement agencies is strongly facilitated and is, as such, always preferred. An argument can be made that a whistle-blower is protected by the law and should, therefore, have no fear. However, whilst the law theoretically protects whistle-blowers from being victimised, the practical side is far from safe. Anonymity must therefore be viewed as a pre-requisite for whistle-blowers that choose to remain anonymous. Crime lines (that is, telephone lines) are currently being used as the preferred mechanism to report suspicious incidents. Unfortunately, telephone reporting provides very little in terms of anonymity. Most companies keep logs of telephone calls being made from internally and if someone is suspected to have reported an incident, such logs can, with relative ease, be accessed by perpetrators to determine where the call was made from. The aim of this conceptual paper is to investigate the potential of Information and communications technology (ICT) to enable a safe alternative to facilitate anonymous whistle-blowing. As is the case with concept papers, the research ----- 10800 Afr. J. Bus. Manage. methodology is primarily based on the limited literature available on the topic. The author made use of newspaper reports to provide situational context. **WHISTLE-BLOWING** The term whistle-blowing originated from the days when the whistle was blown by a police officer when witnessing a criminal deed or by a referee when witnessing someone contradicting the rules of the sports game. There is, however, as metaphors regarding whistle-blowing, a distinct difference between the use of a policeman or a referee. Policemen and referees have the authority to enforce their actions whereas present day whistleblowers do not enjoy such authority (Ellison, 1982). The whistle-bloweris dependent on someone else with authority. The idea behind the term „whistle-blower‟ is, however, very positive. It occurs when someone violates an accepted rule or law and should be stopped from doing so. The whistle-blower does so in the public interest. The whistle-blower is, in fact, trying “…to enlist the support of _others to achieve social objectives” (Ellison, 1982)._ Whistle-blowing is, in general, defined as “raising a concern about malpractice within an organisation”. This definition is attributed to the UK Committee on Standards in Public Life (Camerer, 2001; Martin, 2010). Another definition is the “pursuit of a concern about wrongdoing that does damage to a wider public interest” (Public Concern at Work, 2005). This second definition puts whistle-blowing specifically in the context of the public interest as opposed to the first one which is more specifically about the „organisation‟. Near and Miceli (1985) define whistle-blowing as “the _disclosure by organization members (former or current) of_ _illegal, immoral, or illegitimate practices under the control_ _of their employers, to persons or organizations that may_ _be able to effect action”._ Definitions identify that something illegal or un acceptable is taking place and the person raising the alarm (the „whistle-blower‟) is drawing attention to this fact. The ultimate intention is to avoid a repetition. The person blowing the whistle is, therefore, simply the „messenger‟ who acts in the best interest of the organisation, or in the best interest of society (Camerer, 2001). It is understandable that those people actively parti cipating in the deed will not appreciate their actions being exposed. What comes as a surprise, is that, often, even innocent „onlookers‟ are also critical of the whistleblower‟s action. The result is that whistle-blowers have been getting a poor reputation from some quarters (Camerer, 2001). In an organisational sense, a whistle-blower can be either one of the employees (that is, internally to the organisation), or externally to the organisation. Famous internal whistle-blowers were Sherron Watkins (in the Enron case) and Cynthia Cooper (Worldcom) (Colvin, 2002). The auditing profession, for instance, can be seen as mandatedwhistle-blowers. They are mandated by the shareholders to look for anything outside of the rules, regulations or laws and formally report it to the management and the shareholders. There can be no doubt that whistle-blowing plays a very important role in the fight against corruption (Martin, 2010). Not only does it play a deterring role, but it also assists in bringing criminals to book by making law enforcement agencies aware of criminal activities for further investigation. Estimates of the scale of corruption in South Africa vary considerably, but it is safe to say that it runs into the billions of Rands every year. The South African Minister of Finance (speaking of Income Tax revenue), reported that R13 Billion less would be collected than what was budgeted. This „loss‟ is probably far less than what is lost through corruption, let alone what is lost on wasteful expenditure and, what the Minister calls „extravagance in _public administration‟ (Treasury, 2011). South Africa can_ certainly not afford the losses suffered through corruption. Whistle-blowing, therefore, plays an important role to combat such losses. **LEGAL PROTECTION FOR WHISTLE-BLOWERS** Identified whistle-blowers face many risks (Martin, 2010). These risks can take the form of being victimised in the workplace, and/or the very real possibility of retaliation by those involved in the criminal deeds. To protect whistle-blowers in the workplace, legal protection is available in many countries, including South Africa. Martin (2011) makes the point that “[p]rotection for _whistle-blowers is essential to create a culture of_ _disclosure of wrong-doing”._ The primary South African Act addressing the phenol mena of corruption, is the Prevention and Combating of Corrupt activities 2003, Act 12 of 2004 (Republic of South Africa, 2004). The Protected Disclosures Act, 26 of 2000 (Republic of South Africa, 2000) was designed specifically to provide such protection. Martin (2010, 2011) and Camerer (2001) deal extensively with the legal intentions of the Act in their papers. Both the Companies Act, Act 71 of 2008 (Republic of South Africa, 2008) and the Companies Amendment Act, 3 of 2011 (Republic of South Africa, 2011) also provide legal protection for whistle-blowers. These Acts do not specifically refer to anonymous whistle-blowing, and seem to assume that the identity of the whistle-blower is known. ----- What also needs to be noted is that the legal protection provided by these Acts, is limited to whistle-blowers with known identities. This is to be expected as protection can only be provided to someone who is known. This point is further discussed in the summary to the area dealing with legal protection. Despite the legal protection provided in the Acts, cases are still reported where whistle-blowers did not enjoy the protection promised in the Acts. The (London) Guardian newspaper (Syal, 2010) reports that “…[employment tribunal statistics show that the total _number of people using whistleblowing legislation, which_ _aims to protect workers from victimisation if they have_ _exposed wrongdoing, increased from 157 cases in 1999_ _to 1,791 10 years later”. The article quotes many cases of_ whistle-blowers being dismissed or being „gagged‟. Statistics for the South African situation are not available, but there is plenty evidence of whistle-blowers being dismissed or victimised (Martin, 2010). In an article appearing in the (South African) Mail and Guardian (Calland, 2011) a case is described where a municipal manager was dismissed and her house burnt down when she initiated an investigation into fraud. The article indicates that, although her dismissal had been ruled unfair in the court, her employers still refused to give her back her job. In the same article, it is reported that “…14 government _officials_ _or_ _politicians_ _have_ _been_ _murdered_ _[in_ _Mpumalanga province] since 1998‟ and that there was a_ _“…twelve-fold increase in wasteful expenditure since_ _2007, but a sharp decrease in the number of whistle-_ _blowers coming forward to report malfeasance” (Calland,_ 2011). In yet another article, the (SA) Mail and Guardian Online (2007) reported a case where the medical superintendent of a hospital in the Eastern Cape was dismissed when _“speaking out against [the hospital‟s]_ _handling of the Frere Hospital maternity saga” (Mail and_ Guardian Online, 2007). The superintendent alleged that “200 babies were dying every month at East London‟s _two largest hospitals” (Mail and Guardian Online, 2007)._ Martin (2010, 2011) raises a number of concerns about the adequacy of the protection provided by the legal framework in South Africa. Adv. Madonsela, The Public Protector in South Africa, echoed the same concern (Martin, 2011; Public Protector, 2010). Miceli et al. (1999) point out that, whilst lawmakers generally want to believe that protecting whistle-blowers from retaliation will encourage the practice of whistleblowing, the contrary is true. They quote several research papers with supporting evidence that _“legal protections_ _neither reduce the incidence of retaliation nor increase_ _the incidence of whistle-blowing” (Miceli et al., 1999). In_ their research (covered in the 1999 paper) they again tested, _inter alia, two hypotheses. The first was that an_ effective law (that is, protecting whistle-blowers) was Loggerenberg 10801 likely to cause whistle-blowers to identify themselves rather than do so anonymously. The second was that an effective law was likely to cause perceived retaliation to be less likely to follow identified whistle-blowing. In both hypotheses they found the opposite to what they expected. In the first hypothesis, the number of identified whistle-blowers reduced from 74 to 60%. In the second hypothesis they found that the percentage of identified whistle-blowers who suffered retaliation increased from 15 to 33%. In summary: Despite the legal protection found in the Acts, whistle-blowers experience that such protection is only partly effective at best. This, in itself, should discourage whistle-blowers from disclosing their identities when blowing the whistle, thereby leading to anonymous whistle-blowing. Blowing the whistle anonymously, however, disqualifies them from the legal protection they would have enjoyed as an identified whistle-blower. As an anonymous whistle-blowers, _they would not need_ _protection, on condition that they remain anonymous._ **ANONYMOUS VS IDENTIFIED WHISTLE-BLOWING** Anonymity means that the person‟s _“…identity is not_ _publicly known” (Ellison, 1982). The question is: Is it_ acceptable for someone to remain anonymous when blowing the whistle or is the person obliged to reveal his identity? As Ellison (1982) points out, one has to distinguish between anonymity and two other, closely related terms, namely, secrecy and privacy. He argues that secrecy requires a _“conspiracy of silence”, thereby implying that_ more than one person knows the secret (Ellison, 1982). Something only known to one person is, according to Ellison, the _“extreme form of secrecy”. This type of_ secrecy seems to fall outside of the scope of whistleblowing as it is highly unlikely that one will blow the whistle on oneself. In the context of whistle-blowing, it is the _denial of_ _access to information to others that makes it a secret_ (Ellison, 1982). Privacy, according to Ellison (1982), occurs when one can justify why others are not allowed to share information that one has. He quotes the example of one‟s sex life. One has the right to exclude others from such _private_ information, unless someone else can invoke a higher right than one‟s own to force one to disclose such information. Ellison (1982) makes an important point that, regarding _privacy, the burden of the proof rests with the other party_ who wants to have access to such information. Regarding secrecy, the burden of proof is reversed in that the person with the information has to justify why it should remain secret. This raises the question of whether anonymity in the ----- 10802 Afr. J. Bus. Manage. context of whistle-blowing is more on the side of secrecy or more on the side of privacy. Ellison (1982) argues that the kind of information which is being withheld is not about the deed itself, but about the whistle-blower‟s identity. The question becomes whether the public has a right to know the whistle-blower‟s identity or whether the whistleblower has the right to withhold it (Ellison, 1982). Ellison (1982) points out that many members of the public would consider an anonymous whistle-blower to be _“saying nasty things behind people‟s backs” and, as a_ result, would argue that anonymous whistle-blowing should be discouraged. However, when one considers the risks associated with identified whistle-blowing, the case for anonymity gets much stronger (and the case for speaking behind people‟s backs, weaker). Ellison (1982) argues in favour of identified whistle blowing but concedes that many factors influence the argument. On the one hand, anonymous whistle-blowing _“impedes the pursuit of truth” because it makes law_ enforcement considerably more difficult. Another factor to take into consideration in favour of anonymity is that fear or retaliation by the whistle-blower may result in keeping quiet if not allowed to report it anonymously. Clearly, blowing the whistle anonymously is infinitely more valuable than not blowing it at all (Ellison, 1982). Ellison concludes that _“…blanket condemnation on_ _anonymity is not warranted”. He proposes that the justifi-_ cation must take three factors into consideration: the seriousness of the offense, the probability of retaliation and the social relationships. One can conclude that identified whistle-blowing is definitely the preferred way of blowing the whistle, but, given the risks that have been outlined, there is a strong case to be made for anonymous whistle-blowing. **CURRENT WAYS OF BLOWING THE WHISTLE** There are many ways by which a person can report a suspicious incident. In an organisational context, the first option is to share the concern with a colleague or, more likely, with a supervisor. This unequivocally means that the person who reports the deed is known (referred to as an identified whistle-blower). An identified whistlebloweris, as minimum, known to the person reporting it to, and may also be publicly known. If the supervisor does not respond in a way acceptable to the whistle-blower, the person can report it to higher levels of management, or, report it externally to, say, a newspaper. It could also be reported by means of a „crime line‟ to an agency put in place by the organisation. Crime lines are commonly services procured from outside the organisation, such as ones offered by auditing firms. Reporting a suspicious incident to an external source allows the whistle-blower to either be an identified whistle-blower, or to report it without mentioning his[1] identity. By using technology (as opposed to face-to-face communication), the whistle-blower is given the choice to remain anonymous. Reporting a suspicious incident by not revealing one‟s identity, _assumes anonymity but when one analyses the_ mechanisms facilitating such reporting (the channels), the identity of the caller may be revealed through the channel used. In some cases it may be very simple and in others, whilst more difficult, still very possible and feasible. This has the unpleasant surprise to the whistle-blower that he may be identified - despite his intention to remain anonymous. **Internal telephones** Consider the case where a whistle-blower makes a call from inside his organisation by using the telephone exten-sion assigned to him. Most organisations, as normal good practice, keep logs of all calls made from extensions. These logs, commonly only records the event and not the content itself (although there are some cases where the entire conversation is recorded). By simply analysing the logs of calls made to the „crime line‟ would reveal the extension from which the call was made and the identity of caller can be revealed. This is, obviously, the worst way of trying to be an anonymous whistleblower. We are of the opinion that many whistle-blowers use this option without realising the risk of being identified as a standard feature of technology. Making the call from someone else‟stelephone exten sion will cause the wrong person to be suspected of making the call. This will make it more difficult for a pursuer to identify the true whistle-blower but, the organisation, and the extension from where the call was made, is known. Because pursuers, typically, have good suspicions as to who may possess the relevant information to report them, a good guess can lead them to the whistle-blower. In the worst case, an innocent person may targeted. **Public telephones** What if the whistle-blower goes to a public telephone and makes the call to the crime line? The only way for the pursuer to detect such reporting would be to constantly monitor the crime line number, in other words, to eavesdrop. This is technically quite feasible, albeit illegal. It may also require voice recognition to identify the caller. This may be easy in some cases and more difficult in 1 When referring to the term ‘his’ in respect of a whistle-blower, pursuer or criminal both the male and the female gender is implied. ----- others. Public telephones, therefore, still hold risks to the whistle-blower. **Cellular telephones** Blowing the whistle by using a cellular phone is equally dangerous. Firstly, cellular service providers keep logs of calls made (excluding the content). Normally one needs a court order to get access to such logs and it is likely to be challenging for a pursuer to obtain such permission. Of course, the pursuer can always „persuade‟ an employee of the cellular network to - illegally - obtain the data on his behalf. Even if the pursuer is prevented from getting access to the cellular call logs, it is quite possible and feasible to „eavesdrop‟ on the crime line number (illegal) and „listen in‟ to all the calls being made to the crime line. Callers can be identified through voice recognition techniques or, if the whistle-blower is known to the pursuer, he can quite easily be recognised. **Postal services** Another option open to a whistle-blower is to use the postal service. For instance, a whistle-blower can easily obtain the postal address of the Public Protector and send her an anonymous letter detailing the incident. In this case the „strength‟ of the anonymity is strong, but only the public is not encouraged to use this way of whistle-blowing. Once the public is encouraged to use this mechanism, pursuers only have „persuade‟ someone where mail is received, to intercept suspicious mail. **Email** Some organisations have made facilities available to whistle-blowers to report incidents via email. In some cases, these emails are encrypted. Sometimes these are addressed to a recipient internally to the organisation (for example, Internal audit), or, sometimes externally (for example, an Auditing Firm). The origin for such email messages are easy to trace for the organisation concerned through logs being kept as a standard feature of email platforms. If the message is encrypted, it may be difficult to decipher the content, but the origin would be simple to trace as it would normally have the originator‟s name appearing in the message. The point is this: All the aforementioned mechanisms described have weaknesses regarding the anonymity of the whistle-blower. Weaknesses create risks to the whistle-blower and prospective whistle-blowers will assess such risks when deciding to blow the whistle or not. One needs to acknowledge that the seriousness (or Loggerenberg 10803 scale) of the case also plays an important role. The extent to which the pursuer is prepared to go to prevent or detect whistle-blowers, is directly related to the seriousness of the offence and the severity of consequences the pursuer faces. There will be a world of difference between a traffic official taking a R100 bribe and someone taking a R10 million bribe in, say, an arms deal. Equally, a prospective whistle-blower in an arms deal involving billions of Rands, will expect a much higher level of anonymity than a prospective whistle-blower involving a few thousand Rands in traffic fines. To make it more real: If large scale corruption did indeed take place in South Africa‟s much publicised arms deal, it is a reasonable assumption that „someone out there‟ is in possession of, or have access to documentary evidence to prove such corruption. It is also reasonable to assume that such „someone‟ will think twice before making such evidence available without adequate protection. Such protection has to be in terms of the workplace but, even more importantly, in terms of his personal life and those of his family. The perception of the adequacy and sufficiency of the protection, we argue, will be a deciding factor to the prospective whistle-blower when deciding (a) to blow the whistle or to remain silent and (b) to reveal his identity or to remain anonymous. Anonymity – guaranteed anonymity – is essential, especially when the stakes are high. The current ways of providing anonymity to whistle-blowers have built-in weaknesses and, as a result, seriously jeopardise the safety of whistle-blowers and their families. If a way can be found to guarantee anonymity, we suspect that more people will be prepared to volunteer information about wrongdoings, including, and especially, about corruption. The point has to be made that the current ways of facilitating whistle-blowing (for example, crime lines and all of the others) must remain in place. An additional way of blowing the whistle is required; a way to _guarantee_ anonymity. **TECHNICALLY SAFE ENVIRONMENT** The idea is to create an additional channel to the existing ones for whistle-blowing, but a channel which guarantees anonymity. It is, however, doubtful if the ideal of a 100% guarantee will ever be achieved, simply because of the very nature of technology. Technology advances at a rapid rate and newer technology is always available not only to legitimate users, but also to those wanting to exploit it for selfish or illegitimate purposes. This makes a „100% guarantee‟ a theoretical impossibility. Despite this, it is still, in our opinion, meaningful to try and get as close as possible to the goal so that, when the prospective whistle-blower does his risk assessment, an outcome in favour of blowing the whistle is still achieved. It is hoped that such a safe environment will encourage ----- 10804 Afr. J. Bus. Manage. prospective whistle-blowers to deliver their messages and evidence to law enforcement authorities. **Email** Telephone based technologies for providing the safe environment do not hold much promise, whether land-line based, public or wireless. One therefore has to look for a different kind of technology and email seems to be the next logical choice. Apart from technology constantly changing, it is quite a daunting task to design an environment where the originator of an email message cannot be traced. Every computer which a log on to a data communications network anywhere in the world gets a unique number assigned to it at the moment of logging on. This „number‟ is referred to as the Internet Protocol (IP) address. This IP address is - unknown to most email users - always transmitted along with the message, irrespective of where the message originates or where it terminates. Even when logging onto a website, the website is „aware‟ of the IP address of the computer logging on[2]. The IP address is typically (and deliberately) not under the control of the user of the computer with the result that a user will not be able to hide this unique identifier. An ordinary whistleblower, for instance, would certainly not have the technical skills to send a message by hiding the IP address assigned. In trying to design a technically safe environment, one cannot, therefore, simply advertise an email address as a means for whistle-blowers to inform organisations or law enforcement authorities of suspicious incidents. Just like hackers are traced despite their attempts to hide their identities, a non-technical user sending an email would be relatively easy to trace for someone with the necessary technical skills. **Encryption** A commonly used technique to protect the content of an email message (or even data attached as a file) is to encrypt the message and/or the data. This would make it impossible for a pursuer to read the content of the message even if he gets access to it, but the IP address could be identified and that, in most cases, would reveal thewhistle-blower‟s location and, perhaps, his name. 2 It is for this reason why one is sometimes surprised to see that a website has identified you as originating from, for instance, South Africa when logging on. [A typical example is when trying to log onto www.google.com (the Google](http://www.google.com/) service based in the US), one is automatically rerouted to the local website in South Africa (www.google.co.za). **Internet café** One way to overcome the problem posed by the IP address, is for the whistle-blower to send the email from an internet café and, of course, not revealing anything else about his identity in the message. This would make it more difficult for the pursuer to identify the originator despite tracing it back to the originating internet café. Internet café‟s, typically, do not keep records of the identities of their clients with the result that the whistleblower is reasonably safe from that perspective. However, many internet café‟s record the activities of the clients on video camera, so the whistle-blower could still be identified if the pursuer can trace the message or email back to a particular internet café and then getting access to the video recordings to look for suspects. This, however, will have to be done in a relatively short period of time as the video recordings are typically overwritten after a few days or weeks. There is another danger that the whistle-blower must avoid when making use of an internet café, namely, the data contained in any attachments to the message. When one creates a document in Microsoft Word, for example, it automatically creates a profile of the user and stores it with the rest of the document. Many users are not even aware that this is the case. Depending on how the computer was set up, the creator of the document may well be easily identified by name and surname without him being aware of it. (This is easily seen by simply clicking on „File‟, then „Properties‟, then „Summary‟ of any document created using MS WORD). It is easy enough to delete any such identifiers before attaching the document, but whistle-blowers must, firstly, be made aware of the danger and, secondly, remember to do so – consistently - before sending the attachment. Both aspects pose risks to the kind of technically sound environment one ideally would like to see. **TOWARDS A SAFE ENVIRONMENT** To get closer to the vision of a „100% guarantee‟, one should not have rely on the whistle-blower to, firstly, remove all of the identifying information on attachments, then encrypting the message and data using a robust encryption technique, and then using the internet café to send the email and, even then, run the risk of being traced back to a particular internet cafe. There are simply too many unacceptable risks in the scenario and something more robust and more reliable must be designed. **Onion routing** The IP address poses a challenging problem. However, ----- there is a way of getting rid of that in a relatively simple, but safe, way. This opportunity is provided through making use of a so-called „Onion Routing‟ facility (Feigenbaum et al., 2007). This facility was originally developed by the United States Naval Laboratory for the purpose of _„protecting government communications‟_ [(TOR Project 2011).](https://www.torproject.org/about/overview.html.en) This facility makes use of several websites situated around the globe at undisclosed locations. The user „drops‟ a message or data into a „drop box‟ typically provided by the advertised website for whistle-blowers. The message is then automatically routed in a random route from one location (“node”) to several others (called a „tunnel‟), at the same time, automatically stripping away originating IP addresses. After passing through a random number of nodes, the message is eventually delivered to the receiving party but the receiving party only sees the IP address of the last node sending the message. This IP address is simply one of the many nodes in the Onion Network and of no use to anyone, hence guaranteeing anonymity to the originating node and, the originator. Ironically, the Onion Network was developed by the US Navy to safeguard government communications, but this very same network was used by the Wikileaks[3] organisation to publish government cables and other documentation which caused such an embarrassment to the US Government (and others). Not even the US Military or Navy was able to trace the originator of the documents published on the Wikileaks website. The fact that Bradley Manning was eventually identified as the whistle-blower happened as a result of Manning revealing his identity to someone else whom he thought he could trust and who then disclosed it to the US government (Leigh and Harding, 2011). The claim is made that the Onion Routing facility provides „provable anonymity‟ (Feigenbaum et al., 2007) and this facility is available, free-of-charge, to anyone caring to use it. Of course, this claim only applies to the technical environment used to facilitate anonymity. From what can be gathered in the literature, using the Onion Network is relatively simple as the user is isolated from the technical complexities associated with the network. **Potential solution** Our solution involves a combination of the aforementioned, in the following process: 3 We must point out that we do not necessarily endorse or support any of the actions of the Wikileaks organisation. The Wikileaks organisation has its own objectives and we have our own. Yet, we do not want to pass judgement on what the Wikileaks organisation set out to do. The fact that we are proposing to use some of the same network technology (which technology does not belong to the Wikileaks organisation) must be seen as purely coincidental. Loggerenberg 10805 1. The whistle-blower goes with the evidence to an internet café 2. The whistle-blower drops the message and/or data into an electronic drop-box provided by the whistle-blowing organisation 3. Any data in the message or in the documents that could identify the sender is automatically deleted by the website when dropped in the box 4. The message and data get encrypted automatically by the website 5. The message and data is sent to the whistle-blowing organisation through the TOR network 6. The whistle-blowing websitedeliberately does not keep any logs of email received so that they would not be able to provide information whatsoever about the originator, even when forced to do so with a court order 7. The whistle-blowing organisation waits at least 14 days before making it available to law enforcement authorities so that video recordings at the originating internet café are likely to have been overwritten. **CONCLUSION** The scale of corruption in South Africa, and, for that matter, everywhere else in the World, is unacceptably large. Many African and other countries have poverty problems of immense magnitudes and cannot afford to waste billions of currencies to enrich a few corrupt individuals at the expense of the majority of the citizens. This money could go a long way to improve living conditions, healthcare to and education of the poor. In this respect whistle-blowing plays a very important role. Detection of corrupt deeds is in the hands of people of integrity to observe such corrupt deeds and report them to the relevant authorities. Such reporting carries huge risks, including loss of life, damage to property and/or dismissal or victimisation in the workplace. Whistle-blowers deserve to be protected. Such pro tection must be rooted in the legal framework but it needs to be complemented by mechanisms that allow whistleblowers to remain anonymous if they choose. Such anonymity must get as close to a 100% guarantee as one could possibly get. **RECOMMENDATIONS FOR FURTHER RESEARCH** This paper aimed at stimulating academic research on the topic. The topic can broadly be defined as the use of ICT in the fight against corruption. This paper only looked at one aspect of such broad topic, namely, anonymous whistle-blowing. It is recommended that further academic research gets initiated. For instance, a theoretical framework describing the use of ICT in the fight against corruption could be useful. An Actor Network Research (ANT) approach to ----- 10806 Afr. J. Bus. Manage. the topic is currently being investigated by the author to provide insight into the actors and the roles played by the actors in the corruption phenomenon. **ACKNOWLEDGEMENT** The initial research was partially supported by the German Development Cooperation (GIZ) awarded to Citizens against Corruption, a non-profit company. **REFERENCES** Calland R (2011). Blow the whistle at your peril. Mail and Guardian Online, Oct 17. Camerer L (2001). Protecting whistle-blowers in South Africa: The Protected Disclosures Act, no 26 of 2000. Occasional Paper no 47, Institute for Security Studies. Colvin G (2002). Wonder Women of Whistleblowing. Is it significant that the prominent heroes to emerge from the two great business scandals of recent years were women? Fortune Magazine, August. Ellison FA (1982). Anonymity and Whistleblowing. J. Bus. Ethics p. 1. Feigenbaum J, Johnson A, Syverson P (2007). A Model of Onion Routing with Provable Anonymity. Financial Cryptography and Data Security, 11th International Conference, FC. Leigh D, Harding L (2011). Wikileaks. Inside Julain Assange‟s War on secrecy.Guardian Books, London. Mail and Guardian Online (2007). Frere Hospital whistle-blower fired. 28 September. Cited on 23 October 2011 at mg.co.za/article/2007-0928-frere-hospital-whistleblower-fired. Martin P (2010). The Status of Whistle-Blowing in South Africa: Taking Stock. Open Democracy Advice Centre, June. Martin P (2011). Corruption.Towards A Comprehensive Societal Response. CASAC March. Miceli MP, Rehg M, Near JL, Ryan KC (1999). Can Laws Protect Whistle-Blowers?: Results of a Naturally Occurring Field Experiment. Work Occup. 26:129. Near JP, Miceli MP (1985). Organizational Dissidence: The Case of Whistle-Blowing. J. Bus. Ethics p. 4. Public Protector (2010). Address by Public Protector AdvThuliMadonsela during the Open Democracy Advice Center (ODAC) Conference on whistle-blowing held in Johannesburg, Wednesday, 17 November 2010. Cited at http://www.pprotect.org/media_gallery/2010/17112010_sp.asp. Republic of South Africa (2000). Protected Disclosures Act 2000, Act 26 of 2000. 7 August. Government Gazette p.422. Republic of South Africa (2004). Prevention and combating of corrupt activities 2003, Act 12 of 2004. 28 April. Government Gazette p. 466. Republic of South Africa (2008). Companies Act, Act 71 of 2008. 9 April 2009. Government Gazette p.526. Republic of South Africa (2011). Companies Amendment Act, Act 3 of 2011., 26 April 2011. Government Gazette p.550. [Syal R (2010). Tenfold rise in whistleblower cases taken to tribunal.](http://www.guardian.co.uk/profile/rajeev-syal) Campaigners fear workers deliberately undermined despite repeated [promises to protect them. The Guardian, Monday, 22 March 2010.](http://www.guardian.co.uk/theguardian) TOR Project (2011). Cited at https://wwwtorproject.org/about/overview.html.en. Treasury (2011). Medium Term Budget Policy Statement 2011: Speech by the Minister of Finance, Mr PravinGordhan. Cited at http://www.info.gov.za/speech/DynamicAction?pageid=461&sid=2268 5&tid=47198. -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.5897/AJBM12.992?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5897/AJBM12.992, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "HYBRID", "url": "https://academicjournals.org/journal/AJBM/article-full-text-pdf/1CB625D15993.pdf" }
2,012
[]
true
2012-11-07T00:00:00
[ { "paperId": "5695cc8555799d9e9d49e9cb02b025a2cee18e2f", "title": "WikiLeaks" }, { "paperId": "d9185cf4cf0b3acb3b119a81cc9ec77f2b211d44", "title": "Capital and capital maintenance rules under the Companies Act, Act 61 of 1973 and the Companies Act, Act 71 of 2008" }, { "paperId": "d313d294e142a0de16bdc9c79e4e39c10d6a0ca6", "title": "A Model of Onion Routing with Provable Anonymity" }, { "paperId": "43cf96b32f029285de45a7a269257981cbd644de", "title": "Can Laws Protect Whistle-Blowers?" }, { "paperId": "5d93ab4f1a88e1628217f16961eea529993570c2", "title": "Organizational dissidence: The case of whistle-blowing" }, { "paperId": "c224eaba41c12b28f6ca6e0ea3d38f5d049cbc56", "title": "Anonymity and whistleblowing" }, { "paperId": "d926e48fdb95ee9eb2c17f74670a01ce64d87da6", "title": "Protecting Whistle Blowers in South Africa: The Protected Disclosures Act, No 26 of 2000" }, { "paperId": null, "title": "Blow the whistle at your peril" }, { "paperId": null, "title": "Address by Public Protector AdvThuliMadonsela during the Open Democracy Advice Center (ODAC) Conference on whistle-blowing held in Johannesburg, Wednesday, 17 November 2010" }, { "paperId": null, "title": "The Status of Whistle-Blowing in South Africa: Taking Stock" }, { "paperId": null, "title": "Corruption.Towards A Comprehensive Societal Response" }, { "paperId": null, "title": "Frere Hospital whistle-blower fired" }, { "paperId": null, "title": "Prevention and combating of corrupt activities 2003, Act 12 of 2004" }, { "paperId": null, "title": "Tenfold rise in whistleblower cases taken to tribunal" }, { "paperId": null, "title": "Protected Disclosures Act 2000, Act 26 of 2000" }, { "paperId": null, "title": "Wonder Women of Whistleblowing" }, { "paperId": null, "title": "Wonder Women of Whistleblowing. Is it significant that the prominent heroes to emerge from the two great business scandals of recent years were women? Fortune Magazine" }, { "paperId": null, "title": "Address by Public Protector AdvThuliMadonsela during the Open Democracy Advice Center (ODAC) Conference on whistle-blowing held in Johannesburg, Wednesday" }, { "paperId": null, "title": "Medium Term Budget Policy Statement 2011: Speech by the Minister of Finance, Mr PravinGordhan" }, { "paperId": null, "title": "The Status of Whistle-Blowing in South Africa: Taking Stock. Open Democracy Advice Centre" }, { "paperId": null, "title": "Blow the whistle at your peril. Mail and Guardian Online" }, { "paperId": null, "title": "Tenfold rise in whistleblower cases taken to tribunal. Campaigners fear workers deliberately undermined despite repeated promises to protect them. The Guardian" }, { "paperId": null, "title": "Wikileaks. Inside Julain Assange\"s War on secrecy.Guardian Books, London. Mail and Guardian Online" } ]
9,414
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Business", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/015ed213c9398c10a4fea0eb09a29b7bd9816d81
[ "Computer Science" ]
0.864164
An Introduction to Decentralized Finance (DeFi)
015ed213c9398c10a4fea0eb09a29b7bd9816d81
Complex Systems Informatics and Modeling Quarterly
[ { "authorId": "81268997", "name": "Johannes Rude Jensen" }, { "authorId": "2047997256", "name": "Victor von Wachter" }, { "authorId": "48199469", "name": "Omri Ross" } ]
{ "alternate_issns": null, "alternate_names": [ "Complex Syst Informatics Model Q" ], "alternate_urls": null, "id": "a234316e-a138-4d28-9486-28de0ad963e5", "issn": "2255-9922", "name": "Complex Systems Informatics and Modeling Quarterly", "type": null, "url": "https://csimq-journals.rtu.lv/" }
. Decentralized financial applications (DeFi) are a new breed of consumer-facing financial applications composed as smart contracts, deployed on permissionless blockchain technologies. In this article, we situate the DeFi concept in the theoretical context of permissionless blockchain technology and provide a taxonomical overview of agents, incentives and risks. We examine the key market categories and use-cases for DeFi applications today and identify four key risk groups for potential stakeholders contemplating the advantages of decentralized financial applications. We contribute novel insights into a rapidly emerging field, with far-reaching implications for the financial services.
Complex Systems Informatics and Modeling Quarterly (CSIMQ) eISSN: 2255-9922 [Published online by RTU Press, https://csimq-journals.rtu.lv](https://csimq-journals.rtu.lv/) Article 150, Issue 26, March/April 2021, Pages 46–54 https://doi.org/10.7250/csimq.2021-26.03 # An Introduction to Decentralized Finance (DeFi) Johannes Rude Jensen[1,2*], Victor von Wachter[1], and Omri Ross[1,2 ] 1 Department of Computer Science, University of Copenhagen, Copenhagen, Denmark 2 eToroX Labs, Copenhagen, Denmark ``` johannesrudejensen@gmail.com, victor.vonwachter@di.ku.dk, omri@di.ku.dk ``` **Abstract. Decentralized financial applications (DeFi) are a new breed of** consumer-facing financial applications composed as smart contracts, deployed on permissionless blockchain technologies. In this article, we situate the DeFi concept in the theoretical context of permissionless blockchain technology and provide a taxonomical overview of agents, incentives and risks. We examine the key market categories and use-cases for DeFi applications today and identify four key risk groups for potential stakeholders contemplating the advantages of decentralized financial applications. We contribute novel insights into a rapidly emerging field, with far-reaching implications for the financial services. **Keywords: Blockchain, Decentralized Finance, DeFi, Smart Contracts.** ## 1 Introduction Decentralized financial applications, colloquially referred to as ‘DeFi’, are a new type of open financial applications deployed on publicly accessible, permissionless blockchains. A rapid surge in the popularity of these applications saw the total value of the assets locked in DeFi applications (TVL) grow from $675mn at the outset of 2020 to an excess of $40bn towards the end of first quarter in the following year[†]. While scholars within the information systems and management disciplines recognize the novelty and prospective impact of blockchain technologies, theoretical or empirical work on DeFi remains scarce [1]. In this short article, we provide a conceptual introduction to ‘DeFi’ situated in the theoretical context of permissionless blockchain technology. We introduce a taxonomy of agents, roles, incentives, and risks in DeFi applications and present four potential sources of complexity and risk. This article extends the previous publication on managing risk in DeFi[‡] and is structured as follows. Section 2 introduces the permissionless blockchain technology and decentralized finance. Section 3 presents DeFi application taxonomy. An overview of popular DeFi application - Corresponding author © 2021 Johannes Rude Jensen, Victor von Wachter, and Omri Ross. This is an open access article licensed under the Creative [Commons Attribution License (http://creativecommons.org/licenses/by/4.0).](http://creativecommons.org/licenses/by/4.0) Reference: J. R. Jensen, V. von Wachter, and O. Ross, “An Introduction to Decentralized Finance (DeFi),” Complex Systems Informatics and Modeling Quarterly, CSIMQ, no. 26, pp. 46–54, 2021. Available: https://doi.org/10.7250/csimq.2021-26.03 [Additional information. Author ORCID iD: J. R. Jensen – https://orcid.org/0000-0002-7835-6424, V. von Wachter –](https://orcid.org/0000-0002-7835-6424) [https://orcid.org/0000-0003-4275-3660, and O. Ross – https://orcid.org/0000-0002-0384-1644. PII S225599222100150X.](https://orcid.org/0000-0003-4275-3660) Received: 18 February 2021. Accepted: 15 April 2021. Available online: 30 April 2021. [† https://defipulse.com/](https://defipulse.com/) [‡ http://ceur-ws.org/Vol-2749/short3.pdf](http://ceur-ws.org/Vol-2749/short3.pdf) ----- categories is given in Section 4. The risks in decentralized finance are discussed in Section 5. Section 6 concludes the paper. ## 2 Permissionless Blockchain Technology and Decentralized Finance The implications and design principles for blockchain and distributed ledger technologies have generated a growing body of literature in the information systems (IS) genre [2]. Primarily informed by the commercial implications of smart contract technology, scholars have examined the implications for activities in the financial services such as the settlement and clearing of ‘tokenized’ assets [3] the execution and compilation of financial contracts [4]–[6], complexities in supply-chain logistics [7] and beyond. A blockchain is a type of distributed database architecture in which a decentralized network of stakeholders maintains a singleton state machine. Transactions in the database represent state transitions disseminated amongst network participants in ‘blocks’ of data. The correct order of the blocks containing the chronological overview of transactions in the database is maintained with the use of cryptographical primitives, by which all stakeholders can manually verify the succession of blocks. A network consensus protocol defines the rules for what constitutes a legitimate transaction in the distributed database. In most cases, consensus protocols are rigorous game-theoretical mechanisms in which network participants are economically incentivized to promote network security through rewards and penalties for benevolent or malicious behavior [8]. Scholars typically differentiate between ‘permissioned’ and ‘permissionless’ blockchains. Permissionless blockchains are open environments accessible by all, whereas permissioned blockchains are inaccessible for external parties not recognized by a system administrator [2]. Recent implementations of the technology introduces a virtual machine, the state of which is maintained by the nodes supporting the network. The virtual machine is a simple stack-based architecture, in which network participants can execute metered computations denominated in the native currency format. Because all ‘nodes’ running the blockchain ‘client’ software must replicate the computations required for a program to run, computational expenditures are priced on the open market. This design choice is intended to mitigate excessive use of resources leading to network congestion or abuse. Network participants pass instructions to the virtual machine in a higher-level programming language, the most recent generations of which is used to write programs, referred to as _smart_ _contracts. Because operations in the virtual machine are executed in a shared state, smart_ contracts are both transparent and _stateful, meaning that any application deployed as a smart_ contract executes deterministically. This ensures that once a smart contract is deployed, it will execute exactly as instructed. ## 3 DeFi Agent Taxonomy We denote the concept: ‘DeFi application’ as an arrangement of consumer-facing smart contracts, executing a predefined business logic within the transparent and deterministic computational environment afforded by a permissionless blockchain technology. Blockchain technology is the core infrastructure layer (see Figure 1) storing transactions securely and providing game-theoretic consensus through the issuance of a native asset. As a basic financial function, standardized smart contracts are utilized to create base assets in the asset layer. These assets are utilized as basis for more complex financial instruments in the application layer. In the application layer, DeFi applications are deployed as sophisticated smart contracts and thus execute a given business logic deterministically. Contemporary DeFi applications provide a range of financial services within trading, lending, derivatives, asset management and insurance services. Aggregators source services from multiple applications, largely to provide the best rates across the ecosystem. Finally, user friendly frontends combine the applications and build a service similar to today’s banking apps. In contrast to traditional banking services, in a 47 ----- blockchain-based technology stack, users interact directly with the application independent of any intermediary service provider. **Figure 1. DeFi applications on permisionless blockchain** The metered pricing of computational resources on permissionless blockchains means that DeFi applications are constrained by the computational resources they can use. Application designers seek to mitigate the need for the most expensive operations, such as storing big amounts of data or conducting sophisticated calculations, in the effort of reducing the level of complexity required to execute the service that their application provides. Because the resources required for interacting with a smart contract are paid by the user, DeFi application designers employ an innovative combination of algorithmic financial engineering and game theory to ensure that all stakeholders of their application are sufficiently compensated and incentivized. In Table 1, we introduce a taxonomy for the different types of agents and their roles in contemporary DeFi applications. We highlight the incentives for participation and key risks associated with each role. Owing to the original open-source ethos of blockchain technology, application designers are required to be transparent and build ‘open’ and accessible applications, in which users can take ownership and participate in decision-making processes, primarily concerning new features or changes to the applications. As a reaction to these demands, application designers often issue and distribute so-called governance tokens. Governance tokens are fungible units held by users, which allocates voting power in majority voting-schemes [9]. Much like traditional equities, governance tokens trade on secondary markets which introduces the opportunity for capital 48 ----- formation for early stakeholders and designers of successful applications. By distributing governance tokens, application designers seek to disseminate value to community members while retaining enough capital to scale development of the application by selling inventory over multiple years. **Table 1. Agent classification, incentives, and key risks** **Incentives for** **Agent:** **Role:** **Key risk:** **participation:** **Users** Utilizing the application Profits, credit, exposure and Market risk, technical risk governance token **Liquidity** Supply capital to the Protocol fees, governance Systemic economic risk, **Providers** application in order to ensure token technical risk, regulatory risk, liquidity for traders or opportunity costs of capital borrowers **Arbitrageurs** Return the application to an Arbitrage profits Market risk, network equilibrium state through congestion and transaction strategic purchasing and selling fees of assets **Application** Design, implement and Governance token Software bugs **Designers** maintain the application appreciation **(Team and** **Founders)** The generalized agent classification demonstrated in Table 1 is applicable to a wide area of DeFi applications providing peer-to-peer financial services on blockchain technology including, trading, lending, derivatives and asset management. In the following section, we dive into a number of recent use cases, examining the most recently popular categories of applications. ## 4 An Overview of Popular DeFi Application Categories The development principles presented above have been implemented in a number of live applications to date. In this section, we provide a brief overview of the main categories of DeFi applications. **4.1 Decentralized Exchanges and Automated Market Makers** Facilitating the decentralized exchange of assets requires an efficient solution for matching counterparties with the desire to sell or purchase a given asset for a certain price, a process known as price-discovery. Early implementations of decentralized exchanges on permissionless blockchain technologies successfully demonstrated the feasibility of executing decentralized exchange of assets on permissionless blockchain technology, by imitating the conventional central limit order book (CLOB) design. However, for reasons stipulated below, this proved infeasible and expensive at scale. First, in the unique cost structure of the blockchain based virtual machine format [10], traders engaging with an application, pay fees corresponding to the complexity of the computation and the amount of storage required for the operation they wish to compute. Because the virtual machine is replicated on all active nodes, storing even small amounts of data is exceedingly expensive. Combined with a complex matching logic required to maintain a liquid orderbook, computing fees rapidly exceeded users’ willingness to trade. Second, as ‘miners’ pick transactions for inclusion in the next block by the amount of computational fees attached to the transaction, it is possible to front-run state changes to the decentralized orderbook by attaching a large computational fee to a transaction including a trade, 49 |Agent:|Role:|Incentives for participation:|Key risk:| |---|---|---|---| |Users|Utilizing the application|Profits, credit, exposure and governance token|Market risk, technical risk| |Liquidity Providers|Supply capital to the application in order to ensure liquidity for traders or borrowers|Protocol fees, governance token|Systemic economic risk, technical risk, regulatory risk, opportunity costs of capital| |Arbitrageurs|Return the application to an equilibrium state through strategic purchasing and selling of assets|Arbitrage profits|Market risk, network congestion and transaction fees| |Application Designers (Team and Founders)|Design, implement and maintain the application|Governance token appreciation|Software bugs| ----- which pre-emptively exploits the next state change of the orderbook, thus profiting through arbitrage on a deterministic future state [11]. Subsequent iterations of decentralized exchanges addressed these issues by storing the state of the orderbook separately, using the blockchain only to compute the final settlement [12]. Nevertheless, problems with settlement frequency persisted, as these implementations introduced complex coordination problems between orderbook storage providers, presenting additional risk vectors to storage security. Motivated by the shortcomings of the established CLOB design a generation of blockchain specific ‘automated’ market makers (AMMs) presents a new approach to blockchain enabled market design. By pooling available liquidity in trading pairs or groups, AMMs eliminate the need for the presence of buyers and sellers at the same time, facilitating relatively seamless trade execution without compromising the deterministic integrity of the computational environment afforded by the blockchain. Trading liquidity is provided by ‘liquidity providers’ which lock crypto assets in the pursuit of trading fee returns. **Figure 2. AMM Price Discovery Function** While the primary context for the formal literature on blockchain based AMM has been provided by Angeris and Chitra _et al. [13]–[15] the field has attracted new work on adjacent_ topics such as liquidity provisioning [16]–[18] and token weighted voting systems [19]. **4.2 Peer-to-Peer Lending and Algorithmic Money Markets** The ‘money markets’ to borrow and lend capital with corresponding interest payments occupy an important role in the traditional financial service. Within DeFi, borrowing and lending applications are amongst the largest segments of financial applications with $7bn total value locked[§] at the end of 2020. In borrowing/lending protocols agents with excess capital can lend crypto assets (‘liquidity providers’) to a peer-to-peer protocol receiving continuous interest payments. Consequently, a borrower can borrow crypto assets and pays an interest rate. Given the pseudonymous nature of blockchain technology, it is not possible to borrow funds purely on credit. To borrow funds, the borrowing agent has to ‘overcollateralize’ a loan, by providing another crypto assets exceeding the dollar value of the loan to the smart contract. The smart contract then issues a loan relative to 70–90% of the value of the collateral assets. Should the [§ https://defipulse.com/](https://defipulse.com/) 50 ----- value of the collateral assets drop below the value of the outstanding loan, the smart contract automatically auctions away the collateral on a decentralized exchange at a profit. The interest rate is algorithmically set by the relative supply and demand for each specific crypto asset. Initially pioneered by the MakerDAO [**] application, several protocols are now accessible providing similar services with novel interests rate calculations or optional insurance properties, currently presiding over a $7bn crypto assets under management. **4.3 Derivatives** Blockchain-based financial contracts (derivatives) are one of the fastest growing market segments in DeFi. Here, application designers seek to make traditional financial derivatives such as _options, futures and other kinds of_ _synthetic contracts available to the broader DeFi_ ecosystem. A futures contract stipulates a sale of an asset at a specified price with an expiry date, an option contract stipulates the _right_ but not the obligation to sell or purchase an asset at a specific price. As in traditional finance both financial services can be used as insurance against market movements as well as speculation on prices. Recently, a new segment of ‘synthetic’ assets has entered the market in the form of tokens pegged to an external price, commonly tracking the price of commodities (e.g., gold) or stocks (e.g., Tesla). A user can create such synthetic asset by collateralized crypto assets in a smart contract similar to how a decentralized lending is computed. The synthetic asset tracks an external price feed (‘oracle’) which is provided to the blockchain. However, external price feeds are prone to technical issues and coordination problems leading to staleness in case of network congestions or fraudulent manipulation [20]. **4.4 Automated Asset Management** The traditional practice of ‘asset management’ in the financial services industry consists primarily of the practice of allocating financial assets such as to satisfy the long-term financial objectives of an institution or an individual. As the reader will have noted above, there are an increasing number of DeFi applications, all of which operate algorithmically without human intervention. This means that the DeFi markets operate around the clock and are impossible to manage The two main use cases for automated asset managers are ‘yield aggregators’ and traditional crypto asset indices. Utilizing the interoperability and automation of blockchain technology, ‘yield aggregators’ are smart contract protocols allocating crypto assets according to predefined rules, often with the goal of maximizing yield whilst controlling risk. Users typically allocate assets to a protocols, which automatically allocates assets across applications in order to optimize the aggregate returns, while rebalancing capital allocations on an ongoing basis. Indices, on the other hand, offer a broad exposure to crypto assets akin to the practice of ‘passive’ investing. These applications track a portfolio of crypto assets by automatically purchasing these assets and holding them within the smart contract. Equivalent to exchange traded funds (ETFs), stakeholders purchase ownership of the indices by buying a novel token, granting them the algorithmic rights over a fraction of the total assets held within the smart contract[††]. ## 5 Identifying and Managing Risk in Decentralized Finance In this section, we identify and evaluate four risk factors which are likely to introduce new complexities for stakeholders involved with DeFi applications. [** https://makerdao.com/](https://makerdao.com/) †† blockchain-in-asset-management.pdf (pwc.co.uk) 51 ----- **5.1 Software Integrity and Security** Owing to the deterministic nature of permissionless blockchain technology, applications deployed on as smart contracts are subject to excessive security risks, as any signed transaction remains permanent once included in a block. The irreversible or, ‘immutable’ nature of transactions in a blockchain network has led to significant loss of capital on multiple occasions, most frequently as a result of coding errors, sometimes relating to even the most sophisticated aspects virtual machine and programming language semantics [21]. DeFi applications rely on the integrity of smart contracts and the underlying blockchain. Risk is further enforced through uncertainties in future developments and the novelty of the technology. **5.2 Transaction Costs and Network Congestion** To mitigate abusive or excessive use of the computational resources available on the network, computational resources required to interact with smart contracts are metered. This creates a secondary market for transactions, in which users can outbid each other by attaching transaction fees in the effort of incentivizing miners to select their transaction for inclusion in the next block [11]. In times of network congestion, transactions can remain in a pending state, which ultimately results in market inefficiency and information delays. Furthermore, in these times, complex transactions can cost up to hundreds of dollars, making potential adjustments to the state costly.[‡‡] While intermediary service providers occasionally choose to subsidize protocol transaction fees[§§], application fees are in near all cases paid by the user interacting with the DeFi application. Because application designers seek to lower the aggregate transaction costs, protocol fees, slippage or impermanent loss through algorithmic financial modelling and incentive alignment, stakeholders must carefully observe the state of the blockchain network. If a period of network congestion coincides with a period of volatility, the application design may suddenly impose excessive fees or penalties on otherwise standard actions such as withdrawing or adding funds to a lending market [20]. **5.3 Participation in Decentralized Governance** Responding to implications of the historically concentrated distribution of native assets amongst a small minority of stakeholders, DeFi application designers increasingly rely on a gradual distribution of fungible governance-tokens in the attempt at adequately ‘decentralizing’ decisionmaking processes [9]. While the distribution of governance tokens remains fairly concentrated amongst a small group of colluding stakeholders, the gradual distribution of voting-power to liquidity providers and users will result in an increasingly long-tailed distribution of governance tokens. Broad distributions of governance tokens may result in adversarial implications of a given set of governance outcomes, for stakeholders who are not sufficiently involved in monitoring the governance process [19]. **5.4 Application Interoperability and Systemic Risks** A key value proposition for DeFi applications is the high level of interoperability between applications. As most applications are deployed on the Ethereum blockchain, users can transact seamlessly between different applications with settlement times rarely exceeding a few minutes. This facilitates rapid capital flows between old and new applications on the network. While interoperability is an attractive feature for any set of financial applications, tightly coupled and [‡‡ https://etherscan.io/gastracker](https://etherscan.io/gastracker) [§§ Coinbase.com](https://www.coinbase.com/) 52 ----- complex liquidity systems can generate an excessive degree of financial integration, resulting in systemic dependencies between applications [22]. This factor is exacerbated by the often complex and heterogeneous methodologies for the computation of exposure, debt, value, and collateral value that DeFi application designers have used to improve their product. An increasing degree of contagion between applications may introduce systemic risks, as a sudden failure or exploit in one application could ripple throughout the network, affecting stakeholders across the entire ecosystem of applications. The primary example of this dynamic can be demonstrated by the computation of ownership in so-called liquidity pools used by traders utilizing AMM smart contracts. When providing liquidity in the form of crypto assets to a decentralized exchange, liquidity providers receives ‘liquidity shares’ redeemable for a proportional share of the liquidity pool, together with the accumulated fees generated through trading. As liquidity shares are typically transferable and fungible IOU tokens representing fractional ownership of a liquidity pool, this has led to the emergence of secondary markets for liquidity shares. Providing liquidity in the form of IOU tokens, to these secondary market creates additional (3rd generation) liquidity shares, generating additional fees for the liquidity provider. As a consequence of the increasingly integrated market for liquidity shares, a rapid depreciation of the source asset for the liquidity shares may trigger a sequence of cascading liquidations, as the market struggles to price in any rapid changes in the price of the source asset [20], [22], [23]. ## 6 Conclusion: Is DeFi The Future of Finance? In this article, we have examined the potential implications, complexities and risks associated with the proliferation of consumer facing DeFi applications. While DeFi applications deployed on permissionless blockchains present a radical potential for transforming consumer facing financial services, the risks associated with engaging with these applications remain salient. Future stakeholder contemplating an engagement with these applications ought to consider and evaluate key risks prior to committing or allocating funds to DeFi applications. Scholars interested in DeFi applications may approach the theme from numerous angles, extending early research on the market design of DeFi applications [14] or issues related to governance tokens [9], [19] and beyond. ## Acknowledgments This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 801199. ## References [1] J. Kolb, M. Abdelbaky, R. H. Katz, and D. E. Culler, “Core Concepts, Challenges, and future Directions in Blockchain: A centralized Tutorial,” _ACM Comput. Surv., vol. 53, no. 1, pp. 1–39, 2020. Available:_ [https://doi.org/10.1145/3366370](https://doi.org/10.1145/3366370) [2] O. Labazova, “Towards a Framework for Evaluation of Blockchain Implementations,” in _Conference_ _Proceedings of ICIS (2019), 2019._ [3] O. Ross, J. Jensen, and T. Asheim, “Assets under Tokenization: Can Blockchain Technology Improve Post Trade Processing?” in _Conference_ _Proceedings_ _of_ _ICIS_ _(2019),_ 2019. Available: [https://doi.org/10.2139/ssrn.3488344](https://doi.org/10.2139/ssrn.3488344) [4] J. R. Jensen and O. Ross, “Settlement with Distributed Ledger Technology,” in _Conference Proceedings of_ _ICIS (2020), 2020._ [5] B. Egelund-Müller, M. Elsman, F. Henglein, and O. Ross, “Automated Execution of Financial Contracts on Blockchains,” _Bus._ _Inf._ _Syst._ _Eng.,_ vol. 59, no. 6, pp. 457–467, 2017. Available: [https://doi.org/10.1007/s12599-017-0507-z](https://doi.org/10.1007/s12599-017-0507-z) 53 ----- [6] O. Ross and J. R. Jensen, “Compact Multiparty Verification of Simple Computations,” in _CEUR Workshop_ _[Proceedings, 2018. Available: https://doi.org/10.2139/ssrn.3745627](https://doi.org/10.2139/ssrn.3745627)_ [7] B. Düdder and O. Ross, “Timber Tracking: reducing Complexity of Due Diligence by using Blockchain [Technology,” SSRN, 2017. Available: https://doi.org/10.2139/ssrn.3015219](https://doi.org/10.2139/ssrn.3015219) [8] A. Antonopoulos and G. Wood, Mastering Ethereum: Building Smart Contracts and DApps. Sebastopol, CA: O’Reilly Media, 2018. [9] V. von Wachter, J. R. Jensen, and O. Ross, “How Decentralized is the Governance of Blockchain-based Finance? Empirical Evidence from four Governance Token Distributions,” 2020. Available: [https://arxiv.org/abs/2102.10096](https://arxiv.org/abs/2102.10096) [10] G. Wood, “Ethereum: A secure decentralized generalized Transaction Ledger EIP 150,” in _Ethereum Project_ _Yellow Paper, 2014, pp. 1–32._ [11] P. Daian _et al., “Flash Boys 2.0: Frontrunning, Transaction Reordering, and Consensus Instability in_ [Decentralized Exchanges,” 2019. Available: https://arxiv.org/abs/1904.05234](https://arxiv.org/abs/1904.05234) [12] W. Warren and A. Bandeali, “0x : An open Protocol for decentralized Exchange on the Ethereum Blockchain.” [Available: https://github.com/0xProject](https://github.com/0xProject) [13] G. Angeris, A. Evans, and T. Chitra, “When does the Tail wag the Dog? Curvature and Market Making,” 2020. [Available: https://arxiv.org/abs/2012.08040](https://arxiv.org/abs/2012.08040) [14] G. Angeris, H.-T. Kao, R. Chiang, C. Noyes, and T. Chitra, “An Analysis of Uniswap Markets,” _[Cryptoeconomic Systems, vol. 1, no. 1, 2019. Available: https://doi.org/10.21428/58320208.c9738e64](https://doi.org/10.21428/58320208.c9738e64)_ [15] T. Chitra, “Competitive Equilibria between Staking and on-chain Lending,” Cryptoeconomic Systems, vol. 1, [no. 1, 2021. Available: https://doi.org/10.21428/58320208.9ce1cd26](https://doi.org/10.21428/58320208.9ce1cd26) [16] J. Aoyagi, “Liquidity Provision by Automated Market Makers,” _SSRN,_ 2020. Available: [https://doi.org/10.2139/ssrn.3674178](https://doi.org/10.2139/ssrn.3674178) [17] M. Tassy and D. White, “Growth Rate of A Liquidity Provider’s Wealth in XY = c Automated Market [Makers,” 2020. Available: https://math.dartmouth.edu/~mtassy/articles/AMM_returns.pdf](https://math.dartmouth.edu/~mtassy/articles/AMM_returns.pdf) [18] M. Bartoletti, J. H. Chiang, and A. Lluch-Lafuente, “SoK: Lending Pools in Decentralized Finance,” 2020. [Available: https://arxiv.org/abs/2012.13230](https://arxiv.org/abs/2012.13230) [19] G. Tsoukalas and B. H. Falk, “Token-Weighted Crowdsourcing,” Manag. Sci., vol. 66, no. 9, pp. 3843–3859, [2020. Available: https://doi.org/10.1287/mnsc.2019.3515](https://doi.org/10.1287/mnsc.2019.3515) [20] D. Perez, S. M. Werner, J. Xu, and B. Livshits, “Liquidations: DeFi on a Knife-edge,” 2020. Available: [https://arxiv.org/abs/2009.13235](https://arxiv.org/abs/2009.13235) [21] L. Luu, D.-H. Chu, H. Olickel, P. Saxena, and A. Hobor, “Making Smart Contracts Smarter,” in _Proceedings_ _of the 2016 ACM SIGSAC Conference on Computer and Communications Security (CCS’16), pp.254–269,_ [2016. Available: https://doi.org/10.1145/2976749.2978309](https://doi.org/10.1145/2976749.2978309) [22] L. Gudgeon, D. Perez, D. Harz, B. Livshits, and A. Gervais, “The Decentralized Financial Crisis,” _Crypto_ _Valley_ _Conference_ _on_ _Blockchain_ _Technology_ _(CVCBT),_ pp. 1–15, 2020. Available: [https://doi.org/10.1109/CVCBT50464.2020.00005](https://doi.org/10.1109/CVCBT50464.2020.00005) [23] V. von Wachter, J. R. Jensen, and O. Ross, “Measuring Asset Composability as a Proxy for Ecosystem Integration,” in DeFi _[Workshop Proceedings of FC'21, 2021. Available: https://arxiv.org/abs/2102.04227](https://arxiv.org/abs/2102.04227)_ 54 -----
{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.7250/csimq.2021-26.03?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.7250/csimq.2021-26.03, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://csimq-journals.rtu.lv/article/download/csimq.2021-26.03/2570" }
2,021
[ "JournalArticle", "Review" ]
true
2021-04-30T00:00:00
[ { "paperId": "ccc89f268ae2c69384e4bfaeca463aa38d56d7f1", "title": "How Decentralized is the Governance of Blockchain-based Finance: Empirical Evidence from four Governance Token Distributions" }, { "paperId": "d102780639e4d91e6549b23f621da00e961447bf", "title": "SoK: Lending Pools in Decentralized Finance" }, { "paperId": "3ef3e42c148961a92be129d2ed8f630c0e7fcf7b", "title": "When does the tail wag the dog? Curvature and market making." }, { "paperId": "fa07b0d228edd081d548da718474ec8fcc9b9d13", "title": "Settlement with Distributed Ledger Technology" }, { "paperId": "8cfea92862743cca95fbe48b9f4acf89ce16511d", "title": "Liquidations: DeFi on a Knife-Edge" }, { "paperId": "44f28c4f4e302d9584aa32f1c20c367b2e05c6cd", "title": "Token-Weighted Crowdsourcing" }, { "paperId": "d798034b2e65d8afe2d2ead977731aceb55c9012", "title": "Liquidity Provision by Automated Market Makers" }, { "paperId": "c876d76a22b23ae28fe6d9c5589f857768d543fc", "title": "The Decentralized Financial Crisis" }, { "paperId": "be28d28b546e5f0904cd1a56485511b706363587", "title": "Core Concepts, Challenges, and Future Directions in Blockchain" }, { "paperId": "5f9b892a11df3c3dcb20b7c0ded3a0bf2724f32c", "title": "Competitive equilibria between staking and on-chain lending" }, { "paperId": "5fcbdeabead3daed43816bb714992df1e866b628", "title": "An Analysis of Uniswap Markets" }, { "paperId": "393ab84a86631d5fda128c3aac0bf5476da07791", "title": "Flash Boys 2.0: Frontrunning, Transaction Reordering, and Consensus Instability in Decentralized Exchanges" }, { "paperId": "f63636d729899ce2f070adf9a7af4b1e0892dd9a", "title": "Timber Tracking: Reducing Complexity of Due Diligence by Using Blockchain Technology" }, { "paperId": "1f155d4c196954aa8951ed7c769ddec64631c132", "title": "Automated Execution of Financial Contracts on Blockchains" }, { "paperId": "7968129a609364598baefbc35249400959406252", "title": "Making Smart Contracts Smarter" }, { "paperId": "66b285002b6864019a644217c697b494b0215aa1", "title": "Measuring Asset Composability as a Proxy for Ecosystem Integration" }, { "paperId": "ded2d1d492efc89fee394f50b8716f1c7d6a2370", "title": "Assets under Tokenization: Can Blockchain Technology Improve Post-Trade Processing?" }, { "paperId": "0e5bfb12a8ebf718b2ee41d25957f7c66651b3ab", "title": "Towards a Framework for Evaluation of Blockchain Implementations" }, { "paperId": "7e8e1aeb7b973d3ed99409fe0ce0605d9293eabe", "title": "Compact Multiparty Verification of Simple Computations" }, { "paperId": null, "title": "Mastering Ethereum: Building Smart Contracts and DApps" }, { "paperId": null, "title": "“Ethereum: A secure decentralized generalized Transaction Ledger EIP 150,”" }, { "paperId": null, "title": "“0x : An open Protocol for decentralized Exchange on the Ethereum Blockchain.”" } ]
7,479
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01614001dfc81c1d899402168314ff0c3ef7ce1e
[ "Computer Science" ]
0.842226
Private and efficient set intersection protocol for RFID-based food adequacy check
01614001dfc81c1d899402168314ff0c3ef7ce1e
IEEE Wireless Communications and Networking Conference
[ { "authorId": "3388391", "name": "Zakaria Gheid" }, { "authorId": "1756893", "name": "Y. Challal" }, { "authorId": "2118537403", "name": "Lin Chen" } ]
{ "alternate_issns": null, "alternate_names": [ "IEEE Wirel Commun Netw Conf", "WCNC", "Wireless Communications and Networking Conference", "Wirel Commun Netw Conf" ], "alternate_urls": null, "id": "27235614-bd3e-4d6b-be38-5ede18f4e209", "issn": null, "name": "IEEE Wireless Communications and Networking Conference", "type": "conference", "url": "http://www.ieee-wcnc.org/" }
null
# Private and Efficient Set Intersection Protocol For RFID-Based Food Adequacy Check ## Zakaria Gheid[†], Yacine Challal[∗†], Lin Chen[‡] _∗Centre de Recherche sur l’Information Scientifique et Technique, Algiers, Algeria_ _†Ecole nationale sup´erieure d’informatique, Laboratoire des M´ethodes de Conception des Syst`emes, Algiers, Algeria_ _‡Lab. Recherche Informatique (LRI-CNRS UMR 8623), Univ. Paris-Sud, 91405 Orsay, France_ Email: z gheid@esi.dz, y challal@esi.dz, chen@lri.fr **_Abstract—Radio Frequency Identification (RFID) is a technol-_** **ogy for automatic object identification that has been implemented** **in several real-life applications. In this work, we expand a novel** **relevant application of RFID tags for grocery stores, which aims** **to check the adequacy of food items with respect to the shoppers’** **personal preferences. Unlike similar works, we focus on shoppers’** **privacy and running time efficiency. For this aim, we propose** **a novel private set intersection (PSI) protocol to be used in** **matching the shoppers’ personal preferences with the set of each** **item’s adequate profiles that are held by the back-end server of** **the store. We provide a standard security proof against curious** **stores and malicious customers. For efficiency concern, we build** **our protocol without cryptographic operations, and we achieve** **a linear asymptotic complexity of O(v + c) for communications** **and store-side computations, where v and c are the numbers** **of profiles in the store’s back-end server and the shopper’s list** **of preferences respectively. Moreover, experimental results and** **comparisons with state-of-the art solutions reveal the scalability** **of our novel PSI protocol for big market stores.** **_Index Terms—Radio Frequency Identification (RFID), Profile_** **Matching, Private Set Intersection.** I. INTRODUCTION Radio Frequency Identification (RFID) is a wireless technology that uses radio waves to identify and track objects. RFID systems consist of tags attached to the objects to be identified and readers that communicate with the tags to collect information. Owing to its advantages over barcode systems, RFID market is gaining increasing values that are expected to exceed US$18 billion by 2026 [1]. This rapid proliferation has allowed a wide range of endless applications, where commerce comes in first. For instance, Large-scale supermarkets like Walmart are attaching RFID tags to their goods to increase business revenue lost by theft or inaccurate accounting of goods [2]. Amazon Inc. has launched Amazon Go, a hightech retail store currently in a private beta testing in Seattle, and filled a patent [3] in which they use RFID to detect when a shopper takes an item from the shelf. Then, the system adds up the item and charges the shopper’s Amazon account without requiring going through a traditional check-out line. This, should improve shopping experience for customers with easier item returns. Following this technology adoption, we introduce a novel relevant RFID-based application that we coin as FAC: Food Adequacy Check, which provides customers with information on food items whether they match their preferences or not. For instance, customers suffering from diabetes or needing a gluten-free diet, should have a detailed insight about items before buying them. Traditionally, a shopper can easily make such a check relying on product labels; nevertheless, matching a complex profile that involves several information such as age, weight, several diseases, and follows a special program as Low Carb diet [4], is a tedious task. This is highly true, if the shopper wants to match several preferences including his/her one and those of his/her family. Accordingly, we propose (Π-FAC), a novel private and efficient set intersection protocol that we use to match customers personal preferences with items adequate profiles. We design (Π-FAC) as a multi-party computation (MPC) protocol that is implemented on customers’ smartphones and the supermarket back-end server. We use passive RFID tags with no computational capability, which is the cheapest type of tag, to allow the deployment of our application with affordable cost. We address the privacy concern of the shopper preferences against curious server, besides the privacy of the database of item profiles held by the back-end server against malicious shoppers as this may be a paid service. We provide a simulation-based security proof under the standard real/ideal paradigm [5]. For the efficiency concern, we build our protocol upon efficient matrix algebra without cryptographic operations to ensure its scalability for large supermarkets. We achieve a linear asymptotic complexity of O(v + c) in communications and server side computations, where v and c are, respectively, number of profiles within the server database and the customer list of preferences. Finally, we make experimental evaluations to confirm the efficiency of our (Π-FAC) protocol compared to the hash-based private set intersection solution used in practice. The rest of this paper is organised as follows. In Section II, we review recent literature works in Private Set Intersection field and we discuss them. In Section III, we introduce a novel RFID-based application that we name Food Adequacy Check (FAC). Then, in Section IV, we detail our novel private set intersection protocol used within FAC application for the private profile matching purpose. Next, we provide a standard security analysis of our protocol, in Section V, using the Real/Ideal paradigm. After that, we devote Section VI to evaluate the efficiency of our protocol compared to the hash ----- based solution used in practice. Finally, we conclude this work by summarizing our contribution. II. RELATED WORK In this section, we provide a literature survey on the private set intersection (PSI) functionality that we use to implement the Food Adequacy Check application. We focus on PSI protocols that work in the standard (plain) model, where security is only based on complexity assumptions. Assume a client (C) and a server (V ) having private sets of profiles X and Y of sizes c and v respectively. Two main approaches were used to solve PSI(X,Y), namely Oblivious Polynomial Evaluation (OPE) [8] and Oblivious Pseudo-Random Functions (OPRF) evaluation [9]. _1) OPE based-PSI: In this approach, C defines a poly-_ nomial P (.) such that P (x) = 0 for each x _X, and_ _∈_ sends to V homomorphic encryptions of the coefficients of _P_ (.). Then, V computes the encryption of (r.P (y) + y) for each y _Y, using homomorphic properties of the encryption_ _∈_ system and a fresh random r. Finally, C decrypts the received cyphertexts and gets either elements of the intersection (if plaintexts match an element of X) or random values. In this approach, we find works of Freedman et al. [10], Kissner and Song [11], Dachman-Soled et al. [14] and Hazay [15]. They targeted semi-honest and malicious settings, where the most efficient construction [15] incurs O(v+c) communications and _O(c + v log log c) computations, under the strong Decisional_ Diffie-Hellman assumption (strong-DDH). _2) OPRF-based PSI: In which V defines a random key (k)_ for a pseudo random function (PRF) fk(.) and computes the set fky = {fk(y) : y ∈ _Y }. Then, V and C executes an OPRF_ protocol where V inputs fk(.) and C inputs the set X and gets the set fkx = {fk(x) : x ∈ _X}. At the end, V sends the set_ _fky to C that evaluates fkx_ _fky. This approach was used_ _∩_ by Hazay and Lindell [16], Jarecki and Liu [17], Hazay and Nissim [19] and Hazay [15] to propose PSI protocols secure in the semi-honest and malicious settings. The most efficient protocol that does not require non standard assumptions [15] costs O(v+c) computations under the strong-DDH assumption and O((v + c) log (v + c)) under the DDH assumption. Contrary to existing PSI protocols that rely on cryptographic schemes, we propose a novel PSI protocol based on efficient matrix algebra and secure under the mixed model of adversaries. Our protocol incurs O(v + c) communications and server computations while maintaining fairness. III. FAC: A NOVEL RFID-BASED FOOD ADEQUACY CHECK SYSTEM In this section, we present a novel RFID-application that aims to check the adequacy of foods to shoppers’ personal preferences. _A. FAC Overview_ To illustrate the FAC application, we consider a supermarket that tagged its items with RFID tags, and provides shopping carts with embedded RFID reader devices for its clients. Each Application Communication RFID System Figure 1. Architecture & Infrastructure requirements for FAC application client will be provided a mobile application that he/she sets upSmart **(Π-SI)** Backend Phones Server on his/her smartphone to enter information about personal foodPrivate preferences that he/she wants to match. When a client entersPersonal Profiles Intersection Set All Item Profiles the supermarket he/she uses the provided mobile application to connect to the supermarket wireless gateway. Then, each time the shopper takes an item from the shelf and passes it through the embedded RFID reader device, the latter reads the item tag and passes its information to the mobile application of the shopper. Once the application handles a novel arriving tag information, it sends it to the back-end server with a profile matching request. The shopper’s smartphone and the back-end server start running a profile matching process using our novel private set intersection protocol (Π-FAC). This application ends-up by showing the shopper which profiles match the taken item among the set of profiles that he/she entered. _B. FAC Architecture_ To implement the FAC application, we propose the following architecture model that is based on three layers, namely RFID system, communication, and application (Figure 1). _• RFID system. This is the basic layer. It consists of_ an RFID system with standard components. It involves passive tags put on each item of the supermarket, reader devices that can be either fixed on the shelves or embedded on shopping carts, and an RFID middleware. This latter component is not required by our FAC application, it aims to recover each tag read by a device to enable the integration of other application using the same RFID infrastructure. _• Communication layer. It involves a wireless commu-_ nication gateway that covers the supermarket surface. It aims to interconnect the upper-layer components and ensures the communication with the RFID reader devices and the middleware. _• Application layer. It involves the FAC application set up_ on the clients’ smartphones and the back-end server of the supermarket. The mobile application allows the user to input its personal preferences and connect to the backend to run the private profile matching process using our built-in private set intersection protocol (Π-FAC). Smart Phones Personal Profiles ----- IV. A NOVEL PRIVATE AND EFFICIENT SET INTERSECTION PROTOCOL In this section, we present our novel private set intersection protocol as well as its design model. _A. Our Methodology_ In this work, we use a matrix-based approach in which we represent the private sets of profiles as row matrices (each matrix corresponds to a private set of profiles and each row within it corresponds to a profile in the set). Then, each party obfuscates its matrix by performing a multiplication with a random matrix chosen independently from the input domain. Next, each party sends its resultant matrix to the other party to be multiplied by the other random matrix. Since, matrix product is not commutative, which is required for the correctness of the scheme, the two parties will interchange the side of the matrix product (left multiplication and right multiplication). At the end, the two resulting matrices will be checked for rows equality as each row corresponds to an original element in the set. In what follows, we give a detailed implementation of the Π-FAC protocol. _B. Protocol Design_ To introduce our novel private set intersection protocol (Π-FAC), we consider a client denoted C and a back-end server denoted V having respectively X = {x1, ..., xc} and _Y = {y1, ..., yv} sets of profiles and want to securely get the_ intersection between their sets. Assume for 1 _i_ _c and_ _≤_ _≤_ 1 ≤ _j ≤_ _v: xi and yj ∈_ R[n]. Let M(m, n) denote the set of all m-by-n matrices and denote the matrix multiplication _⊗_ operator. Let M1 and M2 denote random invertible matrices used by C and V respectively to obfuscate their sets, where **M1 ∈** M(c, c) and M2 ∈ M(n, n). Let MX and ∪i>1MYi denote the private sets X and Y respectively, represented as row matrices, where MX ∈ M(c, n) and MYi ∈ M(c, n). We present the detail of Π-FAC protocol in Algorithm 1. V. SECURITY ANALYSIS In this section, we give a security proof of our protocol using the Real/Ideal security model [5]. _A. Security Model_ Let Π denote a multi-party protocol executed by m participants (P1,...,Pm) in order to evaluate a function f . Let B denote the class of adversary that may corrupt participants in Π. Let R and D denote respectively the real and the ideal executions of Π on the set of inputs w and the set of security parameters sec. **Notation 1. Let viewE[Π][(][w][,][sec][)][i][ denote the set of messages]** _received by the party Pi∈{1,...,m} along with its inputs and_ _outputs during the execution E of Π on the set of inputs w_ _and security parameters sec._ **Notation 2. Let out[Π]E[(][w][,][sec][)][i][ denote the output of the party]** _Pi∈{1,...,m} by the execution E of the protocol Π on the set of_ _inputs w and security parameters sec. Let out[Π]E[(][v][,][sec][) denote]_ **Algorithm 1: Π-FAC, a Private and Efficient Set Intersec-** tion Protocol **Input : X = {x1, ..., xc} C’s set of personal profiles** _Y = {y1, ..., yv} V ’s set of all item profiles_ **Output: (For C only) Ψ(X, Y ): the private set** intersection between X and Y **Require: (c, n) ∈** N[2]: 0 < c < n **Step 1 by C** 1: Generates a random invertible M1 ∈ M(c, c) 2: Creates MX ∈ M(c, n) with X’s elements as rows 3: Computes M1X = M1 ⊗ **MX** 4: Sends M1X to V **Step 2 by V** 5: Generates a random invertible M2 ∈ M(n, n) 6: Computes M1X2 = M1X ⊗ **M2** 7: for (i = 1; i < (v/c) + 1; i + +) do 8: Creates MYi ∈ M(c, n) with Y ’s elements as rows 9: Computes MY2i = MYi **M2** _⊗_ 10: end for 11: Sends M1X2 and ∪i>1MY2i to C **Step 3 by C** 12: Computes M1Y2i = M1 ⊗ **MY2i, for each** received MY2i 13: For each (m, n, i) if M1X2[m,*] = M1Y2i[n,*] and M1X2[m,*] ̸∈ Ψ(X, Y ) then puts M1X2[m,*] in Ψ(X, Y ) _the global output of all collaborating parties from the same_ _execution of Π, where_ _out[Π]E[(][w, sec][) =][ ∪]i[m]=1[out]E[Π][(][w, sec][)][i]_ During a real execution (R) we consider the presence of an adversary denoted A that behaves according to the class B while corrupting a set of participants Pi(1≤i≤m). At the end of R, uncorrupted parties output whatever was specified in Π and the corrupted Pi outputs any random functions of their _viewR[Π][(][w][,][sec][)][i][.]_ During an ideal execution (D) we consider the presence of a trusted incorruptible party denoted T, which receives the set of inputs w from all participants in order to evaluate the function f in the presence of an adversary denoted S. We assume S corrupts the same Pi as the correspondent adversary A of real execution, and behaves according to the same class B before sending inputs to T . By the end of D, uncorrupted participants output what was received from T and the corrupted Pi output any random functions of their _viewD[Π]_ [(][w][,][sec][)][i][.] **Definition 1. Let Π and f be as above. We consider Π a** _secure multi-party protocol if for any real adversary A having_ _a class B and attacks the protocol Π during its execution on_ _the set of inputs w and the set of security parameters sec,_ _there exists an adversary S in the ideal execution having the_ _same class B and that can emulate any effect achieved by_ ----- _d_ _A. Let_ _denote the distribution equality. We formalize the_ _≡_ _definition of a secure multi-party protocol Π as follows_ _{out[Π]R[(][w, sec][)][}]_ _≡{d_ _outΠD[(][w, sec][)][}]_ (1) _B. Security Proof_ In what follows, we give security simulations of Π-FAC protocol using Real/Ideal paradigm. The allowed behavioural class of adversary is the mixed one, where the client (C) having a set of inputs X is actively corrupted and the server (V ) having the set of inputs Y is passively corrupted. Let A, S and T denote respectively a real adversary, an ideal adversary and a trusted third party, where A and S have the same class. Let Π denote the Π-FAC protocol (Algorithm 1), _sec denote security parameters that will be presented below_ (Theorem 1), w denote the set of inputs {MX, ∪i>1MYi}, which are the matrix representation of the sets X and Y respectively, and Ψ(X, Y ) denote the private set intersection between X and Y . **Theorem 1. Given a set of security conditions (sec) defined as** _sec = {(n, c) ∈_ N[2] : 0 < c < n}. Under these conditions, the _protocol Π-FAC defined in Algorithm 1 is a secure multi-party_ _protocol against an active corruption of C._ _Proof: Assume C is actively corrupted by A. Then, it_ can only inject fake inputs (MA) since aborting the protocol untimely will have no meaning. Assume C sends a fake MA. In this case, S can emulate A by just handling the fake MA and sends it to T, which performs the required computation and sends back Ψ(X, Y ) to C. Thereby, completing the simulation. At the end, the views of C in Ideal and Real executions are as follows _viewD[Π]_ [(][w, sec][)][C] [=][ {][MX][,][ Ψ(][X, Y][ )][}] (2) _viewR[Π][(][w, sec][)][C]_ [=][ {][MX][,][ M1X2][,][ ∪][i>][1][MY2][i][,][ Ψ(][X, Y][ )][}][ (3)] Otherwise, M1X2 = M1X ⊗ **M2, where M1X ∈** M(c, n) and **M2 ∈** M(n, n). According to security parameters (sec), we have c < n. This preserves well the privacy of M2. Thereby, **M1X2 that contains (c × n) equations opposite to (n × n)** unknowns for C, will not involve meaningful information for it and can be reduced from its view. Likewise, ∪i>1MY2i = _∪i>1MYi ⊗_ **M2, where MYi ∈** M(c, n) and M2 ∈ M(n, n). Then, ∪i>1MY2i will contain α(c × n) equations opposite to (α(c _n)+(n_ _n)) unknowns for C, where 0 < α < (v/c)+1._ _×_ _×_ This, does not involve meaningful information for it and can be so, reduced from its view. After these reductions, the view of C in real execution will be defined as follows _viewR[Π][(][w, sec][)][C]_ [=][ {][MX][,][ Ψ(][X, Y][ )][}] (4) Thus, relying on (2) and (4) we get _{out[Π]R[(][w, sec][)][C][}]_ _≡{d_ _outΠD[(][w, sec][)][C][}]_ (5) On the other hand, the uncorrupted V can not be affected by the corruption of C since V does not require any output in real execution. Thus, T will simply not send it any output during ideal execution. This, means that _{out[Π]R[(][w, sec][)][V]_ _[}]_ _≡{d_ _outΠD[(][w, sec][)][V]_ _[}]_ (6) Through (5) and (6), we proved by simulation that all effects achieved by a real active adversary corrupting C can also be achieved in an ideal execution. Then, Π-FAC is a secure multiparty protocol against active corruption of C (Definition 1). **Theorem 2. Given a set of security conditions (sec) defined as** _sec = {(n, c) ∈_ N[2] : 0 < c < n}. Under these conditions, the _protocol Π-FAC defined in Algorithm 1 is a secure multi-party_ _protocol against a passive corruption of V ._ _Proof: Assume V is passively corrupted. In this case,_ _V should follow the specification of the protocol Π-FAC,_ yet, it is allowed to analyse all data gathered during the execution. Then, S will just handle V ’s input and sends it to T, which performs the required computation and sends Ψ(X, Y ) to C while sending nothing to V . Thereby, completing the simulation. At the end, the views of V in Ideal and Real executions are as follows _viewD[Π]_ [(][w, sec][)][V] [=][ {∪][i>][1][MY][i][}] (7) _viewR[Π][(][w, sec][)][V]_ [=][ {∪][i>][1][MY][i][,][ M1X][}] (8) Moreover, M1X = M1 ⊗ **MX, where, M1 ∈** M(c, c) and **MX ∈** M(c, n). Then, since we defined (0 < c) as security parameter (sec), we get (c _×_ _n)<((c_ _×_ _n)+(c_ _×_ _c)). Thus, M1X_ that contains (c _n) opposite to ((c_ _n)+(c_ _c)) unknowns_ _×_ _×_ _×_ for V will not involve meaningful information for it and can be, so, reduced from its view. After reduction, we obtain _viewR[Π][(][w, sec][)][V]_ [=][ {∪][i>][1][MY][i][}] (9) Thus, relying on (7) and (9) we get _{out[Π]R[(][w, sec][)][V]_ _[}]_ _≡{d_ _outΠD[(][w, sec][)][V]_ _[}]_ (10) On the other hand, the uncorrupted C outputs what was received from T in ideal execution, which is Ψ(X, Y ) according to the simulation given above and outputs what was specified in the protocol Π-FAC in real execution, which is Ψ(X, Y ) (Algorithm 1, Output section) . Then, we have _{out[Π]R[(][w, sec][)][C][}]_ _≡{d_ _outΠD[(][w, sec][)][C][}]_ (11) Through (10) and (11) we proved by simulation that all effects achieved by a real passive adversary corrupting V can also be achieved in an ideal execution. Then, Π-FAC is a secure multiparty protocol against passive corruption of V (Definition 1). **Corollary 1. Given a set of security conditions (sec) defined** _as sec = {(n, k) ∈_ N[2] : 0 < k < n}. Under these conditions, _the protocol Π-FAC defined in Algorithm 1 is a secure multi-_ _party protocol in the mixed model of adversary, where C is_ _actively corrupted and V is passively corrupted._ ----- _Proof: Corollary 1 relies heavily on the Theorem 1 and_ Theorem 2 proved above, while considering separately the case when the client (C) is corrupted and the case when the server (V ) is corrupted. We assume that if both parties are corrupted we are not required to provide security guarantees. VI. PERFORMANCE ANALYSIS In this section, we simulate the performance of our protocol (Π-FAC) using a queueing theory model. We make experiments on the same data sets with a custom simulator built in Python and an Intel i5-2557M CPU running at 1.70 GHz and having a 4 GB of RAM. _1) Experimental scenario: We consider a large-scale su-_ permarket that wants to provide their clients with a Food Adequacy Check (FAC) service using our private set intersection protocol (Π-FAC). In order to assess the scalability of Π-FAC, we consider the back-end server (V ) receiving (N ) FAC requests from different clients (C) at rate (λ) request(s) per minute according to a Poisson process: N _P_ (λ). _∼_ Assume the requests processing times (ti) have an exponential distribution with rate (µ) requests per minute ti ∼ _exp(µ)._ Without loss of generality, we consider fixing the number of client personal profiles to 10 profiles per client, where each profile involves 20 attributes (∈ R[20]). Besides, we consider _V having 1000 profiles of 20 attributes per item. This, results_ in an average range of 50 possible values for each attribute, which is highly sufficient in real applications. For comparison purpose, we model the same above scenario, while V uses the hash-based private set intersection protocol used in practice (Section IV-A) instead of our Π-FAC protocol. For this, we use an efficient commutative hash function _Hk(x) = x[k]_ _mod p, where k is a 32-bit security parameter_ and p is a 32-bit random prime. Assume V having a FIFO service discipline with unlimited access, and operating all day long. Let M/M/1 denote this system using Kendall’s notation [23]. We evaluate this system by varying the the number of clients requesting for FAC service (λ) in the range 10, 20, 50, 100, 200 clients _{_ _}_ (requests) per minute. Let mult, add, exp and mod denote respectively one multiplication, one addition, one exponentiation and one modulo operations. Let v and c denote the number of profiles of V and C respectively, where each profile involves n attributes. To measure µ parameter, we evaluate the computational costs required by V when using Π-FAC and the hashing scheme by the following equations. _Cost[(Π]V_ _[−][F AC][)]_ = n[2](v + c) mult + n(n − 1)(v + c) add _Cost[(]V[hash][)]_ = n(v + c) exp + n(v + c) mod _2) Results & discussion: We have made experimental eval-_ uations by simulating two back-end servers of a supermarket handling FAC requests, while one was running Π-FAC protocol and the other was running a hash-based protocol. We used the model described in Table 1 and we evaluated the system performance for each protocol according to the number of requests (λ) through the following metrics: the usability rate (U ) of the back-end server, its response time (R), the average number of clients (N ) in the system, and the mean length of waiting queue (Q). Let ρ denote the intensity traffic rate. We assess the previous metrics according to the following equations and we present the results in Table 1 and Figure 4. _ρ_ _ρ =_ _[λ]_ _U = ρ_ _N =_ _Q = N_ _ρ_ _R =_ _[N]_ _−_ _µ_ 1 _ρ_ _λ_ _−_ For low arrival rates (λ < 100) results show that the server running Π-FAC was undergoing a slow intensity traffic (ρ < 0.1), which results in a very low probability of server overload. This claim may be confirmed by looking the low server utilization rate (U < 10%), besides, the zero queue length (Q = 0). On the hand, the server running the hash protocol was less efficient with a usability rate of U > 70% for 50 clients per minute. This high usability tends to overload the server if more clients arrive (λ > 50), which may be confirmed by looking the increase in the number of clients waiting in the queue (Q > 0). Regarding response time, ΠFAC provided a high efficient and stable response (R = 2.x ms) compared to the hash protocol, which was less efficient and had a significant delay each time there was an increase in the arrival rate. For high arrival rates (λ > 100), the server running Π-FAC remained efficient with a usability rate of U < 50% for 200 clients per minute while providing a high efficient response time (R < 4 ms). In contrast, the server running the hash protocol was undergoing a very high intensity traffic (ρ > 1), which leads the system to a non-steady state and results in overloading the server with a utilization rate of U > 100%. This, tended to an infinite queue length (Q = ) and an _∞_ infinite response time (R = ). _∞_ Experimental results revealed the efficiency of our Π-FAC protocol compared to the hash-based solution used in practice. This efficiency raises from the fact that our protocol involves efficient arithmetic operations (addition and multiplication) and does not require any expensive computations (modulo, exponentiation), which are involved in cryptographic methods. These performance results show the adequacy of our protocol to be used by large scale supermarkets. VII. CONCLUSION In this paper, we expanded a novel RFID-based application that aims to check whether food items matches the preferences of the shoppers, according to their personal profiles. For this, we proposed Π-FAC, a novel set intersection protocol that targets privacy and efficiency concerns while matching shoppers’ preferences with item profiles that are held by the back-end server of the store. Through security analysis conducted with the standard Real/Ideal paradigm, we showed the privacy guarantees provided by Π-FAC against curious stores and malicious clients. Besides, across empirical performance analysis, we demonstrated the high efficiency of our protocol compared to the hash-based private set intersection used in practice. Evaluation results revealed the adequacy of Π-FAC to provide a private and efficient Food Adequacy Check service for large-scale stores. ----- Table I EVALUATION OF A BACK-EN SERVER STORE USING M/M/1 MODEL **Fixed** **Used** **Parameters** **Protocol** Π-FAC _v= 1000_ _c= 10_ _n= 20_ **Hash** **Client** **Running** **Processing** **Intensity** **Usability** **Number of** **Queue** **Response** **Requests** **Time** **Rate** **Traffic** **Rate** **Client** **Length** **Time** **(λ)/min** **s** **(µ)** **(ρ)** **(U) %** **(N) × 10[2]** **(Q)** **(R) ms** 10 1.29 0.02 2 2 0 2 20 2.54 0.04 4 4 0 2 50 6.30 470 0.10 10 11 0 2.2 100 12.79 0.21 21 26 0 2.6 200 25.53 0.42 42 72 0 3.6 10 8.59 0.14 14 16 0 16 20 17.28 0.28 28 39 0 19.5 50 42.80 70 0.71 71 245 2 49 100 86.53 1.43 143 _∞_ _∞_ _∞_ 200 171.66 2.86 286 _∞_ _∞_ _∞_ (a) Usability Rate (U) (b) Queue length (Q) (c) Response time (R) Figure 2. Evaluation of a back-en server store using M/M/1 model REFERENCES [1] M. R. Das, RFID Forecasts, Players and Opportunities 2016-2026. IDTechEx. [2] Walmart stores, inc. wal-mart continues rfid expansion. [Online]. Available: http://corporate.walmart.com/ [visited 04/30/2017] [3] e. a. Puerini, Gianna Lise, “Transitioning items from a materials handling facility,” US Patent, 01 08, 2015. [4] M. L. Dansinger, J. A. Gleason, J. L. Griffith, H. P. Selker, and E. J. Schaefer, “Comparison of the atkins, ornish, weight watchers, and zone diets for weight loss and heart disease risk reduction: a randomized trial,” Jama, vol. 293, no. 1, pp. 43–53, 2005. [5] R. Canetti, “Security and composition of multiparty cryptographic protocols,” Journal of CRYPTOLOGY, vol. 13, no. 1, pp. 143–202, 2000. [6] The best of global digital marketing. case study: Hellmann’s recipe cart. [Online]. Available: http://www.best-marketing.eu/casestudy-hellmanns-recipe-cart/ [visited 04/30/2017] [7] The wall street journal. whole foods aims for younger shoppers with new stores. [Online]. Available: http://www.wsj.com/articles/wholefoods-to-launch-new-outlets-1431041549 [visited 04/30/2017] [8] M. Naor and B. Pinkas, “Oblivious transfer and polynomial evaluation,” in Proceedings of the Thirty-first Annual ACM Symposium on Theory _of Computing, ser. STOC ’99._ New York, NY, USA: ACM, 1999, pp. 245–254. [9] M. J. Freedman, Y. Ishai, B. Pinkas, and O. Reingold, Keyword Search _and Oblivious Pseudorandom Functions._ Berlin, Heidelberg: Springer Berlin Heidelberg, 2005, pp. 303–324. [10] M. J. Freedman, K. Nissim, and B. Pinkas, Efficient Private Matching _and Set Intersection._ Berlin, Heidelberg: Springer Berlin Heidelberg, 2004, pp. 1–19. [11] L. Kissner and D. Song, “Privacy-preserving set operations,” in Pro_ceedings of the 25th Annual International Conference on Advances in_ _Cryptology, ser. CRYPTO’05._ Berlin, Heidelberg: Springer-Verlag, 2005, pp. 241–257. [12] E. De Cristofaro, J. Kim, and G. Tsudik, Linear-Complexity Private Set _Intersection Protocols Secure in Malicious Model._ Berlin, Heidelberg: Springer Berlin Heidelberg, 2010, pp. 213–231. [13] C. Hazay and M. Venkitasubramaniam, Scalable Multi-party Private _Set-Intersection._ Berlin, Heidelberg: Springer Berlin Heidelberg, 2017, pp. 175–203. [14] D. Dachman-Soled, T. Malkin, M. Raykova, and M. Yung, Efficient _Robust Private Set Intersection._ Berlin, Heidelberg: Springer Berlin Heidelberg, 2009, pp. 125–142. [15] C. Hazay, Oblivious Polynomial Evaluation and Secure Set-Intersection _from Algebraic PRFs._ Berlin, Heidelberg: Springer Berlin Heidelberg, 2015, pp. 90–120. [16] C. Hazay and Y. Lindell, “Efficient protocols for set intersection and pattern matching with security against malicious and covert adversaries,” _Journal of Cryptology, vol. 23, no. 3, pp. 422–456, 2010._ [17] S. Jarecki and X. Liu, Efficient Oblivious Pseudorandom Function with _Applications to Adaptive OT and Secure Computation of Set Intersection._ Berlin, Heidelberg: Springer Berlin Heidelberg, 2009, pp. 577–594. [18] R. Canetti and M. Fischlin, Universally Composable Commitments. Berlin, Heidelberg: Springer Berlin Heidelberg, 2001, pp. 19–40. [19] C. Hazay and K. Nissim, “Efficient set operations in the presence of malicious adversaries,” Journal of Cryptology, vol. 25, no. 3, pp. 383– 433, 2012. [20] Y. Lindell and B. Pinkas, “Secure multiparty computation for privacypreserving data mining,” Journal of Privacy and Confidentiality, vol. 1, no. 1, p. 5, 2009. [21] J. Vaidya and C. Clifton, “Secure set intersection cardinality with application to association rule mining,” J. Comput. Secur., vol. 13, no. 4, pp. 593–622, Jul. 2005. [22] B. Pinkas, T. Schneider, and M. Zohner, “Scalable private set intersection based on ot extension,” 2016. [23] E. Gelenbe, G. Pujolle, and J. Nelson, Introduction to queueing net_works._ John Wiley & Sons, Inc., 1987, vol. 2. -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/WCNC.2018.8377207?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/WCNC.2018.8377207, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "other-oa", "status": "GREEN", "url": "https://hal.archives-ouvertes.fr/hal-01786000/file/Article%20WCNC%202018.pdf" }
2,018
[ "JournalArticle", "Conference" ]
true
2018-04-15T00:00:00
[]
9,454
en
[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/016143d542f7249ee1a1a2aafc45f11fc61a0bc0
[]
0.860276
Blockchain Theory and Applications - Welcome and Committees
016143d542f7249ee1a1a2aafc45f11fc61a0bc0
2022 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops)
[]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }
null
# BRAIN 2022: Third Workshop on Blockchain Theory and Applications - Welcome and Committees Welcome Message BRAIN 2022, the 3rd Workshop on Blockchain theoRy and ApplicatIoNs, aims to provide a venue for researchers from both academy and industry to present and discuss important topics on blockchain technology. In particular, blockchain provides an innovative solution to address challenges in pervasive environments, e.g. distributed computing, smart devices, and device-to-device coordination, in terms of decentralization, data privacy, and network security, while pervasive environments offer elasticity and scalability characteristics to improve the efficiency of blockchain operations. The workshop's goal is to present results on both theoretical and more applicative open challenges, as well as to provide a venue to showcase the current state of existing proposals. The members of the workshop's international program committee refereed the submitted papers by both the quality, far sight, and fit with the workshop topics. This year's workshop program includes the 7 selected high quality papers. The submissions have been divided in two sessions, one covering blockchain applications and the other blockchain analysis and theory. The program is integrated with two blockchain related keynotes, one about layer 2 payment channels and the other about blockchain distributed consensus protocols evolution. We would like to thank the members of the program committee for providing detailed and rigorous reviews, allowing us to select the most engaging papers for this edition of the workshop. We would also like to thank the PerCom organizers, in particular Frank Dürr and Antinisca Di Marco, the PerCom workshop co-chairs, for constantly supporting the workshop and assisting us along the way. Finally, we thank all attendees, keynote speakers, and authors of both accepted and rejected papers for their contributions and participation. Damiano Di Francesco Maesa (University of Pisa, Italy), Laura Ricci (University of Pisa, Italy), Nishanth Sastry (University of Surrey, United Kingdom) ## BRAIN 2022 Organizing Committee ### Co-Chairs Damiano Di Francesco Maesa (University of Pisa & University of Cambridge, United Kingdom (Great Britain)) Laura Ricci (University of Pisa, Italy) Nishanth Sastry (University of Surrey, United Kingdom (Great Britain)) ----- ## Technical Program Committee Andrea Bracciali Stirling University United Kingdom (Great Britain) Andrea Bracciali University of Stirling United Kingdom (Great Britain) Antorweep Chakravorty University of Stavanger Norway Mauro Conti University of Padua Italy Andrea De Salve National Research Council (CNR) Italy Damiano Di Francesco MaesaUniversity of Pisa United Kingdom (Great Britain) Zeki Erkin Delft University of Technology The Netherlands Tooba Faisal Kings College London United Kingdom (Great Britain) Barbara Guidi University of Pisa Italy Mohammad Hammoudeh Manchester Metropolitan UniversityUnited Kingdom (Great Britain) Paolo Mori IIT, CNR Italy Remo Pareschi University of Molise Italy Radu Prodan University of Klagenfurt Austria Laura Ricci University of Pisa Italy Nishanth Sastry University of Surrey United Kingdom (Great Britain) Claudio Schifanella Università di Torino Italy Luca Spalazzi Università Politecnica delle Marche Italy Frank Tietze University of Cambridge United Kingdom (Great Britain) Sara Tucci-Piergiovanni CEA LIST France -----
{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/percomworkshops53856.2022.9767389?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/percomworkshops53856.2022.9767389, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "BRONZE", "url": "https://ieeexplore.ieee.org/ielx7/9766442/9767206/09767389.pdf" }
2,022
[ "Conference" ]
true
2022-03-21T00:00:00
[]
804
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/016320200c600c00a351586477cfcf8f3df423d1
[ "Computer Science", "Medicine" ]
0.852325
Decentralized Real-Time Anomaly Detection in Cyber-Physical Production Systems under Industry Constraints
016320200c600c00a351586477cfcf8f3df423d1
Italian National Conference on Sensors
[ { "authorId": "2198956885", "name": "Christian Goetz" }, { "authorId": "1808059", "name": "B. Humm" } ]
{ "alternate_issns": null, "alternate_names": [ "SENSORS", "IEEE Sens", "Ital National Conf Sens", "IEEE Sensors", "Sensors" ], "alternate_urls": [ "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-142001", "http://www.mdpi.com/journal/sensors", "https://www.mdpi.com/journal/sensors" ], "id": "3dbf084c-ef47-4b74-9919-047b40704538", "issn": "1424-8220", "name": "Italian National Conference on Sensors", "type": "conference", "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-142001" }
Anomaly detection is essential for realizing modern and secure cyber-physical production systems. By detecting anomalies, there is the possibility to recognize, react early, and in the best case, fix the anomaly to prevent the rise or the carryover of a failure throughout the entire manufacture. While current centralized methods demonstrate good detection abilities, they do not consider the limitations of industrial setups. To address all these constraints, in this study, we introduce an unsupervised, decentralized, and real-time process anomaly detection concept for cyber-physical production systems. We employ several 1D convolutional autoencoders in a sliding window approach to achieve adequate prediction performance and fulfill real-time requirements. To increase the flexibility and meet communication interface and processing constraints in typical cyber-physical production systems, we decentralize the execution of the anomaly detection into each separate cyber-physical system. The installation is fully automated, and no expert knowledge is needed to tackle data-driven limitations. The concept is evaluated in a real industrial cyber-physical production system. The test result confirms that the presented concept can be successfully applied to detect anomalies in all separate processes of each cyber-physical system. Therefore, the concept is promising for decentralized anomaly detection in cyber-physical production systems.
# sensors _Article_ ## Decentralized Real-Time Anomaly Detection in Cyber-Physical Production Systems under Industry Constraints **Christian Goetz *** **and Bernhard Humm** Hochschule Darmstadt— Department of Computer Science, University of Applied Sciences, 64295 Darmstadt, Germany *** Correspondence: christian.goetz@yaskawa.eu** **Abstract: Anomaly detection is essential for realizing modern and secure cyber-physical production** systems. By detecting anomalies, there is the possibility to recognize, react early, and in the best case, fix the anomaly to prevent the rise or the carryover of a failure throughout the entire manufacture. While current centralized methods demonstrate good detection abilities, they do not consider the limitations of industrial setups. To address all these constraints, in this study, we introduce an unsupervised, decentralized, and real-time process anomaly detection concept for cyber-physical production systems. We employ several 1D convolutional autoencoders in a sliding window approach to achieve adequate prediction performance and fulfill real-time requirements. To increase the flexibility and meet communication interface and processing constraints in typical cyber-physical production systems, we decentralize the execution of the anomaly detection into each separate cyber-physical system. The installation is fully automated, and no expert knowledge is needed to tackle data-driven limitations. The concept is evaluated in a real industrial cyber-physical production system. The test result confirms that the presented concept can be successfully applied to detect anomalies in all separate processes of each cyber-physical system. Therefore, the concept is promising for decentralized anomaly detection in cyber-physical production systems. **Keywords: anomaly detection; cyber-physical production systems; cyber-physical systems; deep** learning; unsupervised learning **Citation: Goetz, C.; Humm, B.** Decentralized Real-Time Anomaly Detection in Cyber-Physical Production Systems under Industry Constraints. Sensors 2023, 23, 4207. [https://doi.org/10.3390/s23094207](https://doi.org/10.3390/s23094207) Academic Editors: Jun Wu, Zhaojun Steven Li, Yi Qin and Carman K.M. Lee Received: 13 February 2023 Revised: 17 April 2023 Accepted: 21 April 2023 Published: 23 April 2023 **Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons [Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/) [creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/) 4.0/). **1. Introduction** Due to the rising complexity of modern processes in manufacturing, the application of cyber-physical systems (CPS) is increasing. A CPS can be described as a combination of an embedded system with sensors and actuators. The system interacts with these to monitor and control physical processes (Figure 1) [1]. Typically, the embedded system requires a communication interface to exchange data with other systems or a cloud. Many of these CPSs are networked to realize complex physical processes in the real world [2]. CPSs combine powerful information technology to monitor and control engineered systems [3]. Modern production systems, which include CPSs, are defined as cyber-physical production systems (CPPS) [4]. These systems are based on two main functionalities, advanced connectivity to ensure real-time data acquisition from the physical world and feedback from cyberspace. CPPSs break with the structure of the typical automation hierarchy to enable intelligent data management, real-time analytics, and enhanced computational capabilities. The control and field levels still exist to ensure the highest performance for critical loops, while the higher levels are more dynamic and decentralized [5]. ----- _Sensors 2023, 23, 4207_ 2 of 19 **Figure 1. Abstract concept of a CPS [1].** Such a CPPS can be seen in Figure 2. The rotary table dispenser system consists of different CPSs working together to realize several physical processes, e.g., transportation or pick-and-place operations. The overall process involves picking small items from a rotating table and putting them into several containers which are moving around the machine on conveyor belts. After the container is filled and reaches the end position, it gets picked up by the production robot and emptied back onto the rotating table. Thereafter, the container is put in a central location from which the sliding robot places it into the container tray. When the container tray is full, both sliders move to the left side of the system, and the sliding robot sets the container back on the conveyor belt. The described system acts as a simulation of a similar real industrial process and is used as a demonstration unit in Yaskawa. In total, there are nine CPSs, each combining a mechanical and an embedded system. Seven CPSs are based on servomotors and servo controllers. Two CPSs consists of an industrial robot with a robot controller. A central control unit collects data from the different CPSs and regulates the main production process. Additional computational units provide the opportunity to integrate higher functions, e.g., resource planning, production analysis, and process control handling. **Figure 2. Rotary table dispenser system.** ----- _Sensors 2023, 23, 4207_ 3 of 19 In such a connected structure, even a single failure in one CPS can influence the entire production, resulting in a faulty product, a breakdown of the complete process, or a carryover of the failure through the whole system. Therefore, it is necessary to ensure an error-free operation to realize a secure and modern CPPS [6]. Anomalies can be taken as essential failure indicators, such as a rising vibration at a bearing of the rotating table or an unexpected torque increase on the motor of the conveyor belt. Anomaly detection (AD) in CPPSs refers to the identification of behavior that is not shown under the regular operations of the system. Consequently, by detecting anomalies, there is the possibility to recognize, react early, and in the best case, fix the anomaly to prevent the rise or the carryover of the failure throughout the entire manufacture [7]. Techniques for anomaly detection in CPPS can be distinguished into model-based [8] and data-driven approaches [9]. Model-based methods work based on precise and engineered models of the complete system. Creating such models over the complex structure of CPPS is time-consuming while simultaneously requiring deep expert knowledge. Datadriven approaches establish models only on collected data. Through the high amount of monitored and available data in CPPSs, these approaches are more appropriate for such systems, while additionally, no proper expert knowledge is needed [10]. Recent developments in machine learning and deep learning for anomaly detection have improved the detection performance on complex data sets [11]. By following the scheme to deliver all data from the control and field-level device to one CPS at a higher level to process, analyze, and detect anomalies, current centralized data-driven AD approaches in industrial CPPSs demonstrate better detection abilities than decentralized ones. While this is a significant advantage, it first requires a fully connected and high-performance unit for monitoring all integrated CPSs. Subsequently, adding such a unit increases costs and installation time. Additionally, a centralized concept creates a communication delay between the different stations to exchange the enormous amount of data produced in a CPPS. This can result in a delayed response after detecting an anomaly. Furthermore, it slows down the execution, evaluation, and detection of the anomalies in the individual CPS [12]. The structure of a CPPS is highly dynamic. Often single components, such as motors and sensors, are exchanged, replaced, or modified due to predictive or preventive maintenance. In a centralized approach, this results in a complete recreation of the AD due to the changed characteristics. In contrast, a decentralized concept addresses these drawbacks by establishing the AD directly in each CPS. While this allows monitoring of the whole system by combining each separate AD, the need for a high-performance unit can be reduced, and the execution and response time can be increased. Furthermore, changes in a single CPS result in only the retraining of the associated AD. By establishing adequate prediction performances in each single CPS, comparable performance to a centralized AD can be reached. The contribution of this paper is a novel unsupervised, decentralized, and real-time process anomaly detection concept for CPPS under industry constraints. We focus on industrial production processes and common constraints in CPPSs, including real-time requirements, asynchronous signals, prediction quality, configurable design, data-driven limitations, processing limitations, and communication interface constraints. We employ several 1D convolutional autoencoders (1D-ConvAE) in a sliding window approach to achieve adequate prediction performance and fulfill real-time requirements. Current methods do not consider the limitations and constraints of industrial setups and mainly follow a centralized approach. By executing the installation process on an external, removable device, we increase the flexibility of our concept while considering processing limitations. To meet communication interface and processing constraints in typical CPPSs, we decentralize the execution of the AD into each separate CPS. The installation is fully automated to tackle data-driven limitations. Thereby, no expert knowledge about explicit anomalies is needed. Adjustments to the data collection routine were made to optimize the external sampling procedure and improve the installation process. ----- _Sensors 2023, 23, 4207_ 4 of 19 This paper is structured as follows. Section 2 summarizes related work about anomaly detection for industrial CPS and CPPS. The problem statement is specified in Section 3. Section 4 presents a concept for fast and decentralized unsupervised anomaly detection in CPPS. Information about a prototypical implementation is provided in Section 5. In Section 6, the evaluation of the approach is presented based on an industrial setup. Finally, a conclusion and an outlook for future work are given in Section 7. **2. Related Work** Surveys on anomaly detection techniques can be found in [13–15]. More industrialrelated AD methods are described in [16,17]. Overall, these techniques can be differentiated into model-based and data-driven approaches. Model-based techniques detect anomalies by manually creating precise models about the underlying system. This requires a deep prior knowledge of the individual CPPS. While data-driven approaches are also based on models, those models are generated automatically from data and not manually by domain experts. Furthermore, data-driven approaches can be split into supervised and unsupervised techniques. Anomalous data in CPPS is associated with the undefined behavior of the system. Creating such anomalous data can be hazardous for the CPPS itself, while defining all possible anomalies in advance is nearly impossible. Based on the points mentioned above, we focus on unsupervised data-based methods. Common approaches in unsupervised data-based AD are one-class classification methods, such as deep one-class networks [18] and one-class support vector machine [19,20]. While multi-class classification techniques typically require labeled datasets, these approaches focus on the normal samples by learning a discriminative hyperplane surrounding them. Other frequently used techniques are unsupervised clustering methods such as Gaussian Mixture Models [21], k-nearest neighbor methods [22], or random isolation forests [23]. These models can identify anomalies by building a detailed representation of the normal data. While the resulting models are generally lightweight and computationally fast, they lack performance when processing high-dimensional data. Deep learning methods for AD have recently improved the state of the art in detection performance on complex and large datasets [24]. The standard techniques in this field are generative adversarial networks (GAN). GANs consist of a generator combined with a discriminator as the base structure. By teaching the discriminator to distinguish between real and fake samples while the generator tries to generate new data based on the input, GANs can detect anomalies even in large multivariate data streams. Concepts of GANs differ mainly in the models used as the base structure, such as long-short-term-memory (LSTM) recurrent neural networks (RNN) [25], two-dimensional convolutional autoencoder [26], and one-dimensional convolutional autoencoder [27]. While the described approaches achieve good outcomes, they result in highly complex and large models that cannot be applied to a CPS with limited computational resources, which is a common industry constraint. Reconstruction-based methods in AD combine techniques that rely on the assumption that a model trained only on normal data cannot reconstruct abnormal or unseen data. Typical techniques of these fields are PCA methods [28] or sparse representations [29]. A widely used approach for reconstruction-based anomaly detection in CPS is using autoencoders [30,31] or variants thereof [32,33]. By learning the latent features of the input data, autoencoders can reconstruct their input as output. While these models can be applied to analyze the spatial characteristics of the input data, they miss considering the temporal dependencies, which are necessary indicators for anomalies in the time series data of industrial CPS. While convolutional neural networks (CNN) were initially developed for solving image classification tasks, they can also be successfully applied for AD in time series data of a CPS through the ability to extract temporal dependencies [34]. Several industrial applications of CNNs in CPS, such as fault detection in motors [35], AD in wheelset bearings [22], and rolling bearings [36], can be found. Additionally, ref. [37] pointed out that CNNs have ----- _Sensors 2023, 23, 4207_ 5 of 19 lower parameters than other network structures while performing comparably or better, resulting in reduced complexity, needed storage capacity, and computing power [38]. Convolutional autoencoders (ConvAE) combine the ability to detect temporal anomalies with the help of convolutions and spatial anomalies by the autoencoder structure while being also resource-efficient. This results in ideal models for AD in multivariate time series data [39–41]. Using a 2D variational ConvAE, the authors in [42] detect anomalies from unseen abnormal patterns in industrial robots. In [43], a ConvAE based on channel-wise reconstruction in combination with a local outlier factor is used to detect anomalies in automobile sensors. Several approaches for decentralized AD can be found [44–46]. In [12], different decentralized AD techniques are analyzed and compared in complexity and performance. A decentralized approach for real-time AD in transportation networks is introduced in [47]. The authors of [48] presented spatial anomaly detection in sensor networks using neighborhood information. While these are promising approaches for decentralized AD, no work considers all the different industrial constraints simultaneously, which is important for integration into a CPPS. Different automated frameworks for anomaly detection can be found. In [49], a framework for automatic time series anomaly detection is introduced. The study focuses on large-scale time series data in a centralized AD approach, which cannot be applied to a CPPS with limited resources. The authors of [50] introduce an unsupervised framework for anomaly detection in CPS. Furthermore, ref. [51] presents a high-performance unsupervised anomaly detection for CPS networks. Both approaches are developed for CPS, but mainly focus on adversarial attacks and not on the process of the CPS and, respectively, of the CPPS. In our previous work [52], we introduced an unsupervised anomaly detection concept for CPSs under industry constraints while focusing on repetitive tasks with a fixed duration for a single CPS. In this contribution, we improved the concept for CPPS with multiple CPSs, while still considering all industrial constraints. We adapted the technology to a sliding window approach to simultaneously handle processes with variable durations and meet real-time instead of near-time requirements. In summary, there are several approaches for centralized and decentralized datadriven unsupervised anomaly detection. Only a few are evaluated in real CPSs, and even fewer are applied to real production data of a CPPS. Overall, no work considers all the different industrial limitations of a CPPS while following a decentralized and fast approach to realize anomaly detection in industrial production data. In this work, we propose a concept that addresses all the requirements that must be considered to realize a usable decentralized, real-time anomaly detection in CPPS under industrial constraints. Our contribution in this paper is summarized as follows. We employ several 1D-ConvAEs for unsupervised anomaly detection in a CPPS to monitor the different processes. We introduce a novel concept to decentralize the different models in each single CPS of the CPPS by splitting the installation and execution of the anomaly detection to meet industrial requirements. While the concept is fully automated, no expert knowledge about explicit known anomalies is needed to meet the defined requirements. **3. Problem Statement** This article aims at a decentralized concept for real-time unsupervised anomaly detection for production processes under industrial constraints. The problem statement can be described by the different industrial requirements that must be considered to implement such a concept. Several conditions are adapted and extended from [52]. 1. **Anomaly detection: An anomaly detection for a CPPS, such as an industrial pro-** duction system, shall be performed. The CPPS consists of multiple CPSs producing multivariate time series data over variable process lengths, for example, the sliding robot from the CPPS in Figure 2, combining a robot with several axes and a robotic controller to move containers on a conveyor belt. ----- _Sensors 2023, 23, 4207_ 6 of 19 2. **Real-time: To cover all different kinds of anomalies and react even in time-critical** scenarios, such as detecting collisions in the production system, the result and reaction of the anomaly detection should be available as quickly as possible. Therefore, the execution of the anomaly should be performed during production, and the results must be immediately provided after new data from sensors and actors are available, e.g., a few milliseconds after the data is received. 3. **Prediction quality: For an AD application in an industrial environment, adequate** prediction performance is required. This depends on the different use cases for which the anomaly detection is applied, e.g., an F1 score of 0.95 or better for each CPS in the CPPS. 4. **Configurable: To apply AD on different CPPSs in different applications, the anomaly** detection should be adaptable to various CPSs and use cases. The possibility of using the technique for varied time series data with different variable types and diverse time lengths should be given, for instance, robots or transportation systems with features such as torque, position, and speed. 5. **Data-driven: As mentioned before, manually creating models is time-consuming and** requires deep expert knowledge. Simultaneously recording anomalous data from CPPS can be dangerous for the system itself. Therefore, the AD should only be trained with regular production data and without expert knowledge. 6. **Feasible: The AD should be compatible with current technological standards in** industrial environments to realize a generalist integration for various scenarios. This includes constraints and limitations of commonly used CPPSs in production settings: (a) Process limitations, due to the design of CPSs in industry, that are unable to execute process-intensive tasks in parallel to control and monitor the physical process, e.g., limited available RAM and processing power. (b) Communication interface constraints of commonly available CPSs in industry, e.g., OPC UA Communication, to transfer the high amount of production data at a sample rate of 2 ms during the sampling process to a database. **4. A Concept for a Fast, Decentralized, and Unsupervised Anomaly Detection in CPPSs** _4.1. Overview_ This section describes a fast and decentralized process anomaly detection concept based on several 1D-ConvAEs, which fulfills the requirements specified in the problem statement. Figure 3 shows the sequence of the different steps that are carried out. The concept consists of one AD Installation Cycle, which triggers the creation of several anomaly detection pipelines (AD pipelines) through the parallel execution of AD Generation Cycles, as shown in Figure 4. A detailed description of the AD Generation Cycle can be found in Section 4.3 and in Figure 5. The number of different AD Pipelines depends on the number of included CPSs in the CPPS. In the AD Production Cycles, located in every CPS in Figure 4, each pipeline is directly implemented and executed as part of the CPS. Explanations about the AD Production Cycle can be found in Section 4.4 and in Figure 6. The processing unit backend, an external device that can be removed after the installation process is finished, performs all heavy processing tasks of the AD Installation Cycle to meet the previously explained industrial constraints of the CPPS. The concept is developed to be executed automatically, enabling AD implementation without deep expert knowledge. In addition, a direct explanation of the individual components of the diagrams can be found in Appendix A. ----- _Sensors 2023, 23, 4207_ 7 of 19 **Figure 3. Sequence of steps performed in the concept.** **Figure 4. Overview of AD Installation Cycle.** **Figure 5. Overview of AD Generation Cycle.** ----- _Sensors 2023, 23, 4207_ 8 of 19 **Figure 6. Overview of AD Production Cycle.** _4.2. AD Installation_ The AD Installation Cycle consists of four parts: data collection, data analysis, AD generation, and deployment (see Figure 4). **Data collection: The operator triggers the data collection at the processing unit back-** end to record regular process data. Process data samples, single packages of time series data from the individual CPS, are collected at a high sample rate and sent to the control device. Over a defined period of time, the individual data of the various CPSs are recorded and then combined. The resulting package, named regular process data, is then sent to the processing unit backend. This procedure is required to meet the communication interface limitations in the installation process and enable the use of the high sample rates at the AD Production Cycles directly in the CPSs. The data packages are saved inside the processing unit backend until a specified number of records is reached. Regular process data consist of different features like position, torque, and speed sampled in the form of time series data from the various CPSs. This data can be defined as multiple data streams containing the features of the physical process recorded by the different sensors and actors. **Data Analysis: Depending on the diverse CPSs, different features with different** ranges are provided. In the analysis step, unnecessary features are automatically removed, and configuration files are accordingly generated. Each configuration file contains the necessary information for the following AD generation cycle, e.g., feature ranges, types, and default hyperparameters. The operator can manually tune this information, or the default values can be used. **AD Generation: In the installation step for each included CPS, an AD Generation** Cycle (Figure 5) is triggered. The different AD Generation Cycles can be executed in parallel to speed up the installation process. A detailed description of the AD generation cycle can be found in Section 4.3. **Deployment: After the generation of the AD pipelines, each pipeline is exported and** deployed to the separate CPS. This terminates the AD Installation Cycle. _4.3. AD Generation Cycle_ The AD Generation Cycle consists of preprocessing, model initialization, training, evaluation, optimization, and export, as shown as a BPMN diagram in Figure 5. First, the provided regular process data, the combined collected data samples of all CPS, are preprocessed with the information received from the configuration files. This transforms the data, which consist of different ranges and units, into an equal numerical range. The type of the desired preprocessor is defined in the configuration file. This enhances a configurable setup, which can handle various variables with different units and ranges. Next, the model is initialized, trained, evaluated, and optimized. Additional hyperparameters set in the configuration file are, e.g., the number of layers, filters per layer, used loss function, and type of optimizer. After initialization, the model is trained on the preprocessed ----- _Sensors 2023, 23, 4207_ 9 of 19 data. The method specified in the configuration file is used to evaluate the model. In the optimization step, the hyperparameters are changed, influenced by the defined ranges and tuning parameters. The search algorithm declared in the configuration file searches over a generated search space for the best possible parameters. These steps are executed iteratively until the specified reconstruction performance (e.g., the desired MAE Value) is reached. After the tuning is finished, the AD pipeline, a combination of preprocessor and model, is exported to the deployment step. This terminates the AD generation cycle. _4.4. AD Production Cycle_ After the AD Generation Cycle is finished and the AD pipeline is deployed in the CPS, the AD Production Cycle, shown as a BPMN diagram in Figure 6, starts. Process data samples, single packages of time series data from the CPS, are collected at a high sample rate and stored in an in-memory data storage. When the required amount of data packages to execute the AD process step is reached, the data are preprocessed and evaluated by the AD pipeline. After the execution, the previously collected data in the in-memory data storage will be released to limit the needed memory capacity. The AD process step will be executed again immediately after enough data is available. In case of an anomaly, the detection can be delivered to the control unit, or the operator can be directly notified. Additionally, the AD process can be terminated, and therefore the AD Production Cycle. _4.5. Sliding Window Convolutional Autoencoder_ To achieve adequate prediction performance and meet real-time requirements, we choose a sliding-window-based 1D-ConvAE as the model type (see Figure 7). Autoencoders are reconstruction-based neural networks that reconstruct their input as output. By only learning the reconstruction of the regular pattern, every datum consisting of unseen, abnormal patterns cannot be correctly reconstructed, which will result in a higher reconstruction error. To gain adequate prediction performance and meet the processing limitations, 1D convolutional layers are used. Adding these layers to the autoencoder allows the model to learn spatially invariant features and capture spatially local correlations from the data. This means it can recognize patterns of high-dimensional data without requiring feature engineering. At the same time, the required parameters and the computational complexity of a 1D convolutional layer are significantly lower than the comparable 2D convolutional layers. The 1D-ConvAE can be trained without expert knowledge or explicitly known anomalies, only with regular process data. This fulfills the requirement 5, data driven. A detailed comparison with other methods can be found in Appendix B. **Figure 7. Convolutional autoencoder.** ----- _Sensors 2023, 23, 4207_ 10 of 19 _4.6. Anomaly Detection_ Regular process data can be defined as a data stream containing several time series of sampled features F = ( f1, f2, ..., fn), where n defines the number of different features such as position, speed, and torque from the various sensors of the mechanical system. The data stream is split into several windows depending on the chosen window size m and step size s. Each window consists of several time series equal to n different features in the data stream over a time period corresponding to the window size m. These generated sliding windows act as the input to the model. The output of the model is each separated reconstructed sliding window. With the help of an aggregation function (e.g., arithmetic mean), the reconstructed sliding windows can be merged into a reconstructed data stream. To calculate the reconstruction error matrix E, the reconstructed error eit of each feature fi at every time step t can be calculated as the Absolute-Error (AE) e fit =| fi,t − _f[ˆ]i,t | (1 ≤_ _i ≤_ _n)_ between the input and output. This results in a matrix E representing each feature at each time step as a value of the differentiation between the input and reconstructed data stream. Threshold values must be defined to evaluate which value a reconstruction error indicates if an anomaly is detected. We employ the following method for automatically computing and tuning the threshold values. After training the model, the described method re-evaluates all training data. This results in an error matrix over the whole training data stream. The maximum reconstruction error of each feature is taken from this matrix to construct a threshold vector θ. This vector can be adapted when the model is integrated directly into the CPS by automatically tuning the values in the live testing stage. In the AD production cycle, after enough high sample data are collected in the in-memory storage, each column of the reconstruction matrix E is evaluated with the threshold vector θ. The number of collected data samples can be flexibly chosen but must be at least twice as large as the window size to allow the reconstruction concept to be applied. Suppose a value of e f _it,_ where i is the considered feature of the total features and n exceeds the associated threshold value of this feature θ fi (1 ≤ _i ≤_ _n). In that case, the data point in the input data stream_ is declared anomalous. Therefore, anomalies in the input data stream can be detected by applying the threshold vector θ to each timestep t of the reconstruction matrix E. **5. Prototype Implementation** The concept has been implemented prototypically. As programming language, [Python 3.9 is used. A MongoDB https://www.mongodb.com/ (accessed on 12 February](https://www.mongodb.com/) 2023) is established on the processing unit backend to save and export the regular process data. As a preprocessor, a MinMaxScaler was generated. The model is implemented using [the Keras library https://keras.io/ (accessed on 12 February 2023), running on top of](https://keras.io/) [Tensorflow https://www.tensorflow.org/ (accessed on 12 February 2023) [53]. For hyper-](https://www.tensorflow.org/) [parameter tuning, the python library Ray Tune https://ray.io/ (accessed on 12 February](https://ray.io/) [2023) [54] is used. Finally, a tracking server based on the library Mlflow https://mlflow.org/](https://mlflow.org/) (accessed on 12 February 2023) [55] was established to track the training results. The communication between the motion controller and the processing unit backend was realized through an OPC UA server–client model based on publish–subscribe routines. Several function blocks for buffering the high sample process data from the CPS at the motion controller were developed to establish this concept. This enables an intelligent communication pattern, where only minor changes on the motion controller must be performed to allow the described data exchange. The configuration files are written in YAML and can be accessed and changed by the operator. For each CPS, a separate configuration file is created. These files are also tracked to enable a traceable process at a later stage. The preprocessor and model integrations are developed as interfaces to satisfy the configurable requirement. Therefore, various considered models and preprocessors can be implemented as long as they follow the abstract class structure, making it easy to exchange, adapt, or evolve the described technique. To visualize the detection results and allow the user to interact with the system, a dashboard for bi-directional communication between the CPPS and the operator was implemented. ----- _Sensors 2023, 23, 4207_ 11 of 19 **6. Evaluation** _6.1. Experimental Setup_ The rotary table dispenser system shown in Figure 2 was used to evaluate the decentralized concept. The CPPS consists of different CPSs and a control unit working together to realize several processes, e.g., transport and pick-and-place operations. The overall process involves picking small items from a rotating table and putting them into several containers which are moving on conveyor belts around the machine. During the process, the different time series data of each CPS is collected in the motion controller. Several buffers are written in the motion controller to adapt the high sample rate of 2 ms of each CPS to the minimal data exchange cycle time of 50ms at the OPC UA server. After one buffer is filled, the data package is sent to the OPC UA Server running on the processing unit backend, a pc type NUC8i5BEK. The collected data are saved in the established database after each import cycle. In total, 33 different time series over a period of 6 min were recorded. Based on the high sample rate, each time series consists of around 176,000 samples, resulting in approximately 5,808,000 data points as training data. _6.2. Data Recording_ Realistic fault data were generated by forcing different anomalies into the normal process to evaluate the performance of the used models. The resulting deviations from the normal process were manually classified as anomalous areas in the resulting data stream to rate the performance. Five error cases were defined, and at least one error case was generated for each CPS. Additionally, long-term tests, including several complete processes without anomalies, were carried out to control the resulting models in the normal industrial setup. 1. **Friction: To simulate friction, which can result from abrasion of used mechanical** components, delayed maintenance, or broken parts, external forces were applied to the mechanical systems of the different CPSs, e.g., against the rotation direction of the conveyor belt or the movement of the linear sliders. This results in increased torque values at the applied CPS. 2. **Vibration: Undefined vibration, which can be caused by broken bearings or loose** attachments, was applied to the mechanical system of the CPS. The simulation was done by manually applying shocks to the rotating table. 3. **Defect components: Another industry-related anomaly can be caused by defect** components in the production process, such as a broken container. To examine this type of anomaly, different containers were manipulated in such a way that they could not be picked by the robots anymore, resulting in an undefined status of the whole production line. 4. **Incorrect process: In addition, external manipulations can influence industrial pro-** duction lines. These injections in the normal process can result in some undefined behavior of the system, which can cause damage to the products or the system itself. To simulate this kind of anomaly, the placement of the containers on the belt was changed in the running process. Therefore, the real positions differ from the fixed pre-defined positions in the machine scope. 5. **Collision: Due to external influences or process errors, even in modern indus-** trial systems, collisions may occur. The system typically detects heavy collisions, whereas smaller collisions resulting in damaged products or fragile components are mostly not recognized by the internal system. This can be, for example, a collision with an obstacle in the moving path of the linear sliders or a displaced product on the conveyor. _6.3. Model Configuration_ The sampled data from the regular process was used to train the model. A MinMax Scaler was chosen to preprocess the data by scaling the time series between zero and one. An Adam optimizer was used, and the loss function was set to MAE. The default ----- _Sensors 2023, 23, 4207_ 12 of 19 hyperparameter tuning results in 80 different decoder and encoder structures for each CPS. The best model was automatically picked by evaluating the number of parameters and the resulting loss value. Detailed information about the different considered parameters can be found in Table 1. The focus was on realizing small and efficient model architectures to meet the computational limitations (Section 3, point 6). Therefore, shallow structures with a limited amount of parameters were preferred. By comparing all achieved loss values and the resulting model structures, the smallest structure that achieved a low loss value and, thus, a good reconstruction capability was automatically selected. Due to the unsupervised setup, no anomalous data are available in the training process. Therefore, an immediate evaluation of the detection performance is not possible; consequently, only the reconstruction capability can be taken as an additional selection criterion in this process. Additional discussion on the selection process can be found in future work. As activation function, the rectified linear unit was chosen. Dropout layers were applied as regularization between the convolutional layers, and max pooling layers were used to reduce the dimensionality. Different step sizes in the training process were tested. The best results were reached with a step size of one. **Table 1. Summary table of all parameters taken in the process of automatically selecting the models.** **Model Parameter** **Range** **Definition** Number of Layers [4, 8] Total number of layers used in the model. Number of Filters in the first Layer [32, 128] The number of filters used in the first layer of the model. To realize the dimensionality reduction, the inner layers have fewer filters. (In the automated concept, half of the previous layer). Window size [32, 128] Number of time steps of the sliding window. The length of the sequence shifted between the individual Step size [1, 64] windows. Number of epochs with no improvement after which training Patience [1, 10] will be stopped. Total number of parameters [12,642, 208,614] Total number of parameters of the resulting model. Achieved mean absolute error between input and output at the Mean absolute error [0.002, 0.3] end of training. _6.4. Experimental Results_ This section validates the described concept applied in the experimental setup against the requirements defined in the problem statement. 1. **Anomaly detection: Figure 8 shows some of the forced anomalies in the experimental** setup, illustrating the detection performance of the generated models. In the pictures, the detected anomalies are marked with red points, while the pre-defined anomalous areas are indicated by the red background color of the figure. Combined with the results in Table 2, this confirms that the different models can be successfully applied to detect anomalies in the CPSs. 2. **Real-time: The evaluated sliding window sizes from the hyperparameter tuning were** between 32–64, resulting in comparably small windows. To ensure a fast detection in the real process, each generated sliding window was treated as a data stream and evaluated immediately. With a sample rate of 2 ms, the overall time to collect one window as input data for the model is between 64 and 128 ms. The average execution time per reconstruction and verification for anomalies was around 34 ms, with a maximum of 49 ms and a minimum of 22 ms. Therefore, anomaly detection can be carried out with a maximum delay of 177 ms at our setup, which allows an immediate reaction of the system on detected anomalies. ----- _Sensors 2023, 23, 4207_ 13 of 19 3. **Prediction quality: The F1 Score is used to evaluate the model performance. The** detailed performance for each CPS is shown in Table 2. To calculate the F1 Score, the manually forced anomalies were classified as anomalous areas. If an anomaly in a window was detected, the used window was assigned as anomalous and evaluated against the area. By reaching high F1 Scores above 0.95, adequate prediction performances for every single CPS are realized. This confirms that the automatically created models for each CPS can reliably detect anomalies in the given CPPS. 4. **Configurable: The described concept and resulting anomaly detection can be config-** urated for various applications. Only minor changes must be made to the motion controller to enable the sampling process. The automatically generated configuration files can be manually changed, or the default values can be used. 5. **Data-driven: The models are trained only with the regular process data. Therefore,** no anomalous data or feature engineering is needed. No values are added or changed. All removed features are automatically declared. Only the data from the sensors and actors of the CPSs are used. The model is created in an automated way by the configuration file without the need for expert knowledge. 6. **Feasible: The method utilized standard communication technologies of common** industrial setups. By outsourcing the process-intensive tasks to the processing unit backend, the concept enables the application of anomaly detection for the CPPS, even with the processing limitation and constraints of each CPS. In our experimental setup, the simulated process reaches a maximum consumption of 350MB while not exceeding a maximum of 12% CPU load. Based on the experimental results, the introduced novel concept fulfills all the defined industrial requirements of the problem statement in Section 3. **Figure 8. Anomalous Samples.** **Table 2. Performance Evaluation.** **Unit** **TP** **TN** **FP** **FN** **Precision** **Recall** **F1-Score** CB 232 3938 7 10 0.0.9707 0.958 0.964 RT 193 3199 9 8 0.955 0.960 0.957 SR 22 3002 2 0 0.916 1 0.956 PR 22 3002 1 1 0.956 0.956 0.956 P&P S 484 2965 21 20 0.958 0.960 0.959 P&P U 484 2967 29 15 0.943 0.969 0.956 P&P L 484 2964 27 22 0.947 0.956 0.951 CTS 75 3199 3 3 0.961 0.961 0.961 SRS 231 3374 8 10 0.966 0.958 0.962 CB = Conveyor Belt; RT = Rotating Table; SR = Sliding Robot; PR = Production Robot; P&P S = Pick & Place Robot S Axis; P&P U = Pick & Place Robot U Axis; P&P L = Pick & Place Robot L Axis; CTS = Container Tray Slider; SRS = Sliding Robot Slider. **7. Conclusions and Future Work** This paper presents a fast and decentralized anomaly detection concept for CPPS under industry constraints. The concept is configurable and feasible to apply anomaly detection in different use cases under the limitations of commonly used CPPSs in industrial environments. Due to the decentralization, no additional computational units must be ----- _Sensors 2023, 23, 4207_ 14 of 19 integrated. The generated models allow a fast and performant integration. The anomaly detection is executed, and evaluations are carried out immediately during production. The model is generated and tuned in a fully automated fashion. No expert knowledge about anomalous data is needed. Overall, the experiments show that each model achieves stable and accurate results. This presents a promising approach for decentralized and fast anomaly detection in CPPSs under industry constraints. However, despite the apparent success of the concept, there are several directions for future research. In this work, the concept was only tested in a single CPPS with a limited amount of CPSs. Therefore, more studies with different models, more CPSs, and under different scenarios will be performed in future work. Secondly, the models are only evaluated against the defined simulated computational resources and data storage limitations of the used CPSs. This is mainly caused by the integration limitations of the available CPSs. To integrate the models, adaptions to the hardware and software of the CPSs must be carried out in the future. Additionally, several anomalies which can emerge in a CPPS cannot be detected, e.g., process anomalies such as changing the overall process to fewer containers as in the learning process. This forces the CPPS not into an undefined state, although the actual process differs from the learned process. Therefore, another research direction in the future is to adapt the concept even to detect this kind of anomaly. Furthermore, the selection of the model is only based on two parameters, the achieved loss value and the resulting model structure. Despite the good results obtained in the tests with the defined anomalies, this method cannot guarantee the selection of the best model. Further approaches and concepts for a better evaluation of the models and a guaranteed choice of the best model must be found. Finally, up to now, the output of the anomaly detection is the identification of the anomaly, defined by the time and feature, in the data stream. Adding more information may be helpful to increase the accuracy of the AD for the operator. Ways to gather and provide this additional context information will be evaluated and investigated. **Author Contributions: Conceptualization, C.G.; methodology, C.G.; software, C.G.; validation, C.G.;** formal analysis, C.G.; investigation, C.G.; resources, C.G.; data curation, C.G.; writing—original draft preparation, C.G.; writing—review and editing, C.G. and B.H.; visualization, C.G.; supervision, C.G. and B.H.; project administration, C.G.; All authors have read and agreed to the published version of the manuscript. **Funding: This research received no external funding.** **Institutional Review Board Statement: Not applicable.** **Informed Consent Statement: Not applicable.** **Data Availability Statement: Limited accessibility to the dataset can be given in single cases.** **Conflicts of Interest: The author declares no conflict of interest.** **Appendix A** Tables A1–A3 lists each individual item of Figure 4–6 with a short description. **Table A1. Definition Table BPMN Diagram AD Installation Cycle Figure 4.** **Item** **Definition** The processing unit backend, an external device that can be removed after the instalProcessing Unit Backend lation process is finished, performs all heavy processing tasks in the AD Installation cycle to meet the previously explained industrial constraints of the CPPS. Control Device Unit which typically controls the industrial process. The interface of the embedded system to exchange data with the control device or Communication Interface the processing unit backend. ----- _Sensors 2023, 23, 4207_ 15 of 19 **Table A1. Cont.** **Item** **Definition** Part of the CPS which interacts with sensors and actors to monitor and control the Embedded System mechanical system. Mechanical System Summarizes all mechanical components of the system. Single packages of time series data from the individual CPS. Process data samples Process data samples consist of features like position, torque, and speed sampled as time series data from the CPS. Combined process data samples of all CPS collected from the normal process sampled Record Regular Process Data over a defined time. Process data samples at a high sample rate are collected from the different CPS, Collect Process Data Samples combined, and sent to the control device as a data package. In the analysis, unnecessary features are automatically removed from the data, and Analysis important information like feature range and data types are collected. Based on the analysis, configuration files are generated. The operator can manually Generate Configurations tune this information, or the default values can be used. AD Generation Cycle Main cycle to create the preprocessor and train the model. Deployment AD Pipelines Each generated AD pipeline is exported and deployed to a separate CPS. AD Production Cycle Live integration and execution of the AD pipeline in the individual CPS. **Table A2. Definition Table BPMN Diagram AD Production Cycle Figure 6.** **Item** **Definition** A fast and effective data store that caches live data until it is passed to the AD In-memory data storage pipeline for processing. Live process data is sampled at a high sample rate to an in-memory data storage to Record Live Process Data collect the needed data to execute the AD pipeline. Execute AD Process Step The collected live data is preprocessed and evaluated by the AD pipeline. Deliver Results to Control Unit The AD output can be delivered from the CPS to the control unit. Depending on the CPS, the Operator can be immediately notified by the separate Notify Operator CPS. In this step, the whole AD production cycle can be switched off to free resources and Shut Down AD stop the anomaly detection. **Table A3. Definition Table BPMN Diagram AD Generation Cycle Figure 5.** **Item** **Definition** Regular Process Data Data collected from the normal process of the CPS over a defined time. Preprocessed Data Transformed and scaled regular process data by the chosen Preprocessor. AD Pipeline A combination of initialised Preprocessor and trained model. Contains necessary parameters for the separate steps of the generation cycle, e.g., Configuration the number of layers, filters per layer, loss function, and type of optimizer. Default parameters are automatically provided but can also be manually changed and tuned. In the preprocessing step, the regular process data is transformed by the chosen Preprocessing preprocessor. This scales the data provided, which normally consists of different ranges and units, to an equal numerical range. ----- _Sensors 2023, 23, 4207_ 16 of 19 **Table A3. Cont.** **Item** **Definition** Here, the model is built based on the configuration. Therefore, the number of layers, Initialize Model filter, and type of each layer and the optimizer and loss function are set. In this step, the initialized model is trained with the preprocessed regular proTrain Model cess data. Depending on the evaluation method defined in the configuration step, the model is Evaluate Model tested, the results are tracked, and the complete experiment is saved. In the optimization step, the hyperparameters are changed, influenced by the defined Optimize Model ranges and tuning parameters. The search algorithm declared in the configuration file searches over a generated search space for the best possible parameters. Normally, after the tuning is finished, the AD pipeline is exported to the deployExport AD Pipeline ment step. **Appendix B** To reach adequate prediction performance and fulfil the requirements defined in Section 3, several models were investigated. All models were tested and evaluated against a reduced test data set, consisting of the time series data of one CPS and a limited number of anomalies. No optimisations were made to the models. The test results can be seen in Table A4 . Decisive characteristics for the choice of the model are the number of required parameters, the recognition rate and the time required for the training and the evaluation of the test data. The table clearly shows that shallow methods, despite their fast evaluation, have significant weaknesses in the recognition rate for dynamic and complex time series, as specified in [11]. The LSTMAE has a higher number of parameters compared to the other models. Due to the focused universal application and the limited process and memory resources, this is a major disadvantage. The CAE shows a slightly improved recognition in our test dataset compared to the AE architecture without convolutional layers. Based on the findings of [37–41,43], which identified the good performance of the ConvAE on industrial time series data, we have chosen the 1D-ConvAE as the model. **Table A4. Model Evaluation.** **Model** **Performance** **Size** **Avg. Time [ms]** **TP** **FP** **TN** **FN** **Recall Precision** **F1** **Compelxity** **Training** **Evaluation** OCSVM 400 25,005 19,633 138 0.7434 0.0157 0.0308 Low 37,041.6 36,811.1 iForest 467 13,709 30,865 71 0.8680 0.0329 0.0634 Low 1874.9 574.1 LSTMAE 8 1 693 1 0.8888 0.8888 0.8888 High 372,029.5 23,491.3 AE 8 2 692 1 0.8888 0.8 0.8421 Medium 13,151.2 18,720 1D-ConvAE 8 0 694 2 0.8 1 0.8888 Medium 113,227.6 20,045.2 OCSVM = One-Class Support Vector Machine [56]; iForest = Isolation Forest [57]; LSTMAE = Long Short-term Memory Autoencoder [58]; AE = Autoencoder [59]; 1D-ConvAE = One dimensional convolutional Autoencoders. **Appendix C** Figure A1 shows additional showcases of anomalous samples. In Figure A1a the collision case is shown. As described in Section 6.2, the system detects typically heavy collisions. The production of light collisions is a difficult task that requires precise interference in the process. Therefore, a pole was prepared with a predetermined breaking point and applied against the rotation direction of the rotary table during regular operation. Figure A1b shows the case defect component. A container was manipulated to force pressure on the moving conveyor belt at a certain point. To create the error case seen in Figure A1c, minor ----- _Sensors 2023, 23, 4207_ 17 of 19 shocks were manually applied to the plate of the rotation table. A specialist carried out all the different tests, considering safety aspects for the person and the system. **Figure A1. Additional Anomalous Samples.** **References** 1. Marwedel, P. Embedded System Design: Embedded Systems Foundations of Cyber-Physical Systems, and the Internet of Things; Springer: Cham, Switzerland, 2021; pp. 1–15. 2. Jazdi, N. Cyber physical systems in the context of Industry 4.0. In Proceedings of the 2014 IEEE International Conference on [Automation, Quality and Testing, Robotics, Cluj-Napoca, Romania, 22–24 May 2014; pp. 14–16. [CrossRef]](http://doi.org/10.1109/AQTR.2014.6857843) 3. Rajkumar, R.; Lee, I.; Sha, L.; Stankovic, J. Cyber-Physical Systems: The Next Computing Revolution. In Proceedings of the [Design Automation Conference, Anaheim, CA, USA, 13–18 June 2010; pp. 731–736. [CrossRef]](http://dx.doi.org/10.1145/1837274.1837461) 4. Müller, T.; Jazdi, N.; Schmidt, J.; Weyrich, M. Cyber-physical production systems: Enhancement with a self-organized reconfigu[ration management. Procedia CIRP 2021, 9, 549–554. [CrossRef]](http://dx.doi.org/10.1016/j.procir.2021.03.075) 5. Monostori, L. Cyber-physical Production Systems: Roots, Expectations and R & D Challenges. Procedia CIRP 2014, 17, 9–13. [[CrossRef]](http://dx.doi.org/10.1016/j.procir.2014.03.115) 6. Ali, N.; Hussain, M.; Hong, J.-E. SafeSoCPS: A Composite Safety Analysis Approach for System of Cyber-Physical Systems. _[Sensors 2022, 22, 4474. [CrossRef]](http://dx.doi.org/10.3390/s22124474)_ 7. Eiteneuer, B.; Hranisavljevic, N.; Niggemann, O. Dimensionality Reduction and Anomaly Detection for CPPS Data using Autoencoder. In Proceedings of the 2019 IEEE International Conference on Industrial Technology (ICIT), Melbourne, VIC, [Australia, 13–15 February 2019; pp. 1286–1292. [CrossRef]](http://dx.doi.org/10.1109/ICIT.2019.8755116) 8. Adepu, S.; Mathur,A. Distributed Attack Detection in a Water Treatment Plant: Method and Case Study. IEEE Trans. Dependable _[Secur. Comput. 2021, 18, 86–99. [CrossRef]](http://dx.doi.org/10.1109/TDSC.2018.2875008)_ 9. [Chandola, V.; Banerjee, A.; Kumar, V. Anomaly detection: A survey. ACM Comput. Surv. 2009, 41, 58. [CrossRef]](http://dx.doi.org/10.1145/1541880.1541882) 10. Stojanovic, L.; Dinic, M.; Stojanovic, N.; Stojadinovic, A. Big-data-driven anomaly detection in industry (4.0): An approach and a case study. In Proceedings of the 2016 IEEE International Conference on Big Data (Big Data), Washington, DC, USA, 5–8 [December 2016; pp. 1647–1652. [CrossRef]](http://dx.doi.org/10.1109/BigData.2016.7840777) 11. Ruff, L.; Kauffmann, J.R.; Vandermeulen, R.A.; Montavon, G.; Samek, W.; Kloft, M.; Dietterich, T.G.; Muller, K.-R. A Unifying [Review of Deep and Shallow Anomaly Detection. Proc. IEEE 2021, 109, 756–795. [CrossRef]](http://dx.doi.org/10.1109/JPROC.2021.3052449) 12. Gerz, F.; Bastürk, T.R.; Kirchhoff, J.; Denker, J.; Al-Shrouf, L.; Jelali, M. Comparative Study and a New Industrial Platform for Decentralized Anomaly Detection Using Machine Learning Algorithms. In Proceedings of the 2022 International Joint Conference [on Neural Networks (IJCNN), Padua, Italy, 18–23 July 2022; pp. 1–8. [CrossRef]](http://dx.doi.org/10.1109/IJCNN55064.2022.9892939) 13. Bulusu, S.; Kailkhura, B.; Li, B.; Varshney, P.K.; Song, D. Anomalous Example Detection in Deep Learning: A Survey. IEEE Access **[2020, 8, 132330–132347. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.3010274)** 14. Mishra, P.; Varadharajan, V.; Tupakula, U.; Pilli, E.S. A Detailed Investigation and Analysis of Using Machine Learning Techniques [for Intrusion Detection. IEEE Commun. Surv. Tutor. 2019, 21, 686–728. [CrossRef]](http://dx.doi.org/10.1109/COMST.2018.2847722) 15. Thudumu, S.; Branch, P.; Jin, J.; Singh, J. A comprehensive survey of anomaly detection techniques for high dimensional big data. _[Big Data 2020, 7, 1–30. [CrossRef]](http://dx.doi.org/10.1186/s40537-020-00320-x)_ ----- _Sensors 2023, 23, 4207_ 18 of 19 16. Cook, A.A.; Mısırlı, G.; Fan, Z. Anomaly Detection for IoT Time-Series Data: A Survey. IEEE Internet Things J. 2020, 7, 6481–6494. [[CrossRef]](http://dx.doi.org/10.1109/JIOT.2019.2958185) 17. Goldstein, M.; Uchida, S. A Comparative Evaluation of Unsupervised Anomaly Detection Algorithms for Multivariate Data. _[PLoS ONE 2016, 11, e0152173. [CrossRef] [PubMed]](http://dx.doi.org/10.1371/journal.pone.0152173)_ 18. [Oza, P.; Patel, V.M. One-Class Convolutional Neural Network. IEEE Signal Process. Lett. 2019, 26, 277–281. [CrossRef]](http://dx.doi.org/10.1109/LSP.2018.2889273) 19. Erfani, S.M.; Rajasegarar, S.; Karunasekera, S.; Leckie, C. High-dimensional and large-scale anomaly detection using a linear [one-class SVM with deep learning. Pattern Recognit. 2016, 58, 121–134. [CrossRef]](http://dx.doi.org/10.1016/j.patcog.2016.03.028) 20. Smets, K.; Verdonk, B.; Jordaan, E.M. Discovering novelty in spatio/temporal data using one-class support vector machines. In Proceedings of the 2009 International Joint Conference on Neural Networks, Atlanta, GA, USA, 14–19 June 2009; pp. 2956–2963. [[CrossRef]](http://dx.doi.org/10.1109/IJCNN.2009.5178801) 21. Zong, B.; Song, Q.; Min, M.R.; Cheng, W.; Lumezanu, C.; Cho, D.; Chen, H. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In Proceedings of the International Conference on Learning Representations, Vancouver, BC, Canada, 30 April–3 May 2018. 22. Xiaoyi, G.; Akoglu, L.; Rinaldo, A. Statistical analysis of nearest neighbor methods for anomaly detection. _arXiv 2019,_ arXiv:1907.03813. 23. Elnour, M.; Meskin, N.; Khan, K.; Jain, R. A dual-isolation-forests-based attack detection framework for industrial control systems. _[IEEE Access 2020, 8, 36639–36651. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.2975066)_ 24. Pang, G.; Shen, C.; Cao, L.; Van Den Hengel, A. Deep Learning for Anomaly Detection: A Review. ACM Comput. Surv. 2022, 54, [1–38. [CrossRef]](http://dx.doi.org/10.1145/3439950) 25. Li, D.; Chen, D.; Jin, B.; Shi, L.; Goh, J.; Ng, S.K. MAD-GAN: Multivariate Anomaly Detection for Time Series Data with Generative Adversarial Networks. In Artificial Neural Networks and Machine Learning—ICANN 2019: Text and Time Series ICANN 2019 Lecture _[Notes in Computer Science; Springer: Cham, Switzerland, 2019; Volume 11730, pp. 703–716. [CrossRef]](http://dx.doi.org/doi.org/10.1007/978-3-030-30490-4-56)_ 26. Choi, Y.; Lim, H.; Choi, H.; Kim, I.-J. GAN-Based Anomaly Detection and Localization of Multivariate Time Series Data for Power Plant. In Proceedings of the 2020 IEEE International Conference on Big Data and Smart Computing (BigComp), Busan, Republic [of Korea, 19–22 February 2020; pp. 71–74. [CrossRef]](http://dx.doi.org/10.1109/BigComp48618.2020.00-97) 27. Jiang, W.; Hong, Y.; Zhou, B.; He, X.; Cheng, C. A GAN-Based Anomaly Detection Approach for Imbalanced Industrial Time [Series. IEEE Access 2019, 7, 143608–143619. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2944689) 28. Heiko, H. Kernel PCA for novelty detection. Pattern Recognit. 2007, 40, 863–874. 29. Zhao, Y.; Deng, B.; Shen, C.; Liu, Y.; Lu, H.; Hua, X.S. Spatio-temporal autoencoder for video anomaly detection. In Proceedings of the 25th ACM International Conference on Multimedia, Mountain View, CA, USA, 23–27 October 2017. 30. Gong, D.; Liu, L.; Le V.; Saha, B.; Mansour, M.R.; Venkatesh, S.; Hengel, A.V.D. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Republic of Korea, 27 October–2 November 2019. 31. Meidan, Y.; Bohadana, M.; Mathov, Y.; Mirsky, Y.; Shabtai, A.; Breitenbacher, D.; Elovici, Y. N-baiot-network-based detection of iot [botnet attacks using deep autoencoders. IEEE Pervasive Comput. 2018, 17, 12–22. [CrossRef]](http://dx.doi.org/10.1109/MPRV.2018.03367731) 32. Park, S.; Adosoglou, G.; Pardalos, P.M. Interpreting rate-distortion of variational autoencoder and using model uncertainty for [anomaly detection. Ann. Math. Artif. Intell. 2022, 90, 735–752. [CrossRef]](http://dx.doi.org/10.1007/s10472-021-09728-4) 33. Jinwon, A.; Cho, S. Variational autoencoder based anomaly detection using reconstruction probability. Spec. Lect. IE 2015, 2, 1–18. 34. Munir, M.; Siddiqui, S.A.; Dengel, A.; Ahmed, S. DeepAnT: A Deep Learning Approach for Unsupervised Anomaly Detection in [Time Series. IEEE Access 2019, 7, 1991–2005. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2018.2886457) 35. Gong, W.; Chen, H.; Zhang, Z.; Zhang, M.; Gao, H. A Data-Driven-Based Fault Diagnosis Approach for Electrical Power DC-DC Inverter by Using Modified Convolutional Neural Network With Global Average Pooling and 2-D Feature Image. IEEE Access **[2020, 8, 73677–73697. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.2988323)** 36. Gong, W.; Wang, Y.; Zhang, M.; Mihankhah, E.; Chen, H.; Wang, D. A Fast Anomaly Diagnosis Approach Based on Modified [CNN and Multisensor Data Fusion. IEEE Trans. Ind. Electron. 2022, 69, 13636–13646. [CrossRef]](http://dx.doi.org/10.1109/TIE.2021.3135520) 37. Qu, C.; Zhou, Z.; Liu, Z.; Jia, S. Predictive anomaly detection for marine diesel engine based on echo state network and [autoencoder. Energy Rep. 2022, 8 (Suppl. 4), 998–1003. [CrossRef]](http://dx.doi.org/10.1016/j.egyr.2022.01.225) 38. Malviya, V.; Mukherjee, I.; Tallur, S. Edge-Compatible Convolutional Autoencoder Implemented on FPGA for Anomaly Detection [in Vibration Condition-Based Monitoring. IEEE Sens. Lett. 2022, 6, 1–4. [CrossRef]](http://dx.doi.org/10.1109/LSENS.2022.3159972) 39. Guo, X.; Liu, X.; Zhu, E.; Yin, J. Deep Clustering with Convolutional Autoencoders. In Neural Information Processing: 24th _International Conference, ICONIP 2017, Guangzhou, China, 14–18 November 2017; Proceedings, Part II 24; Springer: Cham, Switzerland,_ [2017; p. 10635. [CrossRef]](http://dx.doi.org/10.1007/978-3-319-70096-0-39) 40. Lee, G.; Jung, M.; Song, M.; Choo, J. Unsupervised anomaly detection of the gas turbine operation via convolutional auto-encoder. In Proceedings of the 2020 IEEE International Conference on Prognostics and Health Management (ICPHM), Detroit, MI, USA, [8–10 June 2020; pp. 1–6. [CrossRef]](http://dx.doi.org/10.1109/ICPHM49022.2020.9187054) 41. Yu, J.; Zhou, X. One-Dimensional Residual Convolutional Autoencoder Based Feature Learning for Gearbox Fault Diagnosis. _[IEEE Trans. Ind. Inform. 2020, 16, 6347–6358. [CrossRef]](http://dx.doi.org/10.1109/TII.2020.2966326)_ 42. Chen, T.; Liu, X.; Xia, B.; Wang, W.; Lai, Y. Unsupervised Anomaly Detection of Industrial Robots Using Sliding-Window [Convolutional Variational Autoencoder. IEEE Access 2020, 8, 47072–47081. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.2977892) ----- _Sensors 2023, 23, 4207_ 19 of 19 43. Kwak, M.; Kim, S.B. Unsupervised Abnormal Sensor Signal Detection With Channelwise Reconstruction Errors. IEEE Access 2021, _[9, 39995–40007. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2021.3064563)_ 44. Lai, Y.; Liu, Z.; Song, Z.; Wang, Y.; Gao, Y. Anomaly detection in Industrial Autonomous Decentralized System based on time [series. Simul. Model. Pract. Theory 2016, 65, 57–71. [CrossRef]](http://dx.doi.org/10.1016/j.simpat.2016.01.013) 45. Sanjith, S.L.; Prakash Raj, E.G.D. Decentralized Time-Window Based Real-Time Anomaly Detection Mechanism (DTRAD) in Iot. _[Int. J. Recent Technol. Eng. 2019, 8, 1619–1625. [CrossRef]](http://dx.doi.org/10.35940/ijrte.B2350.078219)_ 46. Gupta, K.; Sahoo, S.; Mohanty, R.; Panigrahi B.K.; Blaabjerg, F. Decentralized Anomaly Identification in Cyber-Physical DC Microgrids. In Proceedings of the 2022 IEEE Energy Conversion Congress and Exposition (ECCE), Detroit, MI, USA, 9–13 October [2022; pp. 1–6. [CrossRef]](http://dx.doi.org/10.1109/ECCE50734.2022.9947581) 47. Wilbur, M.; Dubey, A.; Leão B.; Bhattacharjee, S. A Decentralized Approach for Real Time Anomaly Detection in Transportation Networks. In Proceedings of the 2019 IEEE International Conference on Smart Computing (SMARTCOMP), Washington, DC, [USA, 12–15 June 2019; pp. 274–282. [CrossRef]](http://dx.doi.org/10.1109/SMARTCOMP.2019.00063) 48. Bosman, H.; Iacca, G.; Tejada, A.; Wörtche H.J.; Liotta, A. Spatial anomaly detection in sensor networks using neighborhood [information. Inf. Fusion 2017, 33, 41–56. [CrossRef]](http://dx.doi.org/10.1016/j.inffus.2016.04.007) 49. Nikolay, L.; Amizadeh, S.; Flint, I. Generic and scalable framework for automated time-series anomaly detection. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Sydney, Australia, 10–13 August 2015. 50. Mayra, M.; Wu, C. An unsupervised framework for anomaly detection in a water treatment system. In Proceedings of the 2019 18th IEEE International Conference on Machine Learning And Applications (ICMLA), Boca Raton, FL, USA, 16–19 December 2019. 51. Schneider, P.; Böttinger, K. High-Performance Unsupervised Anomaly Detection for Cyber-Physical System Networks. In Proceedings of the 2018 Workshop on Cyber-Physical Systems Security and PrivaCy, Toronto, ON, Canada, 15–19 October 2018. 52. Goetz, C.; Humm, G.B. Unsupervised Process Anomaly Detection under Industry Constraints in Cyber-Physical Systems using Convolutional Autoencoder. In Computational Intelligence for Engineering and Management Applications, Select Proceedings of CIEMA _2022; Springer: Singapore, 2023, to be published._ 53. Abadi, M.; Agarwal, A.; Barham, P.; Brevdo, E.; Chen, Z.; Citro, C.; Corrado, G.; Davis, A.; Dean, J.; Devin, M.; et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems. 2015. Available online: tensorflow.org (accessed on 13 February 2023). 54. Moritz, P.; Nishihara, R.; Wang, S.; Tumanov, A.; Liaw, R.; Liang, E.; Elibol M.; Yang, Z.; Paul, W.; Jordan M.; et al. Ray: A distributed framework for emerging AI applications. In Proceedings of the 13th USENIX Symposium on Operating Systems Design and Implementation (OSDI 18), Carlsbad, CA, USA, 8–10 October 2018; pp. 561–577. 55. Zaharia, M.A.; Chen, A.; Davidson, A.; Ghodsi, A.; Hong, S.A.; Konwinski, A.; Murching, S.; Nykodym, T.; Ogilvie, P.; Parkhe, M.; et al. Accelerating the Machine Learning Lifecycle with MLflow. IEEE Data Eng. Bull. 2018, 41, 39–45. 56. Schölkopf, B.; Williamson, R.; Smola, A.; Shawe-Taylor, J.; Platt, J. Support Vector Method for Novelty Detection. NIPS 1999, 12, 582–588. 57. [Liu, F.; Ting, K.; Zhou, Z. Isolation-Based Anomaly Detection. ACM Trans. Knowl. Discov. Data 2012, 6, 3. [CrossRef]](http://dx.doi.org/10.1145/2133360.2133363) 58. [Hochreiter, S.; Schmidhuber, J. Long Short-term Memory. Neural Comput. 1997, 9, 1735–1780. [CrossRef]](http://dx.doi.org/10.1162/neco.1997.9.8.1735) 59. Baldi, P. Autoencoders, unsupervised learning and deep architectures. In Proceedings of the 2011 International Conference on Unsupervised and Transfer Learning Workshop, Bellevue, DC, USA, 2 July 2011; Volume 27, pp. 37–50. **Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual** author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. -----
{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC10181007, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://www.mdpi.com/1424-8220/23/9/4207/pdf?version=1682227452" }
2,023
[ "JournalArticle" ]
true
2023-04-23T00:00:00
[ { "paperId": "f662bb9ae32e13fafc2a9783cd6b787aec06e4a7", "title": "A Fast Anomaly Diagnosis Approach Based on Modified CNN and Multisensor Data Fusion" }, { "paperId": "833ca3806caafced84fa9ba8e5ff40ac1b5a84c0", "title": "Decentralized Anomaly Identification in Cyber-Physical DC Microgrids" }, { "paperId": "ff4831857071965395f01dc4eab46a50275fe7f9", "title": "A Comparative Study and a New Industrial Platform for Decentralized Anomaly Detection Using Machine Learning Algorithms" }, { "paperId": "8ab87df3efc998a4901ee67da2c128f448820155", "title": "SafeSoCPS: A Composite Safety Analysis Approach for System of Cyber-Physical Systems" }, { "paperId": "c4e8bec1a6d75c503f81fef28ccfebff600791b4", "title": "Edge-Compatible Convolutional Autoencoder Implemented on FPGA for Anomaly Detection in Vibration Condition-Based Monitoring" }, { "paperId": "b461fe6c747e7a68c75f4941238160b58c0f06a4", "title": "One-Dimensional Residual Convolutional Autoencoder Based Feature Learning for Gearbox Fault Diagnosis" }, { "paperId": "30b99ae0682d42a2010be401dd1d8f7baca9bb5f", "title": "A Unifying Review of Deep and Shallow Anomaly Detection" }, { "paperId": "775247047d0b56950ba5ea77d4a29772eca95c1b", "title": "Deep Learning for Anomaly Detection" }, { "paperId": "aca6d5f3866372a4506cf15773ae298f18c3f453", "title": "A comprehensive survey of anomaly detection techniques for high dimensional big data" }, { "paperId": "30d65fbc8d71ac202ee8d7d5f0e58c63a6c6a957", "title": "Anomaly Detection for IoT Time-Series Data: A Survey" }, { "paperId": "3c9bf9cefe2c365097132c2aa853adc305224806", "title": "Unsupervised anomaly detection of the gas turbine operation via convolutional auto-encoder" }, { "paperId": "7cf5dddc831dacdbb9a1ab5553c0a6e99ef64623", "title": "Interpreting rate-distortion of variational autoencoder and using model uncertainty for anomaly detection" }, { "paperId": "76d98534d3ec7b716806ffa0f84acd1e347f8748", "title": "A Data-Driven-Based Fault Diagnosis Approach for Electrical Power DC-DC Inverter by Using Modified Convolutional Neural Network With Global Average Pooling and 2-D Feature Image" }, { "paperId": "f53edf46eeb9fb110391c6c10ea27e439cf91ec8", "title": "Unsupervised Anomaly Detection of Industrial Robots Using Sliding-Window Convolutional Variational Autoencoder" }, { "paperId": "a5342b260737a8088af20250d82fc2d28589959a", "title": "A Dual-Isolation-Forests-Based Attack Detection Framework for Industrial Control Systems" }, { "paperId": "0c9997586868c163e6888a761df950a6f6c7eb70", "title": "GAN-Based Anomaly Detection and Localization of Multivariate Time Series Data for Power Plant" }, { "paperId": "ea478a48c863649daba5d3394431e5ea77a4449b", "title": "An Unsupervised Framework for Anomaly Detection in a Water Treatment System" }, { "paperId": "ff382aed6fda4c7de23349629138b09263be456a", "title": "Decentralized Time-Window based Real-Time Anomaly Detection Mechanism (DTRAD) in Iot" }, { "paperId": "d2fe8bdc0d638ce6f082b4a0cd08d999f4e9e8d3", "title": "Statistical Analysis of Nearest Neighbor Methods for Anomaly Detection" }, { "paperId": "b1f284c2fb68a609e707c93ea0ace48b1d833063", "title": "A Decentralized Approach for Real Time Anomaly Detection in Transportation Networks" }, { "paperId": "d65eb30e5f0d2013fd5e4f45d1413bc2969ee803", "title": "Memorizing Normality to Detect Anomaly: Memory-Augmented Deep Autoencoder for Unsupervised Anomaly Detection" }, { "paperId": "8dff7bb49b7dd83cbf89f20b2d58fe15a9b12ea8", "title": "A GAN-Based Anomaly Detection Approach for Imbalanced Industrial Time Series" }, { "paperId": "9cf47b757bae8f4b8a3b0ec833c4be610304efdd", "title": "Dimensionality Reduction and Anomaly Detection for CPPS Data using Autoencoder" }, { "paperId": "00695a31a80221c7125e49885a4767896ec2c4f7", "title": "One-Class Convolutional Neural Network" }, { "paperId": "261e81cba749f70271fa4b7e230328fc1a4a6c96", "title": "MAD-GAN: Multivariate Anomaly Detection for Time Series Data with Generative Adversarial Networks" }, { "paperId": "a31912c8684fdb334a62950f45300d42349ffe74", "title": "Distributed Attack Detection in a Water Treatment Plant: Method and Case Study" }, { "paperId": "6b8aec2a36d28f32b85163c7b1107767405d678a", "title": "N-BaIoT—Network-Based Detection of IoT Botnet Attacks Using Deep Autoencoders" }, { "paperId": "dbc7401e3e75c40d3c720e7db3c906d48bd742d7", "title": "Deep Autoencoding Gaussian Mixture Model for Unsupervised Anomaly Detection" }, { "paperId": "a531ba97b05fc3a16f5fb023fb220236e01601d0", "title": "High-Performance Unsupervised Anomaly Detection for Cyber-Physical System Networks" }, { "paperId": "f83a207712fd4cf41aded79e9e6c4345ba879128", "title": "Ray: A Distributed Framework for Emerging AI Applications" }, { "paperId": "1003462fd2d360542ea712430b13cf0c43fc7f9d", "title": "Deep Clustering with Convolutional Autoencoders" }, { "paperId": "fef6f1e04fa64f2f26ac9f01cd143dd19e549790", "title": "Spatio-Temporal AutoEncoder for Video Anomaly Detection" }, { "paperId": "f076e4355c0facf111716dcab2837803367dd2d8", "title": "High-dimensional and large-scale anomaly detection using a linear one-class SVM with deep learning" }, { "paperId": "5fbe96e183ca492c0175e07d4b7c90c77de2469b", "title": "Anomaly detection in Industrial Autonomous Decentralized System based on time series" }, { "paperId": "355692eb86b06a0a23af45c106cfb02c95bf380e", "title": "A Comparative Evaluation of Unsupervised Anomaly Detection Algorithms for Multivariate Data" }, { "paperId": "435c170edd031b65cf9087cf3c8473f071477df5", "title": "Generic and Scalable Framework for Automated Time-series Anomaly Detection" }, { "paperId": "1b7b3df1e711c9bb20f77cb947f53fac21017595", "title": "Cyber physical systems in the context of Industry 4.0" }, { "paperId": "dcd9f5d61bc9a70c40be84a8d78fbee822ffcd9e", "title": "Isolation-Based Anomaly Detection" }, { "paperId": "bb1d6215f0cfd84b5efc7173247b016ade4c976e", "title": "Autoencoders, Unsupervised Learning, and Deep Architectures" }, { "paperId": "4dd03ed9f0821c6663a9f08898b9a97184a171de", "title": "Cyber-physical systems: The next computing revolution" }, { "paperId": "71d1ac92ad36b62a04f32ed75a10ad3259a7218d", "title": "Anomaly detection: A survey" }, { "paperId": "69c784da3839474fc7274f53609e0e3049cff5ef", "title": "Discovering novelty in spatio/temporal data using one-class support vector machines" }, { "paperId": "27222787908c3a1c6fb6c4b5cb5ef8b2542f1b3c", "title": "Kernel PCA for novelty detection" }, { "paperId": "bf206bad6a74d27b40c8ea77ee54e98e492fb7f9", "title": "Support Vector Method for Novelty Detection" }, { "paperId": "2e9d221c206e9503ceb452302d68d10e293f2a10", "title": "Long Short-Term Memory" }, { "paperId": null, "title": "Large-Scale Machine Learning on Heterogeneous Systems" }, { "paperId": "cc0225276effb2219c8e4579be6c4719824fc1e0", "title": "Predictive anomaly detection for marine diesel engine based on echo state network and autoencoder" }, { "paperId": "5affaee79c34c0e299fcf4e0246565ef63f48d2c", "title": "Cyber-physical production systems: enhancement with a self-organized reconfiguration management" }, { "paperId": "0141d6a31c5cd369e362b1d26972000552e2388b", "title": "Unsupervised Abnormal Sensor Signal Detection With Channelwise Reconstruction Errors" }, { "paperId": "37eb32915b7767685ec3c9e9728ca9a50a379b8d", "title": "Anomalous Example Detection in Deep Learning: A Survey" }, { "paperId": "d294d5246e0dd8ed8bd9ec9d24a01fd4ece4fb3c", "title": "A Detailed Investigation and Analysis of Using Machine Learning Techniques for Intrusion Detection" }, { "paperId": "f37bc75aa1833e1330c39c4f04b131baca08d67b", "title": "DeepAnT: A Deep Learning Approach for Unsupervised Anomaly Detection in Time Series" }, { "paperId": "b2e0b79e6f180af2e0e559f2b1faba66b2bd578a", "title": "Accelerating the Machine Learning Lifecycle with MLflow" }, { "paperId": "20ba792ed3646949ff85d72785683a1766196fab", "title": "Spatial anomaly detection in sensor networks using neighborhood information" }, { "paperId": "061146b1d7938d7a8dae70e3531a00fceb3c78e8", "title": "Variational Autoencoder based Anomaly Detection using Reconstruction Probability" }, { "paperId": "df54a028da5d47d8458780c84afd0b38b3ddb012", "title": "Embedded System Design" }, { "paperId": "5eb0cca4c9520bf00e37e4706dd0d1759c21d1ff", "title": "2016 Ieee International Conference on Big Data (big Data) Big-data-driven Anomaly Detection in Industry (4.0): an Approach and a Case Study" } ]
17,226
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/016337f0d33ffcc75d194f57abc788161eaec927
[ "Computer Science" ]
0.8917
On the Convergence of Artificial Intelligence and Distributed Ledger Technology: A Scoping Review and Future Research Agenda
016337f0d33ffcc75d194f57abc788161eaec927
IEEE Access
[ { "authorId": "7259180", "name": "Konstantin D. Pandl" }, { "authorId": "2244544", "name": "Scott Thiebes" }, { "authorId": "1403054338", "name": "Manuel Schmidt-Kraepelin" }, { "authorId": "1697247", "name": "A. Sunyaev" } ]
{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "http://ieeexplore.ieee.org/servlet/opac?punumber=6287639" ], "id": "2633f5b2-c15c-49fe-80f5-07523e770c26", "issn": "2169-3536", "name": "IEEE Access", "type": "journal", "url": "http://www.ieee.org/publications_standards/publications/ieee_access.html" }
Developments in artificial intelligence (AI) and distributed ledger technology (DLT) currently lead to lively debates in academia and practice. AI processes data to perform tasks that were previously thought possible only for humans. DLT has the potential to create consensus over data among a group of participants in untrustworthy environments. In recent research, both technologies are used in similar and even the same systems. This can lead to a convergence of AI and DLT, which in the past, has paved the way for major innovations of other information technologies. Previous work highlights several potential benefits of a convergence of AI and DLT but only provides a limited theoretical framework to describe upcoming real-world integration cases of both technologies. In this research, we review and synthesize extant research on integrating AI with DLT and vice versa to rigorously develop a future research agenda on the convergence of both technologies. In terms of integrating AI with DLT, we identified research opportunities in the areas of secure DLT, automated referee and governance, and privacy-preserving personalization. With regard to integrating DLT with AI, we identified future research opportunities in the areas of decentralized computing for AI, secure data sharing and marketplaces, explainable AI, and coordinating devices. In doing so, this research provides a four-fold contribution. First, it is not constrained to blockchain but instead investigates the broader phenomenon of DLT. Second, it considers the reciprocal nature of a convergence of AI and DLT. Third, it bridges the gap between theory and practice by helping researchers active in AI or DLT to overcome current limitations in their field, and practitioners to develop systems along with the convergence of both technologies. Fourth, it demonstrates the feasibility of applying the convergence concept to research on AI and DLT.
Received February 4, 2020, accepted February 27, 2020, date of publication March 17, 2020, date of current version March 31, 2020. _Digital Object Identifier 10.1109/ACCESS.2020.2981447_ # On the Convergence of Artificial Intelligence and Distributed Ledger Technology: A Scoping Review and Future Research Agenda KONSTANTIN D. PANDL, SCOTT THIEBES, MANUEL SCHMIDT-KRAEPELIN, AND ALI SUNYAEV Institute of Applied Informatics and Formal Description Methods, Karlsruhe Institute of Technology, 76131 Karlsruhe, Germany Corresponding author: Ali Sunyaev (sunyaev@kit.edu) This work was supported by the Karlsruhe Institute of Technology through the KIT-Publication Fund. **ABSTRACT Developments in artificial intelligence (AI) and distributed ledger technology (DLT) currently** lead to lively debates in academia and practice. AI processes data to perform tasks that were previously thought possible only for humans. DLT has the potential to create consensus over data among a group of participants in untrustworthy environments. In recent research, both technologies are used in similar and even the same systems. This can lead to a convergence of AI and DLT, which in the past, has paved the way for major innovations of other information technologies. Previous work highlights several potential benefits of a convergence of AI and DLT but only provides a limited theoretical framework to describe upcoming real-world integration cases of both technologies. In this research, we review and synthesize extant research on integrating AI with DLT and vice versa to rigorously develop a future research agenda on the convergence of both technologies. In terms of integrating AI with DLT, we identified research opportunities in the areas of secure DLT, automated referee and governance, and privacy-preserving personalization. With regard to integrating DLT with AI, we identified future research opportunities in the areas of decentralized computing for AI, secure data sharing and marketplaces, explainable AI, and coordinating devices. In doing so, this research provides a four-fold contribution. First, it is not constrained to blockchain but instead investigates the broader phenomenon of DLT. Second, it considers the reciprocal nature of a convergence of AI and DLT. Third, it bridges the gap between theory and practice by helping researchers active in AI or DLT to overcome current limitations in their field, and practitioners to develop systems along with the convergence of both technologies. Fourth, it demonstrates the feasibility of applying the convergence concept to research on AI and DLT. **INDEX TERMS Artificial intelligence, blockchain, convergence, distributed ledger technology, machine** learning. **I. INTRODUCTION** Artificial intelligence (AI) and distributed ledger technology (DLT) are among today’s most actively debated developments in information technology with potential for tremendous impact on individuals, organizations, and societies over the next decades. A 2018 report from the McKinsey Global Institute estimates that the application of AI in various industries could deliver an additional global economic output of around USD 13 trillion by 2030 [1]. Similarly, a study by the World Economic Forum predicts that by 2025, up to 10 % The associate editor coordinating the review of this manuscript and approving it for publication was Jiafeng Xie. of the world’s GDP may be stored on a blockchain [2], which is the most commonly used concept of DLT today [3]. AI can perform complex tasks that were previously thought possible only for humans to perform. In some application domains, AI can nowadays already exceed human capabilities. Research in the health care domain has, for instance, shown that AI can analyze echocardiograms faster and more accurate than medical professionals [4]. Furthermore, advancements in AI are also expected to be key enablers of important upcoming innovations such as autonomous driving [5] or intelligent robots [6], to name but a few. DLT, on the other hand, can provide consensus over a shared ledger in untrustworthy networks containing, for example, unreachable or maliciously behaving nodes [3]. It became known ----- to the public due to the emergence of the cryptocurrency Bitcoin [7]. Following the success of Bitcoin, further DLT applications are emerging in application domains beyond finance that are often corresponding to those of AI. DLT may, for example, be used to manage access control for electronic health records [8], or to secure the IT systems of autonomous cars [9]. As one result of these developments, we now increasingly see the emergence of applications using both information technologies in close integration. Recent work, for example, uses deep reinforcement learning to explore attacks on blockchain incentive mechanisms [10]. In doing so, the AI system can detect new attack strategies and provide security insights, even for well-studied DLT protocols like Bitcoin. Another recent work uses a DLT-based platform to exchange data and computing resources in order to enable AI applications [11]. The platform gives data providers the opportunity to share their data while keeping it confidential and maintaining the right to manage data access. Data consumers can then train algorithms on the provided data and compensate data providers for use of their data. The examples above demonstrate that the integration of AI and DLT yields great potential to advance the capabilities of both technologies, and, ultimately, to increase the positive impact of AI and DLT on individuals, organizations, and societies. Yet, in order to make meaningful contributions, researchers and practitioners alike will have to keep up with the latest developments in both fields as well as the most recent developments and innovations related to their integration. Owing to the fast pace and interdisciplinary nature of both research areas, assessing the current state of research on AI and DLT and especially their integration in its entirety is a difficult task. In attempt to provide guidance to researchers enticed by the integration of AI and DLT, previous research has either focused on partial aspects of the integration of AI and DLT such as the use of blockchain for AI (e.g., [12]) or on deriving conceptual ideas of how both technologies might be integrated with each other (e.g., [13]). Despite the invaluable contributions that these publications have made to the nascent stream of literature concerning the integration of AI and DLT, we presently lack in-depth knowledge about the current state of research on the integration of AI and DLT that does (a) not only focus on a specific DLT concept (i.e., blockchain), (b) considers the reciprocal integration of both technologies (as opposed to the one-way integration of, for example, DLT into AI), and (c) goes beyond a purely conceptual level. In particular, we still lack a comprehensive overview of the most pressing research challenges that must be overcome in order to unleash the full potential of integrating both technologies. With this research, we aim to address this apparent knowledge gap by asking and answering the following two research questions: _RQ 1: What is the current state of research on the techno-_ _logical integration of AI and DLT?_ _RQ 2: What are open research challenges on the techno-_ _logical integration of AI and DLT?_ To address our research questions, we draw on the concept of convergence (see section II.C) and conduct a systematic literature review on the current state of research on the convergence (i.e., integration) of AI and DLT and develop a future research agenda. The contribution of our work is thereby fourfold. First, in contrast to extant research in this area, we also include non-blockchain distributed ledgers in our analysis of the literature. Prior work has highlighted several current shortcomings of blockchain in the application for AI [12]. Other DLT concepts (e.g., BlockDAG or TDAG), may be more promising for solving some of these shortcomings [3] and thus potentially better suited for certain AI applications. Second, we consider the reciprocal nature of convergence and investigate both perspectives: the usage of AI for DLT, and the usage of DLT for AI. To the best of our knowledge, no holistic review on the convergence of AI and DLT exists today, which considers the large variety of interaction cases from both perspectives. Third, we aim to bridge the gap between theory and practice by drawing theoretical conclusions from practical research, as well as outlining future potential for practical and theoretical research from the theory in these fields. Fourth, we apply the concept of convergence [14], [15] as a theoretical lens for our article. In doing so, we contribute to the understanding of how convergence may drive product innovations and create economic value in the information technology (IT) industry. In addition, we demonstrate how convergence can be applied as a lens to tackle research questions in interdisciplinary, innovative, and emerging research fields. The remainder of this article is organized as follows: In section 2, we discuss related work on AI and DLT, and introduce the concept of convergence. Afterward, we describe our methods in section 3. In section 4, we analyze the current literature on AI usage for DLT, and in chapter 5, provide our future research agenda on AI for DLT. In section 6, we analyze the current literature on DLT usage for AI, and a corresponding research agenda in section 7. In section 8, we discuss our results, before we conclude this article in section 9. **II. RELATED WORK** _A. ARTIFICIAL INTELLIGENCE_ AI enables computers to execute tasks that are easy for people to perform but difficult to describe formally. Such tasks typically occur in complex or uncertain environments [16]. Despite ongoing debates in society about Artificial General _Intelligence, which describes computer programs that can_ control themselves and solve tasks in a variety of different domains [17], most deployed AI-based systems solve tasks in narrow application domains, and are referred to as _Narrow Artificial Intelligence. Several approaches to design_ such narrow AI-based systems exist. For example, knowledge bases have seen a lot of attention by researchers in the past. Nowadays, Machine Learning (ML) seems to be the most well-spread approach toward building AI-based systems [16]. ----- **FIGURE 1. Overview of artificial intelligence.** ML-based systems consist of a model that represents a function between input data and output data. In most cases, ML models have to be trained. In this training phase, an optimization algorithm tweaks the model parameters in order to minimize a loss or maximize a reward. Depending on the application, different types of training exist. In the case of supervised machine learning, the input data and the corresponding output data are known during the training phase. In the case of unsupervised machine learning, only the input data is known but no output data. In a reinforcement learning setting, a learning agent executes actions that result in an immediate reward, whereas the agent’s goal is to maximize a future, cumulative reward. In general, the training phase can require large amounts of data and, thus, is often computationally intensive. This is especially the case for deep neural networks, which are complex ML models with many parameters that have paved the way for many of the recent advancements in ML [18]. In Figure 1, we present a high-level overview of different types of approaches toward designing AI-based systems. The execution of a (trained) ML model is called inference. It is usually computationally less expensive than the initial training phase. Some models can only be described as black boxes, meaning that their inner functionalities are difficult to explain. Among many research streams, some cutting-edge research in ML aims to better explain the inner functioning of ML models in order to guarantee their robustness [19]. Other streams aim to increase ML systems’ capabilities [20], [21], or to ensure training data confidentiality when creating ML-based systems [16]. Besides these ML approaches introduced above, many variations exist. For example, for some ML algorithms, there is no explicit model building phase at all [22]. For a detailed overview of AI and ML, in particular, we refer to Russell and Norvig [23]. _B. DISTRIBUTED LEDGER TECHNOLOGY_ DLT enables the operation of a highly available, appendonly, peer-to-peer database (i.e., the distributed ledger) in **FIGURE 2. Overview of distributed ledger technology and its** characteristics by Kannengießer et al. [3]. untrustworthy environments characterized by Byzantine failures, where separated storage devices (i.e., the nodes) maintain a local replication of the data stored on the ledger. A distributed ledger can either be deployed as a public ledger, if a new node can directly join a new network, or a private ledger, in case a node first needs permission to join the network. Another aspect to distinguish distributed ledgers are the write permissions: In the case of permissionless ledgers, all nodes have equal write permissions. In the case of permissioned ledgers, nodes first require granted permission to validate and commit new data to the ledger [3]. Today, several concepts for DLT exist with different characteristics, for example, regarding the transaction throughput or fault tolerance. The most widespread concept of DLT is blockchain [3], which became known to the public due to the emergence of the cryptocurrency Bitcoin [7]. Other concepts, for example, rely on directed acyclic graphs [3], [24], [25]. We provide an overview of DLT concepts with different characteristics in Figure 2. Nowadays, applications of DLT and especially blockchain beyond financial transactions are emerging, such as the management of medical data [26], autonomous driving [9], or decentralized games [27]. In a blockchain system such as Bitcoin [7], transactions of network participants are stored in blocks of a fixed size. Every block also includes a timestamp, and a hash of the previous block. Thus, the system can be regarded as a chain of blocks. Cryptographic techniques are used to ensure that only legitimate participants holding a cryptographic key can perform transactions, which are stored in the block. Bitcoin [7] was the first blockchain created and has mostly financial applications. The main networks of the widely known Bitcoin and Ethereum [28] are public, unpermissioned blockchains. To secure the network, only a selected node can propose a new block that includes the cryptographically signed transactions. This node has to find a block candidate with a hash below a certain, network-defined threshold. As this hash calculation is impossible to reverse, a node has to make large computational efforts to find that new block, competing against other nodes. In return, the node that successfully found a block gets a reward in cryptocurrency payment. Since there are at any time many miners aiming to find the next block, the chances of finding the next block are relatively low for an individual miner. As a result, the variance of the mining payoff for an individual miner is relatively large. Therefore, most ----- miners nowadays mine together in so-called mining pools. If a mining pool finds a block, the pool collective gets the block reward and distributes it among its miners according to their share of hash calculations. This reduces the payoff variance for individual miners participating in the pool. Besides Bitcoin, several implementations of blockchains exist with different characteristics [3]. Some implementations provide high transaction confidentiality guarantees (e.g., Zcash [29]), or enable a universally usable, decentralized Turing complete computing platform (e.g., Ethereum [28]). In the latter case, programs can be stored and executed upon the blockchain system. These programs are referred to as smart contracts. _C. CONVERGENCE OF AI AND DLT_ First described by Nathan Rosenberg in the 1960s, convergence describes a phenomenon in which two or more initially separate items move toward unity and become increasingly integrated with each other. Convergence typically occurs in four phases [15], [30]. The first phase, which is termed scientific convergence, occurs when distinct scientific fields begin citing each other, leading to an increase in cross-scientific research. It is followed by the technology convergence phase in which previously distinct technological fields increasingly overlap and new technology platforms arise. As technology convergence continues to blur existing technological and market boundaries, market convergence, the third phase, takes place, resulting in the emergence of new product-market combinations. In some cases, the convergence of technologies and markets may impact existing industries such that a fourth phase in form of the convergence of entire industries occurs. A typical and relatively recent example of convergence is the emergence of the smartphone, which nowadays combines initially separate technologies such as phones, cameras, and portable computing devices. It eventually created an entirely new market, completely transformed the mobile phone industry, and even overthrew the compact camera industry. Another example is Bitcoin [7], which was created by combining techniques from various computer science domains such as distributed systems, cryptography, security, and game theory. Concerning a potential convergence of DLT and AI, we currently see the emergence of the first scientific publications that do not simply apply AI in the context of DLT (e.g., deep learning for prediction of the Bitcoin price) or vice versa, but instead discuss a deeper integration of both technologies. Extant literature that provides a consolidated overview of the integration possibilities of AI and DLT, however, is scarce. Dinh and Thai [13] provide a viewpoint with conceptual ideas of how an integration of AI and blockchain technology might look like. They outline possibilities for the reciprocal integration of AI and DLT, highlighting that AI can be used for blockchain, just as blockchain can be used for AI. Specifically, AI can support blockchain systems by increasing their security and scalability, by acting as an automated referee and governance mechanism, and for privacy-preserving personalized systems. Blockchain, on the other hand, can serve AI systems by enabling decentralized computing, providing data sharing infrastructures, serving as a trail for explainable AI, or by coordinating untrusting devices. As the article is a viewpoint, the authors remain on a rather conceptual level and do not provide an in-depth review of extant literature. Salah et al. [12], on the other hand, provide a review and open research challenges on one of these two perspectives, the use of blockchain for AI. The authors identify enhanced data security, improved trust on robotic decisions, business process efficiency increases, collective decision making and decentralized intelligence as main drivers for the usage of blockchain for AI applications. Interestingly, the latter category has similarities with the application of privacy-preserving personalization, which Dinh and Thai [13] classify as an AI for blockchain perspective. Furthermore, Salah et al. [12] develop a taxonomy that provides an overview of different relevant technical characteristics, such as the consensus protocol or the blockchain type. Their review surveys research articles, as well as industry whitepapers. As such, it delivers valuable insights into how blockchain can be used for AI in vertical industry use case cases, such as banking, finance, or agriculture. The article concludes with several open research questions on technologies that are relevant in the context of blockchain for AI applications. These include topics of increasing blockchain systems’ scalability and data confidentiality, interoperability, or quantum cryptography. Lastly, Karafiloski and Mishev [31] provide an overview of how blockchain can be used for storing and organizing large amounts of data. Although the authors do not consider AI use cases in much detail, the presented work is representative for some fields of the convergence of AI and DLT, such as data sharing. In summary, prior research lacks an understanding how AI and DLT are reciprocally integrated in systems today, and how future research can advance the convergence of these technologies. Specifically, prior research focuses on blockchain, a specific DLT concept, and not on other potential DLT concepts that may be better suited in the context of AI [12], [13]. Furthermore, prior research only provides purely conceptual ideas how AI may be used for DLT, and does not evaluate today’s technical feasibility of such systems [13]. **III. METHODS** _A. DATA COLLECTION_ For the identification of articles addressing the convergence of AI and DLT, we systematically searched scientific databases. To cover a wide range of journal and conference publications, we queried IEEE Xplore, ACM Digital Library, AIS Electronic Library, Science Direct, Scopus, Ebsco Business Source, and ArXiv. The reason for including ArXiv preprints in our search is the fact that AI and DLT are fast moving fields of research, where new, potentially groundbreaking research results and insights may be available as preprints that are not finally published yet. Our search string required the publications to have a DLT-specific term and an ----- AI-specific term in either their title, abstract, or keywords. The search string, we employed was: _TIKEAB ((Blockchain OR „Distributed Ledger‘‘ OR DLT_ _OR Bitcoin OR Ethereum OR Cryptocurrency OR „Crypto_ _currency‘‘ OR „Crypto-currency‘‘ OR „block chain‘‘ OR_ _„Smart contract‘‘) AND (AI OR ML OR „Artificial Intelli-_ _gence‘‘ OR „Machine Learning‘‘ OR „Deep Learning‘‘ OR_ _Clustering OR Classification OR „Neural Network‘‘ OR „Big_ _data ‘‘OR „Data mining‘‘ OR „Intelligent system[∗]_ ‘‘OR „Sta_tistical model‘‘, OR „Statistic model‘‘))._ We excluded articles published before 2008, since the concept of Bitcoin emerged that year. For DLT, the search string included its most common concept of blockchain, as well as blockchain’s most frequent implementations (i.e., Bitcoin and Ethereum) and other technical terms. For AI, the search string included ML and the most common application forms of it. We searched on December 19th, 2019, this resulted in 2,411 unique articles. In a first step, we first removed 140 articles published before 2008 from our set, as not all of the databases provided the option to include this as a search criterion. Afterward, we analyzed the titles, abstracts, and keywords for the remaining articles and removed 17 that were not written in English. For the remaining 2,254 articles, we not only analyzed the titles, abstracts, and keywords, but also their full texts. From this set, we first removed 431 articles that turned out to not be research articles. Most of these articles were either the title of a conference proceeding or book chapters. We, then, checked the remaining 1,823 articles for whether they actually covered both, AI and DLT. 1,544 turned out to not cover AI and DLT, the largest group among these consists of 141 articles from the medical field where DLT is an abbreviation for other terms such as dose limiting toxicity or double lumen endobronchial tube. From the remaining 279 articles, we excluded further 213 articles that did not cover the close integration of these technologies according to the concept of convergence [14], [15]. Out of these, the largest group of 92 articles covered AI-based cryptocurrency price prediction or trading. In the remaining set of 66 articles, we excluded another 34 articles that did not clearly answer why they used DLT. This resulted in a set of 32 articles that eventually turned out to be relevant for further analysis. _B. DATA ANALYSIS_ Following the data collection, which resulted in the identification of 32 relevant articles, we categorized all 32 articles into groups. The groups were thereby derived from Dinh and Thai [13]’s viewpoint and adapted to our review, where necessary. Toward this end, we expanded the concept of blockchain to DLT, and reframed the focus and name of some categories to better suit the extant literature (e.g. secure DLTs instead of secure and scalable DLTs, and coordination of devices instead of coordination of untrusting devices). Table 1, below, provides an overview of the adapted coding scheme and the number of articles in each category. Note that we categorized some articles into multiple groups. Therefore, the sum of **TABLE 1. Classification of the identified articles that cover the** integration of AI and DLT into groups. articles in Table 1 is higher than 32. We also added another type of perspective, Both, to the coding scheme. It consists of consolidating works, such as literature reviews or high-level articles covering both aspects of the convergence of AI and DLT. Several of these articles contain concepts from multiple groups, especially articles covering DLT for AI. In Table 6 in the appendix, we present an overview of the 32 analyzed articles and their classification into different groups. Our future research agenda is an extension from our review and draws from multiple sources. On the one hand, the findings, outlook and conclusion of the extant articles discussed in our review. On the other hand, our own assessment of further, recent developments in literature in the fields of AI and DLT. **IV. REVIEW ON AI FOR DLT** Drawing on the general distinction between AI for DLT and DLT for AI proposed by Dinh and Thai [13], this section describes our findings in terms of how extant research has applied AI for DLT. We identified three different groups of use contexts, which we detail on below, and summarize our findings in Table 2. _A. SECURE DLTS_ 1) DLT PROTOCOL SECURITY Within this category, extant literature solely applies AI methods of reinforcement learning to explore and develop strategies in game-theoretic settings concerning the DLT protocol. Specifically, these articles analyze the fairness of mining activities. The learning agents get rewarded with a cryptocurrency-based miner reward. In the category of articles using reinforcement learning for DLT protocol security, two subcategories exist. _a: SELFISH MINING_ Articles in this subcategory analyze selfish mining strategies for blockchain systems [10], [32]. The learning agent ----- **TABLE 2. Overview of concepts in the literature on AI for DLT.** thereby is a blockchain miner, which can strategically delay the publication of found blocks. By doing so, the selfish miner aims to waste mining power of honest miners that work on another fork of the blockchain [33]. Through generating these attacks on blockchain systems in a testing environment using reinforcement learning, researchers can find new insights on the security of these blockchain protocols. For example, Hou et al. [10] performed the simulation with protocols such as Bitcoin [7], Ethereum [28], or GHOST [34]. In their research, an agent learns mining strategies where it can gain disproportionately large mining rewards relative to its hash rate. While some of these adversarial mining strategies are known from theoretical studies of the blockchain protocols [33], AI has the potential to detect new, previously unknown attacks in complex scenarios, for example, in cases with multiple partially cooperating agents. Additionally, the simulation may contribute to a better understanding of the feasibility of these mining strategies [10]. _b: POOL OPERATION_ In research within this subcategory, the agent learns strategies for operating a mining pool. In our review, we identified only one article that fits this subcategory: Haghighat and Shajari [35] specifically analyze a block withholding game, where a pool operator can either decide to mine honestly, or attack another pool by mining (at least with a fraction of its power) for the other pool, but not actually submitting a solution, in case a block is found. As a consequence, the revenue of the attacked pool drops, and it gets less attractive for other, honest miners to mine as a member of this attacked pool. The authors train an agent, which represents a pool operator, with reinforcement learning methods. It decides between mining honestly and attacking other pools at each step in the game. One insight from the results is that it may be more likely than widely assumed that one pool operator at a certain point in time was in control of more than 51 % of Bitcoin’s network mining power [35]. 2) SMART CONTRACT SECURITY Articles within this category aim to protect users from interacting with insecure or malicious smart contracts. Two subcategories with different approaches exist in literature: First, the analysis and detection of such smart contracts. Second, the active interaction with such smart contracts in order to manipulate and invalidate them. _a: DETECTION_ Three articles in this subcategory aim to detect security issues in smart contracts using data analysis techniques. As the smart contract opcode is usually published on the distributed ledger, two articles [37], [38] analyze the opcode with neural networks. Since a compiled opcode is challenging for humans to read and understand, Kim, et al. [37] use neural networks to estimate the functionality of a given smart contract (e.g., whether it is intended for a marketplace or for gaming). Tann, et al. [38], in contrast, aim at detecting security ----- vulnerabilities in smart contracts. To do so, they run long short-term memory neural networks on a sequence of opcodes. In a performance comparison with a symbolic analysis approach, the machine-learning-based analysis method can outperform the formal verification-based method with regards to classification accuracy and speed. Another article by Camino et al. [36] aims at detecting honeypot smart contracts. These types of smart contracts appear to contain free, withdrawable funds. However, once a user aims to invoke the honeypot contract to redeem free tokens, the honeypot actually does not release the funds, resulting in the user losing funds that they may have used to invoke the honeypot contract. Contrary to the other articles in this subcategory, Camino et al. [36] do not use opcode analysis techniques, but analyze the smart contract metadata (i.e., values derived from related transactions and fund flows) and off-chain data (i.e., the availability of a source code and the number of source code lines). In this way, data science techniques can classify more than 85 % of unseen smart contracts correctly. This approach especially works well in cases where there was a minimum of smart contract on-chain activity, and, as a result, on-chain metadata is available. _b: MANIPULATION_ The only article identified within this subcategory [39] uses an active reinforcement learning approach to invalidate criminal smart contracts. Criminal smart contracts can be used for illegal activities, such as selling confidential information [46]. In this research, the agent learns to manipulate the contracts’ data feed and thereby successfully invalidates a substantial share of the studied criminal smart contracts. While doing so, the system aims at invalidating these given smart contracts, but not at detecting possible criminal smart contracts in the first place. _B. AUTOMATED REFEREE AND GOVERNANCE_ Another small group of articles uses AI for DLT-based system governance. Dinh and Thai [13] present the vision that people, devices, and smart contracts record transactions on a distributed ledger. An AI can then solve potential disputes of events happening on- or off-chain, and record the results on a distributed ledger. This automated arbitration could be data driven, unbiased, and, as a result, more consistent, justified, and accepted than arbitrations today [47]. In extant literature, articles toward this vision appear to be in an early stage and fall within two categories: AI-based DLT protocol governance, and AI-based smart contract governance. 1) DLT PROTOCOL GOVERNANCE This category currently consists of two articles. Lundbæk et al. [40] propose an adjusted proof-of-work-based consensus mechanism. Only a subset of nodes participates in the required hash calculation and the DLT protocol uses ML to regularly update system governance parameters, such as the ideal number of miners or the level of mining difficulty. In this scenario, AI is tightly integrated with the DLT protocol. However, the authors do not discuss in detail the functionality and security aspects of their ML-based governance. Gladden [41], on the other hand, discusses ethical aspects about cryptocurrencies governed by an AI system. The author claims that such a system can positively influence the ethos and values of societies. However, the article has a sociotechnical focus and does not describe the technical architecture of the system. 2) SMART CONTRACT GOVERNANCE Only one article in our review fits this category. Liu et al. [42] propose to implement voting mechanisms for participants in smart contracts to alter smart contract parameters. As such, these voting mechanisms can govern smart contracts in complex situations and can alter the smart contracts towards an adaptive behavior. An ML-based system assists the users in voting, based on their past voting behavior. As a result, the ML-based system pre-chooses the users’ selections and eases their tasks. In this scenario, the ML-based system runs outside of the actual DLT system. _C. PRIVACY-PRESERVING PERSONALIZATION_ Many internet platforms nowadays collect data about their users and apply AI-based recommender systems to personalize their content. Examples include Facebook, Netflix, or Taobao in China. A recommendation by this AI-based system thereby is not only influenced by the individual user’s data, but also by all the other users’ data. Netflix, for example, can recommend new movies to users based on their individual and other users’ watching behavior. However, this comes at the risk of impeding users’ privacy. In several cases, private data from such platforms has been leaked to the public [48] or misused [49]. A group of articles envisions AI-based personalization for DLT-based data sharing platforms. This could, for example, include a social network built on a DLT infrastructure [13]. DLT can thereby serve as a transparent trail of data flows, and give users control over their data. Two categories of articles with different approaches toward designing such systems exist in literature. First, a category with articles that aim to use local computation. Second, a category with articles that use distributed ledgers based on hardware-assisted Trusted Execution Environments (TEEs). 1) LOCAL COMPUTATION An ML model inference intended to personalize content on a platform requires data about the user to personalize its recommendation for them. If such a model inference is executed locally, there is no need for the user to share their data, while they can still get a personalized recommendation. However, if no user shares their data with the platform, it is challenging for a platform operator to train ML models using traditional methods. Therefore, articles in this category describe systems that use federated learning. With this distributed ML technique, an ML model gets trained locally on a user’s device and only ML model parameter updates leave the device and are shared with the platform or potentially other ----- parties. Model parameter updates from many users are then aggregated into a new model. Federated learning is already successfully applied on a large scale on smartphones, for example, to predict the next word a user may want to type on their keyboard [50]. The distributed ledger provides an infrastructure to share data on an auditable and immutable ledger. Some articles describe systems that only store hashes of model gradient updates or hashes of aggregated models on the distributed ledger, in order to save storage space on the ledger and to preserve data confidentiality [51]. Other articles describe systems that store the complete gradient updates and aggregated models themselves on the ledger [52]. First applications are illustrated for the Internet of Things [43] or the taxi industry [44]. 2) TRUSTED EXECUTION ENVIRONMENTS The second category uses TEE-based DLTs. TEEs are typically located on a CPU and provide a special enclave for secure computation. In this enclave, applications are executed such that other applications executed on the same CPU but outside the enclave cannot interfere with the state or control flow of the application shielded by the TEE. As a result, the enclave appears as a black box to the outside and secures the application running inside. Furthermore, the hardware of the TEE can generate a proof to the application running inside the TEE to attest that the software is executing on this specific, trusted hardware. This feature is called remote attestation. Extant research uses DLT protocols to coordinate TEEs, thus enabling confidential smart contracts that execute inside the TEE. Since the TEE ensures data confidentiality, this category does not necessarily require local computation. First articles aim at providing solutions for the medical industry [11], [45]. To further increase data security, some articles describe systems that combine TEEs with federated learning [45] or differential privacy [11]. Simplified, differential privacy is a mechanism that adds randomly distributed noise to a set of data points. This protects the information privacy of users that provide data. As the random noise has zero mean (or a predefined mean), an aggregator can still draw meaningful conclusions from the aggregated data set without threatening anyone’s information privacy. **V. FUTURE RESEARCH AGENDA ON AI FOR DLT** In this section, we present our analysis of future research opportunities on the advancement of DLT using AI. Again, we use the categories proposed by Dinh and Thai [13] to structure the future research agenda. We summarize our findings in Table 3. _A. SECURE DLTS_ Dinh and Thai [13] have drawn a futuristic picture with farreaching real-time analysis and decision possibilities of an AI as part of the DLT system. Current AI-based systems, however, do not provide the required security and robustness guarantees necessary to govern a DLT system. Precisely, while AI-based systems can detect software vulnerabilities, they cannot (yet) guarantee that all available security vulnerabilities have been identified. If an AI detects no vulnerability, this does not necessarily mean that there are none [10]. This open research problem of AI robustness is also highly relevant in other AI application domains, such as autonomous driving, where first results may be transferable to software and DLT systems’ security [53]. In general, methods from the field of explainable AI (XAI) [19] could help to better understand how reliable an AI-based decision is. We, therefore, see this as a fruitful field for future, foundational research. While AI is a promising technology to detect DLT security vulnerabilities, further research that aims to resolve these vulnerabilities and to build secure DLT-based systems is required in parallel [35]. Subsequently, we present future research opportunities in the two categories identified in our review in section IV. Furthermore, we see possible research opportunities in the convergence of these two categories, which we also present. 1) DLT PROTOCOL SECURITY The security analysis of DLT protocols using AI is currently dominated by reinforcement learning. Extant literature analyzes game-theoretic incentive mechanisms in widely adopted protocols such as Bitcoin or Ethereum. Possible future research opportunities include the expansion of previous work with settings such as partially cooperating miners [10], or the analysis of strategies other than selfish mining [32]. In our view, beyond analyzing the fairness of mining settings and its DLT security implications, another additional interesting field of AI-based security analysis is the DLT source code itself. This is a highly relevant field in practice. For example, the deployed and widely-used cryptocurrency Zcash, which uses a novel form of zero knowledge proof cryptography, has been subject to a vulnerability that would have allowed an attacker to generate an infinite amount of cryptocurrency tokens [54], thus potentially inflating the assets of other users. Extant literature already uses AI-based analysis methods for software bug identification in general [55], and smart contract opcode analysis in particular [37], [38], thereby often outperforming formal verification-based methods [38]. We, therefore, see a promising future research avenue in further developing and applying such AI methods for analyzing DLT protocol code security. 2) SMART CONTRACT SECURITY Extant literature in this subcategory mostly aims to protect users from insecure [37], [38] or malicious [36], [39] smart contracts. Practical experience, however, has shown that developers struggle to develop secure smart contracts in the first place [56]. Therefore, an interesting perspective for future research would be in the field of using AI to assist developers in developing secure smart contracts. This could include the development of early security warning tools for developers that automatically check developed code for security vulnerabilities. Recent work in other fields of software engineering suggests that such AI-based systems may be feasible [55], [57]. ----- **TABLE 3. Overview of future research opportunities in the field of AI for DLT.** Beyond this suggestion to expand the usage perspective of future research, we see further research opportunities in refining and applying the methods identified in our review. Reinforcement learning, which has already shown promising results for DLT protocol security analysis in simulation settings [10], [32], [35], appears as a promising method to analyze the security of game-theoretically complex smart contracts [39] in simulations. This could include, for example, decentralized token exchanges. The other identified approach is to analyze security vulnerabilities in smart contracts by supervised learning. This appears as a promising field of AI application, as AI-based methods have outperformed classical, formal verificationbased methods [38]. Future research could expand this work by considering further classes of smart contracts [36] and further details, such as different smart contract compiler versions [37]. Furthermore, future research that seeks to support software developers could aim at not only detecting security vulnerabilities in a smart contract, but also at localizing this vulnerability (e.g., by expressing the information which portions of the bytecode cause the vulnerabilities [38]) or even at providing suggestions to fix it [60], [61]. 3) INTERACTION OF DLT PROTOCOLS WITH SMART CONTRACTS AND SECURITY IMPLICATIONS In addition, we see a convergence of these two categories— AI for DLT protocol security, and AI for smart contract security—as a promising avenue for future research. Recent research has investigated the extent to which game-theoretical aspects in DLT protocols and smart contracts influence each other and can lead to unfair conditions for regular users of the overall system [58]. For example, miners who may participate in the trading of assets on blockchain-based decentralized exchanges have the sovereignty to decide which transaction is included in a block and which one is not. This can incentivize other users to pay higher transaction costs for these exchange transactions than they would without the miner participation. Prior research considers this as an unfair setting [58]. The outlined scenario is only one example of the complex interplay of smart contracts and DLT protocol incentives, which is also relevant in decentralized applications other than token exchanges [62], [63]. As it appears to be an even more complex topic than the standalone security of DLT protocols or smart contracts, an AI-based system could provide valuable insights. Prior research uses reinforcement learning for gametheoretic analyses of DLT protocols and smart contracts separately, therefore, reinforcement learning also appears as a promising method to analyze the interplay between both. The results could be used to support the development of fair and secure DLT-based systems. _B. AUTOMATED REFEREE AND GOVERNANCE_ In general, both identified categories—AI for smart contract and for DLT protocol governance—face the reality of ----- current AI’s capabilities. To automatically make far-reaching decisions on a distributed ledger with a potential influence on financial or data confidentiality circumstances, most of today’s AI’s robustness and explainability guarantees are not strong enough. The robustness guarantees of many modern AI-based systems are weak [64] compared to the complexity of the actions that agents perform when interacting with DLT-based systems. Therefore, in our view, breakthroughs in the robustness, explainability, and ultimately security of AI-based systems are required before they can automatically govern DLT protocols and smart contracts on a large scale. However, in the case of humanin-the-loop based AI smart contract governance [42], an AI may assist a human decision maker which can intervene at any time in case an AI-based system faces and detects an irregular situation. This human-in-the-loop model, therefore, appears to provide good practicality with today’s AI technology and appears to be a good starting point for future research [42]. _C. PRIVACY-PRESERVING PERSONALIZATION_ DLT-based platforms for data sharing are starting to get more attention in research but are not deployed and scaled in practice yet. As such, research on such platforms and their AI-based personalization is still in an early stage. Articles in the group of privacy-preserving personalization using AI for DLT often incorporate aspects from multiple other groups, such as decentralized computing for AI, or secure data sharing and marketplaces for AI. In our view, developments in the field of privacy-preserving per_sonalization depend on advancements in the underlying_ secure computation and privacy-preserving data sharing technologies. An interesting future research opportunity which is practical with today’s technologies is the further development and deployment of DLT-based data sharing platforms and an AI deployment on top of these platforms. TEEs appear as a promising technology to realistically build such systems today, due to their relatively small computational overhead when compared with other secure computation technologies, as well as their ability to enforce policies [65]. TEEs can be combined with federated learning [45] or differential privacy techniques [11] for stronger data confidentiality guarantees. A reasonable avenue for deployment appears to be within consortia of (at least partially) trusting participants. This could, for example, include a consortium of hospitals aiming to deploy personalized, AI-based treatments while complying with strong privacy requirements [45]. **VI. REVIEW ON DLT FOR AI** Drawing, again, on the general distinction between AI for DLT and DLT for AI proposed by Dinh and Thai [13], this section describes our findings in terms of how extant research has applied DLT for AI. We identified four different groups of use contexts and summarize our findings in Table 4. _A. DECENTRALIZED COMPUTING FOR AI_ Due to the large amount of processed data, ML model training is often computationally expensive. Nowadays, graphics processing units (GPUs) are typically used to train ML models, as they provide more computational power for most ML training tasks. For complex optimization tasks, many GPUs may be necessary to train ML models in a reasonable amount of time [66]. However, many GPUs and CPUs in computers around the world are only slightly loaded or even unused. Previously, distributed computing has been applied to utilize these unused resources in communities at scale, for example, to perform protein folding calculations for medical research [67]. Articles within this group describe systems that use distributed computation methods to train AI models. DLT can serve as a trail to organize this decentralized computation, as well as to provide a ledger for rewards paid in cryptocurrency [13]. Looking at the literature, three different approaches for DLT-enabled distributed computing for AI exist: First, the computation within the DLT protocol. Second, the computation in smart contracts. Third, the computation outside the distributed ledger. 1) DLT PROTOCOL Especially public, permissionless distributed ledgers, such as the Bitcoin or Ethereum main network, are nowadays secured by a proof-of-work-based consensus mechanism. It refers to the search for a number such that the hash of a blockchain block candidate, which includes that number, is below a certain threshold. This mechanism is required to limit the number of nodes that can propose a new block, and ultimately, to ensure blockchain security. By adjusting the threshold, the network can adjust the difficulty of this search problem, and adjust the average time to find a new block. The search problem is computationally expensive. However, the verification of a potential solution is computationally inexpensive. The proof-of-work mechanism consumes large amounts of energy for protocols such as Bitcoin, which some authors see critical [68]. At the same time, the calculated hashes cannot be meaningfully used for purposes other than securing the network. Some authors therefore propose a proof-ofuseful-work mechanism, in which miners train an ML model instead of finding a number for a certain block hash. The block which contains an ML model with the least test error is then accepted as the new block by the other nodes [69]. For such a system to work, several challenges have to be overcome. On the one hand, the computational difficulty of optimizing an ML model is hard to adjust [70]. This can also cause blockchain forks to occur on a very regular basis. On the other hand, the proof-of-useful-work mechanism for AI requires the secure provision of an ML model architecture, training data, and test data. Extant literature that describes such systems uses selected groups of participants and committees with reputation mechanisms to govern this data provision [69]–[71]. ----- **TABLE 4. Overview of concepts in the literature on DLT for AI.** 2) COMPUTATION IN SMART CONTRACTS In some DLT networks, such as Ethereum, a virtual machine can execute Turing-complete smart contracts. As such, DLT could provide a substrate for ML model training [72]. However, traditional ledgers like Ethereum do not support intensive computations that would be necessary to perform ML model training [74]. To overcome this limitation, some articles propose to extend distributed ledger smart contract execution with TEEs such that these ledgers can support computationally more intensive smart contracts. This TEE can then train simple ML models [11], [59]. 3) OFF-CHAIN COMPUTATION Another subcategory of articles performs computations not on the distributed ledger, but off-chain. Unless using sophisticated cryptographic techniques [79], the integrity of the computation cannot be guaranteed. DLT-based federated learning systems, for example, train an ML model on the edge [43], [44], [52], [73]. The main motivation of most articles focusing on DLT-based federated learning, however, is not the usage of decentralized edge computing resources per se. Instead, the main motivation is secure data sharing. Some of the DLT-based federated learning articles also propose financial rewards for both in combination, model training and data sharing [80], [81]. Besides federated learning, another article by Lu et al. [74] proposes a crowdsourced computation approach to offload heavy computation tasks from a blockchain, such as in ML model training. Multiple offloaded computing engines audit each other’s work and game-theoretic incentive mechanisms are used to build a protocol. The authors also present a security analysis of their protocol. _B. SECURE DATA SHARING AND MARKETPLACES FOR AI_ In addition to the increasingly available computing power, another fundamental reason for the recent advancement in ----- AI is the strong growth of available and digitized data. ML-based systems generally perform better the more data they are trained on, for example, with regards to classification accuracy [21], [82]. Some authors see recent developments in IT, in which few companies are in control of large amounts of personal data, critical and propose DLT-based data markets to democratize such data silos [13], [83]. As such, a group of articles proposes solutions based on DLT to build data sharing infrastructures, thus enabling the deployment of AI. These articles differ in their technical approaches toward designing such systems. Articles which incorporate advanced privacy-preserving mechanisms in their systems mostly focus on the health care industry. We suspect the reasons for this being strong confidentiality requirements for the handling of health data. 1) SMART CONTRACTS Two identified articles use traditional smart contracts on Ethereum to build a data sharing infrastructure and marketplace. These articles differ in their focus, either on sharing training data itself [75], or ML models [76]. Especially the sharing of training data requires the storage and handling of large files. Therefore, Özyilmaz et al. [75] connect their system to the decentralized file protocol SWARM and manage data access rights through the blockchain. These marketplaces allow the payment of data providers with units of the cryptocurrency Ether. To incentivize participation with high quality data, Harris and Waggoner [76] propose staking mechanisms in which malicious participants sharing spam models lose their stake. Such systems can be applied in the Internet of Things industry in general [75]. 2) TRUSTED EXECUTION ENVIRONMENTS TEEs enable the computation of relatively intensive tasks, while preserving data confidentiality and integrity throughout the computation. Therefore, systems described in some articles execute computationally intensive smart contracts offchain in a TEE [11], [45], [59], [84], [85]. These articles use the blockchain concept with different DLT designs [3]. Several articles use the Oasis blockchain [11], [84], [85] Hynes et al. [11], for example, present a privacy-preserving data market that provides solutions for a large share of the ML pipeline. Through smart contracts, data providers can define policies to share their data. These policies, for example, include the asking for a reward and differential privacy requirements. Data consumers can choose to fulfill these policies in order to train an ML model on the providers’ data. Since a TEE ensures confidential computation, the training data does not get leaked and only data consumers can get access to ML model inference. The ML model itself is shielded inside a smart contract and inference executions count toward the provider’s policies, which increases data provider’s privacy against potential inference attacks. This type of attacks aims at executing the ML model in order to extract the underlying training data or the model itself [86]. Other articles, which describe a system with similar capabilities, use an Ethereum-based blockchain [45], [59]. As a special feature, a virtual machine in the enclave can allow the training of proprietary AI models while not comprising the training data security [59]. PasseratPalmbach, et al. [45] incorporate federated learning into the TEE-enabled blockchain. This protects data providers from potential TEE side channel attacks and, therefore, further reduces data privacy risks. 3) FEDERATED LEARNING A large subcategory of articles describes systems that use federated learning with distributed ledgers that are not TEEenabled [43], [44], [51], [52], [81]. DLT thereby serves as a provenance record of data. This data can describe a large share of the ML pipeline —training data origin, training data, ML model modifications, or testing data [51]. In most systems described in these articles, the data itself is not stored on the blockchain, but hashes of the data [51]. In some cases, the systems use relatively simple ML models and store plain model updates on the blockchain [52]. As a result, every participating server can audit and compute the aggregated ML model weight updates. A blockchain-based solution can be well-suited for certain use cases and only provides a small performance overhead of ca. 5 % to 15 %, while enabling transparency and accountability [52]. Some systems go further and replace the traditionally centralized aggregator with a smart contract-based one [44]. As federated learning is potentially vulnerable to inference attacks [86], some articles describe systems that further use differential privacy techniques [43], [73] to increase data confidentiality guarantees. Further systems incorporate financial incentives for participants sharing model updates [43], [81]. _C. EXPLAINABLE AI_ Complex ML models, such as deep neural networks, are nowadays often used in a black box manner [19]. This means that users or even system creators do not have the information how these models come to a certain prediction. However, obtaining this information can be desirable in certain cases, for example, to verify the system’s robustness or to comply with legislation [19]. Dinh and Thai [13] outline DLT as a technology to increase AI-based systems’ explainability. Their vision is that DLT provides an immutable trail to track the data flow of AI-based systems. Looking into the extant literature on DLT for AI, the explainability of the black-box model itself is out of focus for most researchers. The DLT for XAI literature mainly covers data provenance or computational integrity aspects for model training or inference. Sarpatwar et al. [51] design a DLT-based federated learning system for trusted AI and present five requirements of blockchain for trusted AI. First, guarantees of AI model ownership and track of use are important. Second, the confidential sharing of AI assets as they are often created using sensitive data. Third, auditability regarding the AI training process. Fourth, the traceability of ----- the AI pipeline. Fifth, to keep a record in order to recognize a potential bias and model fairness issues. In general, several extant articles aim to use DLT-based federated learning to ensure such aspects of AI model explainability and trustworthiness [51], [52], [76]. Other articles use TEEs to also cover the model inference with their system, and thus, provide stronger explainability and restrictions in order to track data flows for data providers [11], [45], [59]. Hynes et al. [11], for example, use XAI methods to compensate data providers relative to their influence their data had on an ML model’s inference. _D. COORDINATION OF DEVICES_ A final group of articles aims to coordinate devices with DLT. These devices, such as Internet of Things devices, generate data, which AI-based systems can analyze. A distributed ledger can serve as a registrar for these devices, and store different types of data that these devices generate. This includes, for example, metadata (such as hashes), or the generated data itself. In extant literature, articles that use DLT also connect devices to the distributed ledger. Using cryptographic principles, transactions on a distributed ledger (such as smart contracts) are signed, which ensures that only legitimate participants can perform certain transactions. Kang et al. [77] propose to use DLT for reputation management in a federated learning setting. Beyond identification through asymmetric cryptographic mechanisms on a distributed ledger, TEEs provide mechanisms to self-identify and remote-attest [78], such as physically unclonable functions [87]. **VII. FUTURE RESEARCH AGENDA ON DLT FOR AI** In this section, we present our analysis on open, future research fields on DLT for AI. Again, we use the categories based on Dinh and Thai [13] to structure the future research agenda. We present an overview of our results in Table 5. _A. DECENTRALIZED COMPUTING FOR AI_ As we identified three categories of approaches using DLT for decentralized computing for AI, the future research opportunities in these approaches differ. Subsequently, we present our identified research opportunities in all of these. 1) DLT PROTOCOL First, extant research presents blockchain designs with proofof-useful-work mechanisms that could be used for ML model training. Proof-of-useful-work itself has a long history with several blockchain-based cryptocurrency systems deployed [88], for example, for number theoretic research [89]. While this prior work is interesting from an academic point of view, the deployed cryptocurrencies so far had little practical relevance based on their market capitalization. Some blockchain designs, such as the public Ethereum mainnet, actually aim to not use consensus mechanisms based on proof-of-work in the future [90]. Despite that, we see promising avenues for future research on proof-of-useful-work for AI, as well as for actually deploying and evaluating such systems. On the one hand, future research could analyze the economics of such systems and their utility for ML model requesters as well as regular DLT users. On the other hand, future research could analyze the practical security guarantees of such systems. Aspects include the security of incentive mechanisms, as well as the security of DLT transactions. This is particularly interesting, since an attacker with large computational power would not only benefit from rearranging transactions on the ledger, but also potentially from the useful work itself. 2) COMPUTATION IN SMART CONTRACTS Extant research has aimed to compute smart contracts off-chain in TEEs. Out of several approaches to off-chain computation while ensuring integrity [79], TEEs are interesting due to the relatively high computational power they provide [91]. TEEs have seen prior research from both fields separately, AI and DLT. In the field of AI, extant research provides frameworks for ML model training and inference on TEEs [92]. It has shown that TEEs are in general capable of simple practical machine learning tasks, such as speech processing [93], or the processing of small images [91]. In the field of DLT, researchers have sought to increase the low performance of smart contracts using TEEs while ensuring confidentiality [78]. From our point of view, TEEs generally provide a promising trade-off between computational performance and confidentiality. Prior research that provides an ML pipeline with TEE-enabled DLTs does not evaluate its performance [11], [45], [59]. As such, a natural extension would be a practical performance evaluation. This could help researchers aiming to deploy such systems, for example, for privacy-preserving personalization. By knowing what amounts of data and what complexity of computation the system can handle, other researchers could select use cases and deploy and evaluate TEE-enabled DLTs for their use cases. Another avenue for future research is the further development of computationally powerful TEEs, and to potentially even enable TEEs on GPUs [106], which are, for many ML model training tasks, better suited than CPUs. At the same time, research has identified security vulnerabilities in TEEs and potential security measures [94]. Therefore, future research could aim to develop secure TEEs and mechanisms to prevent attacks on TEEs, as well as study the feasibility and implications of TEEs on distributed ledgers. To prevent TEE security vulnerabilities, some researchers combine them with other privacy-enabling technologies, such as federated learning [45] or differential privacy [11], [91]. The further analysis of such combined methods with regards to data confidentiality guarantees, computational overhead, impact on machine learning quality, and practicality for use cases would be another promising avenue for further research. In addition, TEEs with a strong integration into a blockchain protocol enable new mechanisms for block selection. These mechanisms use special functions of TEEs, such as the secure generation of random numbers, and are known ----- **TABLE 5. Overview of future research opportunities in the field of DLT for AI.** as proof of luck [96] or proof-of-elapsed-time [97]. Future research could aim to further integrate such mechanisms into the DLT protocol and evaluate the practical benefits and system security implications. 3) OFF-CHAIN COMPUTATION Most presented DLT-based federated learning protocols do not ensure the integrity of the model training calculations. As such, we see future research potential for such systems in at least partially trusting consortia. Federated learning is already successfully applied in systems with a large number of users [50], therefore, future applied research could provide further promising results when combined with DLT. A further avenue for future research is the use of gametheoretic mechanisms to incentivize honest computation, for example, for ML model training [74]. In general, there is a lot of current research that aims to scale smart contract executions outside a distributed ledger while maintaining integrity [98]–[102]. Much of this research is at the conceptual level and does not yet deal in detail with the application of concrete computationally intensive applications, such as the training of ML models. In this regard, the advancement of cryptographic technologies for secure computation is another avenue for future research. Homomorphic encryption, for example, enables data consumers to perform computations on encrypted data. ----- However, this technology and other technologies (such as secure multiparty computation or zero knowledge proofs) are not yet practicable for ML model computations [11], [78], [84], [91] due to the strong computational overhead. If future research achieves breakthroughs to enable intensive computations with these cryptographic technologies, they would provide an alternative to TEEs. _B. SECURE DATA SHARING AND_ _MARKETPLACES FOR AI_ Articles within this field already present systems that could be further developed and deployed for practical use. Future research could, therefore, evaluate the practicality and user acceptance of such systems. For this practical deployment, health care use cases may be particularly suitable, because the deployment of ML in this context is especially challenging due to high privacy and security requirements [11], [84], [85]. Beyond the deployment and evaluation, a further avenue for future research is the rigorous security analysis of such systems against different types of adversaries. Potential adversaries could, for example, aim to get rewards for sharing spam data, aim to extract training data by training a proprietary ML model, or seek to extract training data of a shared ML model through inference attacks or the shared ML model itself. Only some articles of in our body of literature study such attacks and security measures to protect the system [11], [52], [59]. Especially staking mechanisms have seen little attention from researchers so far [76]. Future research could possibly transfer insights from proof-of-stake-based DLT consensus protocols [107] toward staking in DLT-based data marketplaces. We see differential privacy as a promising technique for ensuring users’ privacy [43], [73], especially in large datasets with low dimensionality. In such cases, its negative effect on the aggregated model quality is small while improving the individual data sharer’s privacy. Previous research on differential privacy for ML studies its theoretical implications on test data sets [16]. In our view, future research could build on these and study differential privacy’s practicality in certain vertical use cases for both, classic ML (relevant in the context of DLT; e.g., through TEEs), and federated learning. Federated learning, for example, can be vulnerable to inference attacks [86]. Differential privacy can help impeding the practical feasibility of such attacks [103]. Articles describing DLT-based federated learning use DLT as an immutable trail and, in some cases, even for the data communication and storage. Furthermore, DLT serves as a ledger for the reward payment using cryptocurrencies. We, therefore, see plain federated learning approaches as particularly promising for applications in (at least partially) trusting consortia. TEE-based approaches, however, not only use DLT for the same reasons, but go further. Smart contracts on a distributed ledger can enforce policies for both, data providers and data consumers. From this point of view, the TEE-based approach is more powerful in terms of covering a large share of the ML pipeline, and ultimately, in ensuring system security. Accordingly, it may have better utility for future research that aims to deploy DLT-based secure data sharing and marketplace systems. Nevertheless, TEE implementations have a rich history of security vulnerabilities [108]. A detailed security analysis of the practical security of TEEs and their implications for TEE-based DLT systems, possibly with further security mechanisms in smart contracts could provide a valuable contribution to future research. Such an analysis could also include combined security measures in smart contracts, like federated learning [45], differential privacy [11], or secure multiparty computation [11]. We see such combinations of privacy techniques [65], [84] with TEEs as an interesting avenue for future research, although some of these cryptography-based privacy techniques need breakthroughs in practicability to be usable for ML [84]. _C. EXPLAINABLE AI_ In our view, DLT is well-suited as a trail for AI model metadata and hashes of data generated around the AI model training and inference phase. As such, DLT can help to enable XAI applications that aim to describe the black box behavior of AI models for both, off-chain, as well as on-chain [11] applications. This can increase an AI-based system’s security and, ultimately, enable the deployment of trustworthy AI [51], [104]. In many DLT-based federated learning articles, only AI model hashes and metadata [51] or relatively small AI models [52] are stored on the distributed ledger. TEE-based systems, on the other hand, can securely write and read external memory while preserving data confidentiality [59]. Thus, we see TEE-based systems [11], [45], [59] as specifically suited for large, data intensive ML environments. Future research could aim to deploy such systems and evaluate the practical explainability of the AI model. _D. COORDINATION OF DEVICES_ One of the core functions of DLT is to provide an immutable ledger where only cryptographically legitimate participants can perform transactions. The usage of DLT for coordinating devices and participants is therefore essential and basic at the same time. Reputation management of participants on a distributed ledger is one field that has only started to see attention and appears to be of potential for future research. Kang et al. [77] specifically mention a dynamic system with variable thresholds as a future research opportunity. With regard to self-identifying TEEs, prior research already uses DLT to coordinate devices [78], [109]. We consider the further dissemination and use of self-identifying functionalities through physically unclonable functions [105] as an interesting future research opportunity. Such selfidentification functionalities could potentially be built in Internet of Things devices, such as sensors or actuators. However, mechanisms for hardware-based authentication have been identified as insecure in the past [110], [111]. Therefore, we see an analysis of their practical security guarantees as another open research problem. ----- **VIII. DISCUSSION** _A. PRINCIPAL FINDINGS_ Both fields, AI and DLT, currently experience a lot of hype within research and practice. Even though our initial database query listed a few articles that did not clearly explain what AI or especially what DLT was used for, a substantial number of extant articles provided profound insights for our review and future research agenda. One analysis aspect in our work is that we do not limit our work to blockchain as the only DLT concept, but also consider other concepts such as directed acyclic graphs. This decision was based on the expectation that other DLT concepts with distinct characteristics might be better suited for some AI applications than the concept of blockchain [3], [12]. Yet, only one out of 32 articles within our review considered DLT concepts other than blockchain in the context of AI [75]. This result is in line with previous research, which has called for more research on other DLT concepts [3]. Improving and deploying DLT concepts other than blockchain is, therefore, a highly interesting and relevant avenue for future research. Overall, the framework provided by Dinh and Thai [13] served as a helpful tool to classify extant literature with regard to the convergence of AI and DLT. However, several articles cover aspects from multiple groups in the subsections of AI for DLT or DLT for AI. This is particularly the case for articles in our review that cover multiple subsections of DLT for AI or the subsection privacy-preserving personalization from the section AI for DLT. We have slightly modified the framework of Dinh and Thai [13] in two ways. First, extant literature applies AI for DLT in the context of secure distributed ledgers, but not for scalable distributed ledgers as well. A possible reason for this is the lack of AI’s robustness and security guarantees, as discussed in our future research agenda. Second, we renamed the subsection coordination of _untrusting devices to coordination of devices, because DLT_ is not the element that establishes trust in all of the articles. For example, TEEs can use physically unclonable functions for remote attestation [87]. In our view, some of the research fields are mature enough to transfer systems into practice and evaluate their influence and user acceptance. This includes, for example, the fields of security analysis of smart contracts or marketplace systems based on TEE-enabled DLTs. Research particularly focus on the application of these marketplaces in the health care industry [11], [45], [59], [84], possibly due to the presence of data lakes and strong data confidentiality requirements. Therefore, further research-based deployment and evaluation of DLTbased marketplaces for AI in health care settings may deliver a promising contribution to their real-world deployment. At the same time, other research fields, such as AI-based automated referee and governance for DLT protocols, appear to require substantial progress in fundamental research (e.g., robust and secure AI), before transferring scientific knowledge into practice and the establishing of real-world systems. By applying convergence as a theoretical lens on extant literature, we were able to focus our research on innovative articles that closely integrate AI and DLT. Furthermore, we were able to exclude research that does not closely integrate AI and DLT, such as AI-based cryptocurrency price prediction or trading. Drawing on the definition of convergence [30], many articles on AI and DLT’s convergence fall into the first phase with cross-scientific research on their integration. Some articles already pave the way for the second phase with new platforms arising that could, for example, accelerate health care research [11], [45], [84]. During our review, we noticed that the concept of convergence has received relatively little attention by IT researchers in the past. This came to our surprise, as convergence has been a main driver of IT innovations over the recent years [14]. Therefore, we consider convergence as a promising theoretical lens to explore interdisciplinary technological settings. _B. LIMITATIONS_ Both fields, AI and DLT, are moving very fast and breakthroughs are regularly achieved. In this respect, we cannot rule out the possibility that in some subfields of the convergence of AI and DLT, future research will achieve innovative breakthroughs that may enable use cases not identified in our review or future research agenda. We have taken several steps to minimize the chances of this outcome. First, our analysis includes ArXiv preprints of research articles that may only soon be published in journals or presented at conferences. Second, we included insights from foundational research on the topics of AI, DLT, and secure computation into our future research agenda. In doing so, we sought to consider aspects that may not be covered in the reviewed body of literature but are nevertheless highly relevant in the individual fields (e.g., the practical security guarantees of TEEs). Third, we have also included articles that cover little researched technologies, often with little practicality today, but that may see breakthroughs in the future. This includes Artificial General Intelligence, DLT which is not based on the concept of blockchain, and cryptographic protocols for computationally intensive tasks. **IX. CONCLUSION** In this research, we investigated the current research status and future research opportunities on the convergence of AI and DLT. In order to assess the current state of convergence, we conducted a systematic literature review and analyzed extant literature through the lens of convergence. Our findings include several different ways that describe how AI can advance DLT applications, as well as how DLT can advance AI applications. In order to develop a future research agenda, we built on the structure of our literature review and linked the ongoing research with other research in the separate fields of AI and DLT, as well as our own view on future research opportunities. Our results reveal multiple future research opportunities in this interdisciplinary field for both, theory- as well as practice-oriented research. ----- **TABLE 6. Overview of relevant and analyzed papers and their classification into different groups.** With our article, we contribute to the current state of research in four ways. First, we expand prior research, which did not consider DLT concepts other than blockchain in the integration with AI. Second, we consider both perspectives, AI for DLT, and DLT for AI and the many different concepts of their integration. Third, we bridge the gap ----- **TABLE 6. (Continued) Overview of relevant and analyzed papers and their classification into different groups.** between theory and practice by drawing theoretical conclusions from practical research and outlining future practical research opportunities from theory. Fourth, we describe how convergence creates innovation in an emerging field. This article provides insights for researchers and practitioners interested in deepening their knowledge for interdisciplinary applications of any of the fields: AI, DLT, their convergence, or convergence in general. By providing these insights with an overview on the upcoming convergence of DLT and AI, we contribute to the development of future innovations in this fast-paced field problem. **APPENDIX** _A. RELEVANT AND ANALYZED PAPERS_ See Table 6. ----- **ACKNOWLEDGMENT** The authors would like to thank Mikael Beyene and Niclas Kannengießer for taking the time to discuss the structure of this article. **REFERENCES** [1] J. Bughin, J. Seong, J. Manyika, M. Chui, and R. Joshi. Notes From _the AI Frontier: Modeling the Impact of AI on the World Economy._ Accessed: Mar. 4, 2020. [Online]. Available: https://www.mckinsey.com/ ~/media/McKinsey/Featured%20Insights/Artificial%20Intelligence/ Notes%20from%20the%20frontier%20Modeling%20the%20impact% 20of%20AI%20on%20the%20world%20economy/MGI-Notes-fromthe-AI-frontier-Modeling-the-impact-of-AI-on-the-world-economySeptember-2018.ashx [2] World Economic Forum. Building Block(chain)s for a Better Planet. Accessed: Mar. 4, 2020. [Online]. Available: http://www3.weforum. org/docs/WEF_Building-Blockchains.pdf [3] N. Kannengießer, S. Lins, T. Dehling, and A. Sunyaev, ‘‘Mind the gap: Trade-offs between distributed ledger technology characteristics,’’ _ACM Comput. Surv., to be published. [Online]. Available: https://_ arxiv.org/abs/1906.00861v1 [4] A. Madani, R. Arnaout, M. Mofrad, and R. Arnaout, ‘‘Fast and accurate view classification of echocardiograms using deep learning,’’ NPJ Digit. _Med., vol. 1, no. 1, p. 6, Dec. 2018._ [5] M. Bojarski, D. Del Testa, D. Dworakowski, B. Firner, B. Flepp, P. Goyal, L. D. Jackel, M. Monfort, U. Müller, J. Zhang, X. Zhang, J. Zhao, and K. Zieba, ‘‘End to end learning for selfdriving cars,’’ 2016, arXiv:1604.07316v1. [Online]. Available: https:// arxiv.org/abs/1604.07316v1 [6] N. Sünderhauf, O. Brock, W. Scheirer, R. Hadsell, D. Fox, J. Leitner, B. Upcroft, P. Abbeel, W. Burgard, M. Milford, and P. Corke, ‘‘The limits and potentials of deep learning for robotics,’’ Int. J. Robot. Res., vol. 37, nos. 4–5, pp. 405–420, Apr. 2018. [7] S. Nakamoto. Bitcoin: A Peer-to-Peer Electronic Cash System. Accessed: Mar. 4, 2020. [Online]. Available: https://bitcoin.org/bitcoin.pdf [8] G. G. Dagher, J. Mohler, M. Milojkovic, and P. B. Marella, ‘‘Ancile: Privacy-preserving framework for access control and interoperability of electronic health records using blockchain technology,’’ Sustain. Cities _Soc., vol. 39, pp. 283–297, May 2018._ [9] L. Davi, D. Hatebur, M. Heisel, and R. Wirtz, ‘‘Combining safety and security in autonomous cars using blockchain technologies,’’ in Proc. _Int. Conf. Comput. Saf., Rel., Secur. Cham, Switzerland: Springer, 2019,_ pp. 223–234. [10] C. Hou, M. Zhou, Y. Ji, P. Daian, F. Tramer, G. Fanti, and A. Juels, ‘‘SquirRL: Automating attack discovery on blockchain incentive mechanisms with deep reinforcement learning,’’ 2019, arXiv:1912.01798v1. [Online]. Available: http://arxiv.org/abs/1912.01798v1 [11] N. Hynes, D. Dao, D. Yan, R. Cheng, and D. Song, ‘‘A demonstration of sterling: A privacy-preserving data marketplace,’’ Proc. VLDB Endow_ment, vol. 11, no. 12, pp. 2086–2089, Aug. 2018._ [12] K. Salah, M. H. U. Rehman, N. Nizamuddin, and A. Al-Fuqaha, ‘‘Blockchain for AI: Review and open research challenges,’’ IEEE _Access, vol. 7, pp. 10127–10149, 2019._ [13] T. N. Dinh and M. T. Thai, ‘‘AI and blockchain: A disruptive integration,’’ _Computer, vol. 51, no. 9, pp. 48–53, Sep. 2018._ [14] S. Jeong, J.-C. Kim, and J. Y. Choi, ‘‘Technology convergence: What developmental stage are we in?’’ Scientometrics, vol. 104, no. 3, pp. 841–871, Sep. 2015. [15] G. Duysters and J. Hagedoorn, ‘‘Technological convergence in the IT industry: The role of strategic technology alliances and technological competencies,’’ Int. J. Econ. Bus., vol. 5, no. 3, pp. 355–368, Nov. 1998. [16] M. Abadi, A. Chu, I. Goodfellow, H. B. Mcmahan, I. Mironov, K. Talwar, and L. Zhang, ‘‘Deep learning with differential privacy,’’ in Proc. ACM _SIGSAC Conf. Comput. Commun. Secur. (CCS), 2016, pp. 308–318._ [17] B. Goertzel and C. Pennachin, Artificial General Intelligence. Berlin, Germany: Springer-Verlag, 2007. [18] Y. LeCun, Y. Bengio, and G. Hinton, ‘‘Deep learning,’’ Nature, vol. 521, no. 7553, p. 436, 2015. [19] W. Samek and K.-R. Müller, ‘‘Towards explainable artificial intelligence,’’ in Explainable AI: Interpreting, Explaining Visualizing Deep _Learning. Cham, Switzerland: Springer, 2019, pp. 5–22._ [20] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, and D. Hassabis, ‘‘Mastering the game of go with deep neural networks and tree search,’’ Nature, vol. 529, no. 7587, pp. 484–489, Jan. 2016. [21] D. Mahajan, ‘‘Exploring the limits of weakly supervised pretraining,’’ in _Proc. Eur. Conf. Comput. Vis. (ECCV), 2018, pp. 181–196._ [22] B. W. Silverman and M. C. Jones, ‘‘E. Fix and J.L. Hodges (1951): An important contribution to nonparametric discriminant analysis and density estimation: Commentary on Fix and Hodges (1951),’’ Int. Stat. _Rev., vol. 57, no. 3, pp. 233–238, 1989._ [23] S. J. Russell and P. Norvig, Artificial Intelligence: A Modern Approach. Upper Saddle River, NJ, USA: Prentice-Hall, 2016. [24] L. G. M. E. Eykholt and J. Denman. _RChain_ _Architecture_ _Documentation. Accessed: Mar. 4, 2020. [Online]. Available: https://_ buildmedia.readthedocs.org/media/pdf/rchain-architecture/stable/rchainarchitecture.pdf [25] S. Popov. The Tangle, IOTA Whitepaper. Accessed: Mar. 4, 2020. [Online]. Available: https://assets.ctfassets.net/r1dr6vzfxhev/ 2t4uxvsIqk0EUau6g2sw0g/45eae33637ca92f85dd9f4a3a218e1ec/iota1 _4_3.pdf. [26] S. Thiebes, N. Kannengießer, M. Schmidt-Kraepelin, and A. Sunyaev, ‘‘Beyond data markets: Opportunities and challenges for distributed ledger technology in genomics,’’ in Proc. 53rd Hawaii Int. Conf. Syst. _Sci., 2020, pp. 1–10._ [27] T. Min, H. Wang, Y. Guo, and W. Cai, ‘‘Blockchain games: A survey,’’ 2019, arXiv:1906.05558v1. [Online]. Available: http://arxiv. org/abs/1906.05558v1 [28] G. Wood, ‘‘Ethereum: A secure decentralised generalised transaction ledger,’’ Ethereum Project Yellow Paper, vol. 151, pp. 1–32, Apr. 2014. [29] E. B. Sasson, A. Chiesa, C. Garman, M. Green, I. Miers, E. Tromer, and M. Virza, ‘‘Zerocash: Decentralized anonymous payments from bitcoin,’’ in Proc. IEEE Symp. Secur. Privacy, May 2014, pp. 459–474. [30] C.-S. Curran, S. Bröring, and J. Leker, ‘‘Anticipating converging industries using publicly available data,’’ Technological Forecasting Social _Change, vol. 77, no. 3, pp. 385–395, Mar. 2010._ [31] E. Karafiloski and A. Mishev, ‘‘Blockchain solutions for big data challenges: A literature review,’’ in Proc. 17th Int. Conf. Smart Technol. _(EUROCON), Jul. 2017, pp. 763–768._ [32] T. Wang, S. Chang Liew, and S. Zhang, ‘‘When blockchain meets AI: Optimal mining strategy achieved by machine learning,’’ 2019, _arXiv:1911.12942v1._ [Online]. Available: http://arxiv.org/ abs/1911.12942v1s [33] I. Eyal and E. G. Sirer, ‘‘Majority is not enough: Bitcoin mining is vulnerable,’’ in Proc. Int. Conf. Financial Cryptogr. Data Secur. Berlin, Germany: Springer, 2014, pp. 436–454. [34] Y. Sompolinsky and A. Zohar, ‘‘Secure high-rate transaction processing in bitcoin,’’ in Proc. Int. Conf. Financial Cryptogr. Data Secur. Berlin, Germany: Springer, 2015, pp. 507–527. [35] A. Toroghi Haghighat and M. Shajari, ‘‘Block withholding game among bitcoin mining pools,’’ Future Gener. Comput. Syst., vol. 97, pp. 482–491, Aug. 2019. [36] R. Camino, C. Ferreira Torres, M. Baden, and R. State, ‘‘A data science approach for honeypot detection in ethereum,’’ 2019, _arXiv:1910.01449v2._ [Online]. Available: http://arxiv.org/abs/1910. 01449v2 [37] Y. Kim, D. Pak, and J. Lee, ‘‘ScanAT: Identification of bytecodeonly smart contracts with multiple attribute tags,’’ IEEE Access, vol. 7, pp. 98669–98683, 2019. [38] W. J.-W. Tann, X. J. Han, S. S. Gupta, and Y.-S. Ong, ‘‘Towards safer smart contracts: A sequence learning approach to detecting security threats,’’ 2018, arXiv:1811.06632v3. [Online]. Available: http://arxiv.org/abs/1811.06632v3 [39] L. Zhang, Y. Wang, F. Li, Y. Hu, and M. H. Au, ‘‘A game-theoretic method based on Q-learning to invalidate criminal smart contracts,’’ Inf. Sci., vol. 498, pp. 144–153, Sep. 2019. [40] L.-N. Lundbæk, D. Janes Beutel, M. Huth, S. Jackson, L. Kirk, and R. Steiner, ‘‘Proof of kernel work: A democratic low-energy consensus for distributed access-control protocols,’’ Roy. Soc. Open Sci., vol. 5, no. 8, Aug. 2018, Art. no. 180422. ----- [41] M. E. Gladden, ‘‘Cryptocurrency with a conscience: Using artificial intelligence to develop money that advances human ethical values,’’ Annales. _Etyka W Życiu Gospodarczym, vol. 18, no. 4, pp. 85–98, 2015._ [42] S. Liu, F. Mohsin, L. Xia, and O. Seneviratne, ‘‘Strengthening smart contracts to handle unexpected situations,’’ in Proc. IEEE Int. Conf. _Decentralized Appl. Infrastruct. (DAPPCON), Apr. 2019, pp. 182–187._ [43] Y. Zhao, J. Zhao, L. Jiang, R. Tan, and D. Niyato, ‘‘Mobile edge computing, blockchain and reputation-based crowdsourcing IoT federated learning: A secure, decentralized and privacy-preserving system,’’ 2019, arXiv:1906.10893v1. [Online]. Available: http://arxiv. org/abs/1906.10893v1 [44] P. Ramanan, K. Nakayama, and R. Sharma, ‘‘BAFFLE: Blockchain based aggregator free federated learning,’’ 2019, arXiv:1909.07452v1. [Online]. Available: http://arxiv.org/abs/1909.07452v1 [45] J. Passerat-Palmbach, T. Farnan, R. Miller, M. S. Gross, H. Leigh Flannery, and B. Gleim, ‘‘A blockchain-orchestrated federated learning architecture for healthcare consortia,’’ 2019, arXiv:1910.12603v1. [Online]. Available: http://arxiv.org/abs/1910.12603v1 [46] A. Juels, A. Kosba, and E. Shi, ‘‘The ring of gyges: Investigating the future of criminal smart contracts,’’ presented at the ACM SIGSAC Conf. Comput. Commun. Secur., Vienna, Austria, Oct. 2016, doi: [10.1145/2976749.2978362.](http://dx.doi.org/10.1145/2976749.2978362) [47] J. Kang, M. Bennett, D. Carbado, and P. Casey, ‘‘Implicit bias in the courtroom,’’ UCLA Law Rev., vol. 59, p. 1124, Jun. 2012. [48] A. Narayanan and V. Shmatikov, ‘‘Robust de-anonymization of large sparse datasets,’’ in Proc. IEEE Symp. Secur. Privacy (SP), May 2008, pp. 111–125. [49] J. Isaak and M. J. Hanna, ‘‘User data privacy: Facebook, Cambridge analytica, and privacy protection,’’ Computer, vol. 51, no. 8, pp. 56–59, Aug. 2018. [50] T. Yang, G. Andrew, H. Eichner, H. Sun, W. Li, N. Kong, D. Ramage, and F. Beaufays, ‘‘Applied federated learning: Improving Google keyboard query suggestions,’’ 2018, arXiv:1812.02903v1. [Online]. Available: http://arxiv.org/abs/1812.02903v1 [51] K. Sarpatwar, R. Vaculin, H. Min, G. Su, T. Heath, G. Ganapavarapu, and D. Dillenberger, ‘‘Towards enabling trusted artificial intelligence via blockchain,’’ in Policy-Based Autonomic Data Governance. Cham, Switzerland: Springer, 2019, pp. 137–153. [52] D. Preuveneers, V. Rimmer, I. Tsingenopoulos, J. Spooren, W. Joosen, and E. Ilie-Zudor, ‘‘Chained anomaly detection models for federated learning: An intrusion detection case study,’’ Appl. Sci., vol. 8, no. 12, p. 2663, 2018. [53] X. Huang, M. Kwiatkowska, S. Wang, and M. Wu, ‘‘Safety verification of deep neural networks,’’ in Proc. Int. Conf. Comput. Aided Verification. Springer, 2017, pp. 3–29. [54] B. W. Josh Swihart and S. Bowe. Zcash Counterfeiting Vulnerability _Successfully Remediated. Accessed: Mar. 4, 2020. [Online]. Available:_ https://electriccoin.co/blog/zcash-counterfeiting-vulnerabilitysuccessfully-remediated/ [55] M. Pradel and K. Sen, ‘‘DeepBugs: A learning approach to name-based bug detection,’’ ACM Program. Lang., vol. 2, pp. 1–25, Oct. 2018. [56] M. Bartoletti and L. Pompianu, ‘‘An empirical analysis of smart contracts: Platforms, applications, and design patterns,’’ in Proc. Int. Conf. _Financial Cryptogr. Data Secur. Cham, Switzerland: Springer, 2017,_ pp. 494–509. [57] M. Allamanis, E. T. Barr, P. Devanbu, and C. Sutton, ‘‘A survey of machine learning for big code and naturalness,’’ ACM Comput. Surv., vol. 51, no. 4, pp. 1–37, Jul. 2018. [58] P. Daian, ‘‘Flash boys 2.0: Frontrunning in decentralized exchanges, miner extractable value, and consensus instability,’’ in Proc. IEEE Symp. _Secur. Privacy (SP), May 2020, pp. 566–583._ [59] S. Zhangy, A. Kim, D. Liu, S. C. Nuckchadyy, L. Huangy, A. Masurkary, J. Zhangy, L. P. Karnatiz, L. Martinezx, T. Hardjono, M. Kellis, and Z. Zhang, ‘‘Genie: A secure, transparent sharing and services platform for genetic and health data,’’ 2018, arXiv:1811.01431v1. [Online]. Available: http://arxiv.org/abs/1811.01431v1 [60] V. Raychev, M. Vechev, and E. Yahav, ‘‘Code completion with statistical language models,’’ ACM SIGPLAN Notices, vol. 49, no. 6, pp. 419–428, Jun. 2014. [61] A. Hindle, E. T. Barr, Z. Su, M. Gabel, and P. Devanbu, ‘‘On the naturalness of software,’’ in Proc. 34th Int. Conf. Softw. Eng. (ICSE), Jun. 2012, pp. 837–847. [62] S. Eskandari, S. Moosavi, and J. Clark, ‘‘SoK: Transparent dishonesty: Front-running attacks on blockchain,’’ in Financial Cryptography and _Data Security. Cham, Switzerland: Springer, 2020, pp. 170–189._ [63] A. Judmayer, ‘‘Pay-to-win: Incentive attacks on proof-of-work cryptocurrencies,’’ Cryptol. ePrint Arch., 2019. [Online]. Available: https://eprint.iacr.org/2019/775/20200217:161450 [64] H. Bae, J. Jang, D. Jung, H. Jang, H. Ha, and S. Yoon, ‘‘Security and privacy issues in deep learning,’’ 2018, arXiv:1807.11655v3. [Online]. Available: http://arxiv.org/abs/1807.11655v3 [65] UN Global Working Group on Big Data. UN Handbook on Privacy_Preserving Computation Techniques. Accessed: Mar. 4, 2020. [Online]._ Available: http://publications.officialstatistics.org/handbooks/privacypreserving-techniques-handbook/UN%20Handbook%20for%20PrivacyPreserving%20Techniques.pdf [66] H. Cui, H. Zhang, G. R. Ganger, P. B. Gibbons, and E. P. Xing, ‘‘GeePS: Scalable deep learning on distributed GPUs with a GPU-specialized parameter server,’’ in Proc. 11th Eur. Conf. Comput. Syst. (EuroSys), 2016, pp. 1–16. [67] A. L. Beberg, D. L. Ensign, G. Jayachandran, S. Khaliq, and V. S. Pande, ‘‘Folding@home: Lessons from eight years of volunteer distributed computing,’’ in Proc. IEEE Int. Symp. Parallel Distrib. Process., May 2009, pp. 1–8. [68] J. Truby, ‘‘Decarbonizing bitcoin: Law and policy choices for reducing the energy consumption of blockchain technologies and digital currencies,’’ Energy Res. Social Sci., vol. 44, pp. 399–410, Oct. 2018. [69] C. Chenli, B. Li, Y. Shi, and T. Jung, ‘‘Energy-recycling blockchain with proof-of-deep-learning,’’ 2019, arXiv:1902.03912v1. [Online]. Available: http://arxiv.org/abs/1902.03912v1 [70] H. Turesson, M. Laskowski, A. Roatis, and H. M. Kim, ‘‘Privacypreserving blockchain mining: Sybil-resistance by proof-of-usefulwork,’’ 2019, arXiv:1907.08744v2. [Online]. Available: http://arxiv.org/ abs/1907.08744v2 [71] A. Baldominos and Y. Saez, ‘‘Coin.AI: A proof-of-useful-work scheme for blockchain-based distributed deep learning,’’ Entropy, vol. 21, no. 8, p. 723, 2019. [72] M. Swan, ‘‘Blockchain thinking: The brain as a decentralized autonomous corporation [commentary],’’ IEEE Technol. Soc. Mag., vol. 34, no. 4, pp. 41–52, Dec. 2015. [73] M. Shayan, C. Fung, C. J. M. Yoon, and I. Beschastnikh, ‘‘Biscotti: A ledger for private and secure peer-to-peer machine learning,’’ 2018, _arXiv:1811.09904v4._ [Online]. Available: http://arxiv.org/ abs/1811.09904v4 [74] Y. Lu, Q. Tang, and G. Wang, ‘‘On enabling machine learning tasks atop public blockchains: A crowdsourcing approach,’’ in Proc. IEEE Int. Conf. _Data Mining Workshops (ICDMW), Nov. 2018, pp. 81–88._ [75] K. R. Özyilmaz, M. Do?an, and A. Yurdakul, ‘‘IDMoB: IoT data marketplace on blockchain,’’ in Proc. Crypto Valley Conf. Blockchain Technol. _(CVCBT), Jun. 2018, pp. 11–19._ [76] J. D. Harris and B. Waggoner, ‘‘Decentralized and collaborative AI on blockchain,’’ in Proc. IEEE Int. Conf. Blockchain (Blockchain), Jul. 2019, pp. 368–375. [77] J. Kang, Z. Xiong, D. Niyato, Y. Zou, Y. Zhang, and M. Guizani, ‘‘Reliable federated learning for mobile networks,’’ 2019, arXiv:1910.06837v1. [Online]. Available: http://arxiv.org/abs/1910.06837v1 [78] R. Cheng, F. Zhang, J. Kos, W. He, N. Hynes, N. Johnson, A. Juels, A. Miller, and D. Song, ‘‘Ekiden: A platform for confidentialitypreserving, trustworthy, and performant smart contracts,’’ in Proc. IEEE _Eur. Symp. Secur. Privacy (EuroS&P), Jun. 2019, pp. 185–200._ [79] J. Eberhardt and J. Heiss, ‘‘Off-chaining models and approaches to offchain computations,’’ in Proc. 2nd Workshop Scalable Resilient Infras_truct. Distrib. Ledgers (SERIAL), 2018, pp. 7–12._ [80] X. Chen, J. Ji, C. Luo, W. Liao, and P. Li, ‘‘When machine learning meets blockchain: A decentralized, privacy-preserving and secure design,’’ in _Proc. IEEE Int. Conf. Big Data (Big Data), Dec. 2018, pp. 1178–1187._ [81] H. Kim, J. Park, M. Bennis, and S.-L. Kim, ‘‘Blockchained on-device federated learning,’’ IEEE Commun. Lett., to be published. [82] M. Banko and E. Brill, ‘‘Scaling to very very large corpora for natural language disambiguation,’’ in Proc. 39th Annu. Meeting Assoc. Comput. _Linguistics-ACL, 2001, pp. 26–33._ [83] G. A. Montes and B. Goertzel, ‘‘Distributed, decentralized, and democratized artificial intelligence,’’ Technol. Forecasting Social Change, vol. 141, pp. 354–358, Apr. 2019. ----- [84] M. Jones, M. Johnson, M. Shervey, J. T. Dudley, and N. Zimmerman, ‘‘Privacy-preserving methods for feature engineering using blockchain: Review, evaluation, and proof of concept,’’ J. Med. Internet Res., vol. 21, no. 8, 2019, Art. no. e13600. [85] M. Johnson, M. Jones, M. Shervey, J. T. Dudley, and N. Zimmerman, ‘‘Building a secure biomedical data sharing decentralized app (DApp): Tutorial,’’ J. Med. Internet Res., vol. 21, no. 10, 2019, Art. no. e13601. [86] L. Melis, C. Song, E. De Cristofaro, and V. Shmatikov, ‘‘Exploiting unintended feature leakage in collaborative learning,’’ in Proc. IEEE _Symp. Secur. Privacy (SP), May 2019, pp. 691–706._ [87] V. Costan and S. Devadas, ‘‘Intel SGX explained,’’ IACR Cryptol. _ePrint Arch., vol. 2016, no. 86, pp. 1–118, 2016. [Online]. Available:_ https://eprint.iacr.org/2016/086/20170221:054353 [88] F. Tschorsch and B. Scheuermann, ‘‘Bitcoin and beyond: A technical survey on decentralized digital currencies,’’ IEEE Commun. Surveys Tuts., vol. 18, no. 3, pp. 2084–2123, 3rd Quart., 2016. [89] S. King. Primecoin: Cryptocurrency With Prime Number Proof-of_Work. Accessed: Mar. 4, 2020. [Online]. Available: https://primecoin.io/_ bin/primecoin-paper.pdf [90] F. Saleh, ‘‘Blockchain without waste: Proof-of-stake,’’ SSRN Elec_tron. J., 2019. [Online]. Available: https://papers.ssrn.com/sol3/papers._ cfm?abstract_id=3183935 [91] N. Hynes, R. Cheng, and D. Song, ‘‘Efficient deep learning on multisource private data,’’ 2018, arXiv:1807.06689v1. [Online]. Available: http://arxiv.org/abs/1807.06689v1 [92] R. Kunkel, D. Le Quoc, F. Gregor, S. Arnautov, P. Bhatotia, and C. Fetzer, ‘‘TensorSCONE: A secure TensorFlow framework using intel SGX,’’ 2019, arXiv:1902.04413v1. [Online]. Available: http://arxiv.org/abs/1902.04413v1 [93] F. Brasser, T. Frassetto, K. Riedhammer, A.-R. Sadeghi, T. Schneider, and C. Weinert, ‘‘VoiceGuard: Secure and private speech processing,’’ in Proc. InterSpeech, Sep. 2018, pp. 1303–1307. [94] D. Lee, D. Kohlbrenner, S. Shinde, D. Song, and K. Asanović, ‘‘Keystone: An open framework for architecting TEEs,’’ 2019, arXiv:1907.10119v2. [Online]. Available: http://arxiv.org/abs/1907.10119v2 [95] F. Brasser, S. Capkun, A. Dmitrienko, T. Frassetto, K. Kostiainen, and A.-R. Sadeghi, ‘‘DR.SGX: Automated and adjustable side-channel protection for SGX using data location randomization,’’ in Proc. 35th Annu. _Comput. Secur. Appl. Conf. (ACSAC), 2019, pp. 788–800._ [96] M. Milutinovic, W. He, H. Wu, and M. Kanwal, ‘‘Proof of luck: An efficient blockchain consensus protocol,’’ in Proc. 1st Workshop Syst. Softw. _Trusted Execution (SysTEX), 2016, pp. 1–6._ [97] L. Chen, L. Xu, N. Shah, Z. Gao, Y. Lu, and W. Shi, ‘‘On security analysis of proof-of-elapsed-time (PoET),’’ in Proc. Int. Symp. Stabilization, Saf., _Secur. Distrib. Syst. Cham, Switzerland: Springer, 2017, pp. 282–297._ [98] A. Miller, I. Bentov, S. Bakshi, R. Kumaresan, and P. McCorry, ‘‘Sprites and state channels: Payment networks that go faster than lightning,’’ in _Proc. Int. Conf. Financial Cryptogr. Data Secur. Cham, Switzerland:_ Springer, 2019, pp. 508–526. [99] J. Poon and V. Buterin. Plasma: Scalable Autonomous Smart Con_tracts. Accessed: Mar. 4, 2020. [Online]. Available: https://www._ plasma.io/plasma.pdf [100] S. Dziembowski, S. Faust, and K. Hostáková, ‘‘General state channel networks,’’ in Proc. ACM SIGSAC Conf. Comput. Commun. Secur., Jan. 2018, pp. 949–966. [101] H. Kalodner, S. Goldfeder, X. Chen, S. M. Weinberg, and E. W. Felten, ‘‘Arbitrum: Scalable, private smart contracts,’’ in 27th USENIX Secur. _Symp. (USENIX Security), 2018, pp. 1353–1370._ [102] J. Teutsch and C. Reitwießner, ‘‘A scalable verification solution for blockchains,’’ 2019, arXiv:1908.04756v1. [Online]. Available: http://arxiv.org/abs/1908.04756v1 [103] R. C. Geyer, T. Klein, and M. Nabi, ‘‘Differentially private federated learning: A client level perspective,’’ 2017, arXiv:1712.07557v2. [Online]. Available: http://arxiv.org/abs/1712.07557v2 [104] High-Level Expert Group on Artificial Intelligence. Ethics Guide_lines for Trustworthy AI. Accessed: Mar. 4, 2020. [Online]. Available:_ https://ec.europa.eu/newsroom/dae/document.cfm?doc_id=60419 [105] S. Schulz, A.-R. Sadeghi, and C. Wachsmann, ‘‘Short paper: Lightweight remote attestation using physical functions,’’ in Proc. 4th ACM Conf. _Wireless Netw. Secur. (WiSec), 2011, pp. 109–114._ [106] S. Volos, K. Vaswani, and R. Bruno, ‘‘Graviton: Trusted execution environments on GPUs,’’ in Proc. 13th USENIX Symp. Oper. Syst. Design _Implement. (OSDI), 2018, pp. 681–696._ [107] S. Bano, A. Sonnino, M. Al-Bassam, S. Azouvi, P. McCorry, S. Meiklejohn, and G. Danezis, ‘‘SoK: Consensus in the age of blockchains,’’ in Proc. 1st ACM Conf. Adv. Financial Technol. (AFT), 2019, pp. 183–198. [108] J. Van Bulck, M. Minkin, O. Weisse, D. Genkin, B. Kasikci, F. Piessens, M. Silberstein, T. F. Wenisch, Y. Yarom, and R. Strackx, ‘‘Foreshadow: Extracting the keys to the Intel SGX kingdom with transient out-oforder execution,’’ in Proc. 27th USENIX Secur. Symp. (USENIX Security), 2018, pp. 991–1008. [109] P. Das, ‘‘FastKitten: Practical smart contracts on bitcoin,’’ in Proc. 28th _USENIX Secur. Symp. (USENIX Security), 2019, pp. 801–818._ [110] A. Vijayakumar and S. Kundu, ‘‘A novel modeling attack resistant PUF design based on non-linear voltage transfer characteristics,’’ in Proc. _Design, Autom. Test Eur. Conf. Exhib. (DATE), 2015, pp. 653–658._ [111] F. Ganji, On the Learnability of Physically Unclonable Functions. Cham, Switzerland: Springer, 2018. KONSTANTIN D. PANDL studied electrical engineering and information technology at the Karlsruhe Institute of Technology (KIT), Germany, Purdue University, USA, and Tongji University, China, and graduated with the master’s degree. He is currently pursuing the Ph.D. degree with the Institute of Applied Informatics and Formal Description Methods, KIT. He also gained experience in industry at Siemens’ Venture Unit Next47, San Francisco Bay Area, and Kearney, Germany. His research interests include machine learning, system security, and distributed systems. His previous research appeared at the IEEE International Conference on Intelligent Transportation Systems. SCOTT THIEBES received the degree in information systems from the University of Cologne. He is currently pursuing the Ph.D. degree with the Institute of Applied Informatics and Formal Description Methods, Karlsruhe Institute of Technology, Germany. His research interests include emerging technologies in health care, information security and privacy, and gamification. His recent research interests include on the implications of applying distributed ledger technology and artificial intelligence in health care settings. His research appeared in journals, including the Journal of Medical Internet Research, BMC Bioinformatics, and the European Journal of Human Genetics. MANUEL SCHMIDT-KRAEPELIN received the degree in information systems from the University of Cologne. He is currently pursuing the Ph.D. degree with the Institute of Applied Informatics and Formal Description Methods, Karlsruhe Institute of Technology, Germany. He is also a Research Associate with the Institute of Applied Informatics and Formal Description Methods, Karlsruhe Institute of Technology. His research interests include gamification and emerging technologies in health care. His research appeared in leading scientific conferences, including ICIS, ECIS, and HICSS. ALI SUNYAEV is currently a Professor of computer science with the Karlsruhe Institute of Technology, Germany. His research interests include trustworthy Internet technologies and complex health IT applications. His research work accounts for the multifaceted use contexts of digital technologies with research on human behavior affecting Internet-based systems and vice versa. His research appeared in journals, including ACM CSUR, JIT, JMIS, the IEEE TRANSACTIONS ON CLOUD COMPUTING, Communications of the ACM, and others. His research work has been appreciated numerous times and is featured in a variety of media outlets. CLOUD COMPUTING, -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2001.11017, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": "CCBY", "status": "GOLD", "url": "https://ieeexplore.ieee.org/ielx7/6287639/8948470/09039606.pdf" }
2,020
[ "JournalArticle", "Review" ]
true
2020-01-29T00:00:00
[ { "paperId": "1b32b77718be22ca31c2a0dba709c815e19d4672", "title": "Ethics Guidelines for Trustworthy AI" }, { "paperId": "03d1b883e9d8474212094e5764646bc6450cf565", "title": "Blockchain Without Waste: Proof-of-Stake" }, { "paperId": "759b10011ea66667345d6954dcfa8eb5192e68d3", "title": "Trade-offs between Distributed Ledger Technology Characteristics" }, { "paperId": "128a5c0ddd09c6d36ef5d483b6d140dfbbcb0ce3", "title": "Flash Boys 2.0: Frontrunning in Decentralized Exchanges, Miner Extractable Value, and Consensus Instability" }, { "paperId": "5df775b74e058d4849f1c18280791a83efcda0fb", "title": "Beyond Data Markets: Opportunities and Challenges for Distributed Ledger Technology in Genomics" }, { "paperId": "af86081d5394948d8570f1baa2a67cd78b57f177", "title": "When blockchain meets AI: Optimal mining strategy achieved by machine learning" }, { "paperId": "618babd71f257f063d6ad2ecf1b2070d34721dda", "title": "Reliable Federated Learning for Mobile Networks" }, { "paperId": "f04b7ebb62aeef8c8a37bf3838871318af01d41c", "title": "A blockchain-orchestrated Federated Learning architecture for healthcare consortia" }, { "paperId": "edc02b24064569dfed1f3aef69e240792a9c0e0d", "title": "A Data Science Approach for Detecting Honeypots in Ethereum" }, { "paperId": "bc488489d20833338d9764406635b224cf82c6bf", "title": "Building a Secure Biomedical Data Sharing Decentralized App (DApp): Tutorial" }, { "paperId": "2ec1f1c4c641b80167f2458968c3c12ea0602b9a", "title": "BAFFLE : Blockchain Based Aggregator Free Federated Learning" }, { "paperId": "21f214165bda8813c95cfa3e471e24aaf8d5776b", "title": "Towards Explainable Artificial Intelligence" }, { "paperId": "758a4f817461315cd1a5e7a49b5c2a9202f54c84", "title": "Combining Safety and Security in Autonomous Cars Using Blockchain Technologies" }, { "paperId": "886c2128a5a7be8e5c2d806f99af5553ef9b70ed", "title": "A game-theoretic method based on Q-learning to invalidate criminal smart contracts" }, { "paperId": "44eb58178387617469d60a482e71b60f1a14117b", "title": "A scalable verification solution for blockchains" }, { "paperId": "443762d49b2481ddd60687eb12cdc4490b31678a", "title": "Privacy-Preserving Methods for Feature Engineering Using Blockchain: Review, Evaluation, and Proof of Concept" }, { "paperId": "c0a53bc7a381e1e315c6b6d00dbe1db0ca5759de", "title": "Keystone: An Open Framework for Architecting TEEs" }, { "paperId": "69ddd8b6df52836d279e00cf379689378eb33975", "title": "Keystone: A Framework for Architecting TEEs" }, { "paperId": "6d39fc36e536494f941314db3fda129466fa9d32", "title": "Proof-of-Useful-Work as Dual-Purpose Mechanism for Blockchain and AI: Blockchain Consensus that Enables Privacy Preserving Data Mining" }, { "paperId": "0efbaec0a4875ad8d88a4e07e87f425d8d102295", "title": "Decentralized and Collaborative AI on Blockchain" }, { "paperId": "7c7ccf8defa9c61eb2487af1249974b3f1213957", "title": "Mobile Edge Computing, Blockchain and Reputation-based Crowdsourcing IoT Federated Learning: A Secure, Decentralized and Privacy-preserving System" }, { "paperId": "c0126e061a58b13d4f52686c9c59a81f815ccdb3", "title": "Blockchain Games: A Survey" }, { "paperId": "393ab84a86631d5fda128c3aac0bf5476da07791", "title": "Flash Boys 2.0: Frontrunning, Transaction Reordering, and Consensus Instability in Decentralized Exchanges" }, { "paperId": "26db4528466b86a942dcdf3ba5fe394657443c55", "title": "Strengthening Smart Contracts to Handle Unexpected Situations" }, { "paperId": "96321435462d9ea1e7971afccdf3752dab9cb4b3", "title": "Distributed, decentralized, and democratized artificial intelligence" }, { "paperId": "c172b3429e70fc8e4489475a2f09da9234189d98", "title": "Coin.AI: A Proof-of-Useful-Work Scheme for Blockchain-Based Distributed Deep Learning" }, { "paperId": "3b9926b55a8e622472b890d9393d8174f3af43a5", "title": "Block withholding game among bitcoin mining pools" }, { "paperId": "aefe8b786cb8e7e4674c4e38962a0a57ae735e7b", "title": "SoK: Transparent Dishonesty: Front-Running Attacks on Blockchain" }, { "paperId": "a7d0dfe415ab572db38f340e61cf4748ca3a9bf0", "title": "FastKitten: Practical Smart Contracts on Bitcoin" }, { "paperId": "a96405df7f3b6318d31f088e3988fe93dbecc949", "title": "TensorSCONE: A Secure TensorFlow Framework using Intel SGX" }, { "paperId": "eb8c673d677e0daa6bb94b3925ce1b9c350128f7", "title": "Energy-recycling Blockchain with Proof-of-Deep-Learning" }, { "paperId": "fbe6553641ce3c6a1914ef5d9f8797fa762c8bad", "title": "Chained Anomaly Detection Models for Federated Learning: An Intrusion Detection Case Study" }, { "paperId": "094e1f0434195c6fa8f3df6e7adad1f3b67ba801", "title": "Off-chaining Models and Approaches to Off-chain Computations" }, { "paperId": "b97047c4dc75cbe8d6fc5cb3dd5a81d36458892d", "title": "APPLIED FEDERATED LEARNING: IMPROVING GOOGLE KEYBOARD QUERY SUGGESTIONS" }, { "paperId": "33c3f816bde8ee63ee9f2e60d4387b9390696371", "title": "Beyond Inferring Class Representatives: User-Level Privacy Leakage From Federated Learning" }, { "paperId": "5883a18e504baff774c01060343c71619f1823b8", "title": "When Machine Learning Meets Blockchain: A Decentralized, Privacy-preserving and Secure Design" }, { "paperId": "1df492149bce34a88c60fa19e01c25c77e3733af", "title": "Biscotti: A Ledger for Private and Secure Peer-to-Peer Machine Learning" }, { "paperId": "4f03d676a3dc4c93dd13bfdd77542acc8b706417", "title": "Towards Safer Smart Contracts: A Sequence Learning Approach to Detecting Security Threats." }, { "paperId": "197be7cf03dbb250130461600c17528b21a0fe69", "title": "Towards Safer Smart Contracts: A Sequence Learning Approach to Detecting Vulnerabilities" }, { "paperId": "e9447839d763937af2753e3623e586a346c10752", "title": "Genie: A Secure, Transparent Sharing and Services Platform for Genetic and Health Data" }, { "paperId": "e9cbe218e5a3ad4719f6e9d512bad06fc6a343f5", "title": "Privacy‐Preserving Blockchain Mining: Sybil-Resistance by Proof‐of‐Useful‐Work" }, { "paperId": "0313cb8bfdd7540686237221bbd9fd919fcb1992", "title": "On Enabling Machine Learning Tasks atop Public Blockchains: A Crowdsourcing Approach" }, { "paperId": "1636594517efaad1d911b4065ede56f1f77bacd2", "title": "Competition, Market Power and Third-Party Tracking" }, { "paperId": "4f394320c64162ea7a5d13db32d3b83e5fe1c3d5", "title": "Graviton: Trusted Execution Environments on GPUs" }, { "paperId": "0625f8ea4149463e0df6648b3349801302d564ae", "title": "General State Channel Networks" }, { "paperId": "15c69c4ece24b82105fc41d270d642593ce4318f", "title": "Decarbonizing Bitcoin: Law and policy choices for reducing the energy consumption of Blockchain technologies and digital currencies" }, { "paperId": "1a5ba159030ad2349b5870b57697fbb45deabaf0", "title": "VoiceGuard: Secure and Private Speech Processing" }, { "paperId": "54a4b505f267d5c27bcaa70d6a259a1035d22395", "title": "AI and Blockchain: A Disruptive Integration" }, { "paperId": "721ba8f893ea219617f4ab4607abbf7d2cee0d54", "title": "Foreshadow: Extracting the Keys to the Intel SGX Kingdom with Transient Out-of-Order Execution" }, { "paperId": "0b96c12fd4e25b949dea4ff57df033bf16b2d100", "title": "Blockchained On-Device Federated Learning" }, { "paperId": "70d2008b0dd9a4d201e62bd25f723d3e6d277c2e", "title": "A Demonstration of Sterling: A Privacy-Preserving Data Marketplace" }, { "paperId": "356bb26086b2bbc36ed099ee4bec4555d7f89c4c", "title": "User Data Privacy: Facebook, Cambridge Analytica, and Privacy Protection" }, { "paperId": "73423b28da891984a7e86b65d0a2148ab9d134b0", "title": "Proof of Kernel Work: a democratic low-energy consensus for distributed access-control protocols" }, { "paperId": "7eb657c0394df220db3013bf7277cbde75dd18cf", "title": "Security and Privacy Issues in Deep Learning" }, { "paperId": "147db0417e95ce8039ce18671260096c1bc046b8", "title": "Efficient Deep Learning on Multi-Source Private Data" }, { "paperId": "8727ed6c89d23d520dffd7781b4f61b92b730ff7", "title": "IDMoB: IoT Data Marketplace on Blockchain" }, { "paperId": "30e0ffeb519a4df2d4a2067e899c5fb5c5e85e70", "title": "Exploiting Unintended Feature Leakage in Collaborative Learning" }, { "paperId": "0f885fd46064d271d4404cf9bb3d758e1a6f8d55", "title": "Exploring the Limits of Weakly Supervised Pretraining" }, { "paperId": "863dff6fea7811e6c2b76b3eb64eee84ee280b33", "title": "Ancile: Privacy-Preserving Framework for Access Control and Interoperability of Electronic Health Records Using Blockchain Technology" }, { "paperId": "c70f11da5a8bb2df36def1d99c4c08df315e2233", "title": "DeepBugs: a learning approach to name-based bug detection" }, { "paperId": "643fb5aeccce0351cac89bb8e446c8d7e48a3d99", "title": "The limits and potentials of deep learning for robotics" }, { "paperId": "c3a2176f4fa079793741966194f3a9a8e7491a24", "title": "Ekiden: A Platform for Confidentiality-Preserving, Trustworthy, and Performant Smart Contracts" }, { "paperId": "ae712766e4a39c69af165ca7d23cc968824d4309", "title": "Fast and accurate view classification of echocardiograms using deep learning" }, { "paperId": "4170c46d6e257358fb6e6ec4ad69c307406a0b45", "title": "Should We Be Concerned About Data-opolies?" }, { "paperId": "b1e538dbf538fd9fdf5f5870c5b7416ae08c9882", "title": "Differentially Private Federated Learning: A Client Level Perspective" }, { "paperId": "a37facbe13d9988ad32816cbcc34962235e11f62", "title": "SoK: Consensus in the Age of Blockchains" }, { "paperId": "60c01f8700f063990da7a0d84adaa7a2d120c7b5", "title": "On Security Analysis of Proof-of-Elapsed-Time (PoET)" }, { "paperId": "5d294a8e8c52485c8010e4bd6fc5ee45e68f390f", "title": "DR.SGX: automated and adjustable side-channel protection for SGX using data location randomization" }, { "paperId": "960ba564e9e598d864dff38d2f3d0bad1b319ead", "title": "A Survey of Machine Learning for Big Code and Naturalness" }, { "paperId": "58e0ca33ae3068fee7005f339bf6c444fc17d55f", "title": "Explainable Artificial Intelligence: Understanding, Visualizing and Interpreting Deep Learning Models" }, { "paperId": "37d9283061bb8057adff53ff4033dd11ccdf2a0c", "title": "Blockchain solutions for big data challenges: A literature review" }, { "paperId": "73fd4caae8b4ce04df63437fb99e17f2cc1b6b39", "title": "Financial Cryptography and Data Security" }, { "paperId": "4dce5b72e1f205e11dcdcc8db68ea1cb9a68bbc5", "title": "Sprites and State Channels: Payment Networks that Go Faster Than Lightning" }, { "paperId": "ac72566bbc7628255002a70ca5bec0874929eba4", "title": "Sprites: Payment Channels that Go Faster than Lightning" }, { "paperId": "d2cb2e69e04e8a22125b6956ab69af045719db0e", "title": "Proof of Luck: an Efficient Blockchain Consensus Protocol" }, { "paperId": "48e34a61f715b791396457345f5a5a28bcf7b41d", "title": "The Ring of Gyges: Investigating the Future of Criminal Smart Contracts" }, { "paperId": "d31655b3f82038c513eed0e4a84c8f1c89d5bfdb", "title": "Safety Verification of Deep Neural Networks" }, { "paperId": "9543fc3670c251bc57c5bcccd9aed332bcec8aba", "title": "The blockchain: a new framework for robotic swarm systems" }, { "paperId": "bd8a307efcffbf57d2e5c3c23577de44d883d865", "title": "MedRec: Using Blockchain for Medical Data Access and Permission Management" }, { "paperId": "e9a986c8ff6c2f381d026fe014f6aaa865f34da7", "title": "Deep Learning with Differential Privacy" }, { "paperId": "0e3cc46583217ec81e87045a4f9ae3478a008227", "title": "End to End Learning for Self-Driving Cars" }, { "paperId": "7707082cc4522a91057fa3a8031d67fe3ff1f5a9", "title": "GeePS: scalable deep learning on distributed GPUs with a GPU-specialized parameter server" }, { "paperId": "fa794a84f08546a630cfd90b0b1d5efd532fd112", "title": "Artificial General Intelligence" }, { "paperId": "8db5d1d7169a1f5391cb184332b95835ae668cf4", "title": "Bitcoin and Beyond: A Technical Survey on Decentralized Digital Currencies" }, { "paperId": "846aedd869a00c09b40f1f1f35673cb22bc87490", "title": "Mastering the game of Go with deep neural networks and tree search" }, { "paperId": "bc577f09b0b84545fc1c88a6099ef1461ab8e2e0", "title": "Blockchain Thinking : The Brain as a Decentralized Autonomous Corporation [Commentary]" }, { "paperId": "4c006a424f230f2fd846146056bf7a2ac5fcaf81", "title": "Cryptocurrency with a Conscience: Using Artificial Intelligence to Develop Money that Advances Human Ethical Values" }, { "paperId": "00fe80e570a54010d2eb3d89aae55c7cb7176e13", "title": "Technology convergence: What developmental stage are we in?" }, { "paperId": "b304c0a3f402e8332a9c4bd3f1b459d4ac1a25fd", "title": "A novel modeling attack resistant PUF design based on non-linear voltage transfer characteristics" }, { "paperId": "728b60c04afb5b87853b59265e49f430dbf631db", "title": "Secure High-Rate Transaction Processing in Bitcoin" }, { "paperId": "1e688be9f4554aa981fe3db9e2a66388b05bd167", "title": "Code completion with statistical language models" }, { "paperId": "3797d924fc5f832b72a84c512e92021906793965", "title": "Zerocash: Decentralized Anonymous Payments from Bitcoin" }, { "paperId": "7bf81e964a7c20d829c1225685ae138bf1489c99", "title": "Majority is not enough" }, { "paperId": "77bdb4c1117d72ad863845f35ba50929cb623776", "title": "On the naturalness of software" }, { "paperId": "5acdfa8829f5e698c62cb11129fb48496626e74c", "title": "Implicit Bias in the Courtroom" }, { "paperId": "56252c731dbbe411caf9825821369ea30bc89e28", "title": "Short paper: lightweight remote attestation using physical functions" }, { "paperId": "8727420aefa8f092b3f1a17f7a1c967d1c55ee3e", "title": "Fifty Years of Moore's Law" }, { "paperId": "42980fe896329cd28d39b4ab09d7613bf4dd8608", "title": "Anticipating converging industries using publicly available data" }, { "paperId": "1c37285eb427f9d5bae45dc0624cb5ee78abe827", "title": "Folding@home: Lessons from eight years of volunteer distributed computing" }, { "paperId": "52d6c4a64bb9ffc8d8052413c3ad69df4c62481d", "title": "Robust De-anonymization of Large Sparse Datasets" }, { "paperId": "7628b62d64d2e5c33a13a5a473bc41b2391c1ebc", "title": "Scaling to Very Very Large Corpora for Natural Language Disambiguation" }, { "paperId": "de49b20ae2d1bd69c754d0ed8eb7055955e64709", "title": "Technological Convergence in the IT Industry: The Role of Strategic Technology Alliances and Technological Competencies" }, { "paperId": "0ccd35653fed0836977432cbe3d9f3caa93842ec", "title": "E. Fix and J.L. Hodges (1951): An Important Contribution to Nonparametric Discriminant Analysis and Density Estimation: Commentary on Fix and Hodges (1951)" }, { "paperId": "62cb06e6d7164d26a1957c374c7669d720831d64", "title": "SquirRL: Automating Attack Analysis on Blockchain Incentive Mechanisms with Deep Reinforcement Learning" }, { "paperId": null, "title": "Primecoin: Cryptocurrency With Prime Number Proof-of- Work" }, { "paperId": null, "title": "Zcash Counterfeiting Vulnerability Successfully Remediated" }, { "paperId": null, "title": "The Tangle, IOTA Whitepaper" }, { "paperId": null, "title": "Notes From the AI Frontier: Modeling the Impact of AI on the World Economy" }, { "paperId": null, "title": "RChain Architecture Documentation" }, { "paperId": "2614a0d0213e8cc457ea62b435a8f43dea54245a", "title": "Blockchain for AI: Review and Open Research Challenges" }, { "paperId": "e0b828a391f7a2c0b17ac6e000184e32e7e77bdf", "title": "ScanAT: Identification of Bytecode-Only Smart Contracts With Multiple Attribute Tags" }, { "paperId": "0d185e5aa2796f0e862b98fa7aa8feada2f5ed13", "title": "Pay-To-Win: Incentive Attacks on Proof-of-Work Cryptocurrencies" }, { "paperId": "881eb0b890402d469c4ceb0ef48b2a57931c8caf", "title": "High-Level Expert Group on Artificial Intelligence – Draft Ethics Guidelines for Trustworthy AI" }, { "paperId": "ff6c3c2505e953d2d82be7f97f12ce8d71442447", "title": "Towards Enabling Trusted Artificial Intelligence via Blockchain" }, { "paperId": "e5559ea78af0685df47b534c4d96c5ae09474501", "title": "Arbitrum: Scalable, private smart contracts" }, { "paperId": null, "title": "Building Block(chain)s for a Better Planet" }, { "paperId": "cbc775e301d62740bcb3b8ec361721b3edd7c879", "title": "Plasma : Scalable Autonomous Smart Contracts" }, { "paperId": "62f81ba2d672d470caf8a14fd8460ac257c9e4ff", "title": "On the learnability of physically unclonable functions" }, { "paperId": null, "title": "Prioriello, EOS" }, { "paperId": "2d7f3f4ca3fbb15ae04533456e5031e0d0dc845a", "title": "Intel SGX Explained" }, { "paperId": null, "title": "Byteball: A decentralized system for storage and transfer of value" }, { "paperId": null, "title": "Artificial intelligence: a modern approach. Malaysia Pearson Education Limited" }, { "paperId": "2913c2bf3f92b5ae369400a42b2d27cc5bc05ecb", "title": "Deep Learning" }, { "paperId": "55234eb4ea806df40cf9d8ff57abf7b14c346af0", "title": "The Ring of Gyges : Using Smart Contracts for Crime" }, { "paperId": "3c50bb6cc3f5417c3325a36ee190e24f0dc87257", "title": "ETHEREUM: A SECURE DECENTRALISED GENERALISED TRANSACTION LEDGER" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": null, "title": "Optimal Mining Strategy Achieved By Machine Learning , \" arXiv preprint arXiv :" }, { "paperId": "f2d1e3ca651d5b23016d333bdaebf15a408aebec", "title": "MultiChain Private Blockchain — White Paper" } ]
26,546
en
[ { "category": "Computer Science", "source": "external" }, { "category": "Engineering", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]
https://www.semanticscholar.org/paper/01633c22086abbdc8f180de4361585c1a731bc93
[ "Computer Science", "Engineering" ]
0.823053
A Distributed Implementation of Steady-State Kalman Filter
01633c22086abbdc8f180de4361585c1a731bc93
IEEE Transactions on Automatic Control
[ { "authorId": "152551086", "name": "Jiaqi Yan" }, { "authorId": "2143732199", "name": "Xu Yang" }, { "authorId": "1760677", "name": "Yilin Mo" }, { "authorId": "3527062", "name": "Keyou You" } ]
{ "alternate_issns": null, "alternate_names": [ "IEEE Trans Autom Control" ], "alternate_urls": null, "id": "1283a59c-0d1f-48c3-81d7-02172f597e70", "issn": "0018-9286", "name": "IEEE Transactions on Automatic Control", "type": "journal", "url": "http://ieeexplore.ieee.org/servlet/opac?punumber=9" }
This article studies the distributed state estimation in sensor network, where <inline-formula><tex-math notation="LaTeX">$m$</tex-math></inline-formula> sensors are deployed to infer the <inline-formula><tex-math notation="LaTeX">$n$</tex-math></inline-formula>-dimensional state of a linear time-invariant Gaussian system. By a lossless decomposition of the optimal steady-state Kalman filter, we show that the problem of distributed estimation can be reformulated as that of the synchronization of homogeneous linear systems. Based on such decomposition, a distributed estimator is proposed, where each sensor node runs a local filter using only its own measurement, alongside with a consensus algorithm to fuse the local estimate of every node. We prove that the average of local estimates from all sensors coincides with the optimal Kalman estimate, and under certain condition on the graph Laplacian matrix and the system matrix, the covariance of local estimation error is bounded and the asymptotic error covariance is derived. As a result, the distributed estimator is stable for each single node. We further show that the proposed algorithm has a low message complexity of <inline-formula><tex-math notation="LaTeX">$\min (m,n)$</tex-math></inline-formula>. Numerical examples are provided in the end to illustrate the efficiency of the proposed algorithm.
# A Distributed Implementation of Steady-State Kalman Filter ### Jiaqi Yan, Xu Yang, Yilin Mo[∗], and Keyou You **_Abstract—This paper studies the distributed state estimation_** **in sensor network, where m sensors are deployed to infer the n-** **dimensional state of a Linear Time-Invariant (LTI) Gaussian sys-** **tem. By a lossless decomposition of optimal steady-state Kalman** **filter, we show that the problem of distributed estimation can** **be reformulated as the synchronization of homogeneous linear** **systems. Based on such decomposition, a distributed estimator is** **proposed, where each sensor node runs a local filter using only** **its own measurement, alongside with a consensus algorithm to** **fuse the local estimate of every node. We prove that the average** **of estimates from all sensors coincides with the optimal Kalman** **estimate, and under certain condition on the graph Laplacian** **matrix and the system matrix, the covariance of estimation error** **is bounded and the asymptotic error covariance is derived. As** **a result, the distributed estimator is stable for each single node.** **We further show that the proposed algorithm has a low message** **complexity of min(m, n). Numerical examples are provided in** **the end to illustrate the efficiency of the proposed algorithm.** **_Index Terms—Distributed estimation, Kalman filter, Linear_** **system synchronization, Consensus algorithm.** I. INTRODUCTION _yi(k)_ _yj(k)_ |Local filter Local filter|Col2|Col3| |---|---|---| |Synchronization ∆(k) i Linear system Linear system ∆ (k) j||| |||| _xˆi(k)_ _xˆj(k)_ Fig. 1: The information flow of most existing algorithms, where sensors i and j are immediate neighbors. The past decades have witnessed remarkable research interests in multi-sensor networked systems. As one of its important focuses, distributed estimation has been widely studied in various applications including robot formation control, environment monitoring, spacecraft navigation (see [1]–[5]). Compared with the centralized architecture, it provides better robustness, flexibility and reliability. One fundamental problem in distributed estimation is to estimate the state of an LTI Gaussian system using multiple sensors, where the well-known Kalman filter provides the optimal solution in a centralized manner [6]. Thus, many research efforts have been devoted to the distributed implementation of Kalman filter. For example, in an early work [7], the authors suggest a fusion algorithm for two-sensor networks, where local estimate of the first sensor is considered as a pseudo measurement of the second one. Due to its ease of implementation, this approach has then inspired the sequential fusion in multisensor networks [8]–[10], where the multiple nodes repeatedly perform the two-sensor fusion in a sequential manner. As the result of serial operation, these algorithms require special communication topology which should be sequentially connected as a ring/chain. In [11], Olfati-Saber et. al consider the more general network topology. They introduce the consensus The authors are with Department of Automation, BNRist, Tsinghua University. Emails: jiaqiyan@tsinghua.edu.cn, xu-yang16@mails.tsinghua.edu.cn, ylmo@tsinghua.edu.cn,youky@tsinghua.edu.cn C di th algorithms into distributed estimation and propose KalmanConsensus Filter (KCF), where the average consensus on local estimates is performed. Since then, various consensusbased distributed estimators have been proposed in literature [12]–[24]. For example, instead of doing consensus on local estimates, [14] suggests achieving consensus respectively on noisy measurements and inverse-covariance matrices. On the other hand, Battistelli et. al [25] find that, by performing consensus on the Kullback-Leibler average of local probability density function, estimation stability is also guaranteed. They further prove that, if the single-consensus step is used, this approach is reduced to the well-known covariance intersection fusion rule [26], [27]. Since the consensus-based estimators usually require multiple consensus steps during each sampling period, they generate better estimation performance. In Fig. 1, we present the general information flow of the existing consensus-based estimation algorithms, where ∆i(k) is the information transmitted by sensor i and to be fused by consensus algorithms, which could be the local estimates ([11], [12]), measurements ([13]–[15], [28]), or information matrices ([25], [29]). It is noticed from the figure that the consensus/synchronization process is usually coupled with the local filter in these works, making it hard to analyze the performance of local estimates. Due to this fact, while the aforementioned algorithms are successful in distributing the fusion task over multiple nodes and providing stable local estimates, i.e. the error covariance is proved to be bounded at each sensor side, the exact calculation of error covariance can hardly be obtained. Moreover, the global optimality (namely, whether performance of the algorithm can converge to that of the centralized Kalman filter) is also difficult to be analyzed and guaranteed in some works ----- It is worth noticing that in theory, the gain of the Kalman filter converges to a steady-state gain exponentially fast [30], which can be calculated off-line. Moreover, in practice, a fixed gain estimator is usually implemented, which has the same asymptotic performance as the time-varying Kalman filter. Hence, this paper focuses on the distributed implementation of the centralized steady-state Kalman filter. In contrast to most of the existing algorithms, we decouple the local filter from the consensus process. Such decoupling enables us to provide a new framework for designing distributed estimators, by reformulating the problem of distributed state estimation into that of linear system synchronization. We, hence, are able to leverage the methodologies from latter field to propose solutions for distributed estimation. To be specific, in the synchronization of linear systems, the dynamics of each agent is governed by an LTI system, the control input to which is generated using the local information within the neighborhood, in order to achieve asymptotic consensus on the local states of agents. Over the past years, lots of research efforts have been devoted to this area (see [31]–[36] for examples) by designing synchronization algorithms that can handle various network constraints. Exploiting the results therein, the distributed estimator in this work is designed through two phases as below: 1) (Local measurement processing) A lossless decomposition of steady-state Kalman filter is proposed, where each sensor node runs a local estimator based on this decomposition using solely its own measurement. 2) (Information fusion via consensus) The sensor infers the local estimates of all the others via a modified consensus algorithm designed for achieving linear system synchronization. The contributions of this paper are summarized as follows: 1) By removing assumptions regarding the eigenvalues of system matrix, this paper extends, in a non trivial way, the results in [37], and thus develops the local filters for losslessly decomposing Kalman filter in estimating the general systems. (Lemma 3) 2) Through the decomposition of Kalman filter, this paper bridges two different fields and makes it possible to leverage a general class of algorithms designed for achieving the synchronization of linear systems to solve the problem of distributed state estimation. By doing so, we can propose stable distributed estimators under different communication constraints, such as time delay, switching topology, random link failures, etc. (Theorem 4) 3) For certain synchronization algorithm, e.g., [31], the stability criterion of the proposed estimator is established. Moreover, in contrast to the existing literature, the covariance of the estimation error can be exactly derived by solving Lyapunov equations. (Theorem 2, Theorem 3, and Corollary 1) 4) The designed estimator enjoys low communication cost, where the size of message sent by each sensor is min _m, n_ _,_ _{_ _}_ with n and m being dimensions of the state and measurement respectively. (Remark 6) Some preliminary results are reported in our previous work [38], where most of the proofs are missing. This paper further extends the results in [38] by computing the exact asymptotic error covariance, instead of only showing the stability of proposed algorithms The extension to the more general random communication topology is also added. Moreover, a model reduction method is further proposed in this work to reduce the message complexity from m to min _m, n_ . _{_ _}_ _Notations: For vectors vi ∈_ R[m][i] _, the vector_ �v1[T] _[, . . ., v]N[T]_ �T is defined by col(v1, . . ., vN ). Moreover, A ⊗ _B indicates_ the Kronecker product of matrices A and B. Throughout this paper, we define a stochastic signal as “stable” if its covariance is bounded at any time. The remainder of this paper is organized as follows. Section II introduces the preliminaries and formulates the problem of interest. A lossless decomposition of optimal Kalman filter is given in Section III, where a model reduction approach is further proposed to reduce the system order. With the aim of realizing the optimal Kalman filter, distributed solutions for state estimation are given and analyzed in Section IV. We then discuss some extensions in Section V and validate performance of the developed estimator through numerical examples in Section VI. Finally, Section VII concludes the paper. II. PROBLEM FORMULATION In this paper, we consider the LTI system as given below: _x(k + 1) = Ax(k) + w(k),_ (1) where x(k) ∈ R[n] is the system state, w(k) ∼N (0, Q) is independent and identically distributed (i.i.d) Gaussian noise with zero mean and covariance matrix Q 0. The initial _≥_ state x(0) is also assumed to be Gaussian with zero mean and covariance matrix Σ 0, and is independent from the process _≥_ noise _w(k)_ . _{_ _}_ A network consisting of m sensors is monitoring the above system. The measurement from each sensor i 1, _, m_ _∈{_ _· · ·_ _}_ is given by [1]: _yi(k) = Cix(k) + vi(k),_ (2) where yi(k) ∈ R is the output of sensor i, Ci is an ndimensional row vector, and vi(k) ∈ R is the Gaussian measurement noise. By stacking the measurement equations, one gets _y(k) = Cx(k) + v(k),_ (3) where _y1(k)_ ... _ym(k)_   _C1_ ... _Cm_   _v1(k)_ ... _vm(k)_   _, C ≜_   _, v(k) ≜_  , (4) _y(k) ≜_   and v(k) is zero-mean i.i.d. Gaussian noise with covariance _R_ 0 and is independent from w(k) and x(0). _≥_ Throughout this paper, we assume that (A, C) is observable. On the other hand, (A, Ci) may not necessarily be observable, i.e., a single sensor may not be able to observe the whole state space. 1The results in this paper can be readily generalized to cases where the sensor outputs a vector measurement, by treating each entry independently as l t ----- _A. Preliminaries: the centralized Kalman filter_ If all measurements are collected at a single fusion center, the centralized Kalman filter is optimal for state estimation purpose, and provides a fundamental limit for all other estimation schemes. For this reason, this part will briefly review the centralized solution given by the Kalman filter. Let us denote by P (k) the error covariance of estimate given by Kalman filter at time k. Since (A, C) is observable, it is well-known that the error covariance will converge to the steady state [6]: _P = lim_ (5) _k→∞_ _[P]_ [(][k][)][.] Since the operation of a typical sensor network lasts for an extended period of time, we assume that the Kalman filter is in the steady state, or equivalently Σ = P, which results in a steady-state Kalman filter with fixed gain[2] _K = PC_ _[T][ �]CPC_ _[T]_ + R�−1 . (6) Accordingly, the optimal Kalman estimate is computed recursively as _xˆ(k + 1) = Axˆ(k) + K(y(k + 1)_ _CAxˆ(k))_ _−_ (7) = (A _KCA)ˆx(k) + Ky(k + 1)._ _−_ It is clear that the optimal estimate (7) requires the information from all sensors. However, in a distributed framework, each sensor is only capable of communicating with immediate neighbors, rendering the centralized solution impractical. Therefore, this paper is devoted to the implementation of Kalman filter in a distributed fashion. III. DECOMPOSITION OF KALMAN FILTER In this section, we shall provide a local decomposition of the Kalman filter (7), where the Kalman estimate can be recovered as a linear combination of the estimates from local filters. This section extends, in a non-trivial way, the results in [37] by removing the assumptions on the eigenvalues of system matrix therein, and thus proposes the local filter for estimating the general systems. The results in this part would further help us to design distributed estimation algorithms in the next sections. Without loss of generality, let the system matrix be _A. Local decomposition of Kalman filter_ To locally decompose Kalman filter, we first introduce the following lemmas, the proofs of which are given in appendix: **Proposition 1. If Λ is a non-derogatory[3]** _Jordan matrix, then_ _both (Λ, 1) and (Λ[T]_ _, 1) are controllable._ **Lemma 1. Let (X, p) be controllable, where X ∈** R[n][×][n] _and_ _p ∈_ R[n]. For any q ∈ R[n], if X + pq[T] _and X do not share any_ _eigenvalues, then (X +pq[T]_ _, q[T]_ ) is observable, or equivalently (X _[T]_ + qp[T] _, q) is controllable._ **Lemma 2. Let (X, p) be controllable, where X ∈** R[n][×][n] _and_ _p ∈_ R[n]. Denote the characteristic polynomial X as ϕ(s) = det(sI − _X). Let Y ∈_ R[m][×][m] _and q ∈_ R[m]. Suppose that _ϕ(Y )q = 0,_ (11) _then there exists T ∈_ R[m][×][n] _which solves the equation below:_ _TX = Y T, Tp = q._ (12) With the above preparations, let us consider the optimal Kalman estimate in (7). For simplicity, we denote by Kj the j-th column of the Kalman gain K. Namely, K = [K1, · · ·, Km]. Accordingly, (7) can be rewritten as _xˆ(k + 1) = (A_ _KCA)ˆx(k) +_ _−_ _m_ � _Kiyi(k + 1)._ (13) _i=1_ �Au _A =_ _A[s]_ Notice that A _KCA is stable. It is clear that we can always_ _−_ find a Jordan matrix Λ ∈ R[n][×][n], such that Λ is strictly stable, non-derogatory and has the same characteristic polynomial of _A_ _KCA. In view of Proposition 1, we conclude that (Λ, 1)_ _−_ is controllable. Therefore, by Lemma 2, we can always find matrices Fi’s, such that the following equalities hold for i = 1, _, m:_ _· · ·_ _FiΛ = (A −_ _KCA)Fi, Fi1n = Ki._ (14) Suppose each sensor i performs the following local filter solely based on its own measurements: _ξˆi(k + 1) = Λˆξi(k) + 1n yi(k + 1),_ (15) where _ξ[ˆ]i(k) is the output of local filter from sensor i, and_ **1n ∈** R[n] is a vector of all ones. Then it is proved the optimal Kalman filter can be decomposed as a weighted sum of local estimates _ξ[ˆ]i(k)’s, as stated below:_ **Lemma 3. Suppose each sensor performs the local filter (15).** _The optimal Kalman estimate (7) can be recovered from the_ _local estimates_ _ξ[ˆ]i(k), i = 1, 2, · · ·, m as_ � _,_ (8) where A[u] _∈_ R[n][u][×][n][u] and A[s] _∈_ R[n][s][×][n][s], such that any eigenvalue of A[u] lies on or outside the unit circle while the eigenvalues of A[s] are strictly within the unit circle. It thus follows from (1) that _x[s](k + 1) = A[s]x[s](k) + Jw(k),_ (9) where J = �0 **1ns** [�] _∈_ R[n][s][×][n] and x(k) = col(x[u](k), x[s](k)). Accordingly, Ci is partitioned as _Ci =_ �Ciu _Ci[s]�_ _,_ (10) where Ci[u] _[∈]_ [R][n][u][×][1][ and][ C]i[s] _[∈]_ [R][n][s][×][1][.] 2Notice that even if Σ ̸= P, the Kalman estimate converges to the steadyt t K l filt i th t d t t ti t i t ti ll ti l _Proof. By multiplying both sides of the recursive equation_ (15) by Fi, we arrive at _Fiξ[ˆ]i(k + 1) = FiΛξ[ˆ]i(k) + Fi 1n yi(k + 1)._ (17) 3A matrix is defined to be non-derogatory if every eigenvalue of it has t i lti li it 1 _xˆ(k) =_ _where Fi is defined in (14)._ _m_ � _Fiξ[ˆ]i(k),_ (16) _i=1_ ----- Then it follows from (14) that _Fiξ[ˆ]i(k + 1) = (A −_ _KCA)Fiξ[ˆ]i(k) + Kiyi(k + 1),_ (18) Summing up the above equation for all i = 1, _, m and_ _· · ·_ comparing it with (13), we can conclude that (16) holds. Notice that the equality in Lemma 3 surely holds. That means the Kalman filter can be perfectly recovered by (16). We hence claim that (16) is a lossless decomposition of optimal Kalman filter. To better illustrate the ideas, the information flow of centralized Kalman filter and local decomposition (16) is given in Fig 2. _y1(k)_ _· · ·_ _ym(k)_ _y1(k)_ _· · ·_ _ym(k)_ Kalman filter _xˆ(k)_ _ξˆ1(k)_ _ξˆm(k)_ Weighted sum _xˆ(k)_ Fig. 2: The information flow of centralized Kalman filter (left hand), and local decomposition of Kalman filter (16) (right hand). _B. A reformulation of (15) with stable inputs_ It is noted that the system matrix A may be unstable which implies that the covariance of measurement y(k) is not necessarily bounded. As a result, we need to redesign (15) using the stable residual zi(k) as an input instead of the raw measurement yi(k). The main reason for this reformulation is to make the consensus algorithm feasible and develop stable distributed estimators, which will be further discussed in the proof of Theorem 3. Towards the end, notice that (Λ, 1) is controllable, Λ is stable and any eigenvalue of Au is unstable. Hence, we can always find a non-zero β ∈ R[n] and compute _S = Λ + 1β[T]_ _,_ (19) such that 1) the characteristic polynomial of _A[u]_ divides _φ(s),_ where φ(s) is the characteristic polynomial of S, and _φ(s)/ det(sI_ _A[u]) has only strictly stable roots;_ _−_ 2) S do not share eigenvalues with Λ. Hence, by the virtue of Lemma 1, (S[T] _, β) is controllable._ **Remark 1. Notice that by using β, we place the eigenvalues** _of S to the locations which consist of two parts: the unstable_ _ones that coincide with the eigenvalues of Au and the stable_ _ones that are freely assigned but cannot be the eigenvalues of_ Λ. This is feasible as (Λ, 1) is controllable. Next, let us consider the filter below: _zi(k) = yi(k + 1) −_ _β[T][ ˆ]ξi(k),_ (20) _ξˆ (k_ 1) _Sξ[ˆ] (k)_ **1** (k) where β and S are calculated through (19). In the following lemma, we shall show that (20) also losslessly decomposes the Kalman filter. Moreover, the covariance of zi(k) is bounded at any time. **Lemma 4. Consider the local filter (20). The following** _statements hold at any instant k:_ _1) (20) has the same input-output relationship with (15)._ _Namely, given the input yi(k), they yield the same output_ _ξˆi(k);_ _2) zi(k) is stable, i.e., the covariance of zi(k) is always_ _bounded._ _Proof. The proof is given in Appendix-C._ **Remark 2. If A has unstable modes, the previous discussions** _show that (15) can be seen as a linear system with stable_ _system matrix Λ but unstable input yi(k + 1). As a contrast,_ (20) has unstable system matrix S but stable input zi(k). _This formulation is essential to guarantee the stability of local_ _estimators, as will be seen in the proof of Theorem 4._ _C. A reduced-order decomposition of Kalman filter when n <_ _m_ To simplify notations, we define the following aggregated matrices: _S˜ ≜_ _Im ⊗_ _S, ˜Li ≜_ _ei ⊗_ **1n, ˜L ≜** [˜L1, · · ·, ˜Lm] = Im ⊗ **1n,** (21) where Im is an m-dimensional identity matrix and ei is the _ith canonical basis vector in R[m]. We thus collect (16) and_ (20) in matrix form as:  _ξˆ1(k + 1)_   _ξˆ1(k)_   _z1(k)_  = ˜S + ˜L _,_ ... ... ...       _ξˆm(k + 1)_ _ξˆm(k)_ _zm(k)_ (22)  _ξˆ1(k)_  _xˆ(k) = F_ _._ ...   _ξˆm(k)_ where F ≜ [F1, F2, · · ·, Fm]. By Lemmas 3 and 4, (22) represents a lossless decomposition of Kalman filter. Notice that the system order of (22) is mn. In this part, we shall show that by performing model reduction, this order can be further reduced to n[2] when the state dimension is less than the number of sensors, namely n < m. These discussions would be useful for us to achieve a low communication complexity in distributed frameworks. To proceed, we regard the input and output of (22) as z(k) and ˆx(k), respectively, where _z(k) ≜_ �z1(k), · · ·, zm(k)�T . (23) Let us introduce the below lemma, the proof of which is given in Appendix-D: **Lemma 5. Any matrix W ∈** R[n][×][n] _can be decomposed as_ _W_ _H ϕ (S) + H ϕ (S) +_ + H ϕ (S) (24) ----- _where Hi ≜_ _eiβ[T]_ _, {ϕj(S)} are certain polynomials of S, and_ _S and β are given in (19)._ As a direct result of Lemma 5, for any Fi in (16), we can always rewrite it by using the polynomials of S, i.e., {pij(S)}: _Fi =_ _n_ � _Hjpij(S)._ (25) _j=1_ For simplicity, we also denote _Ti ≜_ [(pi1(S) 1n)[T] _, · · ·, (pin(S) 1n)[T]_ ][T] _._ (26) It is then proved in the below theorem that system (22) can be reduced with a less order: |yi(k) yj(k)|Col2|Col3| |---|---|---| |Local filter Local filter ξˆ i(k) ξˆ j(k) z i(k) z j(k)||| |Synchronization Linear system ∆ i(k) Linear system η i(k) ∆ j(k) η j(k)||| |||| **Theorem 1. Consider the following system:** θ1(k + 1) θ1(k)   _..._  = (In ⊗ _S)_  _..._  + T  _θn(k + 1)_ _θn(k)_ θ1(k) _x˜(k) = H_ _..._ _,_   _θn(k)_ _z1(k)_ _..._ _zm(k)_   _,_ (27) _x˘i(k)_ _x˘j(k)_ Fig. 3: The information flow of Algorithm 1, where nodes i and j are immediate neighbors. if and only aij > 0. By denoting the degree matrix as D ≜ diag (deg1, . . ., degN ) with degi = [�]j[N]=1 _[a][ij][,][ the Laplacian]_ matrix of G is defined as LG ≜ _D −A. In this paper, a_ connected network is considered. We therefore can arrange the eigenvalues of Laplacian matrix as 0 = µ1 < µ2 _µm._ _≤· · · ≤_ _A. Description of the distributed estimator_ In light of (16), the optimal estimate fuses _ξ[ˆ]i(k) from all_ sensors. However, in a distributed framework, each sensor can only access the information in its neighborhood. Hence, any sensor i needs to, through the communication over network, infer _ξ[ˆ]j(k) for all j ∈V to achieve a stable local estimate._ Let us denote by ηi,j(k) as the inference from sensor i on sensor j. As will be proved later in this section, ηi,j(k), by running a synchronization algorithm, can track _m1_ _[ξ][ˆ][j][(][k][)]_ with bounded error. Hence, every sensor i can make a decent inference on _ξ[ˆ]j(k)._ By collecting its inference on all sensors together, each sensor i keeps a local state as below: _where_ _T = [T1, T2, · · ·, Tm], H = [H1, H2, · · ·, Hn]._ (28) _It holds that system (27) shares the same transfer function with_ (22). _Proof. The proof is presented in Appendix-E._ Therefore, by performing model reduction, we present system (27) which shares the same transfer function with (22) but with a reduced order. As proved previously, the output of (22) is the optimal Kalman estimate. As a result, (27) also has the Kalman estimate as its output and the Kalman filter can be perfectly recovered by (27) as well. We hereby refer both (22) and (27) to lossless decomposition of Kalman fiter. Depending on the size of m and n, one should use a system with smaller dimension to represent the centralized Kalman filter. IV. LOCAL IMPLEMENTATION OF KALMAN FILTER From Fig. 2, it is clear that local decomposition proposed in Section III is still centralized as a fusion center is required for calculating the weighted sum. In this section, we shall provide distributed algorithms for implementing it, where each sensor node performs local filtering by using the results from Section III, and global fusion by exchanging information with neighbors and running consensus algorithm. Based on whether n is greater than m or not, different algorithms will be presented to achieve a low communication complexity. We use a weighted undirected graph = _,_ _,_ to _G_ _{V_ _E_ _A}_ model the interaction among nodes, where = 1, 2, ..., m _V_ _{_ _}_ is the set of sensors, is the set of edges, and _E ⊂V × V_ _A = [aij] is the weighted adjacency matrix. It is assumed_ _aij ≥_ 0 and aij = aji, ∀i, j ∈V. An edge between sensors i and j is denoted by eij ∈E, indicating that these two agents can communicate directly with each other Note that e _∈E_ _ηi,1(k)_ ... _ηi,m(k)_  _∈_ Rmn, (29)  _ηi(k) ≜_   which will be updated by synchronization algorithms. Since _ηi(k) contains the fair inference on all_ _ξ[ˆ]j(k), j ∈V, sensor i_ finally uses it to compute a stable local estimate. To be concrete, let us define the message sent by agent i at time k as ∆i(k) ≜ Γ[˜]ηi(k) ∈ R[m], where Γ =[˜] _Im ⊗_ Γ and Γ is a design parameter to be given later. We are now ready to present the main algorithm. Suppose each node i is initialized with ˆxi(0) = 0 and ηi(0) = 0. At any instant k > 0, its update is outlined in Algorithm 1, the information flow of which is shown in Fig. 3. Compared with Fig. 2, the proposed algorithm is achieved in a distributed manner. **Remark 3. Instead of transmitting the raw estimate ηi(k) ∈** R[mn], each agent sends a “coded” vector ∆i(k), with a _smaller size m._ ----- **Algorithm 1 Distributed estimation algorithm for sensor i** 1: Using the latest measurement from itself, sensor i computes the local residual and update the local estimate by _zi(k) = yi(k + 1) −_ _β[T][ ˆ]ξi(k),_ (30) _ξˆi(k + 1) = S ˆξi(k) + 1n zi(k)._ 2: Compute ∆i(k) = Γ[˜]ηi(k) and collect ∆j(k) from neighbors and fuse the neighboring information with the consensus algorithm as _m_ � _ηi(k + 1) = Sη[˜]_ _i(k) + L[˜]izi(k) + B[˜]_ _aij(∆j(k) −_ ∆i(k)), _j=1_ (31) where _S[˜] and_ _L[˜]i are given in (21), and_ _B[˜] ≜_ _Im_ **1n.** _⊗_ 3: Update the fused estimate on system state as: _x˘i(k + 1) = mFηi(k + 1)._ (32) 4: Transmit the new state ∆i(k + 1) to neighbors. _B. Performance analysis_ This part is devoted to the performance analysis of Algorithm 1. We shall first provide the following theorem: **Theorem 2. With Algorithm 1, the average of fused estimates** _from all sensors coincides with the optimal Kalman estimate_ _at any instant k. That is,_ 1 _m_ _m_ � _x˘i(k) = ˆx(k), ∀k ≥_ 0. (33) _i=1_ _Proof. Summing (31) over all i = 1, 2, ..., m yields_ _m_ � _ηi(k + 1) = S[˜]_ _i=1_ _m_ � _ηi(k) +_ _i=1_ _m_ � _L˜izi(k),_ (34) _i=1_ where we use the fact that aij = aji for any i, j ∈V. Comparing it with (20), it holds for any instant k and any _j_ that: _∈V_ _m_ _ξˆj(k) =_ � _ηi,j(k)._ (35) _i=1_ Therefore, the following equation is satisfied at any k 0: _≥_ _m_ � _Fηi(k) =_ _i=1_ _m_ � � �[m] � _Fj_ _ηi,j(k)_ = _j=1_ _i=1_ _m_ � _i=1_ _m_ � _Fjηi,j(k)_ _j=1_ 1 _m_ _m_ � _x˘i(k) =_ _i=1_ = number (the ratio of the maximum and minimum nonzero eigenvalues of the Laplacian matrix): **Lemma 6. Suppose that the product of all unstable eigenval-** _ues of matrix S meets the following condition:_ � _|λ[u]j_ [(][S][)][|][ <][ 1 +][ µ][2][/µ][m] _,_ (37) _j_ 1 − _µ2/µm_ _where λ[u]j_ [(][S][)][ represents the][ j][th unstable eigenvalue of][ S][. Let] Γ = 2 **_1[T]n_** _[P][S]_ _∈_ R[1][×][n], (38) _µ2 + µm_ **_1[T]n_** _[P][ 1][n]_ _where µ2 and µm are, respectively, the second smallest and_ _largest eigenvalues of LG. Moreover, P > 0 is the solution to_ _the following modified algebraic Riccati inequality_ _S[T]_ _S +_ �1 _ζ_ [2][�] _[S][T][ P][ 1][n][ 1]n[T]_ _[P][S]_ _> 0,_ (39) _P −_ _P_ _−_ **_1[T]n_** _[P][ 1][n]_ _with ζ satisfying_ [�]j ��λuj [(][S][)]�� _< ζ_ _−1 ≤_ 11+−µµ22/µ/µmm _[.][ Then for]_ _any j_ 2, ..., n _, it holds that_ _∈{_ _}_ _ρ(S −_ _µj 1n Γ) < 1._ (40) _Proof. For any j ∈{2, ..., n}, let us denote ζj = 1−2µj/(µ2+_ _µm) ≤_ _ζ. Since (S, 1n) is controllable, there exists some P >_ 0 which solves (39). Together with (38), it holds that (S − _µj 1n Γ)[T]_ _P(S −_ _µj 1n Γ) −P_ =S[T] _PS −_ (1 − _ζj[2][)]_ _[S][T][ P][ 1][n][ 1]n[T]_ _[P][S]_ _−P_ **1[T]n** _[P][ 1][n]_ (41) _S[T]_ _S_ (1 _ζ_ [2]) _[S][T][ P][ 1][n][ 1]n[T]_ _[P][S]_ _< 0._ _≤_ _P_ _−_ _−_ _−P_ **1[T]n** _[P][ 1][n]_ Hence, our proof completes. **Remark 4. Note that, if all the eigenvalues of S lie on or** _outside the unit circle, You et al. [31] prove that (40) holds_ _if and only if (37) is satisfied. In Lemma 6, we further show_ _that, (37) is still a sufficient condition to facilitate (40) if S_ _has stable modes._ **Remark 5. Invoking Remark 1, each λ[u]j** [(][S][)][ corresponds to] _a root of the characteristic polynomial of A[u]. Thus, the_ _condition (37) can be rewritten using the system matrix A[u],_ � _|rj(A[u])| <_ [1 +][ µ][2][/µ][m] _,_ (42) _j_ 1 − _µ2/µm_ _where rj(A[u]_ _is a root of the characteristic polynomial of A[u]._ With the above preparations, we are now ready to analyze the error covariance of local estimator as below: **Theorem 3. Suppose that the Mahler measure of S meets** _condition (37), and Γ is designed based on (38)–(39). With_ _Algorithm 1, the error covariance of each local estimate ˘xi(k)_ _is bounded at any instant k._ _Proof. Due to space limitation, the proof is given in Appendix-_ F _m_ � _Fjξ[ˆ]j(k) = ˆx(k)._ _j=1_ (36) This completes the proof. On the other hand, in order to show the stability of proposed estimator, it is also desired to prove the boundedness of error covariance. Towards this end, we introduce the following lemma, the condition of which is characterized in terms of a certain relation between the Mahler measure (the absolute product of unstable eigenvalues of S) and the graph condition ----- The proof of Theorem 3 implies that we present a distributed estimation scheme with quantifiable performance. **Corollary 1. Suppose that the Mahler measure of S meets** _condition (37), and Γ is designed based on (38)–(39). Let_ _W[˘]_ _be the asymptotic error covariance of local estimates. Namely,_ _W˘_ ≜ lim _k→∞_ [cov(˘][e][(][k][))][,] _where ˘e(k) ≜_ col[(˘x1(k) − _x(k)), · · ·, (˘xm(k) −_ _x(k))]. By_ _using Algorithm 1, it holds that_ _W˘_ = ¯W + (1m 1[T]m[)][ ⊗] _[P,]_ (43) _where_ _W[¯]_ _is the asymptotic error covariance between local_ _estimate and the Kalman estimate, and P is the error covari-_ _ance of Kalman filter as defined in (5). Moreover,_ _W[˘]_ _can be_ _exactly calculated._ As seen from the calculation, _W¯_, i.e., the performance gap between our estimator and the optimal Kalman filter, is purely caused by the consensus error. Therefore, if infinite consensus steps are allowed between two consecutive sampling instants, the consensus error vanishes and the performance of the proposed estimator coincides with that of the Kalman filter. Combining Theorems 2 and 3, the local estimator is stable at each sensor side. Therefore, we conclude that by applying the algorithm designed for linear system synchronization, i.e., (31), the problem of distributed state estimation is resolved. **Remark 6. Note that Algorithm 1 requires each agent to send** _out an m-dimensional vector ∆i(k) at any time. Therefore, in_ _the network with a large number of sensors, i.e., n < m, this_ _solution will cause a high communication cost. To address this_ _issue, this remark, by leveraging the reduced-order estimator_ (27) in Theorem 1, modifies Algorithm 1 to introduce less com_munication complexity. To be specific, we aim to implement the_ _reduced order system (27) with distributed estimators. Similar_ _as before, any agent i stores its estimate on all the others in_ _a variable ϑi(k), where_ **Algorithm 2 Distributed estimation algorithm 2 for sensor i** 1: Using the latest measurement from itself, sensor i computes the local residual and update the local estimate by _zi(k) = yi(k + 1) −_ _β[T][ ˆ]ξi(k),_ _ξˆi(k + 1) = S ˆξi(k) + 1n zi(k)._ 2: Compute ∆i(k) = (In ⊗ Γ)ϑi(k) such that Γ is calculated by (38). Collect ∆j(k) from neighbors and fuse the neighboring information with the consensus algorithm as _ϑi(k + 1) =(In ⊗_ _S)ϑi(k) + Tizi(k)_ _m_ � (45) + (In ⊗ **1n)** _aij(∆j(k) −_ ∆i(k)), _j=1_ _ϑi,1(k)_ _..._ _ϑi,n(k)_ where Ti is defined in (26). 3: Update the fused estimate on system state as: _x˘i(k + 1) = mHϑi(k + 1),_ (46) where H is given in (28). 4: Transmit the new state ∆i(k + 1) to neighbors. V. EXTENSIONS OF PROPOSED SOLUTIONS In the previous sections, we leverage the linear system synchronization algorithm proposed in [31], to solve the problem of distributed state estimation. In this section, we aim to extend such a result and show that any control strategy, which can facilitate the linear system synchronization, can be modified to yield a stable distributed estimator. As a result, we bridge the fields of distributed state estimation and linear system synchronization. Let us consider the synchronization of the following homogeneous LTI system: _ηi(k + 1) = Sη[˜]_ _i(k) + Bu[˜]_ _i(k), ∀i ∈V,_ (47) where ui(k) is the control input of agent i. In literature, a large variety of synchronization algorithms has been proposed with the framework below: _ωi(k + 1) = Aωi(k) + Bηi(k + 1),_ ∆i(k) = Γ[˜]ωi(k), _m_ (48) � _ui(k) =_ _aijγij(k)(∆j(k) −_ ∆i(k)), _j=1_ where ωi(k) is the “hidden state” that is necessary for agent i to yield the communication state ∆i(k) and input ui(k), and Γ[˜] refers to the control gain. Notice that (48) can be used to model the controller with memory. Moreover, γij(k) ∈ [0, 1] models the fading or lossy effect of the communication channel from agent j to agent i. At every time, the agent collects the available information in its neighborhood and synthesizes its communication state and control signal via (48). For simplicity, we denote as the control strategy that can _U_ be represented by (48). Let the average of local states at time _k be_ _m_ � _η¯(k) = [1]_ _ηi(k)._ _m_  _n[2]_ _∈_ R _._ (44)  _ϑi(k) ≜_   _For each sensor i, it is initialized with ˆxi(0) = 0 and ϑi(0) =_ 0. For the case of n < m, the estimation algorithm works as in _Algorithm 2. Following similar arguments, the local estimator_ _at each sensor side is proved to be stable._ _Combining it with Algorithm 1, we conclude the size of_ _message sent by each sensor at any time is min_ _m, n_ _. Com-_ _{_ _}_ _pared with the existing solutions in distributed estimation, e.g.,_ _[12]–[16], our algorithm enjoys lower message complexity._ **Remark 7. Notice that sensor node i has perfect information** _of its own local estimate ξi(k). Therefore, instead of using_ _ηi,i(k) to infer ξi(k)/m, node i can just use ξi(k)/m to_ _replace ηi,i(k) in (32), which potentially improves the per-_ _formance of the estimators._ ----- The network of subsystems (47) reaches strong synchronization under, if the following statements hold at any time: _U_ 1) Consistency: the average of local states keeps consistent throughout the execution, i.e., _η¯(k + 1) = S[˜]η¯(k)._ (49) 2) Exponential Stability: agents exponentially reach consensus in mean square sense, i.e., there exist c > 0 and _ρ_ (0, 1) such that _∈_ E[||ηi(k) − _η¯(k)||[2]] ≤_ _cρ[k], ∀i ∈V._ (50) We now review several existing strategies which facilitate the strong synchronizationand show that they can be represented by (48): 1) Let ∆i(k) = Γ[˜]ηi(k) be the communication state defined in Section IV-A. To facilitate the synchronization of homogeneous linear systems in undirected communication topology, You et al. [31] design the following control law: _yields a stable estimator for each sensor node. Specifically,_ _the following statements hold for any k_ 0: _≥_ _1) the average of local estimates from all sensor coincides_ _with the optimal Kalman estimate;_ _2) the error covariance of each local estimate is bounded._ _Proof. The proof is given in Appendix-H._ **Remark 8. Theorem 4 assumes the independence of the com-** _munication topology and system/measurement noises. There-_ _fore, as for the event-based synchronization algorithms, where_ _the communication relies on the agents’ states, we cannot_ _analyze its efficiency of solving the distributed estimation_ _problem by directly resorting to Theorem 4. In the future work,_ _we will continue to investigate this topic._ In contrast with Fig 1, this work, by using the lossless decomposition of Kalman filter, decouples the local filter from the consensus process, as shown in Fig. 3. The decoupling enables us to leverage the rich results in linear systems synchronization to analyze the performance of local estimators, as proved in Theorem 4. Moreover, following the similar proof arguments as that of Theorem 3, we can show that with our framework, the error covariance of each local estimate actually consists of two orthogonal parts: the inherent estimation error of Kalman filter and the distance from local estimate to Kalman filter, namely: cov(˘ei(k)) = cov(˘xi(k) − _x(k))_ = cov(˘xi(k) − _xˆ(k) + ˆx(k) −_ _x(k))_ = cov(˘xi(k) − _xˆ(k)) + cov(ˆx(k) −_ _x(k))_ _m_ � � � = cov _x˘i(k) −_ [1] _x˘i(k)_ + cov(ˆx(k) − _x(k))_ _m_ _i=1_ _m_ =m[2]F cov �ηi(k) − [1] � _ηi(k)�F_ _[T]_ + cov(ˆx(k) − _x(k)),_ _m_ _i=1_ where the third equality holds due to the optimality of Kalman filter, and the last equality holds by (32). Notice that the second term of RHS is the error covariance of Kalman filter, while first term is the error between local estimate and Kalman filter and purely determined by the consensus process. Therefore, by choosing proper strategy, extensive results on achieving _U_ strong synchronization can be applied to (53) to deal with the consensus error in various settings, such as directed graph, time-varying topologies, etc. Particularly, if infinite consensus steps are allowed between two consecutive sampling instants, the consensus error vanishes, i.e., ηi(k) − _m[1]_ �mi=1 _[η][i][(][k][) = 0][,]_ and the performance of the proposed estimator is optimal since it coincides with that of the Kalman filter. That means the global optimality can be guaranteed. VI. NUMERICAL EXAMPLE In this section, we present numerical examples to verify the theoretical results obtained in previous sections _ui(k) =_ _m_ � _aij(∆j(k) −_ ∆i(k)), (51) _j=1_ which coincides with (48). 2) Another example is the filtered consensus protocol given in [34]. By designing the hidden state as _ωi(k) = F_ (q)ηi(k), (52) where q is the unit advance operator, i.e., q[−][1]s(k) = _s(k_ 1), and F (z) is the transfer function of a square _−_ stable filter, the synchronization of linear systems is achieved by (48) under a more relaxed condition than (37), that is: [�]j _[|][λ]j[u][(][S][)][|][ <]_ 11+−[√][√]µµ22/µ/µmm _[.]_ 3) Instead of focusing on perfect communication channels, the authors in [32] and [33] develop the control protocols to account for the random failure on communication links and Markovian switching topologies, respectively. By modeling the packet loss with the Bernoulli random variable γij(k) ∈{0, 1}, these works complement the results in [31] and prove the mean square stability under the control strategy (48). Notice that Algorithms 1 and 2 utilize (51) for achieving synchronization and producing stable distributed estimators. In what follows, we argue that the optimal Kalman estimate can indeed be distributively implemented using any linear system synchronization algorithms facilitating (49)-(50). To be specific, Algorithm 1 should be modified[4] by replacing (31) with _ηi(k + 1) = Sη[˜]_ _i(k) + Bu[˜]_ _i(k) + L[˜]izi(k),_ (53) where ui(k) is generated by U that facilitates (49)-(50). We then state the stability of local estimators as below: **Theorem 4. Consider any algorithm U which facilitates the** _statements (49) and (50). At any time k, suppose each γij(k)_ _is independent of the noise_ _w(k)_ _and_ _v(k)_ _. Then (53)_ _{_ _}_ _{_ _}_ 4Similarly, in the case of n < m, one can also derive the general form of Al ith 2 ith li t h i ti t t ----- _A. Numerical example when n < m_ Let us consider the case where four sensors cooperatively estimate the system state. The system parameters are listed below: _B. Numerical example when n > m_ In the second example, we simulate the heat transfer process 5 in a planar closed region discussed in [41] and [42]: � _,_ (56) �0.9 0 � �1 0 1 1 �T _A =_ _, C =_ _,_ 0 1.1 0 1 1 _−1_ (54) _Q = 0.25I2, R = 4I4._ In this example, the number of states is smaller than that of sensors, i.e. n < m. We therefore choose Algorithm 2. Moreover, notice that the system is unstable, and sensor 1 cannot observe the unstable state. Suppose that the topology of these four sensors is a ring with weight 1 for each edge. The Laplacian matrix is thus: _∂u_ � _∂2u_ + _[∂][2][u]_ _∂t_ [=][ α] _∂x[2]1_ _∂x[2]2_ with boundary conditions _∂x∂u1_ ���t,0,x2 [=][ ∂u]∂x1 ���t,l,x2 [=][ ∂u]∂x2 ���t,x1,0 [=][ ∂u]∂x2 ���t,x1,l [= 0][,] (57) where x1 and x2 are the coordinates in the region; u(t, x1, x2) indicates the temperature at time t at position (x1, x2), l is the side length of the square region and α adjusts the speed of the diffusion process. With a N _N grid and sample frequency_ _×_ 1Hz, the diffusion process can be discretized as: _u(k + 1, i, j)_ _u(k, i, j) =_ _[α]_ _−_ _h[2][ [][u][(][k, i][ −]_ [1][, j][) +][ u][(][k, i, j][ −] [1)] + u(k, i + 1, j) + u(k, i, j + 1) 4u(k, i, j)], _−_ (58) where h = _N_ _−l_ 1 [denotes the size of each grid and][ u][(][k, i, j][)] indicates the temperature at time k at location (ih, jh). By collecting all the temperature values of each grid, we define the state variable U (k) = [u(k, 0, 0), _, u(k, 0, N_ _· · ·_ _−_ 1), u(k, 1, 0), _, u(k, N_ 1, N 1))][T] . Further, by introduc_· · ·_ _−_ _−_ ing process noise into (58), one derives the following system equation: _U_ (k + 1) = AU (k) + w(k), (59) where w(k) (0, Q) is Gaussian noise. _∼N_ As shown in Fig. 5, m sensors are randomly deployed in this region to monitor the temperature, where the measurement of each sensor is a linear combination of temperature of the grids around it. Specifically, suppose the location of sensor _s is (ˆx1, ˆx2) such that ˆx1 ∈_ [i, i + 1), ˆx2 ∈ [j, j + 1), we define ∆ˆx1 = xi1 _i and ∆ˆx1 = xi2_ _j. We assume that the_ _−_ _−_ measurement of sensor s at time k is 2 1 0 1 _−_ _−_ 1 2 1 0 _−_ _−_ 0 1 2 1 _−_ _−_ 1 0 1 2 _−_ _−_  _._ (55)  _LG =_   It is not difficult to check that the second smallest and the largest eigenvalues of LG are respectively µ2 = 2, µ4 = 4. To fulfill the sufficient condition in Lemma 6, let us choose _ζ = 0.5._ We set the initial state x(0) (0, I) and the initial local _∼N_ estimate ˘xi(0) = 0 for each sensor i ∈{1, 2, 3, 4}. It can be seen that the mean squared local estimate error ei(k) enters steady state and is stable after a few steps (see Fig. 4). 1 KF 0.8 0.6 0.4 0.2 2.5 |Col1|Col2|Col3|s is (ˆx1, define ∆ˆx| |---|---|---|---| ||||define ∆ KF measurem s1 s2 y (k s s3 s4| ||||| ||||| ||||| ||||| We collect the measurements of each sensor at time k and denote it as Y (k), then it follows _Y (k) = CU_ (k) + v(k), (61) (60) _ys(k) = [1]_ �(1 − ∆ˆx1)(1 − ∆ˆx2)u(k, i, j) _h[2]_ + ∆ˆx1(1 − ∆ˆx2)u(k, i + 1, j) + (1 − ∆ˆx1)∆ˆx2u(k, i, j + 1) + ∆ˆx1∆ˆx2u(k, i + 1, j + 1)� + vs(k). 2 0 10 20 30 Time/s 0 10 20 30 1.5 1 |Col1|Time/s|Col3|We collec denote it| |---|---|---|---| ||||KF s1 s2 where v( s3 surement s4 for the si α =| ||||| ||||| ||||| Time/s Fig. 4: Average mean square estimation error of system states under Kalman filter and local estimators in 10000 experiments. where v(k) (0, R) is the measurement noise and the mea_∼N_ surement matrix C can be derived from (60). The parameters for the simulation are listed below: _• α = 0.2;_ _• l = 4 and N = 5, thus the grid size h = 1;_ _• n = N_ [2] = 25 and m = 15. Therefore, n > m, which is different from our first example.; 5State estimation in diffusion process has wide applications in sensor network, e.g., urban CO2 emission monitoring [39], temperature monitoring i d t t [40] t ----- Fig. 5: (a) The position and topology of m sensors in the N _×_ _N grid lines; (b) The estimate variance of centralized Kalman_ filter; (c) The estimate variance of local Kalman filter; (d) The estimate variance of our estimators in 10000 experiments. _• Q = 0.2I25 and R = 3I15._ As discussed in Remark 7, we replace ηi,i(k) with the estimates given by local Kalman filters. The results are shown in Fig. 5. Our algorithm achieves better performance compared with local Kalman filters which merely use the measurement of the sensor itself. The improvement of each sensor can be found in TABLE I. Specifically, for each sensor i, we respectively define the performance of local Kalman filter and our algorithm in terms of: _ϱi1 ≜_ [tr( ˆ][P][i][)] (62) tr(P ) _[, ϱ][i][2][ ≜]_ [tr( ˘]tr([P]P[i])[)] _[,]_ Fig. 6: Comparison of the mean square error of the estimates provided by different algorithms in 10000 experiments. In the example, m = 4 sensors are connected as a ring to infer the system state. Let the measurement equation be 1 0 0 0 1 0 0 0 1 0 0 0 0 0 1 0  _x(k) + v(k),_ (63)  _y(k) =_   where _P[ˆ]i,_ _P[˘]i and P are respectively the steady-state error_ covariance of local Kalman filter, our estimator and centralized Kalman filter. We see that the proposed scheme outperforms the local Kalman filter by at least 50% for each sensor. _C. Comparison with existing algorithms_ We further compare the performance of Algorithm 1 with those of existing algorithms: 1) centralized Kalman filter (CKF), 2) KCF2009 ([13]), and 3) CMKF2018 ([43]), through a numerical example on inverted pendulum. Notice that an inverted pendulum has n = 4 states: _x = [p;_ _p˙; θ;_ _θ˙], namely, the cart position, cart velocity,_ pendulum angle from vertical and pendulum angle velocity, respectively. We consider the system linearized at θ = θ[˙] = 0 and discretized with sampling interval T = 0.01s, where the detailed system equation can be found in [44] with system noise w(k) (0 0 05[2]I ) _N_ where v(k) ∼N (0, 0.3[2]Im). Notice that sensor 4 cannot fully observe the state space. Fig. 6 illustrates the mean square error (MSE) of its estimate on p and θ, respectively. The results show that our algorithm yields better estimation performance. _D. Experiment when the global knowledge on system matrix_ _is unavailable_ Finally, notice that the proposed distributed estimator is based on a lossless decomposition of Kalman filter as developed in Section III, which requires the global knowledge on 1) the system matrix A, 2) the measurement matrix C, and 3) noise covariance matrix Q and R. In the case that certain part of A, C, Q and R are unknown, before running Algorithm 1 or 2, each sensor can broadcast its local parameters. In this way, every sensor can obtain the system parameters it needs within finite steps. To quantify the overhead incurred by this initialization, i.e., broadcasting the parameters, in the third example, we conduct an experiment using m = 15 raspberry pis equipped with temperature sensors which run the proposed distributed estimation algorithm every minute. In our experiment, it is assumed that the sensors do not have global information on _C and R Thus let each of them broadcast its C_ and R ----- TABLE I: Performance Improvement in Comparison with Local Kalman filter Sensor index i 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Local KF performance ϱi1 1.94 1.94 1.96 1.96 1.94 1.93 1.97 1.94 1.95 1.94 1.92 1.94 1.95 1.94 1.95 Our estimator performance ϱi2 1.26 1.35 1.31 1.31 1.26 1.13 1.22 1.21 1.23 1.44 1.12 1.22 1.18 1.35 1.18 Improvement ϱi1 − _ϱi2_ 68% 59% 65% 65% 68% 80% 75% 73% 71% 50% 80% 72% 76% 59% 77% at the starting phase so that every sensor can obtain system parameters it needs. The mean traffic of a sensor with 3 neighbors is shown in Fig. 7. It turns out, compared with the centralized Kalman filter, our solution induces lower communication burden even with the additional effort on initial broadcasting. Obviously, the merits become more apparent with the increasing scale of sensor networks. Fig. 7: Mean network traffic v.s. time. VII. CONCLUSION In this paper, the problem of distributed state estimation has been studied for an LTI Gaussian system. We investigate both cases where m > n and m _n, and propose distributed_ _≤_ estimators for both cases to introduce low communication cost. The local estimator is proved to be stable at each sensor side, in the sense that the covariance of estimation error is proved to be bounded and the asymptotic error covariance can also be derived. Our major merit lays in reformulating the problem of distributed estimation to that of linear system synchronization. APPENDIX A PROOF OF LEMMA 1 We will prove by contradiction. If (X _[T]_ + qp[T] _, q) is not_ controllable, then we can find some s, such that the rank of �X _[T]_ + qp[T] _−_ _sI_ _q�_ is strictly less than n. Therefore, there exists a non-zero v, such that _v[T][ �]X_ _[T]_ + qp[T] _−_ _sI_ _q�_ = 0, which implies that (X + pq[T] )v _sv = 0, q[T]_ _v = 0._ _−_ Therefore (X+pq[T] )v _sv = 0 and Xv_ _sv = 0, implying that_ _−_ _−_ _s is an eigenvalue of both X and X + pq[T]_, which contradicts with the assumption X and X +pq[T] do not share eigenvalues. We thus complete the proof APPENDIX B PROOF OF LEMMA 2 We will prove this lemma by construction. Towards the end, let us next consider the following equation: _T_ [p, Xp, · · ·, X _[n][−][1]p] = TRX = [q, Y q, · · ·, Y_ _[n][−][1]q] = RY,_ (64) where _RX_ = [p, Xp, · · ·, X _[n][−][1]p]_ and _RY_ = [q, Y q, _, Y_ _[n][−][1]q]._ _· · ·_ Since (X, p) is controllable, RX is full rank and thus invertible, and T = RY RX[−][1] solves (64). Clearly Tp = q. In what follows, we shall prove that TX = Y T . To this end, let us denote the characteristic polynomial of X as _ϕ(s) = s[n]_ + αn−1s[n][−][1] + . . . α0. It is noted that _TX_ _[n]p = T_ (−αn−1X _[n][−][1]_ _−_ _αn−2X_ _[n][−][2]_ _−· · · −_ _α0I)p_ = (−αn−1Y _[n][−][1]q −_ _αn−2Y_ _[n][−][2]q −· · · −_ _α0q) = Y_ _[n]q,_ (65) where the first and the last equality is due to Carley-Hamilton and the second equality is from the fact TRX = RY . As a result _TXRX = T_ [Xp, · · ·, X _[n]p] = [Y q, · · ·, Y_ _[n]q] = Y RY,_ Hence, TX = Y RY RX[−][1] [=][ Y T] [, which finishes the proof.] APPENDIX C PROOF OF LEMMA 4 1) From (20), it is easy to verify that _Sξ[ˆ]i(k) + 1n zi(k)_ =(Λ + 1n β[T] )ξ[ˆ]i(k) + 1n[yi(k + 1) − _β[T][ ˆ]ξi(k)]_ (66) =Λξ[ˆ]i(k) + 1n yi(k + 1). As a result, the local filter (20) has the same input-output relationship with (15). 2) By Lemma 2, we know that for any i, we can find _∈V_ _G[u]i_ _[∈]_ [R][n][×][n][u] [, such that] (G[u]i [)][T][ S][T][ = (][A][u][)][T][ (][G]i[u][)][T][,][ (][G]i[u][)][T][ β][ = (][C]i[u][A][u][)][T][,] which implies that _G[u]i_ _[A][u][ −]_ **[1][n][C]i[u][A][u][ =][ SG]i[u]** _[β][T][ G][u]i_ _[−]_ **[1][n]** = (Λ + 1β[T] )G[u]i _[β][T][ G][u]i_ [= Λ][G]i[u][,] _[−]_ **[1][n]** _β[T]_ _G[u]i_ [=][ C]i[u][A][u][.] (67) Furthermore, �Gui 0[�] _A −_ **1nCiA =** �Gui _[A][u]_ 0[�] _−_ **1n** �Ciu[A][u] _Ci[s][A][s][�]_ = Λ �Gui 0[�] _−_ **1n** �0 _Ci[s][A][s][�]_ _,_ _β[T][ �]G[u]i_ 0[�] = �Ciu[A][u] 0[�] = CiA − �0 _Ci[s][A][s][�]_ _,_ (68) ----- where A and Ci are given in (8) and (10), respectively. Now we are ready to prove Lemma 5. Notice that any matrix For simplicity, we denote _W can be decomposed as_ _Gi ≜_ �Gui 0[�] _∈_ R[n][×][n]. (69) w1[T]  _W =_ ... = e1w1T [+][ e][2][w]2[T] [+][ · · ·][ e][n][w]n[T] _[.]_ (75) Moreover, let   _ϵi(k) ≜_ _Gix(k) −_ _ξ[ˆ]i(k)._ (70) _wn[T]_ Since (S[T] _, β) is controllable, (24) can be concluded by_ It follows from (15) that applying Lemma 7 to (75). _ϵi(k + 1) = Gix(k + 1) −_ _ξ[ˆ]i(k + 1)_ = GiAx(k) + Giw(k) − Λξ[ˆ]i(k) − **1n yi(k + 1)** = (Gi − **1n Ci)Ax(k) −** Λξ[ˆ]i(k) + (Gi − **1n Ci)w(k)** APPENDIX E PROOF OF THEOREM 1 _−_ **1n vi(k + 1)** = ΛG − **1n** �0 _Ci[s][A][s][�]_ _x(k) −_ Λξ[ˆ]i(k) + (Gi − **1n Ci)w(k)** To begin with, we note that the following relation holds true _−_ **1n vi(k + 1)** at any k ≥ 0: = Λϵi(k) − **1n Ci[s][A][s][x][s][(][k][) + (][G][i]** _[−]_ **[1][n]** _[C][i][)][w][(][k][)][ −]_ **[1][n]** _[v][i][(][k][ + 1)][,]_ � �n � �n (71) _FiS[k]_ = _Hjpij(S)_ _S[k]_ = _HjS[k]pij(S),_ (76) where the second to last equality holds by (68). Due to the _j=1_ _j=1_ fact that Λ and A[s] are stable, we conclude that ϵi(k) is stable, where the last equality holds as S is commutable with any i.e., cov(ϵi(k)) is bounded. polynomials of itself. Then let us consider the output of system (22): One thus has _zi(k) = yi(k + 1) −_ _β[T][ ˆ]ξi(k)_ = yi(k + 1) − _β[T]_ (Gix(k) − _ϵi(k))_ = Ci(Ax(k) + w(k)) + vi(k + 1) + β[T] _ϵi(k)_ _−_ (CiA − �0 _Ci[s][A][s][�])x(k)_ = β[T] _ϵi(k) + Ci[s][A][s][x][s][(][k][) +][ C][i][w][(][k][) +][ v][i][(][k][ + 1)][.]_ _k_ � _F_ (Im ⊗ _S)[t](Im ⊗_ **1n)z(k −** _t)_ _t=0_ _k_ � _H(In ⊗_ _S)[t]Tz(k −_ _t) = ˜x(k + 1)._ _t=0_ (72) _xˆ(k + 1) =_ = = = = _k_ � _t=0_ _k_ � _t=0_ _k_ � _t=0_ � �[m] � _FiS[t]1nzi(k −_ _t)_ _i=1_ _n_ � � _HjS[t][�]_ �[m] _pij(S)1nzi(k −_ _t)�[�]_ _j=1_ _i=1_ _n_ � � _HjS[t]pij(S)1nzi(k −_ _t)_ _j=1_ As proved in (71), cov(ϵi(k)) is bounded. Moreover, it follows from (9) that Ci[s][A][s][x][s][(][k][)][ is a linear combination of the stable] parts in x(k). Also, the covariance of w(k) and vi(k + 1) are bounded as Q and Ri, respectively. We thus conclude that _zi(k) is stable, i.e., the covariance of zi(k) is always bounded._ APPENDIX D PROOF OF LEMMA 5 For the proof of Lemma 5, we need the following result: **Lemma 7. Given any vector w ∈** R[n]. Suppose (S[T] _, v) is_ _controllable, then there exists a polynomial p of at most n_ 1 _−_ _degree, such that w can be decomposed as_ _w[T]_ = v[T] _ϕ(S)._ (73) _Proof. Suppose ϕ(S) = α0I + α1S + · · · + αn−1S[n][−][1]. We_ thus rewrite (73) as � �[m] _i=1_ (77) Notice that (27) has z(k) as its input and ˜x(k) as its output. As proved, given any z(k), (22) and (27) yield the same output, i.e., ˜x(k) = ˆx(k + 1). Hence, we conclude that the two systems have the identical transfer functions. The proof is thus completed. APPENDIX F PROOF OF THEOREM 3 For simplicity, we first define aggregated vectors and matrices as below: (78) _L˜1_ ... _L˜_   _._   _ξˆ1(k)_ ... _ξˆm(k)_   _,_   _, ˆξ(k) ≜_  � � _w =_ _v_ _S[T]_ _v_ �S[n][−][1][�][T] _v_ _· · ·_  _α0_ ... _αn−1_ _η(k) ≜_ _Lη ≜_   _._ (74)     _η1(k)_ ... _ηm(k)_ Since (S[T] _, v) is controllable, the first matrix on the RHS of the_ equation has a column rank of n and hence the above equation is always solvable We therefore complete the proof ----- Then, we can rewrite (31) in matrix form as: _η(k + 1)_ =(Im ⊗ _S[˜])η(k) + Lηz(k) −_ [Im ⊗ ( B[˜]Γ)]([˜] _LG ⊗_ _In)η(k)_ =[Im ⊗ _S[˜] −LG ⊗_ ( B[˜]Γ)][˜] _η(k) + Lηz(k)._ (79) Next let us denote the average state of all agents as Lemma 6, diag( S[˜]−µ2B[˜]Γ[˜], · · ·, _S[˜]−µmB[˜]Γ)[˜]_ is Schur. Recalling Lemma 4, z(k) is also stable. We therefore conclude that (89) is stable, which further implies the stability of (87). On the other hand, one derives from (72) that _z(k) = Cw(k)+v(k+1)+(Im⊗β[T]_ )ϵ(k)+C _[s]A[s]x[s](k), (90)_ where ϵ(k) ≜ col(ϵ1(k), · · ·, ϵm(k)) and _C1[s]_ ... _Cm[s]_  . _η¯(k) ≜_ [1] _m_ _m_ � _ηi(k) = m[1]_ [(][1]m[T] _[⊗][I][mn][)][η][(][k][)][.]_ (80) _i=1_ _C_ _[s]_ =   Since 1[T]m _[L][G]_ [= 0][, it holds that] _η¯(k + 1)_ � � = m[1] [(][1]m[T] _[⊗][I][mn][)]_ [Im ⊗ _S[˜] −LG ⊗_ ( B[˜]Γ)][˜] _η(k) + Lηz(k)_ = S[˜]η¯(k) + m[1] [(][1]m[T] _[⊗][I][mn][)][L][η][z][(][k][)][.]_ (81) Recalling (71), it follows that    _w(k)_ Furthermore, we define the state deviation of each sensor as _δi(k) ≜_ _ηi(k) −_ _η¯(k) and then stack them as an aggregated_ vector δ(k) ≜ col(δ1(k), · · ·, δm(k)). Combining (79) and (81) yields the dynamic equation of δ(k): _δ(k + 1) = [Im ⊗_ _S[˜] −LG ⊗_ ( B[˜]Γ)][˜] _δ(k) + Lδz(k),_ (82) where _Lδ ≜_ [(Im − _m[1]_ **[1][m][ 1]m[T]** [)][ ⊗] _[I][mn][]][L][η][.]_ (83) Recall that the Laplacian matrix of an undirected graph is symmetric. Therefore, we can always find an unitary matrix Φ ≜ [ _√1m 1m, φ2, · · ·, φm], such that LG is diagonalized as_ diag(0, µ2, · · ·, µm) = Φ[T] _LGΦ._ (84) Using the property of Kronecker product yields that (Φ ⊗ _Imn)[T]_ [Im ⊗ _S[˜] −LG ⊗_ ( B[˜]Γ)](Φ[˜] _⊗_ _Imn)_ (85) = diag( S,[˜] _S[˜] −_ _µ2B[˜]Γ[˜], ...,_ _S[˜] −_ _µmB[˜]Γ)[˜]_ _._ Denote _δ˜(k) ≜_ (Φ ⊗ _Imn)[T]_ _δ(k)._ (86) One has _δ˜(k + 1) = Aδ˜δ˜(k) + Lδ˜z(k),_ (87) where Aδ˜ ≜ _diag( S,[˜]_ _S[˜] −_ _µ2B[˜]Γ[˜], · · ·,_ _S[˜] −_ _µmB[˜]Γ)[˜]_ and Lδ˜ ≜ [(Φ[T] _−_ _m[1]_ [Φ][T][ 1][m][ 1]m[T] [)][ ⊗] _[I][mn][]][L][η][.]_ We next study the stability of above system. To proceed, let us partition the state into two parts, i.e., _δ[˜](k) =_ [δ[˜]1[T] [(][k][)][,][ ˜][δ]2[T] [(][k][)]][T][, where][ ˜][δ][1][(][k][)][ ∈] [R][mn][ is a vector consisting of] the first mn entries of _δ[˜](k) and satisfies_  _G1_ **1n C1**  _−_ _ϵ(k + 1) = (Im ⊗_ Λ)ϵ(k) +  ...  _w(k)_ _Gm_ **1n Cm** _−_ _−(Im ⊗_ **1n)v(k + 1) −** (Im ⊗ **1n)C** _[s]A[s]x[s](k)_ = (Im ⊗ Λ)ϵ(k) + Wϵw(k) + Vϵv(k + 1) + Aϵx[s](k), _Wr = ArWrA[T]r_ [+]  _Wϵ_  _Q_  _Wϵ_  + Vϵ _J_ _J_ 0 where Aδ˜ _Lδ˜(Im ⊗_ _β[T]_ ) _Lδ˜C_ _[s]A[s]_ _Ar =_  _Im ⊗_ Λ _Aϵ_ _A[s]_ where (91)     _,_  _G1_ **1n C1**  _−_ _Wϵ ≜_ ... _,_   _Gm_ **1n Cm** _−_ _Vϵ ≜_ _−(Im ⊗_ **1n), Aϵ ≜** _−(Im ⊗_ **1n)C** _[s]A[s]._ By combining the above dynamics with (9), one derives that  _δ˜(k + 1)_  Aδ˜ _Lδ˜(Im ⊗_ _β[T]_ ) _Lδ˜C_ _[s]A[s]_  _δ˜(k)_   _ϵ(k + 1)_  =  _Im ⊗_ Λ _Aϵ_   _ϵ(k)_  _x[s](k + 1)_ _A[s]_ _x[s](k)_  _w(k) +_   _v(k + 1)_  Lδ˜C Lδ˜ +  _Wϵ_  _w(k) +_ Vϵ  _v(k + 1)_ _J_ 0 (92) Notice that the above system is stable. Hence, we calculate the covariance at both sides and in steady state. It holds that _Wr, the steady state covariance, is the unique solution of below_ Lyapunov equation: Lδ˜ Vϵ 0  _R_  Lδ˜C  _Wϵ_ _J_  _Q_  Lδ˜C  _Wϵ_ _J_   _T_ Lδ˜ Vϵ 0 T  _,_ _[s]_ _._  _m_ � (ηi(k) − _η¯(k)) = 0. (88)_ _i=1_  _δ˜(k)_  _δ(k) =_ �Φ ⊗ _Imn_ 0 0[�]  _ϵ(k)_  = Φδ _x[s](k)_ 1 _δ˜1(k) =_ _√_ _m_ _m_ � 1 _δi(k) =_ _√_ _m_ _i=1_ In view of (86), it holds that  _δ˜(k)_  _ϵ(k)_ _x[s](k)_ (93)  _,_ (94)  Therefore, _δ[˜]1(k) is stable. Moreover, it holds that_ _δ˜2(k_ +1) = diag( ˜S − _µ2 ˜B˜Γ, · · ·, ˜S −_ _µm ˜B˜Γ)˜δ2(k)+ ˜Lδ˜z(k),_ (89) where _L[˜] consists the last (m[2]n_ _mn) rows of_ _L[˜]_ In view of where Φ ≜ �Φ ⊗ _I_ 0 0[�] (95) ----- Moreover, let us denote _e¯i(k) ≜_ _x˘i(k) −_ _xˆ(k),_ (96) which is the bias from local estimate ˘xi(k) to optimal Kalman one. Combining (16) and (35) yields By Cauchy-Schwarz inequality, it holds for any i, j that � E[κ[T]i _[κ][j][]][ ≤]_ � E[κ[T]i _[κ][i][]]_ E[κ[T]j _[κ][j][]][.]_ (106) _xˆ(k) = F_ _m_ � _ηi(k)._ (97) _i=1_ One thus has _e¯i(k) = mF_ (ηi(k) − _η¯(k)) = mFδi(k)._ (98) Stacking such errors from all sensors together yields � _._ (99) The proof is thus completed. We next prove Theorem 4. Applying similar arguments to Theorem 2, it is easy to see from the consistency condition (49) that the average of local estimates coincides with the optimal Kalman filter. We hence focus on the analysis of estimation error covariance. Let us denote δi(k) ≜ _ηi(k)−1/m_ [�]i[m]=1 _[η][i][(][k][)][ and][ ϖ][i][(][k][)][ ≜]_ _ωi(k) −_ 1/m [�]i[m]=1 _[ω][i][(][k][)][. Moreover, we define]_ _δ(k) ≜_ col(δ1(k), · · ·, δm(k)), _ϖ(k) ≜_ col(ϖ1(k), · · ·, ϖm(k)). It hence follows from (48) that �δ(k + 1)� �D(k) _J (k)��_ _δ(k)_ � �Lδ� = + _z(k),_ _ϖ(k)_ _ϖ(k_ 1) 0 _B�_ _A�_ _−_ (107) where Lδ is defined in (83), and _D(k) ≜_ _Im ⊗_ _S[˜] −L(k) ⊗_ ( B[˜]Γ[˜]B), _J (k) ≜_ _−L(k) ⊗_ ( B[˜]Γ[˜]A), _A[�] ≜_ _Im ⊗A,_ _B[�] ≜_ _Im ⊗B,_ _e¯(k) = (Im ⊗_ _mF_ )δ(k) = (Im ⊗ _mF_ )Φδ �δ˜(k) _ϵ(k)_ Therefore, in steady state, the covariance of ¯e(k) can be calculated as _W¯_ = [(Im ⊗ _mF_ )Φδ]Wr[(Im ⊗ _mF_ )Φδ][T] _._ (100) Finally, for any sensor i, let us denote its estimation error as _e˘i(k) = ˘xi(k) −_ _x(k)_ = (˘xi(k) − _xˆ(k)) + (ˆx(k) −_ _x(k))_ (101) = ¯ei(k) + ˆe(k), where ˆe(k) is the estimation error of Kalman filter. Since Kalman filter is optimal, ¯ei(k) is orthogonal to ˆe(k). By defining ˘e(k) ≜ col(˘e1(k), · · ·, ˘em(k)), we therefore have _e˘(k) = ¯e(k) + 1m ⊗eˆ(k)._ (102) Calculating the covariance of both sides yields _W˘_ = ¯W + (1m 1[T]m[)][ ⊗] _[P,]_ (103) where _W[˘]_ is the steady-state covariance of ˘e(k) and P is given in (5). Notice that the above calculation also indicates the boundedness of cov(˘e(k)) at any time. APPENDIX G PROOF OF COROLLARY 1 As proved in Appendix-F, one can exactly calculate _W[¯]_ by solving Lyapunov equations (93) and (100). The result is thus obvious by invoking (103). APPENDIX H PROOF OF THEOREM 4 To proceed, let us introduce the following lemma: **Lemma 8. Given any random variables κ1, · · ·, κτ** _, it follows_ _that_ _τ_ _τ_ � 2[�] � � � �2 E������� _κi������_ _≤_ E[||κi||[2]] _._ (104) _i=1_ _i=1_ _Proof. In order to prove (104), it is equivalent to show that_ with L(k) ≜ _{Li,j(k)} being the (random) Laplacian matrix_ with respect to the weights {aijγij(k)}. Namely, _Li,j(k) ≜_ ��−amlij=1γij[a][il](k[γ])[il],[(][k][)][,] _j =j ̸= i i [.]_ (108) For simplicity, Let � (k) (k) _D_ _J_ _Q(k) ≜_ _B�_ _A�_ � _._ (109) Since δi(0) = 0 and ϖi(0) = 0 hold for any i, it follows that �δ(k + 1)� �k � �Lδ� � = (k, t + 1) _z(t)_ _,_ (110) _ϖ(k)_ _Q_ 0 _t=0_ where the transition matrix is defined as � (k) (k 1) (s), _k_ _s,_ _Q_ _Q_ _−_ _· · · Q_ _≥_ (k, s) = _Q_ _I,_ _k < s._ Then consider the update of any agent i. From the above equation, we conclude that _δi(k + 1) =_ _k_ � Πi(k, t + 1)z(t), (111) _t=0_ _τ_ _τ_ � � E[κ[T]i _[κ][j][]][ ≤]_ _τ_ _τ_ � � [�] � E[κ[T]i _[κ][i][]]_ E[κ[T]j _[κ][j][]][.]_ (105) where Πi(k, t + 1) refers to the i-th row of matrix Q(k, t + 1)[Lδ 0][T] _. Namely, the consensus error of agent i, i.e. δi(k+_ 1), is caused by the sequence of residuals _z(t)_, where t _k._ _{_ _}_ _≤_ For simplicity, we denote _κi(k, t) ≜_ Πi(k, t + 1)z(t). Since cov(z(t)) is bounded at any time, in view of (50), the following statement holds for any t _k:_ _≤_ E[||κ (k t)||[2]] ≤ _cρ[k][−][t]_ (112) ----- Therefore, one has that _cov(δi(k + 1)) = E[||δi(k + 1)||[2]] = E�������_ _k_ � 2[�] _κi(k, t)_ ������ _t=0_ _τ_ _k_ � � � �2 � � _≤_ E[||κi(k, t)||[2]] _≤_ _i=1_ _t=0_ = _[c][(1][ −√][ρ][k][)][2]_ (1 _ρ)[2][,]_ _−_ _[√]_ [14] G. Battistelli, L. Chisci, G. Mugnai, A. Farina, and A. Graziano, “Consensus-based linear and nonlinear filtering,” IEEE Transactions on _Automatic Control, vol. 60, no. 5, pp. 1410–1415, 2014._ [15] W. Li and Y. Jia, “Consensus-based distributed multiple model UKF for jump Markov nonlinear systems,” IEEE Transactions on Automatic _Control, vol. 57, no. 1, pp. 227–233, 2011._ [16] G. Battistelli and L. Chisci, “Stability of consensus extended Kalman filter for distributed state estimation,” Automatica, vol. 68, pp. 169–178, 2016. [17] S. Del Favero and S. Zampieri, “Distributed estimation through randomized gossip Kalman filter,” in Proceedings of the 48th IEEE Conference _on Decision and Control (CDC) held jointly with the 28th Chinese_ _Control Conference._ IEEE, 2009, pp. 7049–7054. [18] S. Kar and J. M. Moura, “Gossip and distributed Kalman filtering: Weak consensus under weak detectability,” IEEE Transactions on Signal _Processing, vol. 59, no. 4, pp. 1766–1784, 2010._ [19] K. Ma, S. Wu, Y. Wei, and W. Zhang, “Gossip-based distributed tracking in networks of heterogeneous agents,” IEEE Communications Letters, vol. 21, no. 4, pp. 801–804, 2016. [20] F. S. Cattivelli and A. H. Sayed, “Diffusion strategies for distributed Kalman filtering and smoothing,” IEEE Transactions on Automatic _Control, vol. 55, no. 9, pp. 2069–2084, 2010._ [21] J. Hu, L. Xie, and C. Zhang, “Diffusion Kalman filtering based on covariance intersection,” IEEE Transactions on Signal Processing, vol. 60, no. 2, pp. 891–902, 2011. [22] F. S. Cattivelli, C. G. Lopes, and A. H. Sayed, “Diffusion strategies for distributed Kalman filtering: Formulation and performance analysis,” _Proc. Cognitive Information Processing, pp. 36–41, 2008._ [23] M. Farina, G. Ferrari-Trecate, and R. Scattolini, “Distributed moving horizon estimation for linear constrained systems,” IEEE Transactions _on Automatic Control, vol. 55, no. 11, pp. 2462–2475, 2010._ [24] A. Haber and M. Verhaegen, “Moving horizon estimation for largescale interconnected systems,” IEEE Transactions on Automatic Control, vol. 58, no. 11, pp. 2834–2847, 2013. [25] G. Battistelli and L. Chisci, “Kullback–Leibler average, consensus on probability densities, and distributed state estimation with guaranteed stability,” Automatica, vol. 50, no. 3, pp. 707–718, 2014. [26] L. Chen, P. O. Arambel, and R. K. Mehra, “Estimation under unknown correlation: Covariance intersection revisited,” IEEE Transactions on _Automatic Control, vol. 47, no. 11, pp. 1879–1882, 2002._ [27] X. He, W. Xue, and H. Fang, “Consistent distributed state estimation with global observability over sensor network,” Automatica, vol. 92, pp. 162–172, 2018. [28] S. Das and J. M. Moura, “Consensus+ innovations distributed Kalman filter with optimized gains,” IEEE Transactions on Signal Processing, vol. 65, no. 2, pp. 467–481, 2016. [29] G. Battistelli, L. Chisci, and D. Selvi, “A distributed Kalman filter with event-triggered communication and guaranteed stability,” Automatica, vol. 93, pp. 75–82, 2018. [30] L. Shi, P. Cheng, and J. Chen, “Sensor data scheduling for optimal state estimation with communication energy constraint,” Automatica, vol. 47, no. 8, pp. 1693–1698, 2011. [31] K. You and L. Xie, “Network topology and communication data rate for consensusability of discrete-time multi-agent systems,” IEEE Trans_actions on Automatic Control, vol. 56, no. 10, pp. 2262–2275, 2011._ [32] K. You, Z. Li, and L. Xie, “Consensus condition for linear multiagent systems over randomly switching topologies,” Automatica, vol. 49, no. 10, pp. 3125–3132, 2013. [33] L. Xu, Y. Mo, and L. Xie, “Distributed consensus over Markovian packet loss channels,” IEEE Transactions on Automatic Control, vol. 65, no. 1, pp. 279–286, 2019. [34] G. Gu, L. Marinovici, and F. L. Lewis, “Consensusability of discretetime dynamic multiagent systems,” IEEE Transactions on Automatic _Control, vol. 57, no. 8, pp. 2085–2089, 2011._ [35] F. Amato, M. Ariola, and P. Dorato, “Finite-time control of linear systems subject to parametric uncertainties and disturbances,” Automatica, vol. 37, no. 9, pp. 1459–1463, 2001. [36] Y. Su and J. Huang, “Two consensus problems for discrete-time multiagent systems with switching network topology,” Automatica, vol. 48, no. 9, pp. 1988–1997, 2012. [37] Y. Mo and E. Garone, “Secure dynamic state estimation via local estimators,” in 2016 IEEE 55th Conference on Decision and Control _(CDC)._ IEEE, 2016, pp. 5073–5078. [38] X. Yang, J. Yan, Y. Mo, and K. You, “A distributed implementation of steady-state Kalman filter,” in 2021 40th Chinese Control Conference _(CCC)_ IEEE 2021 5154 5159 � �2 _cρ[k][−][t]_ (113) where the first inequality holds by using Lemma 8. Since ρ _∈_ (0, 1), combining the above results with (98) and (101) yields that the estimation error is stable. **Remark 9. It is noted that the reformulation (20) with** _stable input zi(k) is essential to establish the stability of_ _local estimators. To be concrete, the stability of (111) is_ _guaranteed under the bounded input, which is key to prove_ _the boundedness of estimation error covariance, as can be_ _observed from (98)-(103). On the other hand, if an unstable_ _input, e.g., yi(k) as in (15), is applied, we cannot conclude_ _on the stability of local estimator even using the exponentially_ _converged synchronization algorithms._ REFERENCES [1] M. V. Subbotin and R. S. Smith, “Design of distributed decentralized estimators for formations with fixed and stochastic communication topologies,” Automatica, vol. 45, no. 11, pp. 2491–2501, 2009. [2] L. Xie, D.-H. Choi, S. Kar, and H. V. Poor, “Fully distributed state estimation for wide-area monitoring systems,” IEEE Transactions on _Smart Grid, vol. 3, no. 3, pp. 1154–1169, 2012._ [3] Z.-Q. Luo, “Universal decentralized estimation in a bandwidth constrained sensor network,” IEEE Transactions on Information Theory, vol. 51, no. 6, pp. 2210–2219, 2005. [4] T. T. Vu and A. R. Rahmani, “Distributed consensus-based Kalman filter estimation and control of formation flying spacecraft: Simulation and validation,” in Proceedings of the AIAA Guidance, Navigation, and _Control Conference, 2015, p. 1553._ [5] B. Jia, K. D. Pham, E. Blasch, D. Shen, Z. Wang, and G. Chen, “Cooperative space object tracking using space-based optical sensors via consensus-based filters,” IEEE Transactions on Aerospace and _Electronic Systems, vol. 52, no. 4, pp. 1908–1936, 2016._ [6] B. D. Anderson and J. B. Moore, Optimal Filtering. Courier Corporation, 2012. [7] Y. Bar-Shalom and L. Campo, “The effect of the common process noise on the two-sensor fused-track covariance,” IEEE Transactions on _Aerospace and Electronic Systems, no. 6, pp. 803–805, 1986._ [8] K. H. Kim, “Development of track to track fusion algorithms,” in _Proceedings of 1994 American Control Conference._ IEEE, 1994, pp. 1037–1041. [9] S.-L. Sun and Z.-L. Deng, “Multi-sensor optimal information fusion Kalman filter,” Automatica, vol. 40, no. 6, pp. 1017–1023, 2004. [10] B. Chen, G. Hu, D. W. Ho, and L. Yu, “Distributed Kalman filtering for time-varying discrete sequential systems,” Automatica, vol. 99, pp. 228–236, 2019. [11] R. Olfati-Saber, “Distributed Kalman filtering for sensor networks,” in _Proceedings of the 46th IEEE Conference on Decision and Control._ IEEE, 2007, pp. 5492–5498. [12] ——, “Distributed Kalman filter with embedded consensus filters,” in _Proceedings of the 44th IEEE Conference on Decision and Control._ IEEE, 2005, pp. 8179–8184. [13] ——, “Kalman-consensus filter: Optimality, stability, and performance,” in Proceedings of the 48th IEEE Conference on Decision and Control _(CDC) held jointly with the 28th Chinese Control Conference._ IEEE, 2009 7036 7042 ----- [39] X. Mao, X. Miao, Y. He, X.-Y. Li, and Y. Liu, “Citysee: Urban co2 monitoring with sensors,” in 2012 Proceedings IEEE INFOCOM. IEEE, 2012, pp. 1611–1619. [40] L. Parolini, B. Sinopoli, B. H. Krogh, and Z. Wang, “A cyber–physical systems approach to data center modeling and control for energy efficiency,” Proceedings of the IEEE, vol. 100, no. 1, pp. 254–268, 2011. [41] Y. Mo, R. Ambrosino, and B. Sinopoli, “Sensor selection strategies for state estimation in energy constrained wireless sensor networks,” _Automatica, vol. 47, no. 7, pp. 1330 – 1338, 2011._ [42] ——, “Network energy minimization via sensor selection and topology control,” IFAC Proceedings Volumes, vol. 42, no. 20, pp. 174 – 179, 2009. [43] W. Li, G. Wei, D. W. Ho, and D. Ding, “A weightedly uniform detectability for sensor networks,” IEEE transactions on neural networks _and learning systems, vol. 29, no. 11, pp. 5790–5796, 2018._ [44] Z. Li and Y. Mo, “Efficient secure state estimation against sparse integrity attack for system with non-derogatory dynamics,” arXiv preprint _arXiv:2106.03066, 2021._ -----
{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2101.10689, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.", "license": null, "status": "GREEN", "url": "https://arxiv.org/pdf/2101.10689" }
2,021
[ "JournalArticle" ]
true
2021-01-26T00:00:00
[ { "paperId": "09e7bf9d0fee7565a9f6a0a7ccdba053818d1bfd", "title": "Efficient Secure State Estimation against Sparse Integrity Attack for System with Non-derogatory Dynamics" }, { "paperId": "8c3e6887cbe510fadb9cf24e9d174674bfb5791f", "title": "Distributed Finite-Time Integral Sliding-Mode Control for Multi-Agent Systems with Multiple Disturbances Based on Nonlinear Disturbance Observers" }, { "paperId": "ef1913c7386ed0c5e986553d617667f41c8a112a", "title": "Distributed Stochastic Source Seeking for Multiple Vehicles over Fixed Topology" }, { "paperId": "ed11f237b14d580ca6dc917f959180dc22500864", "title": "Distributed Kalman filtering for time-varying discrete sequential systems" }, { "paperId": "31ec777792b6d07ac8853e9af23c9eda782a8aff", "title": "Distributed Consensus Over Markovian Packet Loss Channels" }, { "paperId": "f5addbb8dd20a0535446f9944a22e73cd653b84d", "title": "A distributed Kalman filter with event-triggered communication and guaranteed stability" }, { "paperId": "5bfb3af403c6e826a5780bdcc96f26ac91a271d0", "title": "A Weightedly Uniform Detectability for Sensor Networks" }, { "paperId": "4dc31e90d3c14b60522df872efb343146e272e20", "title": "Consistent distributed state estimation with global observability over sensor network" }, { "paperId": "149a15b0af3ca5778101d32b608f8ccbff95fddb", "title": "Gossip-Based Distributed Tracking in Networks of Heterogeneous Agents" }, { "paperId": "1abf5d805d2c8a2b2099784f7f4bb58160d96f1d", "title": "Secure dynamic state estimation via local estimators" }, { "paperId": "3ff60b5aaec9f573a904f68d077a367815ff5b85", "title": "Consensus of Linear Multi-agent Systems with Persistent Disturbances" }, { "paperId": "78b9eb80d61ab8f675bac04f1f9a39402d075082", "title": "Cooperative space object tracking using space-based optical sensors via consensus-based filters" }, { "paperId": "5ff2f5ffd74269830a7574741a25ede58ec9dd30", "title": "Stability of consensus extended Kalman filter for distributed state estimation" }, { "paperId": "20c2239ccab70bbb1b549951290aba883fe2250c", "title": "Consensus+Innovations Distributed Kalman Filter With Optimized Gains" }, { "paperId": "bd52592f98ca0677f4419eb63df13eb0363dfffe", "title": "Consensus-Based Linear and Nonlinear Filtering" }, { "paperId": "83e02cc6e69ba124df9530c77a7b483d8d9190fe", "title": "Distributed Consensus-Based Kalman Filter Estimation and Control of Formation Flying Spacecraft: Simulation and Validation" }, { "paperId": "7f07bd0f2af6882bdb24c763e9fc0e9d762c4f55", "title": "Kullback-Leibler average, consensus on probability densities, and distributed state estimation with guaranteed stability" }, { "paperId": "2a759e85ffaecb06c60fa86e00e1e4bb26339888", "title": "Consensus condition for linear multi-agent systems over randomly switching topologies" }, { "paperId": "a082005bd3e48621ef959f98237a5df7bf112c86", "title": "Moving Horizon Estimation for Large-Scale Interconnected Systems" }, { "paperId": "32c78d11f695d620294655bdfe94d41790163720", "title": "Two consensus problems for discrete-time multi-agent systems with switching network topology" }, { "paperId": "017989d26bca7ad17848f244fbee5e2879e2f959", "title": "Consensusability of Discrete-Time Dynamic Multiagent Systems" }, { "paperId": "ab0f9990c79569765a5e78acf07c47ad406ebddb", "title": "Fully Distributed State Estimation for Wide-Area Monitoring Systems" }, { "paperId": "0bf29d128b87211f60fa60dc1fedd40844f4e306", "title": "CitySee: Urban CO2 monitoring with sensors" }, { "paperId": "8d851384a52d00b06842a7d40dc791ca6aacdff6", "title": "Diffusion Kalman Filtering Based on Covariance Intersection" }, { "paperId": "0122e7e784aae390dcc230f23bc79e8d19073092", "title": "Network Topology and Communication Data Rate for Consensusability of Discrete-Time Multi-Agent Systems" }, { "paperId": "a0cade3d1a9a71e780e0ec6d59279eec4159da14", "title": "Sensor selection strategies for state estimation in energy constrained wireless sensor networks" }, { "paperId": "7489910c640dccc86ab8c60c75d44e28015099ac", "title": "Distributed moving horizon estimation for nonlinear constrained systems" }, { "paperId": "baf83fa0a09085c7b0ed76c99f6e9a5993f05035", "title": "Diffusion Strategies for Distributed Kalman Filtering and Smoothing" }, { "paperId": "887e48278e922b1c258ea14d088c3c867fb1a527", "title": "Kalman-Consensus Filter : Optimality, stability, and performance" }, { "paperId": "c53bf33f401988b4f2a74102ac629f0cc7346e69", "title": "Distributed estimation through randomized gossip Kalman filter" }, { "paperId": "67726517ea954b26e5c1ee561e6cf213fd2d3dee", "title": "Design of distributed decentralized estimators for formations with fixed and stochastic communication topologies" }, { "paperId": "9bd11e99ba0045ce3e2f9526826d0ce2ccded294", "title": "Network Energy Minimization via Sensor Selection and Topology Control" }, { "paperId": "5021b6fc3bba5fb2a69ca9d4aa150e36d9d8702e", "title": "Distributed Kalman filtering for sensor networks" }, { "paperId": "abca3c16c31d20c9d2796c07c5b24b06dee076b2", "title": "Distributed Kalman Filter with Embedded Consensus Filters" }, { "paperId": "f85c47d9487462566c58f8862b6b81af9dd90a2b", "title": "Universal decentralized estimation in a bandwidth constrained sensor network" }, { "paperId": "4d17c39e14a5491adb58a339a541f17bad1aa266", "title": "Multi-sensor optimal information fusion Kalman filter" }, { "paperId": "628bf7b0af173de590e435b13f950d28f3ea0fc0", "title": "Estimation under unknown correlation: covariance intersection revisited" }, { "paperId": "209e79be0d39307846f099a6f4fb4b3bd721ba76", "title": "Technical Communique: Finite-time control of linear systems subject to parametric uncertainties and disturbances" }, { "paperId": "d508522cf62c9a806d44c6e1f4602a4c3316d018", "title": "On optimal ℓ∞ to ℓ∞ filtering" }, { "paperId": "52074c55ae13e19b51131b108bba6bac5f5a0360", "title": "Development of track to track fusion algorithms" }, { "paperId": "f2963642d0e3d9d9f5b015b89f7da084a9e492a0", "title": "The Effect of the Common Process Noise on the Two-Sensor Fused-Track Covariance" }, { "paperId": null, "title": "“A distributed implementation of steady-state Kalman filter,”" }, { "paperId": "2e15eac922143e30cdde0ee2ab8c3e729cbd3cf7", "title": "A Cyber–Physical Systems Approach to Data Center Modeling and Control for Energy Efficiency" }, { "paperId": "3f49bb27995f4b4004d535f9f17002b283a89e32", "title": "Consensus-Based Distributed Multiple Model UKF for Jump Markov Nonlinear Systems" }, { "paperId": "9b5e1d798bfc8a7a367ef4fcf6988e6d9f112c39", "title": "Distributed Kalman Filtering : Weak Consensus Under Weak Detectability" }, { "paperId": "e1c61bf68627ef57cd8dd5e2f279a64e5ec5a87d", "title": "Sensor data scheduling for optimal state estimation with communication energy constraint" }, { "paperId": "05b6fa6eb34fa83a741a304ac6cf6544bc887924", "title": "Diffusion strategies for distributed Kalman filtering: formulation and performance analysis" }, { "paperId": "79fa5e51a519d30a19b3fd3accfa580cbd556b33", "title": "Distributed Algorithms" }, { "paperId": null, "title": "Zeromq, an open-source universal messaging library" }, { "paperId": null, "title": "B.a.t.m.a.n. advanced documentation overview" } ]
27,006